Throw an NSException on Apple platforms when panicking

Replace the num_views for OpenGL multiviwe only when
- The engine is initialized with multiview stereo
- The variant for the material contains STE flag
- The program is for surface
- It's vertex shader (this is already in)

.github/workflows/release.yml

@@ -205,7 +205,7 @@ jobs:
           TAG: ${{ steps.git_ref.outputs.tag }}
         run: |
           build\windows\build-github.bat release
-          move out\filament-windows.tgz out\filament-$Env:TAG-windows.tgz
+          move out\filament-windows.tgz out\filament-%TAG%-windows.tgz
         shell: cmd
       - uses: actions/github-script@v6
         env:

CMakeLists.txt

@@ -45,6 +45,8 @@ option(FILAMENT_ENABLE_TSAN "Enable Thread Sanitizer" OFF)
 
 option(FILAMENT_ENABLE_FEATURE_LEVEL_0 "Enable Feature Level 0" ON)
 
+option(FILAMENT_ENABLE_MULTIVIEW "Enable multiview for Filament" OFF)
+
 set(FILAMENT_NDK_VERSION "" CACHE STRING
     "Android NDK version or version prefix to be used when building for Android."
 )
@@ -531,6 +533,21 @@ else()
     option(FILAMENT_DISABLE_MATOPT "Disable material optimizations" ON)
 endif()
 
+# This only affects the prebuilt shader files in gltfio and samples, not filament library.
+# The value can be either "instanced" or "multiview".
+set(FILAMENT_SAMPLES_STEREO_TYPE "instanced" CACHE STRING
+    "Stereoscopic type that shader files in gltfio and samples are built for."
+)
+string(TOLOWER "${FILAMENT_SAMPLES_STEREO_TYPE}" FILAMENT_SAMPLES_STEREO_TYPE)
+if (NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "instanced" AND NOT FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "multiview")
+    message(FATAL_ERROR "Invalid stereo type: \"${FILAMENT_SAMPLES_STEREO_TYPE}\" choose either \"instanced\" or \"multiview\" ")
+endif ()
+
+# Compiling samples for multiview implies enabling multiview feature as well.
+if (FILAMENT_SAMPLES_STEREO_TYPE STREQUAL "multiview")
+    set(FILAMENT_ENABLE_MULTIVIEW ON)
+endif ()
+
 # ==================================================================================================
 # Material compilation flags
 # ==================================================================================================

NEW_RELEASE_NOTES.md

@@ -7,3 +7,5 @@ for next branch cut* header.
 appropriate header in [RELEASE_NOTES.md](./RELEASE_NOTES.md).
 
 ## Release notes for next branch cut
+
+- engine: Add experimental APIs `Engine::builder::paused()` and `Engine::setPaused()`

README.md

@@ -31,7 +31,7 @@ repositories {
 }
 
 dependencies {
-    implementation 'com.google.android.filament:filament-android:1.50.5'
+    implementation 'com.google.android.filament:filament-android:1.51.0'
 }
 ```
 
@@ -51,7 +51,7 @@ Here are all the libraries available in the group `com.google.android.filament`:
 iOS projects can use CocoaPods to install the latest release:
 
 ```shell
-pod 'Filament', '~> 1.50.5'
+pod 'Filament', '~> 1.51.0'
 ```
 
 ### Snapshots

RELEASE_NOTES.md

@@ -7,6 +7,19 @@ 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).
 
+## v1.51.1
+
+
+## v1.51.0
+
+- materials: add support for post-lighting mix factor (b/328498606) [⚠️ **New Material Version**]
+
+## v1.50.6
+
+- Add new API `SwapChain::getFrameScheduledCallback`
+- vulkan: fixed validation error VUID-vkAcquireNextImageKHR-semaphore-01779
+- opengl: Add support for protected content swapchains and contexts
+
 ## v1.50.5
 
 - android: NDK 26.1.10909125 is used by default

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

@@ -391,6 +391,13 @@ Java_com_google_android_filament_Engine_nFlush(JNIEnv*, jclass,
     engine->flush();
 }
 
+extern "C" JNIEXPORT void JNICALL
+Java_com_google_android_filament_Engine_nSetPaused(JNIEnv*, jclass,
+        jlong nativeEngine, jboolean paused) {
+    Engine* engine = (Engine*) nativeEngine;
+    engine->setPaused(paused);
+}
+
 // Managers...
 
 extern "C" JNIEXPORT jlong JNICALL
@@ -487,7 +494,8 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
         jlong jobSystemThreadCount,
         jlong textureUseAfterFreePoolSize, jboolean disableParallelShaderCompile,
         jint stereoscopicType, jlong stereoscopicEyeCount,
-        jlong resourceAllocatorCacheSizeMB, jlong resourceAllocatorCacheMaxAge) {
+        jlong resourceAllocatorCacheSizeMB, jlong resourceAllocatorCacheMaxAge,
+        jboolean disableHandleUseAfterFreeCheck) {
     Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
     Engine::Config config = {
             .commandBufferSizeMB = (uint32_t) commandBufferSizeMB,
@@ -502,6 +510,7 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
             .stereoscopicEyeCount = (uint8_t) stereoscopicEyeCount,
             .resourceAllocatorCacheSizeMB = (uint32_t) resourceAllocatorCacheSizeMB,
             .resourceAllocatorCacheMaxAge = (uint8_t) resourceAllocatorCacheMaxAge,
+            .disableHandleUseAfterFreeCheck = (bool) disableHandleUseAfterFreeCheck,
     };
     builder->config(&config);
 }
@@ -518,6 +527,12 @@ extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBu
     builder->sharedContext((void*) sharedContext);
 }
 
+extern "C" JNIEXPORT void JNICALL Java_com_google_android_filament_Engine_nSetBuilderPaused(
+        JNIEnv*, jclass, jlong nativeBuilder, jboolean paused) {
+    Engine::Builder* builder = (Engine::Builder*) nativeBuilder;
+    builder->paused((bool) paused);
+}
+
 extern "C" JNIEXPORT jlong JNICALL
 Java_com_google_android_filament_Engine_nBuilderBuild(JNIEnv*, jclass, jlong nativeBuilder) {
     Engine::Builder* builder = (Engine::Builder*) nativeBuilder;

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

@@ -104,6 +104,14 @@ Java_com_google_android_filament_RenderableManager_nBuilderGeometry__JIIJJIIII(J
             (size_t) count);
 }
 
+extern "C"
+JNIEXPORT void JNICALL
+Java_com_google_android_filament_RenderableManager_nBuilderGeometryType(JNIEnv*, jclass,
+        jlong nativeBuilder, int type) {
+    RenderableManager::Builder *builder = (RenderableManager::Builder *) nativeBuilder;
+    builder->geometryType((RenderableManager::Builder::GeometryType)type);
+}
+
 extern "C"
 JNIEXPORT void JNICALL
 Java_com_google_android_filament_RenderableManager_nBuilderMaterial(JNIEnv*, jclass,

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

@@ -34,7 +34,15 @@ Java_com_google_android_filament_SwapChain_nSetFrameCompletedCallback(JNIEnv* en
 }
 
 extern "C" JNIEXPORT jboolean JNICALL
-Java_com_google_android_filament_SwapChain_nIsSRGBSwapChainSupported(JNIEnv *, jclass, jlong nativeEngine) {
+Java_com_google_android_filament_SwapChain_nIsSRGBSwapChainSupported(
+        JNIEnv *, jclass, jlong nativeEngine) {
     Engine* engine = (Engine*) nativeEngine;
-    return (bool)SwapChain::isSRGBSwapChainSupported(*engine);
+    return (jboolean)SwapChain::isSRGBSwapChainSupported(*engine);
+}
+
+extern "C" JNIEXPORT jboolean JNICALL
+Java_com_google_android_filament_SwapChain_nIsProtectedContentSupported(
+        JNIEnv *, jclass, jlong nativeEngine) {
+    Engine* engine = (Engine*) nativeEngine;
+    return (jboolean)SwapChain::isProtectedContentSupported(*engine);
 }

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

@@ -224,7 +224,8 @@ public class Engine {
                     config.jobSystemThreadCount,
                     config.textureUseAfterFreePoolSize, config.disableParallelShaderCompile,
                     config.stereoscopicType.ordinal(), config.stereoscopicEyeCount,
-                    config.resourceAllocatorCacheSizeMB, config.resourceAllocatorCacheMaxAge);
+                    config.resourceAllocatorCacheSizeMB, config.resourceAllocatorCacheMaxAge,
+                    config.disableHandleUseAfterFreeCheck);
             return this;
         }
 
@@ -239,6 +240,20 @@ public class Engine {
             return this;
         }
 
+        /**
+         * Sets the initial paused state of the rendering thread.
+         *
+         * <p>Warning: This is an experimental API. See {@link Engine#setPaused(boolean)} for
+         * caveats.
+         *
+         * @param paused Whether to start the rendering thread paused.
+         * @return A reference to this Builder for chaining calls.
+         */
+        public Builder paused(boolean paused) {
+            nSetBuilderPaused(mNativeBuilder, paused);
+            return this;
+        }
+
         /**
          * Creates an instance of Engine
          *
@@ -400,16 +415,19 @@ public class Engine {
         public long stereoscopicEyeCount = 2;
 
         /*
-         * Size in MiB of the frame graph texture cache. This should be adjusted based on the
-         * size of used render targets (typically the screen).
+         * @Deprecated This value is no longer used.
          */
         public long resourceAllocatorCacheSizeMB = 64;
 
         /*
          * This value determines for how many frames are texture entries kept in the cache.
-         * The default value of 30 corresponds to about half a second at 60 fps.
          */
-        public long resourceAllocatorCacheMaxAge = 30;
+        public long resourceAllocatorCacheMaxAge = 2;
+
+        /*
+         * Disable backend handles use-after-free checks.
+         */
+        public boolean disableHandleUseAfterFreeCheck = false;
     }
 
     private Engine(long nativeEngine, Config config) {
@@ -656,7 +674,7 @@ public class Engine {
      */
     @NonNull
     public SwapChain createSwapChain(@NonNull Object surface) {
-        return createSwapChain(surface, SwapChain.CONFIG_DEFAULT);
+        return createSwapChain(surface, SwapChainFlags.CONFIG_DEFAULT);
     }
 
     /**
@@ -664,15 +682,15 @@ public class Engine {
      *
      * @param surface on Android, <b>must be</b> an instance of {@link android.view.Surface}
      *
-     * @param flags configuration flags, see {@link SwapChain}
+     * @param flags configuration flags, see {@link SwapChainFlags}
      *
      * @return a newly created {@link SwapChain} object
      *
      * @exception IllegalStateException can be thrown if the SwapChain couldn't be created
      *
-     * @see SwapChain#CONFIG_DEFAULT
-     * @see SwapChain#CONFIG_TRANSPARENT
-     * @see SwapChain#CONFIG_READABLE
+     * @see SwapChainFlags#CONFIG_DEFAULT
+     * @see SwapChainFlags#CONFIG_TRANSPARENT
+     * @see SwapChainFlags#CONFIG_READABLE
      *
      */
     @NonNull
@@ -690,21 +708,22 @@ public class Engine {
      *
      * @param width  width of the rendering buffer
      * @param height height of the rendering buffer
-     * @param flags  configuration flags, see {@link SwapChain}
+     * @param flags  configuration flags, see {@link SwapChainFlags}
      *
      * @return a newly created {@link SwapChain} object
      *
      * @exception IllegalStateException can be thrown if the SwapChain couldn't be created
      *
-     * @see SwapChain#CONFIG_DEFAULT
-     * @see SwapChain#CONFIG_TRANSPARENT
-     * @see SwapChain#CONFIG_READABLE
+     * @see SwapChainFlags#CONFIG_DEFAULT
+     * @see SwapChainFlags#CONFIG_TRANSPARENT
+     * @see SwapChainFlags#CONFIG_READABLE
      *
      */
     @NonNull
     public SwapChain createSwapChain(int width, int height, long flags) {
         if (width >= 0 && height >= 0) {
-            long nativeSwapChain = nCreateSwapChainHeadless(getNativeObject(), width, height, flags);
+            long nativeSwapChain =
+                nCreateSwapChainHeadless(getNativeObject(), width, height, flags);
             if (nativeSwapChain == 0) throw new IllegalStateException("Couldn't create SwapChain");
             return new SwapChain(nativeSwapChain, null);
         }
@@ -716,11 +735,12 @@ public class Engine {
      *
      * @param surface a properly initialized {@link NativeSurface}
      *
-     * @param flags configuration flags, see {@link SwapChain}
+     * @param flags configuration flags, see {@link SwapChainFlags}
      *
      * @return a newly created {@link SwapChain} object
      *
-     * @exception IllegalStateException can be thrown if the {@link SwapChain} couldn't be created
+     * @exception IllegalStateException can be thrown if the {@link SwapChainFlags} couldn't be
+     *            created
      */
     @NonNull
     public SwapChain createSwapChainFromNativeSurface(@NonNull NativeSurface surface, long flags) {
@@ -1189,6 +1209,22 @@ public class Engine {
         nFlush(getNativeObject());
     }
 
+    /**
+     * Pause or resume the rendering thread.
+     *
+     * <p>Warning: This is an experimental API. In particular, note the following caveats.
+     *
+     * <ul><li>
+     * Buffer callbacks will never be called as long as the rendering thread is paused.
+     * Do not rely on a buffer callback to unpause the thread.
+     * </li><li>
+     * While the rendering thread is paused, rendering commands will continue to be queued until the
+     * buffer limit is reached. When the limit is reached, the program will abort.
+     * </li></ul>
+     */
+    public void setPaused(boolean paused) {
+        nSetPaused(getNativeObject(), paused);
+    }
 
     @UsedByReflection("TextureHelper.java")
     public long getNativeObject() {
@@ -1263,6 +1299,7 @@ public class Engine {
     private static native void nDestroyEntity(long nativeEngine, int entity);
     private static native void nFlushAndWait(long nativeEngine);
     private static native void nFlush(long nativeEngine);
+    private static native void nSetPaused(long nativeEngine, boolean paused);
     private static native long nGetTransformManager(long nativeEngine);
     private static native long nGetLightManager(long nativeEngine);
     private static native long nGetRenderableManager(long nativeEngine);
@@ -1283,8 +1320,10 @@ public class Engine {
             long minCommandBufferSizeMB, long perFrameCommandsSizeMB, long jobSystemThreadCount,
             long textureUseAfterFreePoolSize, boolean disableParallelShaderCompile,
             int stereoscopicType, long stereoscopicEyeCount,
-            long resourceAllocatorCacheSizeMB, long resourceAllocatorCacheMaxAge);
+            long resourceAllocatorCacheSizeMB, long resourceAllocatorCacheMaxAge,
+            boolean disableHandleUseAfterFreeCheck);
     private static native void nSetBuilderFeatureLevel(long nativeBuilder, int ordinal);
     private static native void nSetBuilderSharedContext(long nativeBuilder, long sharedContext);
+    private static native void nSetBuilderPaused(long nativeBuilder, boolean paused);
     private static native long nBuilderBuild(long nativeBuilder);
 }

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

@@ -175,6 +175,32 @@ public class RenderableManager {
             return this;
         }
 
+        /**
+         * Type of geometry for a Renderable
+         */
+        public enum GeometryType {
+            /** dynamic gemoetry has no restriction */
+            DYNAMIC,
+            /** bounds and world space transform are immutable */
+            STATIC_BOUNDS,
+            /** skinning/morphing not allowed and Vertex/IndexBuffer immutables */
+            STATIC
+        }
+
+        /**
+         * Specify whether this renderable has static bounds. In this context his means that
+         * the renderable's bounding box cannot change and that the renderable's transform is
+         * assumed immutable. Changing the renderable's transform via the TransformManager
+         * can lead to corrupted graphics. Note that skinning and morphing are not forbidden.
+         * Disabled by default.
+         * @param enable whether this renderable has static bounds. false by default.
+         */
+        @NonNull
+        public Builder geometryType(GeometryType type) {
+            nBuilderGeometryType(mNativeBuilder, type.ordinal());
+            return this;
+        }
+
         /**
          * Binds a material instance to the specified primitive.
          *
@@ -964,6 +990,7 @@ public class RenderableManager {
     private static native void nBuilderGeometry(long nativeBuilder, int index, int value, long nativeVertexBuffer, long nativeIndexBuffer);
     private static native void nBuilderGeometry(long nativeBuilder, int index, int value, long nativeVertexBuffer, long nativeIndexBuffer, int offset, int count);
     private static native void nBuilderGeometry(long nativeBuilder, int index, int value, long nativeVertexBuffer, long nativeIndexBuffer, int offset, int minIndex, int maxIndex, int count);
+    private static native void nBuilderGeometryType(long nativeBuilder, int type);
     private static native void nBuilderMaterial(long nativeBuilder, int index, long nativeMaterialInstance);
     private static native void nBuilderBlendOrder(long nativeBuilder, int index, int blendOrder);
     private static native void nBuilderGlobalBlendOrderEnabled(long nativeBuilder, int index, boolean enabled);

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

@@ -68,77 +68,30 @@ public class SwapChain {
     private final Object mSurface;
     private long mNativeObject;
 
-    public static final long CONFIG_DEFAULT = 0x0;
-
-    /**
-     * This flag indicates that the <code>SwapChain</code> must be allocated with an
-     * alpha-channel.
-     */
-    public static final long CONFIG_TRANSPARENT = 0x1;
-
-    /**
-     * This flag indicates that the <code>SwapChain</code> may be used as a source surface
-     * for reading back render results.  This config must be set when creating
-     * any <code>SwapChain</code>  that will be used as the source for a blit operation.
-     *
-     * @see Renderer#copyFrame
-     */
-    public static final long CONFIG_READABLE = 0x2;
-
-    /**
-     * Indicates that the native X11 window is an XCB window rather than an XLIB window.
-     * This is ignored on non-Linux platforms and in builds that support only one X11 API.
-     */
-    public static final long CONFIG_ENABLE_XCB = 0x4;
-
-    /**
-     * Indicates that the SwapChain must automatically perform linear to sRGB encoding.
-     *
-     * This flag is ignored if isSRGBSwapChainSupported() is false.
-     *
-     * When using this flag, post-processing should be disabled.
-     *
-     * @see SwapChain#isSRGBSwapChainSupported
-     * @see View#setPostProcessingEnabled
-     */
-    public static final long CONFIG_SRGB_COLORSPACE = 0x10;
-
-    /**
-     * Indicates that this SwapChain should allocate a stencil buffer in addition to a depth buffer.
-     *
-     * This flag is necessary when using View::setStencilBufferEnabled and rendering directly into
-     * the SwapChain (when post-processing is disabled).
-     *
-     * The specific format of the stencil buffer depends on platform support. The following pixel
-     * formats are tried, in order of preference:
-     *
-     * Depth only (without CONFIG_HAS_STENCIL_BUFFER):
-     * - DEPTH32F
-     * - DEPTH24
-     *
-     * Depth + stencil (with CONFIG_HAS_STENCIL_BUFFER):
-     * - DEPTH32F_STENCIL8
-     * - DEPTH24F_STENCIL8
-     *
-     * Note that enabling the stencil buffer may hinder depth precision and should only be used if
-     * necessary.
-     *
-     * @see View#setStencilBufferEnabled
-     * @see View#setPostProcessingEnabled
-     */
-    public static final long CONFIG_HAS_STENCIL_BUFFER = 0x20;
-
     SwapChain(long nativeSwapChain, Object surface) {
         mNativeObject = nativeSwapChain;
         mSurface = surface;
     }
 
     /**
-     * Return whether createSwapChain supports the SWAP_CHAIN_CONFIG_SRGB_COLORSPACE flag.
+     * Return whether createSwapChain supports the CONFIG_PROTECTED_CONTENT flag.
      * The default implementation returns false.
      *
      * @param engine A reference to the filament Engine
-     * @return true if SWAP_CHAIN_CONFIG_SRGB_COLORSPACE is supported, false otherwise.
+     * @return true if CONFIG_PROTECTED_CONTENT is supported, false otherwise.
+     * @see SwapChainFlags#CONFIG_PROTECTED_CONTENT
+     */
+    public static boolean isProtectedContentSupported(@NonNull Engine engine) {
+        return nIsProtectedContentSupported(engine.getNativeObject());
+    }
+
+    /**
+     * Return whether createSwapChain supports the CONFIG_SRGB_COLORSPACE flag.
+     * The default implementation returns false.
+     *
+     * @param engine A reference to the filament Engine
+     * @return true if CONFIG_SRGB_COLORSPACE is supported, false otherwise.
+     * @see SwapChainFlags#CONFIG_SRGB_COLORSPACE
      */
     public static boolean isSRGBSwapChainSupported(@NonNull Engine engine) {
         return nIsSRGBSwapChainSupported(engine.getNativeObject());
@@ -186,4 +139,5 @@ public class SwapChain {
 
     private static native void nSetFrameCompletedCallback(long nativeSwapChain, Object handler, Runnable callback);
     private static native boolean nIsSRGBSwapChainSupported(long nativeEngine);
+    private static native boolean nIsProtectedContentSupported(long nativeEngine);
 }

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

@@ -0,0 +1,97 @@
+/*
+ * Copyright (C) 2024 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.
+ */
+
+package com.google.android.filament;
+
+// Note: SwapChainFlags is kept separate from SwapChain so that UiHelper does not need to depend
+// on SwapChain. This allows clients to use UiHelper without requiring all of Filament's Java
+// classes.
+
+/**
+ * Flags that a <code>SwapChain</code> can be created with to control behavior.
+ *
+ * @see Engine#createSwapChain
+ * @see Engine#createSwapChainFromNativeSurface
+ */
+public final class SwapChainFlags {
+
+    public static final long CONFIG_DEFAULT = 0x0;
+
+    /**
+     * This flag indicates that the <code>SwapChain</code> must be allocated with an
+     * alpha-channel.
+     */
+    public static final long CONFIG_TRANSPARENT = 0x1;
+
+    /**
+     * This flag indicates that the <code>SwapChain</code> may be used as a source surface
+     * for reading back render results.  This config must be set when creating
+     * any <code>SwapChain</code>  that will be used as the source for a blit operation.
+     *
+     * @see Renderer#copyFrame
+     */
+    public static final long CONFIG_READABLE = 0x2;
+
+    /**
+     * Indicates that the native X11 window is an XCB window rather than an XLIB window.
+     * This is ignored on non-Linux platforms and in builds that support only one X11 API.
+     */
+    public static final long CONFIG_ENABLE_XCB = 0x4;
+
+    /**
+     * Indicates that the SwapChain must automatically perform linear to sRGB encoding.
+     *
+     * This flag is ignored if isSRGBSwapChainSupported() is false.
+     *
+     * When using this flag, post-processing should be disabled.
+     *
+     * @see SwapChain#isSRGBSwapChainSupported
+     * @see View#setPostProcessingEnabled
+     */
+    public static final long CONFIG_SRGB_COLORSPACE = 0x10;
+
+    /**
+     * Indicates that this SwapChain should allocate a stencil buffer in addition to a depth buffer.
+     *
+     * This flag is necessary when using View::setStencilBufferEnabled and rendering directly into
+     * the SwapChain (when post-processing is disabled).
+     *
+     * The specific format of the stencil buffer depends on platform support. The following pixel
+     * formats are tried, in order of preference:
+     *
+     * Depth only (without CONFIG_HAS_STENCIL_BUFFER):
+     * - DEPTH32F
+     * - DEPTH24
+     *
+     * Depth + stencil (with CONFIG_HAS_STENCIL_BUFFER):
+     * - DEPTH32F_STENCIL8
+     * - DEPTH24F_STENCIL8
+     *
+     * Note that enabling the stencil buffer may hinder depth precision and should only be used if
+     * necessary.
+     *
+     * @see View#setStencilBufferEnabled
+     * @see View#setPostProcessingEnabled
+     */
+    public static final long CONFIG_HAS_STENCIL_BUFFER = 0x20;
+
+    /**
+     * The SwapChain contains protected content. Only supported when isProtectedContentSupported()
+     * is true.
+     */
+    public static final long CONFIG_PROTECTED_CONTENT   = 0x40;
+}
+

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

@@ -27,7 +27,7 @@ import android.view.SurfaceHolder;
 import android.view.SurfaceView;
 import android.view.TextureView;
 
-import com.google.android.filament.SwapChain;
+import com.google.android.filament.SwapChainFlags;
 
 /**
  * UiHelper is a simple class that can manage either a SurfaceView, TextureView, or a SurfaceHolder
@@ -538,7 +538,7 @@ public class UiHelper {
      * the options set on this UiHelper.
      */
     public long getSwapChainFlags() {
-        return isOpaque() ? SwapChain.CONFIG_DEFAULT : SwapChain.CONFIG_TRANSPARENT;
+        return isOpaque() ? SwapChainFlags.CONFIG_DEFAULT : SwapChainFlags.CONFIG_TRANSPARENT;
     }
 
     /**

android/gltfio-android/CMakeLists.txt

@@ -81,9 +81,17 @@ set(GLTFIO_SRCS
         ${GLTFIO_DIR}/src/TangentsJob.cpp
         ${GLTFIO_DIR}/src/TangentsJob.h
         ${GLTFIO_DIR}/src/UbershaderProvider.cpp
+        ${GLTFIO_DIR}/src/Utility.cpp
+        ${GLTFIO_DIR}/src/Utility.h
         ${GLTFIO_DIR}/src/Wireframe.cpp
         ${GLTFIO_DIR}/src/Wireframe.h
         ${GLTFIO_DIR}/src/downcast.h
+        ${GLTFIO_DIR}/src/extended/AssetLoaderExtended.h
+        ${GLTFIO_DIR}/src/extended/TangentsJobExtended.cpp
+        ${GLTFIO_DIR}/src/extended/TangentsJobExtended.h
+        ${GLTFIO_DIR}/src/extended/TangentSpaceMeshWrapper.cpp
+        ${GLTFIO_DIR}/src/extended/TangentSpaceMeshWrapper.h
+
 
         src/main/cpp/Animator.cpp
         src/main/cpp/AssetLoader.cpp

android/gradle.properties

@@ -1,5 +1,5 @@
 GROUP=com.google.android.filament
-VERSION_NAME=1.50.5
+VERSION_NAME=1.51.0
 
 POM_DESCRIPTION=Real-time physically based rendering engine for Android.
 

build.sh

@@ -516,7 +516,7 @@ function build_android {
             if [[ "${BUILD_ANDROID_SAMPLES}" == "true" ]]; then
                 for sample in ${ANDROID_SAMPLES}; do
                     echo "Installing out/${sample}-debug.apk"
-                    cp samples/${sample}/build/outputs/apk/debug/${sample}-debug-unsigned.apk \
+                    cp samples/${sample}/build/outputs/apk/debug/${sample}-debug.apk \
                         ../out/${sample}-debug.apk
                 done
             fi

filament/CMakeLists.txt

@@ -252,8 +252,13 @@ set(MATERIAL_SRCS
 )
 
 set(MATERIAL_FL0_SRCS
-        src/materials/defaultMaterial0.mat
-        src/materials/skybox0.mat
+        src/materials/defaultMaterial.mat
+        src/materials/skybox.mat
+)
+
+set(MATERIAL_MULTIVIEW_SRCS
+        src/materials/defaultMaterial.mat
+        src/materials/skybox.mat
 )
 
 # Embed the binary resource blob for materials.
@@ -285,6 +290,11 @@ if (FILAMENT_ENABLE_FEATURE_LEVEL_0)
     add_definitions(-DFILAMENT_ENABLE_FEATURE_LEVEL_0)
 endif()
 
+# Whether to include MULTIVIEW materials.
+if (FILAMENT_ENABLE_MULTIVIEW)
+    add_definitions(-DFILAMENT_ENABLE_MULTIVIEW)
+endif()
+
 # ==================================================================================================
 # Definitions
 # ==================================================================================================
@@ -315,33 +325,42 @@ foreach (mat_src ${MATERIAL_SRCS})
     get_filename_component(localname "${mat_src}" NAME_WE)
     get_filename_component(fullname "${mat_src}" ABSOLUTE)
     set(output_path "${MATERIAL_DIR}/${localname}.filamat")
-
     add_custom_command(
             OUTPUT ${output_path}
             COMMAND matc ${MATC_BASE_FLAGS} -o ${output_path} ${fullname}
             MAIN_DEPENDENCY ${fullname}
             DEPENDS matc
-            COMMENT "Compiling material ${mat_src} to ${output_path}"
+            COMMENT "Compiling material ${fullname} to ${output_path}"
     )
     list(APPEND MATERIAL_BINS ${output_path})
-endforeach()
-
-if (FILAMENT_ENABLE_FEATURE_LEVEL_0)
-    foreach (mat_src ${MATERIAL_FL0_SRCS})
-        get_filename_component(localname "${mat_src}" NAME_WE)
-        get_filename_component(fullname "${mat_src}" ABSOLUTE)
-        set(output_path "${MATERIAL_DIR}/${localname}.filamat")
 
+    list(FIND MATERIAL_FL0_SRCS ${mat_src} index)
+    if (${index} GREATER -1 AND FILAMENT_ENABLE_FEATURE_LEVEL_0)
+        string(REGEX REPLACE "[.]filamat$" "_fl0.filamat" output_path_fl0 ${output_path})
         add_custom_command(
-                OUTPUT ${output_path}
-                COMMAND matc ${MATC_BASE_FLAGS} -o ${output_path} ${fullname}
+                OUTPUT ${output_path_fl0}
+                COMMAND matc ${MATC_BASE_FLAGS} -PfeatureLevel=0 -o ${output_path_fl0} ${fullname}
                 MAIN_DEPENDENCY ${fullname}
                 DEPENDS matc
-                COMMENT "Compiling material ${mat_src} to ${output_path}"
+                COMMENT "Compiling material ${fullname} to ${output_path_fl0}"
         )
-        list(APPEND MATERIAL_BINS ${output_path})
-    endforeach ()
-endif ()
+        list(APPEND MATERIAL_BINS ${output_path_fl0})
+    endif ()
+
+    list(FIND MATERIAL_MULTIVIEW_SRCS ${mat_src} index)
+    if (${index} GREATER -1 AND FILAMENT_ENABLE_MULTIVIEW)
+        string(REGEX REPLACE "[.]filamat$" "_multiview.filamat" output_path_multiview ${output_path})
+        add_custom_command(
+                OUTPUT ${output_path_multiview}
+                COMMAND matc ${MATC_BASE_FLAGS} -PstereoscopicType=multiview -o ${output_path_multiview} ${fullname}
+                MAIN_DEPENDENCY ${fullname}
+                DEPENDS matc
+                COMMENT "Compiling material ${fullname} to ${output_path_multiview}"
+        )
+        list(APPEND MATERIAL_BINS ${output_path_multiview})
+    endif ()
+
+endforeach()
 
 # Additional dependencies on included files for materials
 

filament/backend/CMakeLists.txt

@@ -166,6 +166,8 @@ endif()
 if (FILAMENT_SUPPORTS_VULKAN)
     list(APPEND SRCS
             include/backend/platforms/VulkanPlatform.h
+            src/vulkan/caching/VulkanDescriptorSet.cpp
+            src/vulkan/caching/VulkanDescriptorSet.h
             src/vulkan/platform/VulkanPlatform.cpp
             src/vulkan/platform/VulkanPlatformSwapChainImpl.cpp
             src/vulkan/platform/VulkanPlatformSwapChainImpl.h

filament/backend/include/backend/DriverEnums.h

@@ -81,7 +81,14 @@ static constexpr uint64_t SWAP_CHAIN_CONFIG_SRGB_COLORSPACE     = 0x10;
 /**
  * Indicates that the SwapChain should also contain a stencil component.
  */
-static constexpr uint64_t SWAP_CHAIN_HAS_STENCIL_BUFFER         = 0x20;
+static constexpr uint64_t SWAP_CHAIN_CONFIG_HAS_STENCIL_BUFFER  = 0x20;
+static constexpr uint64_t SWAP_CHAIN_HAS_STENCIL_BUFFER         = SWAP_CHAIN_CONFIG_HAS_STENCIL_BUFFER;
+
+/**
+ * The SwapChain contains protected content. Currently only supported by OpenGLPlatform and
+ * only when OpenGLPlatform::isProtectedContextSupported() is true.
+ */
+static constexpr uint64_t SWAP_CHAIN_CONFIG_PROTECTED_CONTENT   = 0x40;
 
 
 static constexpr size_t MAX_VERTEX_ATTRIBUTE_COUNT  = 16;   // This is guaranteed by OpenGL ES.
@@ -129,6 +136,12 @@ enum class Backend : uint8_t {
     NOOP = 4,     //!< Selects the no-op driver for testing purposes.
 };
 
+enum class TimerQueryResult : int8_t {
+    ERROR = -1,     // an error occurred, result won't be available
+    NOT_READY = 0,  // result to ready yet
+    AVAILABLE = 1,  // result is available
+};
+
 static constexpr const char* backendToString(Backend backend) {
     switch (backend) {
         case Backend::NOOP:
@@ -659,16 +672,17 @@ enum class TextureFormat : uint16_t {
 };
 
 //! Bitmask describing the intended Texture Usage
-enum class TextureUsage : uint8_t {
-    NONE                = 0x00,
-    COLOR_ATTACHMENT    = 0x01,                     //!< Texture can be used as a color attachment
-    DEPTH_ATTACHMENT    = 0x02,                     //!< Texture can be used as a depth attachment
-    STENCIL_ATTACHMENT  = 0x04,                     //!< Texture can be used as a stencil attachment
-    UPLOADABLE          = 0x08,                     //!< Data can be uploaded into this texture (default)
-    SAMPLEABLE          = 0x10,                     //!< Texture can be sampled (default)
-    SUBPASS_INPUT       = 0x20,                     //!< Texture can be used as a subpass input
-    BLIT_SRC            = 0x40,                     //!< Texture can be used the source of a blit()
-    BLIT_DST            = 0x80,                     //!< Texture can be used the destination of a blit()
+enum class TextureUsage : uint16_t {
+    NONE                = 0x0000,
+    COLOR_ATTACHMENT    = 0x0001,            //!< Texture can be used as a color attachment
+    DEPTH_ATTACHMENT    = 0x0002,            //!< Texture can be used as a depth attachment
+    STENCIL_ATTACHMENT  = 0x0004,            //!< Texture can be used as a stencil attachment
+    UPLOADABLE          = 0x0008,            //!< Data can be uploaded into this texture (default)
+    SAMPLEABLE          = 0x0010,            //!< Texture can be sampled (default)
+    SUBPASS_INPUT       = 0x0020,            //!< Texture can be used as a subpass input
+    BLIT_SRC            = 0x0040,            //!< Texture can be used the source of a blit()
+    BLIT_DST            = 0x0080,            //!< Texture can be used the destination of a blit()
+    PROTECTED           = 0x0100,            //!< Texture can be used the destination of a blit()
     DEFAULT             = UPLOADABLE | SAMPLEABLE   //!< Default texture usage
 };
 
@@ -1172,11 +1186,27 @@ struct StencilState {
 
     //! Stencil operations for front-facing polygons
     StencilOperations front = {
-            .stencilFunc = StencilFunction::A, .ref = 0, .readMask = 0xff, .writeMask = 0xff };
+            .stencilFunc = StencilFunction::A,
+            .stencilOpStencilFail = StencilOperation::KEEP,
+            .padding0 = 0,
+            .stencilOpDepthFail = StencilOperation::KEEP,
+            .stencilOpDepthStencilPass = StencilOperation::KEEP,
+            .padding1 = 0,
+            .ref = 0,
+            .readMask = 0xff,
+            .writeMask = 0xff };
 
     //! Stencil operations for back-facing polygons
     StencilOperations back  = {
-            .stencilFunc = StencilFunction::A, .ref = 0, .readMask = 0xff, .writeMask = 0xff };
+            .stencilFunc = StencilFunction::A,
+            .stencilOpStencilFail = StencilOperation::KEEP,
+            .padding0 = 0,
+            .stencilOpDepthFail = StencilOperation::KEEP,
+            .stencilOpDepthStencilPass = StencilOperation::KEEP,
+            .padding1 = 0,
+            .ref = 0,
+            .readMask = 0xff,
+            .writeMask = 0xff };
 
     //! Whether stencil-buffer writes are enabled
     bool stencilWrite = false;

filament/backend/include/backend/PipelineState.h

@@ -29,10 +29,13 @@ namespace filament::backend {
 //! \privatesection
 
 struct PipelineState {
-    Handle<HwProgram> program;
-    RasterState rasterState;
-    StencilState stencilState;
-    PolygonOffset polygonOffset;
+    Handle<HwProgram> program;                                              //  4
+    Handle<HwVertexBufferInfo> vertexBufferInfo;                            //  4
+    RasterState rasterState;                                                //  4
+    StencilState stencilState;                                              // 12
+    PolygonOffset polygonOffset;                                            //  8
+    PrimitiveType primitiveType = PrimitiveType::TRIANGLES;                 //  1
+    uint8_t padding[3] = {};                                                //  3
 };
 
 } // namespace filament::backend

filament/backend/include/backend/Platform.h

@@ -23,6 +23,7 @@
 #include <utils/Invocable.h>
 
 #include <stddef.h>
+#include <stdint.h>
 
 namespace filament::backend {
 
@@ -56,9 +57,14 @@ public:
 
         /**
          * Set to `true` to forcibly disable parallel shader compilation in the backend.
-         * Currently only honored by the GL backend.
+         * Currently only honored by the GL and Metal backends.
          */
         bool disableParallelShaderCompile = false;
+
+        /**
+         * Disable backend handles use-after-free checks.
+         */
+        bool disableHandleUseAfterFreeCheck = false;
     };
 
     Platform() noexcept;
@@ -188,9 +194,40 @@ public:
     size_t retrieveBlob(const void* UTILS_NONNULL key, size_t keySize,
             void* UTILS_NONNULL value, size_t valueSize);
 
+    using DebugUpdateStatFunc = utils::Invocable<void(const char* UTILS_NONNULL key, uint64_t value)>;
+
+    /**
+     * Sets the callback function that the backend can use to update backend-specific statistics
+     * to aid with debugging. This callback is guaranteed to be called on the Filament driver
+     * thread.
+     *
+     * @param debugUpdateStat   an Invocable that updates debug statistics
+     */
+    void setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noexcept;
+
+    /**
+     * @return true if debugUpdateStat is valid.
+     */
+    bool hasDebugUpdateStatFunc() const noexcept;
+
+    /**
+     * To track backend-specific statistics, the backend implementation can call the
+     * application-provided callback function debugUpdateStatFunc to associate or update a value
+     * with a given key. It is possible for this function to be called multiple times with the
+     * same key, in which case newer values should overwrite older values.
+     *
+     * This function is guaranteed to be called only on a single thread, the Filament driver
+     * thread.
+     *
+     * @param key          a null-terminated C-string with the key of the debug statistic
+     * @param value        the updated value of key
+     */
+    void debugUpdateStat(const char* UTILS_NONNULL key, uint64_t value);
+
 private:
     InsertBlobFunc mInsertBlob;
     RetrieveBlobFunc mRetrieveBlob;
+    DebugUpdateStatFunc mDebugUpdateStat;
 };
 
 } // namespace filament

filament/backend/include/backend/Program.h

@@ -116,6 +116,8 @@ public:
 
     Program& cacheId(uint64_t cacheId) noexcept;
 
+    Program& multiview(bool multiview) noexcept;
+
     ShaderSource const& getShadersSource() const noexcept { return mShadersSource; }
     ShaderSource& getShadersSource() noexcept { return mShadersSource; }
 
@@ -143,6 +145,8 @@ public:
 
     uint64_t getCacheId() const noexcept { return mCacheId; }
 
+    bool isMultiview() const noexcept { return mMultiview; }
+
     CompilerPriorityQueue getPriorityQueue() const noexcept { return mPriorityQueue; }
 
 private:
@@ -158,6 +162,11 @@ private:
     utils::FixedCapacityVector<std::pair<utils::CString, uint8_t>> mAttributes;
     std::array<UniformInfo, Program::UNIFORM_BINDING_COUNT> mBindingUniformInfo;
     CompilerPriorityQueue mPriorityQueue = CompilerPriorityQueue::HIGH;
+    // Indicates the current engine was initialized with multiview stereo, and the variant for this
+    // program contains STE flag. This will be referred later for the OpenGL shader compiler to
+    // determine whether shader code replacement for the num_views should be performed.
+    // This variable could be promoted as a more generic variable later if other similar needs occur.
+    bool mMultiview = false;
 };
 
 } // namespace filament::backend

filament/backend/include/backend/TargetBufferInfo.h

@@ -32,6 +32,10 @@ struct TargetBufferInfo {
     // texture to be used as render target
     Handle<HwTexture> handle;
 
+    // starting layer index for multiview. This value is only used when the `layerCount` for the
+    // render target is greater than 1.
+    uint8_t baseViewIndex = 0;
+
     // level to be used
     uint8_t level = 0;
 
@@ -80,7 +84,7 @@ public:
 
     // this is here for backward compatibility
     MRT(Handle<HwTexture> handle, uint8_t level, uint16_t layer) noexcept
-            : mInfos{{ handle, level, layer }} {
+            : mInfos{{ handle, 0, level, layer }} {
     }
 };
 

filament/backend/include/backend/platforms/OpenGLPlatform.h

@@ -22,7 +22,9 @@
 #include <backend/Platform.h>
 
 #include <utils/compiler.h>
+#include <utils/Invocable.h>
 
+#include <stddef.h>
 #include <stdint.h>
 
 namespace filament::backend {
@@ -62,6 +64,22 @@ public:
      */
     virtual void terminate() noexcept = 0;
 
+    /**
+     * Return whether createSwapChain supports the SWAP_CHAIN_CONFIG_SRGB_COLORSPACE flag.
+     * The default implementation returns false.
+     *
+     * @return true if SWAP_CHAIN_CONFIG_SRGB_COLORSPACE is supported, false otherwise.
+     */
+    virtual bool isSRGBSwapChainSupported() const noexcept;
+
+    /**
+     * Return whether protected contexts are supported by this backend.
+     * If protected context are supported, the SWAP_CHAIN_CONFIG_PROTECTED_CONTENT flag can be
+     * used when creating a SwapChain.
+     * The default implementation returns false.
+     */
+    virtual bool isProtectedContextSupported() const noexcept;
+
     /**
      * Called by the driver to create a SwapChain for this driver.
      *
@@ -71,17 +89,9 @@ public:
      * @return              The driver's SwapChain object.
      *
      */
-    virtual SwapChain* UTILS_NONNULL createSwapChain(
+    virtual SwapChain* UTILS_NULLABLE createSwapChain(
             void* UTILS_NULLABLE nativeWindow, uint64_t flags) noexcept = 0;
 
-    /**
-     * Return whether createSwapChain supports the SWAP_CHAIN_CONFIG_SRGB_COLORSPACE flag.
-     * The default implementation returns false.
-     *
-     * @return true if SWAP_CHAIN_CONFIG_SRGB_COLORSPACE is supported, false otherwise.
-     */
-    virtual bool isSRGBSwapChainSupported() const noexcept;
-
     /**
      * Called by the driver create a headless SwapChain.
      *
@@ -116,20 +126,70 @@ public:
     virtual TargetBufferFlags getPreservedFlags(SwapChain* UTILS_NONNULL swapChain) noexcept;
 
     /**
-     * Called by the driver to establish the default FBO. The default implementation returns 0.
-      * @return a GLuint casted to a uint32_t that is an OpenGL framebuffer object.
+     * Returns true if the swapchain is protected
      */
-    virtual uint32_t createDefaultRenderTarget() noexcept;
+    virtual bool isSwapChainProtected(Platform::SwapChain* UTILS_NONNULL swapChain) noexcept;
+
+    /**
+     * Called by the driver to establish the default FBO. The default implementation returns 0.
+     *
+     * This method can be called either on the regular or protected OpenGL contexts and can return
+     * a different or identical name, since these names exist in different namespaces.
+     *
+     * @return a GLuint casted to a uint32_t that is an OpenGL framebuffer object.
+     */
+    virtual uint32_t getDefaultFramebufferObject() noexcept;
+
+
+    /**
+     * Type of contexts available
+     */
+    enum class ContextType {
+        NONE,           //!< No current context
+        UNPROTECTED,    //!< current context is unprotected
+        PROTECTED       //!< current context supports protected content
+    };
+
+    /**
+     * Returns the type of the context currently in use. This value is updated by makeCurrent()
+     * and therefore can be cached between calls. ContextType::PROTECTED can only be returned
+     * if isProtectedContextSupported() is true.
+     * @return ContextType
+     */
+    virtual ContextType getCurrentContextType() const noexcept;
+
+    /**
+     * Binds the requested context to the current thread and drawSwapChain to the default FBO
+     * returned by getDefaultFramebufferObject().
+     *
+     * @param type type of context to bind to the current thread.
+     * @param drawSwapChain SwapChain to draw to. It must be bound to the default FBO.
+     * @param readSwapChain SwapChain to read from (for operation like `glBlitFramebuffer`)
+     * @return true on success, false on error.
+     */
+    virtual bool makeCurrent(ContextType type,
+            SwapChain* UTILS_NONNULL drawSwapChain,
+            SwapChain* UTILS_NONNULL readSwapChain) noexcept = 0;
 
     /**
      * Called by the driver to make the OpenGL context active on the calling thread and bind
-     * the drawSwapChain to the default render target (FBO) created with createDefaultRenderTarget.
+     * the drawSwapChain to the default FBO returned by getDefaultFramebufferObject().
+     * The context used is either the default context or the protected context. When a context
+     * change is necessary, the preContextChange and postContextChange callbacks are called,
+     * before and after the context change respectively. postContextChange is given the index
+     * of the new context (0 for default and 1 for protected).
+     * The default implementation just calls makeCurrent(getCurrentContextType(), SwapChain*, SwapChain*).
+     *
      * @param drawSwapChain SwapChain to draw to. It must be bound to the default FBO.
      * @param readSwapChain SwapChain to read from (for operation like `glBlitFramebuffer`)
+     * @param preContextChange called before the context changes
+     * @param postContextChange called after the context changes
      */
     virtual void makeCurrent(
             SwapChain* UTILS_NONNULL drawSwapChain,
-            SwapChain* UTILS_NONNULL readSwapChain) noexcept = 0;
+            SwapChain* UTILS_NONNULL readSwapChain,
+            utils::Invocable<void()> preContextChange,
+            utils::Invocable<void(size_t index)> postContextChange) noexcept;
 
     /**
      * Called by the driver once the current frame finishes drawing. Typically, this should present

filament/backend/include/backend/platforms/PlatformCocoaGL.h

@@ -57,7 +57,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
     void commit(SwapChain* swapChain) noexcept override;
     OpenGLPlatform::ExternalTexture* createExternalImageTexture() noexcept override;
     void destroyExternalImage(ExternalTexture* texture) noexcept override;

filament/backend/include/backend/platforms/PlatformCocoaTouchGL.h

@@ -45,7 +45,7 @@ public:
 
     void terminate() noexcept override;
 
-    uint32_t createDefaultRenderTarget() noexcept override;
+    uint32_t getDefaultFramebufferObject() noexcept override;
 
     bool isExtraContextSupported() const noexcept override;
     void createContext(bool shared) override;
@@ -53,7 +53,7 @@ public:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
     void commit(SwapChain* swapChain) noexcept override;
 
     OpenGLPlatform::ExternalTexture* createExternalImageTexture() noexcept override;

filament/backend/include/backend/platforms/PlatformEGL.h

@@ -25,6 +25,8 @@
 #include <EGL/eglext.h>
 #include <EGL/eglplatform.h>
 
+#include <utils/Invocable.h>
+
 #include <initializer_list>
 #include <utility>
 #include <vector>
@@ -41,15 +43,11 @@ class PlatformEGL : public OpenGLPlatform {
 public:
 
     PlatformEGL() noexcept;
-    bool isExtraContextSupported() const noexcept override;
-    void createContext(bool shared) override;
-    void releaseContext() noexcept override;
 
     // Return true if we're on an OpenGL platform (as opposed to OpenGL ES). false by default.
     virtual bool isOpenGL() const noexcept;
 
 protected:
-
     // --------------------------------------------------------------------------------------------
     // Helper for EGL configs and attributes parameters
 
@@ -89,13 +87,30 @@ protected:
     // --------------------------------------------------------------------------------------------
     // OpenGLPlatform Interface
 
+    bool isExtraContextSupported() const noexcept override;
+    void createContext(bool shared) override;
+    void releaseContext() noexcept override;
+
     void terminate() noexcept override;
 
+    bool isProtectedContextSupported() const noexcept override;
+
     bool isSRGBSwapChainSupported() const noexcept override;
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool isSwapChainProtected(SwapChain* swapChain) noexcept override;
+
+    ContextType getCurrentContextType() const noexcept override;
+
+    bool makeCurrent(ContextType type,
+            SwapChain* drawSwapChain,
+            SwapChain* readSwapChain) noexcept override;
+
+    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain,
+            utils::Invocable<void()> preContextChange,
+            utils::Invocable<void(size_t index)> postContextChange) noexcept override;
+
     void commit(SwapChain* swapChain) noexcept override;
 
     bool canCreateFence() noexcept override;
@@ -122,16 +137,27 @@ protected:
     static void clearGlError() noexcept;
 
     /**
-     * Always use this instead of eglMakeCurrent().
+     * Always use this instead of eglMakeCurrent(), as it tracks some state.
      */
-    EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept;
+
+    EGLContext getContextForType(ContextType type) const noexcept;
+
+    // makes the draw and read surface current without changing the current context
+    EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept {
+        return egl.makeCurrent(drawSurface, readSurface);
+    }
+
+    // makes context current and set draw and read surfaces to EGL_NO_SURFACE
+    EGLBoolean makeCurrent(EGLContext context) noexcept {
+        return egl.makeCurrent(context, mEGLDummySurface, mEGLDummySurface);
+    }
 
     // TODO: this should probably use getters instead.
     EGLDisplay mEGLDisplay = EGL_NO_DISPLAY;
     EGLContext mEGLContext = EGL_NO_CONTEXT;
-    EGLSurface mCurrentDrawSurface = EGL_NO_SURFACE;
-    EGLSurface mCurrentReadSurface = EGL_NO_SURFACE;
+    EGLContext mEGLContextProtected = EGL_NO_CONTEXT;
     EGLSurface mEGLDummySurface = EGL_NO_SURFACE;
+    ContextType mCurrentContextType = ContextType::NONE;
     // mEGLConfig is valid only if ext.egl.KHR_no_config_context is false
     EGLConfig mEGLConfig = EGL_NO_CONFIG_KHR;
     Config mContextAttribs;
@@ -148,6 +174,7 @@ protected:
             bool KHR_gl_colorspace = false;
             bool KHR_no_config_context = false;
             bool KHR_surfaceless_context = false;
+            bool EXT_protected_content = false;
         } egl;
     } ext;
 
@@ -156,12 +183,29 @@ protected:
         Config attribs{};
         EGLNativeWindowType nativeWindow{};
         EGLConfig config{};
+        uint64_t flags{};
     };
 
     void initializeGlExtensions() noexcept;
 
 protected:
     EGLConfig findSwapChainConfig(uint64_t flags, bool window, bool pbuffer) const;
+
+private:
+    class EGL {
+        EGLDisplay& mEGLDisplay;
+        EGLSurface mCurrentDrawSurface = EGL_NO_SURFACE;
+        EGLSurface mCurrentReadSurface = EGL_NO_SURFACE;
+        EGLContext mCurrentContext = EGL_NO_CONTEXT;
+    public:
+        explicit EGL(EGLDisplay& dpy) : mEGLDisplay(dpy) {}
+        EGLBoolean makeCurrent(EGLContext context,
+                EGLSurface drawSurface, EGLSurface readSurface) noexcept;
+
+        EGLBoolean makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept {
+            return makeCurrent(mCurrentContext, drawSurface, readSurface);
+        }
+    } egl{ mEGLDisplay };
 };
 
 } // namespace filament::backend

filament/backend/include/backend/platforms/PlatformGLX.h

@@ -51,7 +51,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
     void commit(SwapChain* swapChain) noexcept override;
 
 private:

filament/backend/include/backend/platforms/PlatformWGL.h

@@ -53,7 +53,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
     void commit(SwapChain* swapChain) noexcept override;
 
 protected:

filament/backend/include/backend/platforms/PlatformWebGL.h

@@ -46,7 +46,7 @@ protected:
     SwapChain* createSwapChain(void* nativewindow, uint64_t flags) noexcept override;
     SwapChain* createSwapChain(uint32_t width, uint32_t height, uint64_t flags) noexcept override;
     void destroySwapChain(SwapChain* swapChain) noexcept override;
-    void makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
+    bool makeCurrent(ContextType type, SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept override;
     void commit(SwapChain* swapChain) noexcept override;
 };
 

filament/backend/include/private/backend/CommandBufferQueue.h

@@ -50,12 +50,13 @@ class CommandBufferQueue {
     size_t mFreeSpace = 0;
     size_t mHighWatermark = 0;
     uint32_t mExitRequested = 0;
+    bool mPaused = false;
 
     static constexpr uint32_t EXIT_REQUESTED = 0x31415926;
 
 public:
     // requiredSize: guaranteed available space after flush()
-    CommandBufferQueue(size_t requiredSize, size_t bufferSize);
+    CommandBufferQueue(size_t requiredSize, size_t bufferSize, bool paused);
     ~CommandBufferQueue();
 
     CircularBuffer& getCircularBuffer() noexcept { return mCircularBuffer; }
@@ -80,6 +81,9 @@ public:
     // returns from waitForCommands() immediately.
     void requestExit();
 
+    // suspend or unsuspend the queue.
+    void setPaused(bool paused);
+
     bool isExitRequested() const;
 };
 

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

@@ -239,6 +239,7 @@ DECL_DRIVER_API_R_N(backend::RenderTargetHandle, createRenderTarget,
         uint32_t, width,
         uint32_t, height,
         uint8_t, samples,
+        uint8_t, layerCount,
         backend::MRT, color,
         backend::TargetBufferInfo, depth,
         backend::TargetBufferInfo, stencil)
@@ -298,14 +299,16 @@ DECL_DRIVER_API_SYNCHRONOUS_0(bool, isFrameBufferFetchMultiSampleSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isFrameTimeSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isAutoDepthResolveSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isSRGBSwapChainSupported)
+DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedContentSupported)
 DECL_DRIVER_API_SYNCHRONOUS_N(bool, isStereoSupported, backend::StereoscopicType, stereoscopicType)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isParallelShaderCompileSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(bool, isDepthStencilResolveSupported)
+DECL_DRIVER_API_SYNCHRONOUS_0(bool, isProtectedTexturesSupported)
 DECL_DRIVER_API_SYNCHRONOUS_0(uint8_t, getMaxDrawBuffers)
 DECL_DRIVER_API_SYNCHRONOUS_0(size_t, getMaxUniformBufferSize)
 DECL_DRIVER_API_SYNCHRONOUS_0(math::float2, getClipSpaceParams)
 DECL_DRIVER_API_SYNCHRONOUS_N(void, setupExternalImage, void*, image)
-DECL_DRIVER_API_SYNCHRONOUS_N(bool, getTimerQueryValue, backend::TimerQueryHandle, query, uint64_t*, elapsedTime)
+DECL_DRIVER_API_SYNCHRONOUS_N(backend::TimerQueryResult, getTimerQueryValue, backend::TimerQueryHandle, query, uint64_t*, elapsedTime)
 DECL_DRIVER_API_SYNCHRONOUS_N(bool, isWorkaroundNeeded, backend::Workaround, workaround)
 DECL_DRIVER_API_SYNCHRONOUS_0(backend::FeatureLevel, getFeatureLevel)
 
@@ -491,6 +494,17 @@ DECL_DRIVER_API_N(blit,
         math::uint2, srcOrigin,
         math::uint2, size)
 
+DECL_DRIVER_API_N(bindPipeline,
+        backend::PipelineState, state)
+
+DECL_DRIVER_API_N(bindRenderPrimitive,
+        backend::RenderPrimitiveHandle, rph)
+
+DECL_DRIVER_API_N(draw2,
+        uint32_t, indexOffset,
+        uint32_t, indexCount,
+        uint32_t, instanceCount)
+
 DECL_DRIVER_API_N(draw,
         backend::PipelineState, state,
         backend::RenderPrimitiveHandle, rph,
@@ -498,6 +512,7 @@ DECL_DRIVER_API_N(draw,
         uint32_t, indexCount,
         uint32_t, instanceCount)
 
+
 DECL_DRIVER_API_N(dispatchCompute,
         backend::ProgramHandle, program,
         math::uint3, workGroupCount)

filament/backend/include/private/backend/HandleAllocator.h

@@ -37,9 +37,9 @@
 #include <stddef.h>
 #include <stdint.h>
 
-#define HandleAllocatorGL  HandleAllocator<24,  64, 136>    // ~3640 / pool / MiB
-#define HandleAllocatorVK  HandleAllocator<80, 176, 320>    // ~1820 / pool / MiB
-#define HandleAllocatorMTL HandleAllocator<48, 160, 592>    // ~1310 / pool / MiB
+#define HandleAllocatorGL  HandleAllocator<32,  64, 136>    // ~4520 / pool / MiB
+#define HandleAllocatorVK  HandleAllocator<64, 160, 312>    // ~1820 / pool / MiB
+#define HandleAllocatorMTL HandleAllocator<32,  48, 552>    // ~1660 / pool / MiB
 
 namespace filament::backend {
 
@@ -49,7 +49,7 @@ namespace filament::backend {
 template<size_t P0, size_t P1, size_t P2>
 class HandleAllocator {
 public:
-    HandleAllocator(const char* name, size_t size) noexcept;
+    HandleAllocator(const char* name, size_t size, bool disableUseAfterFreeCheck) noexcept;
     HandleAllocator(HandleAllocator const& rhs) = delete;
     HandleAllocator& operator=(HandleAllocator const& rhs) = delete;
     ~HandleAllocator();
@@ -169,11 +169,13 @@ public:
 
         if (isPoolHandle(handle.getId())) {
             // check for use after free
-            uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
-            auto const pNode = static_cast<typename Allocator::Node*>(p);
-            uint8_t const expectedAge = pNode[-1].age;
-            ASSERT_POSTCONDITION(expectedAge == age,
-                    "use-after-free of Handle with id=%d", handle.getId());
+            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
+                uint8_t const age = (tag & HANDLE_AGE_MASK) >> HANDLE_AGE_SHIFT;
+                auto const pNode = static_cast<typename Allocator::Node*>(p);
+                uint8_t const expectedAge = pNode[-1].age;
+                ASSERT_POSTCONDITION(expectedAge == age,
+                        "use-after-free of Handle with id=%d", handle.getId());
+            }
         }
 
         return static_cast<Dp>(p);
@@ -187,7 +189,6 @@ public:
         return handle_cast<Dp>(const_cast<Handle<B>&>(handle));
     }
 
-
 private:
 
     template<typename D>
@@ -210,8 +211,9 @@ private:
         Pool<P1> mPool1;
         Pool<P2> mPool2;
         UTILS_UNUSED_IN_RELEASE const utils::AreaPolicy::HeapArea& mArea;
+        bool mUseAfterFreeCheckDisabled;
     public:
-        explicit Allocator(const utils::AreaPolicy::HeapArea& area);
+        explicit Allocator(const utils::AreaPolicy::HeapArea& area, bool disableUseAfterFreeCheck);
 
         static constexpr size_t getAlignment() noexcept { return MIN_ALIGNMENT; }
 
@@ -237,8 +239,10 @@ private:
             // check for double-free
             Node* const pNode = static_cast<Node*>(p);
             uint8_t& expectedAge = pNode[-1].age;
-            ASSERT_POSTCONDITION(expectedAge == age,
-                    "double-free of Handle of size %d at %p", size, p);
+            if (UTILS_UNLIKELY(!mUseAfterFreeCheckDisabled)) {
+                ASSERT_POSTCONDITION(expectedAge == age,
+                        "double-free of Handle of size %d at %p", size, p);
+            }
             expectedAge = (expectedAge + 1) & 0xF; // fixme
 
             if (size <= mPool0.getSize()) { mPool0.free(p); return; }
@@ -348,6 +352,7 @@ private:
     mutable utils::Mutex mLock;
     tsl::robin_map<HandleBase::HandleId, void*> mOverflowMap;
     HandleBase::HandleId mId = 0;
+    bool mUseAfterFreeCheckDisabled = false;
 };
 
 } // namespace filament::backend

filament/backend/src/CommandBufferQueue.cpp

@@ -39,10 +39,11 @@ using namespace utils;
 
 namespace filament::backend {
 
-CommandBufferQueue::CommandBufferQueue(size_t requiredSize, size_t bufferSize)
+CommandBufferQueue::CommandBufferQueue(size_t requiredSize, size_t bufferSize, bool paused)
         : mRequiredSize((requiredSize + (CircularBuffer::getBlockSize() - 1u)) & ~(CircularBuffer::getBlockSize() -1u)),
           mCircularBuffer(bufferSize),
-          mFreeSpace(mCircularBuffer.size()) {
+          mFreeSpace(mCircularBuffer.size()),
+          mPaused(paused) {
     assert_invariant(mCircularBuffer.size() > requiredSize);
 }
 
@@ -56,6 +57,16 @@ void CommandBufferQueue::requestExit() {
     mCondition.notify_one();
 }
 
+void CommandBufferQueue::setPaused(bool paused) {
+    std::lock_guard<utils::Mutex> const lock(mLock);
+    if (paused) {
+        mPaused = true;
+    } else {
+        mPaused = false;
+        mCondition.notify_one();
+    }
+}
+
 bool CommandBufferQueue::isExitRequested() const {
     std::lock_guard<utils::Mutex> const lock(mLock);
     ASSERT_PRECONDITION( mExitRequested == 0 || mExitRequested == EXIT_REQUESTED,
@@ -115,6 +126,9 @@ void CommandBufferQueue::flush() noexcept {
 #endif
 
         SYSTRACE_NAME("waiting: CircularBuffer::flush()");
+        ASSERT_POSTCONDITION(!mPaused,
+                "CommandStream is full, but since the rendering thread is paused, "
+                "the buffer cannot flush and we will deadlock. Instead, abort.");
         mCondition.wait(lock, [this, requiredSize]() -> bool {
             // TODO: on macOS, we need to call pumpEvents from time to time
             return mFreeSpace >= requiredSize;
@@ -127,7 +141,7 @@ std::vector<CommandBufferQueue::Range> CommandBufferQueue::waitForCommands() con
         return std::move(mCommandBuffersToExecute);
     }
     std::unique_lock<utils::Mutex> lock(mLock);
-    while (mCommandBuffersToExecute.empty() && !mExitRequested) {
+    while ((mCommandBuffersToExecute.empty() || mPaused) && !mExitRequested) {
         mCondition.wait(lock);
     }
 

filament/backend/src/DriverBase.h

@@ -113,7 +113,9 @@ struct HwTexture : public HwBase {
     uint8_t levels : 4;  // This allows up to 15 levels (max texture size of 32768 x 32768)
     uint8_t samples : 4; // Sample count per pixel (should always be a power of 2)
     TextureFormat format{};
+    uint8_t reserved0 = 0;
     TextureUsage usage{};
+    uint16_t reserved1 = 0;
     HwStream* hwStream = nullptr;
 
     HwTexture() noexcept : levels{}, samples{} {}

filament/backend/src/HandleAllocator.cpp

@@ -39,8 +39,10 @@ using namespace utils;
 
 template <size_t P0, size_t P1, size_t P2>
 UTILS_NOINLINE
-HandleAllocator<P0, P1, P2>::Allocator::Allocator(AreaPolicy::HeapArea const& area)
-        : mArea(area) {
+HandleAllocator<P0, P1, P2>::Allocator::Allocator(AreaPolicy::HeapArea const& area,
+        bool disableUseAfterFreeCheck)
+        : mArea(area),
+          mUseAfterFreeCheckDisabled(disableUseAfterFreeCheck) {
 
     // The largest handle this allocator can generate currently depends on the architecture's
     // min alignment, typically 8 or 16 bytes.
@@ -74,8 +76,10 @@ HandleAllocator<P0, P1, P2>::Allocator::Allocator(AreaPolicy::HeapArea const& ar
 // ------------------------------------------------------------------------------------------------
 
 template <size_t P0, size_t P1, size_t P2>
-HandleAllocator<P0, P1, P2>::HandleAllocator(const char* name, size_t size) noexcept
-    : mHandleArena(name, size) {
+HandleAllocator<P0, P1, P2>::HandleAllocator(const char* name, size_t size,
+        bool disableUseAfterFreeCheck) noexcept
+    : mHandleArena(name, size, disableUseAfterFreeCheck),
+      mUseAfterFreeCheckDisabled(disableUseAfterFreeCheck) {
 }
 
 template <size_t P0, size_t P1, size_t P2>

filament/backend/src/Platform.cpp

@@ -53,4 +53,18 @@ size_t Platform::retrieveBlob(void const* key, size_t keySize, void* value, size
     return 0;
 }
 
+void Platform::setDebugUpdateStatFunc(DebugUpdateStatFunc&& debugUpdateStat) noexcept {
+    mDebugUpdateStat = std::move(debugUpdateStat);
+}
+
+bool Platform::hasDebugUpdateStatFunc() const noexcept {
+    return bool(mDebugUpdateStat);
+}
+
+void Platform::debugUpdateStat(const char* key, uint64_t value) {
+    if (mDebugUpdateStat) {
+        mDebugUpdateStat(key, value);
+    }
+}
+
 } // namespace filament::backend

filament/backend/src/Program.cpp

@@ -91,6 +91,11 @@ Program& Program::cacheId(uint64_t cacheId) noexcept {
     return *this;
 }
 
+Program& Program::multiview(bool multiview) noexcept {
+    mMultiview = multiview;
+    return *this;
+}
+
 io::ostream& operator<<(io::ostream& out, const Program& builder) {
     out << "Program{";
     builder.mLogger(out);

filament/backend/src/metal/MetalBuffer.h

@@ -18,7 +18,6 @@
 #define TNT_FILAMENT_DRIVER_METALBUFFER_H
 
 #include "MetalContext.h"
-#include "MetalBufferPool.h"
 
 #include <backend/DriverEnums.h>
 
@@ -28,9 +27,50 @@
 
 #include <utility>
 #include <memory>
+#include <atomic>
 
 namespace filament::backend {
 
+class TrackedMetalBuffer {
+public:
+    TrackedMetalBuffer() noexcept : mBuffer(nil) {}
+    TrackedMetalBuffer(id<MTLBuffer> buffer) noexcept : mBuffer(buffer) {
+        if (buffer) {
+            aliveBuffers++;
+        }
+    }
+    ~TrackedMetalBuffer() {
+        if (mBuffer) {
+            aliveBuffers--;
+        }
+    }
+
+    TrackedMetalBuffer(TrackedMetalBuffer&&) = delete;
+    TrackedMetalBuffer(TrackedMetalBuffer const&) = delete;
+    TrackedMetalBuffer& operator=(TrackedMetalBuffer const&) = delete;
+
+    TrackedMetalBuffer& operator=(TrackedMetalBuffer&& rhs) noexcept {
+        swap(rhs);
+        return *this;
+    }
+
+    id<MTLBuffer> get() const noexcept { return mBuffer; }
+    operator bool() const noexcept { return bool(mBuffer); }
+
+    static uint64_t getAliveBuffers() { return aliveBuffers; }
+
+private:
+    void swap(TrackedMetalBuffer& other) noexcept {
+        id<MTLBuffer> temp = mBuffer;
+        mBuffer = other.mBuffer;
+        other.mBuffer = temp;
+    }
+
+    id<MTLBuffer> mBuffer;
+
+    static std::atomic<uint64_t> aliveBuffers;
+};
+
 class MetalBuffer {
 public:
 
@@ -82,7 +122,7 @@ public:
 
 private:
 
-    id<MTLBuffer> mBuffer = nil;
+    TrackedMetalBuffer mBuffer;
     size_t mBufferSize = 0;
     void* mCpuBuffer = nullptr;
     MetalContext& mContext;
@@ -151,7 +191,7 @@ public:
             // finishes executing.
             mAuxBuffer = [mDevice newBufferWithLength:mSlotSizeBytes options:mBufferOptions];
             assert_invariant(mAuxBuffer);
-            return {mAuxBuffer, 0};
+            return {mAuxBuffer.get(), 0};
         }
         mCurrentSlot = (mCurrentSlot + 1) % mSlotCount;
         mOccupiedSlots->fetch_add(1, std::memory_order_relaxed);
@@ -180,9 +220,9 @@ public:
      */
     std::pair<id<MTLBuffer>, NSUInteger> getCurrentAllocation() const {
         if (UTILS_UNLIKELY(mAuxBuffer)) {
-            return { mAuxBuffer, 0 };
+            return { mAuxBuffer.get(), 0 };
         }
-        return { mBuffer, mCurrentSlot * mSlotSizeBytes };
+        return { mBuffer.get(), mCurrentSlot * mSlotSizeBytes };
     }
 
     bool canAccomodateLayout(MTLSizeAndAlign layout) const {
@@ -191,8 +231,8 @@ public:
 
 private:
     id<MTLDevice> mDevice;
-    id<MTLBuffer> mBuffer;
-    id<MTLBuffer> mAuxBuffer;
+    TrackedMetalBuffer mBuffer;
+    TrackedMetalBuffer mAuxBuffer;
 
     MTLResourceOptions mBufferOptions;
 

filament/backend/src/metal/MetalBuffer.mm

@@ -15,12 +15,15 @@
  */
 
 #include "MetalBuffer.h"
+#include "MetalBufferPool.h"
 
 #include "MetalContext.h"
 
 namespace filament {
 namespace backend {
 
+std::atomic<uint64_t> TrackedMetalBuffer::aliveBuffers = 0;
+
 MetalBuffer::MetalBuffer(MetalContext& context, BufferObjectBinding bindingType, BufferUsage usage,
         size_t size, bool forceGpuBuffer) : mBufferSize(size), mContext(context) {
     // If the buffer is less than 4K in size and is updated frequently, we don't use an explicit
@@ -61,7 +64,7 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
     // Acquire a staging buffer to hold the contents of this update.
     MetalBufferPool* bufferPool = mContext.bufferPool;
     const MetalBufferPoolEntry* const staging = bufferPool->acquireBuffer(size);
-    memcpy(staging->buffer.contents, src, size);
+    memcpy(staging->buffer.get().contents, src, size);
 
     // The blit below requires that byteOffset be a multiple of 4.
     ASSERT_PRECONDITION(!(byteOffset & 0x3u), "byteOffset must be a multiple of 4");
@@ -70,9 +73,9 @@ void MetalBuffer::copyIntoBuffer(void* src, size_t size, size_t byteOffset) {
     id<MTLCommandBuffer> cmdBuffer = getPendingCommandBuffer(&mContext);
     id<MTLBlitCommandEncoder> blitEncoder = [cmdBuffer blitCommandEncoder];
     blitEncoder.label = @"Buffer upload blit";
-    [blitEncoder copyFromBuffer:staging->buffer
+    [blitEncoder copyFromBuffer:staging->buffer.get()
                    sourceOffset:0
-                       toBuffer:mBuffer
+                       toBuffer:mBuffer.get()
               destinationOffset:byteOffset
                            size:size];
     [blitEncoder endEncoding];
@@ -93,7 +96,7 @@ id<MTLBuffer> MetalBuffer::getGpuBufferForDraw(id<MTLCommandBuffer> cmdBuffer) n
         return nil;
     }
     assert_invariant(mBuffer);
-    return mBuffer;
+    return mBuffer.get();
 }
 
 void MetalBuffer::bindBuffers(id<MTLCommandBuffer> cmdBuffer, id<MTLCommandEncoder> encoder,

filament/backend/src/metal/MetalBufferPool.h

@@ -19,6 +19,8 @@
 
 #include <Metal/Metal.h>
 
+#include "MetalBuffer.h"
+
 #include <map>
 #include <mutex>
 #include <unordered_set>
@@ -30,7 +32,7 @@ struct MetalContext;
 
 // Immutable POD representing a shared CPU-GPU buffer.
 struct MetalBufferPoolEntry {
-    id<MTLBuffer> buffer;
+    TrackedMetalBuffer buffer;
     size_t capacity;
     mutable uint64_t lastAccessed;
     mutable uint32_t referenceCount;

filament/backend/src/metal/MetalBufferPool.mm

@@ -45,12 +45,12 @@ MetalBufferPoolEntry const* MetalBufferPool::acquireBuffer(size_t numBytes) {
     id<MTLBuffer> buffer = [mContext.device newBufferWithLength:numBytes
                                                         options:MTLResourceStorageModeShared];
     ASSERT_POSTCONDITION(buffer, "Could not allocate Metal staging buffer of size %zu.", numBytes);
-    MetalBufferPoolEntry* stage = new MetalBufferPoolEntry({
+    MetalBufferPoolEntry* stage = new MetalBufferPoolEntry {
         .buffer = buffer,
         .capacity = numBytes,
         .lastAccessed = mCurrentFrame,
         .referenceCount = 1
-    });
+    };
     mUsedStages.insert(stage);
 
     return stage;

filament/backend/src/metal/MetalContext.h

@@ -99,6 +99,7 @@ struct MetalContext {
     std::array<BufferState, MAX_SSBO_COUNT> ssboState;
     CullModeStateTracker cullModeState;
     WindingStateTracker windingState;
+    Handle<HwRenderPrimitive> currentRenderPrimitive;
 
     // State caches.
     DepthStencilStateCache depthStencilStateCache;

filament/backend/src/metal/MetalDriver.mm

@@ -20,6 +20,7 @@
 #include "metal/MetalDriver.h"
 
 #include "MetalBlitter.h"
+#include "MetalBufferPool.h"
 #include "MetalContext.h"
 #include "MetalDriverFactory.h"
 #include "MetalEnums.h"
@@ -36,6 +37,7 @@
 
 #include <utils/Log.h>
 #include <utils/Panic.h>
+#include <utils/sstream.h>
 
 #include <algorithm>
 
@@ -48,24 +50,26 @@ Driver* MetalDriverFactory::create(MetalPlatform* const platform, const Platform
     // For reference on a 64-bits machine in Release mode:
     //    MetalTimerQuery              :  16       few
     //    HwStream                     :  24       few
+    //    MetalRenderPrimitive         :  24       many
+    //    MetalVertexBuffer            :  32       moderate
+    // -- less than or equal 32 bytes
     //    MetalIndexBuffer             :  40       moderate
     //    MetalFence                   :  48       few
     //    MetalBufferObject            :  48       many
     // -- less than or equal 48 bytes
     //    MetalSamplerGroup            : 112       few
-    //    MetalProgram                 : 144       moderate
+    //    MetalProgram                 : 152       moderate
     //    MetalTexture                 : 152       moderate
-    //    MetalVertexBuffer            : 152       moderate
-    // -- less than or equal 160 bytes
     //    MetalSwapChain               : 184       few
     //    MetalRenderTarget            : 272       few
-    //    MetalRenderPrimitive         : 584       many
-    // -- less than or equal to 592 bytes
+    //    MetalVertexBufferInfo        : 552       moderate
+    // -- less than or equal to 552 bytes
 
     utils::slog.d
            << "\nMetalSwapChain: " << sizeof(MetalSwapChain)
            << "\nMetalBufferObject: " << sizeof(MetalBufferObject)
            << "\nMetalVertexBuffer: " << sizeof(MetalVertexBuffer)
+           << "\nMetalVertexBufferInfo: " << sizeof(MetalVertexBufferInfo)
            << "\nMetalIndexBuffer: " << sizeof(MetalIndexBuffer)
            << "\nMetalSamplerGroup: " << sizeof(MetalSamplerGroup)
            << "\nMetalRenderPrimitive: " << sizeof(MetalRenderPrimitive)
@@ -96,7 +100,9 @@ Dispatcher MetalDriver::getDispatcher() const noexcept {
 MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig& driverConfig) noexcept
         : mPlatform(*platform),
           mContext(new MetalContext(driverConfig.textureUseAfterFreePoolSize)),
-          mHandleAllocator("Handles", driverConfig.handleArenaSize) {
+          mHandleAllocator("Handles",
+                  driverConfig.handleArenaSize,
+                  driverConfig.disableHandleUseAfterFreeCheck) {
     mContext->driver = this;
 
     mContext->device = mPlatform.createDevice();
@@ -178,7 +184,10 @@ MetalDriver::MetalDriver(MetalPlatform* platform, const Platform::DriverConfig&
         mContext->eventListener = [[MTLSharedEventListener alloc] initWithDispatchQueue:queue];
     }
 
-    mContext->shaderCompiler = new MetalShaderCompiler(mContext->device, *this);
+    const MetalShaderCompiler::Mode compilerMode = driverConfig.disableParallelShaderCompile
+            ? MetalShaderCompiler::Mode::SYNCHRONOUS
+            : MetalShaderCompiler::Mode::ASYNCHRONOUS;
+    mContext->shaderCompiler = new MetalShaderCompiler(mContext->device, *this, compilerMode);
     mContext->shaderCompiler->init();
 
 #if defined(FILAMENT_METAL_PROFILING)
@@ -207,6 +216,9 @@ void MetalDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
 #if defined(FILAMENT_METAL_PROFILING)
     os_signpost_interval_begin(mContext->log, mContext->signpostId, "Frame encoding", "%{public}d", frameId);
 #endif
+    if (mPlatform.hasDebugUpdateStatFunc()) {
+        mPlatform.debugUpdateStat("filament.metal.alive_buffers", TrackedMetalBuffer::getAliveBuffers());
+    }
 }
 
 void MetalDriver::setFrameScheduledCallback(Handle<HwSwapChain> sch,
@@ -371,7 +383,7 @@ void MetalDriver::createDefaultRenderTargetR(Handle<HwRenderTarget> rth, int dum
 
 void MetalDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         TargetBufferFlags targetBufferFlags, uint32_t width, uint32_t height,
-        uint8_t samples, MRT color,
+        uint8_t samples, uint8_t layerCount, MRT color,
         TargetBufferInfo depth, TargetBufferInfo stencil) {
     ASSERT_PRECONDITION(!isInRenderPass(mContext),
             "createRenderTarget must be called outside of a render pass.");
@@ -770,6 +782,11 @@ bool MetalDriver::isSRGBSwapChainSupported() {
     return false;
 }
 
+bool MetalDriver::isProtectedContentSupported() {
+    // the SWAP_CHAIN_CONFIG_PROTECTED_CONTENT flag is not supported
+    return false;
+}
+
 bool MetalDriver::isStereoSupported(backend::StereoscopicType stereoscopicType) {
     switch (stereoscopicType) {
     case backend::StereoscopicType::INSTANCED:
@@ -781,13 +798,17 @@ bool MetalDriver::isStereoSupported(backend::StereoscopicType stereoscopicType)
 }
 
 bool MetalDriver::isParallelShaderCompileSupported() {
-    return true;
+    return mContext->shaderCompiler->isParallelShaderCompileSupported();
 }
 
 bool MetalDriver::isDepthStencilResolveSupported() {
     return false;
 }
 
+bool MetalDriver::isProtectedTexturesSupported() {
+    return false;
+}
+
 bool MetalDriver::isWorkaroundNeeded(Workaround workaround) {
     switch (workaround) {
         case Workaround::SPLIT_EASU:
@@ -914,9 +935,10 @@ void MetalDriver::setExternalImagePlane(Handle<HwTexture> th, void* image, uint3
 void MetalDriver::setExternalStream(Handle<HwTexture> th, Handle<HwStream> sh) {
 }
 
-bool MetalDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapsedTime) {
+TimerQueryResult MetalDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapsedTime) {
     auto* tq = handle_cast<MetalTimerQuery>(tqh);
-    return mContext->timerQueryImpl->getQueryResult(tq, elapsedTime);
+    return mContext->timerQueryImpl->getQueryResult(tq, elapsedTime) ?
+           TimerQueryResult::AVAILABLE : TimerQueryResult::NOT_READY;
 }
 
 void MetalDriver::generateMipmaps(Handle<HwTexture> th) {
@@ -1592,11 +1614,13 @@ void MetalDriver::finalizeSamplerGroup(MetalSamplerGroup* samplerGroup) {
     }
 }
 
-void MetalDriver::draw(PipelineState ps, Handle<HwRenderPrimitive> rph,
-        uint32_t const indexOffset, uint32_t const indexCount, uint32_t const instanceCount) {
+void MetalDriver::bindPipeline(PipelineState ps) {
     ASSERT_PRECONDITION(mContext->currentRenderPassEncoder != nullptr,
-            "Attempted to draw without a valid command encoder.");
-    auto primitive = handle_cast<MetalRenderPrimitive>(rph);
+            "bindPipeline() without a valid command encoder.");
+
+    MetalVertexBufferInfo const* const vbi =
+            handle_cast<MetalVertexBufferInfo>(ps.vertexBufferInfo);
+
     auto program = handle_cast<MetalProgram>(ps.program);
     const auto& rs = ps.rasterState;
 
@@ -1610,7 +1634,7 @@ void MetalDriver::draw(PipelineState ps, Handle<HwRenderPrimitive> rph,
         return;
     }
 
-    functions.validate();
+    ASSERT_PRECONDITION(bool(functions), "Attempting to bind an invalid Metal program.");
 
     auto [fragment, vertex] = functions.getRasterFunctions();
 
@@ -1637,7 +1661,7 @@ void MetalDriver::draw(PipelineState ps, Handle<HwRenderPrimitive> rph,
     MetalPipelineState const pipelineState {
         .vertexFunction = vertex,
         .fragmentFunction = fragment,
-        .vertexDescription = primitive->vertexDescription,
+        .vertexDescription = vbi->vertexDescription,
         .colorAttachmentPixelFormat = {
             colorPixelFormat[0],
             colorPixelFormat[1],
@@ -1731,20 +1755,6 @@ void MetalDriver::draw(PipelineState ps, Handle<HwRenderPrimitive> rph,
         mContext->currentPolygonOffset = ps.polygonOffset;
     }
 
-    // Bind uniform buffers.
-    MetalBuffer* uniformsToBind[Program::UNIFORM_BINDING_COUNT] = { nil };
-    NSUInteger offsets[Program::UNIFORM_BINDING_COUNT] = { 0 };
-
-    enumerateBoundBuffers(BufferObjectBinding::UNIFORM,
-            [&uniformsToBind, &offsets](const BufferState& state, MetalBuffer* buffer,
-                    uint32_t index) {
-        uniformsToBind[index] = buffer;
-        offsets[index] = state.offset;
-    });
-    MetalBuffer::bindBuffers(getPendingCommandBuffer(mContext), mContext->currentRenderPassEncoder,
-            UNIFORM_BUFFER_BINDING_START, MetalBuffer::Stage::VERTEX | MetalBuffer::Stage::FRAGMENT,
-            uniformsToBind, offsets, Program::UNIFORM_BINDING_COUNT);
-
     // Bind sampler groups (argument buffers).
     for (size_t s = 0; s < Program::SAMPLER_BINDING_COUNT; s++) {
         MetalSamplerGroup* const samplerGroup = mContext->samplerBindings[s];
@@ -1775,19 +1785,29 @@ void MetalDriver::draw(PipelineState ps, Handle<HwRenderPrimitive> rph,
                                                           atIndex:(SAMPLER_GROUP_BINDING_START + s)];
         }
     }
+}
+
+void MetalDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
+    ASSERT_PRECONDITION(mContext->currentRenderPassEncoder != nullptr,
+            "bindRenderPrimitive() without a valid command encoder.");
 
     // Bind the user vertex buffers.
-
     MetalBuffer* vertexBuffers[MAX_VERTEX_BUFFER_COUNT] = {};
     size_t vertexBufferOffsets[MAX_VERTEX_BUFFER_COUNT] = {};
     size_t maxBufferIndex = 0;
 
+    MetalRenderPrimitive const* const primitive = handle_cast<MetalRenderPrimitive>(rph);
+    MetalVertexBufferInfo const* const vbi =
+            handle_cast<MetalVertexBufferInfo>(primitive->vertexBuffer->vbih);
+
+    mContext->currentRenderPrimitive = rph;
+
     auto vb = primitive->vertexBuffer;
-    for (auto m : primitive->bufferMapping) {
+    for (auto m : vbi->bufferMapping) {
         assert_invariant(
                 m.bufferArgumentIndex >= USER_VERTEX_BUFFER_BINDING_START &&
                 m.bufferArgumentIndex < USER_VERTEX_BUFFER_BINDING_START + MAX_VERTEX_BUFFER_COUNT);
-        size_t vertexBufferIndex = m.bufferArgumentIndex - USER_VERTEX_BUFFER_BINDING_START;
+        size_t const vertexBufferIndex = m.bufferArgumentIndex - USER_VERTEX_BUFFER_BINDING_START;
         vertexBuffers[vertexBufferIndex] = vb->buffers[m.sourceBufferIndex];
         maxBufferIndex = std::max(maxBufferIndex, vertexBufferIndex);
     }
@@ -1802,6 +1822,27 @@ void MetalDriver::draw(PipelineState ps, Handle<HwRenderPrimitive> rph,
     [mContext->currentRenderPassEncoder setVertexBytes:bytes
                                                 length:16
                                                atIndex:ZERO_VERTEX_BUFFER_BINDING];
+}
+
+void MetalDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
+    ASSERT_PRECONDITION(mContext->currentRenderPassEncoder != nullptr,
+            "draw() without a valid command encoder.");
+
+    // Bind uniform buffers.
+    MetalBuffer* uniformsToBind[Program::UNIFORM_BINDING_COUNT] = { nil };
+    NSUInteger offsets[Program::UNIFORM_BINDING_COUNT] = { 0 };
+
+    enumerateBoundBuffers(BufferObjectBinding::UNIFORM,
+            [&uniformsToBind, &offsets](const BufferState& state, MetalBuffer* buffer,
+                    uint32_t index) {
+                uniformsToBind[index] = buffer;
+                offsets[index] = state.offset;
+            });
+    MetalBuffer::bindBuffers(getPendingCommandBuffer(mContext), mContext->currentRenderPassEncoder,
+            UNIFORM_BUFFER_BINDING_START, MetalBuffer::Stage::VERTEX | MetalBuffer::Stage::FRAGMENT,
+            uniformsToBind, offsets, Program::UNIFORM_BINDING_COUNT);
+
+    auto primitive = handle_cast<MetalRenderPrimitive>(mContext->currentRenderPrimitive);
 
     MetalIndexBuffer* indexBuffer = primitive->indexBuffer;
 
@@ -1815,6 +1856,16 @@ void MetalDriver::draw(PipelineState ps, Handle<HwRenderPrimitive> rph,
                                                 instanceCount:instanceCount];
 }
 
+void MetalDriver::draw(PipelineState ps, Handle<HwRenderPrimitive> rph,
+        uint32_t const indexOffset, uint32_t const indexCount, uint32_t const instanceCount) {
+    MetalRenderPrimitive const* const rp = handle_cast<MetalRenderPrimitive>(rph);
+    ps.primitiveType = rp->type;
+    ps.vertexBufferInfo = rp->vertexBuffer->vbih;
+    bindPipeline(ps);
+    bindRenderPrimitive(rph);
+    draw2(indexOffset, indexCount, instanceCount);
+}
+
 void MetalDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGroupCount) {
     ASSERT_PRECONDITION(!isInRenderPass(mContext),
             "dispatchCompute must be called outside of a render pass.");

filament/backend/src/metal/MetalHandles.h

@@ -149,7 +149,21 @@ struct MetalVertexBufferInfo : public HwVertexBufferInfo {
     MetalVertexBufferInfo(MetalContext& context,
             uint8_t bufferCount, uint8_t attributeCount, AttributeArray const& attributes);
 
-    AttributeArray attributes;
+    // This struct is used to create the pipeline description to describe vertex assembly.
+    VertexDescription vertexDescription = {};
+
+    struct Entry {
+        uint8_t sourceBufferIndex = 0;
+        uint8_t stride = 0;
+        // maps to ->
+        uint8_t bufferArgumentIndex = 0;
+
+        Entry(uint8_t sourceBufferIndex, uint8_t stride, uint8_t bufferArgumentIndex)
+                : sourceBufferIndex(sourceBufferIndex),
+                  stride(stride),
+                  bufferArgumentIndex(bufferArgumentIndex) {}
+    };
+    utils::FixedCapacityVector<Entry> bufferMapping;
 };
 
 struct MetalVertexBuffer : public HwVertexBuffer {
@@ -176,22 +190,6 @@ struct MetalRenderPrimitive : public HwRenderPrimitive {
 
     MetalVertexBuffer* vertexBuffer = nullptr;
     MetalIndexBuffer* indexBuffer = nullptr;
-
-    // This struct is used to create the pipeline description to describe vertex assembly.
-    VertexDescription vertexDescription = {};
-
-    struct Entry {
-        uint8_t sourceBufferIndex = 0;
-        uint8_t stride = 0;
-        // maps to ->
-        uint8_t bufferArgumentIndex = 0;
-
-        Entry(uint8_t sourceBufferIndex, uint8_t stride, uint8_t bufferArgumentIndex)
-                : sourceBufferIndex(sourceBufferIndex),
-                  stride(stride),
-                  bufferArgumentIndex(bufferArgumentIndex) {}
-    };
-    utils::FixedCapacityVector<Entry> bufferMapping;
 };
 
 class MetalProgram : public HwProgram {

filament/backend/src/metal/MetalHandles.mm

@@ -19,6 +19,7 @@
 #include "MetalBlitter.h"
 #include "MetalEnums.h"
 #include "MetalUtils.h"
+#include "MetalBufferPool.h"
 
 #include <filament/SwapChain.h>
 
@@ -343,27 +344,9 @@ void MetalBufferObject::updateBufferUnsynchronized(void* data, size_t size, uint
 MetalVertexBufferInfo::MetalVertexBufferInfo(MetalContext& context, uint8_t bufferCount,
         uint8_t attributeCount, AttributeArray const& attributes)
         : HwVertexBufferInfo(bufferCount, attributeCount),
-          attributes(attributes) {
-}
+          bufferMapping(utils::FixedCapacityVector<Entry>::with_capacity(MAX_VERTEX_BUFFER_COUNT)) {
 
-MetalVertexBuffer::MetalVertexBuffer(MetalContext& context,
-        uint32_t vertexCount, uint32_t bufferCount, Handle<HwVertexBufferInfo> vbih)
-    : HwVertexBuffer(vertexCount), vbih(vbih), buffers(bufferCount, nullptr) {
-}
-
-MetalIndexBuffer::MetalIndexBuffer(MetalContext& context, BufferUsage usage, uint8_t elementSize,
-        uint32_t indexCount) : HwIndexBuffer(elementSize, indexCount),
-        buffer(context, BufferObjectBinding::VERTEX, usage, elementSize * indexCount, true) { }
-
-MetalRenderPrimitive::MetalRenderPrimitive()
-    : bufferMapping(utils::FixedCapacityVector<Entry>::with_capacity(MAX_VERTEX_BUFFER_COUNT)) {}
-
-void MetalRenderPrimitive::setBuffers(MetalVertexBufferInfo const* const vbi,
-        MetalVertexBuffer* vertexBuffer, MetalIndexBuffer* indexBuffer) {
-    this->vertexBuffer = vertexBuffer;
-    this->indexBuffer = indexBuffer;
-
-    const size_t attributeCount = vbi->attributes.size();
+    const size_t maxAttributeCount = attributes.size();
 
     auto& mapping = bufferMapping;
     mapping.clear();
@@ -403,8 +386,8 @@ void MetalRenderPrimitive::setBuffers(MetalVertexBufferInfo const* const vbi,
         }
     };
 
-    for (uint32_t attributeIndex = 0; attributeIndex < attributeCount; attributeIndex++) {
-        const auto& attribute = vbi->attributes[attributeIndex];
+    for (uint32_t attributeIndex = 0; attributeIndex < maxAttributeCount; attributeIndex++) {
+        const auto& attribute = attributes[attributeIndex];
 
         // If the attribute is unused, bind it to the zero buffer. It's a Metal error for a shader
         // to read from missing vertex attributes.
@@ -431,6 +414,24 @@ void MetalRenderPrimitive::setBuffers(MetalVertexBufferInfo const* const vbi,
     }
 }
 
+MetalVertexBuffer::MetalVertexBuffer(MetalContext& context,
+        uint32_t vertexCount, uint32_t bufferCount, Handle<HwVertexBufferInfo> vbih)
+    : HwVertexBuffer(vertexCount), vbih(vbih), buffers(bufferCount, nullptr) {
+}
+
+MetalIndexBuffer::MetalIndexBuffer(MetalContext& context, BufferUsage usage, uint8_t elementSize,
+        uint32_t indexCount) : HwIndexBuffer(elementSize, indexCount),
+        buffer(context, BufferObjectBinding::VERTEX, usage, elementSize * indexCount, true) { }
+
+MetalRenderPrimitive::MetalRenderPrimitive() {
+}
+
+void MetalRenderPrimitive::setBuffers(MetalVertexBufferInfo const* const vbi,
+        MetalVertexBuffer* vertexBuffer, MetalIndexBuffer* indexBuffer) {
+    this->vertexBuffer = vertexBuffer;
+    this->indexBuffer = indexBuffer;
+}
+
 MetalProgram::MetalProgram(MetalContext& context, Program&& program) noexcept
     : HwProgram(program.getName()), mContext(context) {
 
@@ -770,13 +771,13 @@ void MetalTexture::loadWithCopyBuffer(uint32_t level, uint32_t slice, MTLRegion
         PixelBufferDescriptor const& data, const PixelBufferShape& shape) {
     const size_t stagingBufferSize = shape.totalBytes;
     auto entry = context.bufferPool->acquireBuffer(stagingBufferSize);
-    memcpy(entry->buffer.contents,
+    memcpy(entry->buffer.get().contents,
             static_cast<uint8_t*>(data.buffer) + shape.sourceOffset,
             stagingBufferSize);
     id<MTLCommandBuffer> blitCommandBuffer = getPendingCommandBuffer(&context);
     id<MTLBlitCommandEncoder> blitCommandEncoder = [blitCommandBuffer blitCommandEncoder];
     blitCommandEncoder.label = @"Texture upload buffer blit";
-    [blitCommandEncoder copyFromBuffer:entry->buffer
+    [blitCommandEncoder copyFromBuffer:entry->buffer.get()
                           sourceOffset:0
                      sourceBytesPerRow:shape.bytesPerRow
                    sourceBytesPerImage:shape.bytesPerSlice

filament/backend/src/metal/MetalShaderCompiler.h

@@ -41,6 +41,11 @@ class MetalShaderCompiler {
     struct MetalProgramToken;
 
 public:
+    enum class Mode {
+        SYNCHRONOUS,    // synchronous shader compilation
+        ASYNCHRONOUS    // asynchronous shader compilation
+    };
+
     class MetalFunctionBundle {
     public:
         using Raster = std::tuple<id<MTLFunction>, id<MTLFunction>>;
@@ -110,7 +115,7 @@ public:
 
     using program_token_t = std::shared_ptr<MetalProgramToken>;
 
-    explicit MetalShaderCompiler(id<MTLDevice> device, MetalDriver& driver);
+    explicit MetalShaderCompiler(id<MTLDevice> device, MetalDriver& driver, Mode mode);
 
     MetalShaderCompiler(MetalShaderCompiler const& rhs) = delete;
     MetalShaderCompiler(MetalShaderCompiler&& rhs) = delete;
@@ -120,15 +125,15 @@ public:
     void init() noexcept;
     void terminate() noexcept;
 
-    // Creates a program asynchronously
+    bool isParallelShaderCompileSupported() const noexcept;
+
+    // Creates a program, either synchronously or asynchronously, depending on the Mode
+    // MetalShaderCompiler was constructed with.
     program_token_t createProgram(utils::CString const& name, Program&& program);
 
     // Returns the functions, blocking if necessary. The Token is destroyed and becomes invalid.
     MetalFunctionBundle getProgram(program_token_t& token);
 
-    // Destroys a valid token and all associated resources. Used to "cancel" a program compilation.
-    static void terminate(program_token_t& token);
-
     void notifyWhenAllProgramsAreReady(
             CallbackHandler* handler, CallbackHandler::Callback callback, void* user);
 
@@ -138,6 +143,7 @@ private:
     CompilerThreadPool mCompilerThreadPool;
     id<MTLDevice> mDevice;
     CallbackManager mCallbackManager;
+    Mode mMode;
 };
 
 } // namespace filament::backend

filament/backend/src/metal/MetalShaderCompiler.mm

@@ -70,22 +70,31 @@ struct MetalShaderCompiler::MetalProgramToken : ProgramToken {
 
 MetalShaderCompiler::MetalProgramToken::~MetalProgramToken() = default;
 
-MetalShaderCompiler::MetalShaderCompiler(id<MTLDevice> device, MetalDriver& driver)
+MetalShaderCompiler::MetalShaderCompiler(id<MTLDevice> device, MetalDriver& driver, Mode mode)
         : mDevice(device),
-          mCallbackManager(driver) {
+          mCallbackManager(driver),
+          mMode(mode) {
 
 }
 
 void MetalShaderCompiler::init() noexcept {
-    const uint32_t poolSize = 2;
-    mCompilerThreadPool.init(poolSize, []() {}, []() {});
+    const uint32_t poolSize = 1;
+    if (mMode == Mode::ASYNCHRONOUS) {
+        mCompilerThreadPool.init(poolSize, []() {}, []() {});
+    }
 }
 
 void MetalShaderCompiler::terminate() noexcept {
-    mCompilerThreadPool.terminate();
+    if (mMode == Mode::ASYNCHRONOUS) {
+        mCompilerThreadPool.terminate();
+    }
     mCallbackManager.terminate();
 }
 
+bool MetalShaderCompiler::isParallelShaderCompileSupported() const noexcept {
+    return mMode == Mode::ASYNCHRONOUS;
+}
+
 /* static */ MetalShaderCompiler::MetalFunctionBundle MetalShaderCompiler::compileProgram(
         const Program& program, id<MTLDevice> device) {
     std::array<id<MTLFunction>, Program::SHADER_TYPE_COUNT> functions = { nil };
@@ -180,14 +189,26 @@ MetalShaderCompiler::program_token_t MetalShaderCompiler::createProgram(
 
     token->handle = mCallbackManager.get();
 
-    CompilerPriorityQueue const priorityQueue = program.getPriorityQueue();
-    mCompilerThreadPool.queue(priorityQueue, token,
-            [this, name, device = mDevice, program = std::move(program), token]() {
-                MetalFunctionBundle compiledProgram = compileProgram(program, device);
+    switch (mMode) {
+        case Mode::ASYNCHRONOUS: {
+            CompilerPriorityQueue const priorityQueue = program.getPriorityQueue();
+            mCompilerThreadPool.queue(priorityQueue, token,
+                    [this, name, device = mDevice, program = std::move(program), token]() {
+                        MetalFunctionBundle compiledProgram = compileProgram(program, device);
+                        token->set(compiledProgram);
+                        mCallbackManager.put(token->handle);
+                    });
 
-                token->set(compiledProgram);
-                mCallbackManager.put(token->handle);
-            });
+            break;
+        }
+
+        case Mode::SYNCHRONOUS: {
+            MetalFunctionBundle compiledProgram = compileProgram(program, mDevice);
+            token->set(compiledProgram);
+            mCallbackManager.put(token->handle);
+            break;
+        }
+    }
 
     return token;
 }
@@ -195,35 +216,23 @@ MetalShaderCompiler::program_token_t MetalShaderCompiler::createProgram(
 MetalShaderCompiler::MetalFunctionBundle MetalShaderCompiler::getProgram(program_token_t& token) {
     assert_invariant(token);
 
-    if (!token->isReady()) {
-        auto job = mCompilerThreadPool.dequeue(token);
-        if (job) {
-            job();
+    if (mMode == Mode::ASYNCHRONOUS) {
+        if (!token->isReady()) {
+            auto job = mCompilerThreadPool.dequeue(token);
+            if (job) {
+                job();
+            }
         }
     }
 
+    // The job isn't guaranteed to have finished yet. We may have failed to dequeue it above,
+    // which means it's currently running. In that case get() will block until it finishes.
+
     MetalShaderCompiler::MetalFunctionBundle program = token->get();
-
     token = nullptr;
-
     return program;
 }
 
-/* static */ void MetalShaderCompiler::terminate(program_token_t& token) {
-    assert_invariant(token);
-
-    auto job = token->compiler.mCompilerThreadPool.dequeue(token);
-    if (!job) {
-        // The job is being executed right now (or has already executed).
-        token->wait();
-    } else {
-        // The job has not executed yet.
-        token->compiler.mCallbackManager.put(token->handle);
-    }
-
-    token.reset();
-}
-
 void MetalShaderCompiler::notifyWhenAllProgramsAreReady(
         CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
     mCallbackManager.setCallback(handler, callback, user);

filament/backend/src/noop/NoopDriver.cpp

@@ -177,6 +177,10 @@ bool NoopDriver::isSRGBSwapChainSupported() {
     return false;
 }
 
+bool NoopDriver::isProtectedContentSupported() {
+    return false;
+}
+
 bool NoopDriver::isStereoSupported(backend::StereoscopicType) {
     return false;
 }
@@ -189,6 +193,10 @@ bool NoopDriver::isDepthStencilResolveSupported() {
     return true;
 }
 
+bool NoopDriver::isProtectedTexturesSupported() {
+    return true;
+}
+
 bool NoopDriver::isWorkaroundNeeded(Workaround) {
     return false;
 }
@@ -244,8 +252,8 @@ void NoopDriver::update3DImage(Handle<HwTexture> th,
 void NoopDriver::setupExternalImage(void* image) {
 }
 
-bool NoopDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapsedTime) {
-    return false;
+TimerQueryResult NoopDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapsedTime) {
+    return TimerQueryResult::ERROR;
 }
 
 void NoopDriver::setExternalImage(Handle<HwTexture> th, void* image) {
@@ -342,6 +350,15 @@ void NoopDriver::blit(
         math::uint2 size) {
 }
 
+void NoopDriver::bindPipeline(PipelineState pipelineState) {
+}
+
+void NoopDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
+}
+
+void NoopDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
+}
+
 void NoopDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> rph,
         uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
 }

filament/backend/src/opengl/OpenGLContext.cpp

@@ -16,10 +16,25 @@
 
 #include "OpenGLContext.h"
 
-#include <backend/platforms/OpenGLPlatform.h>
+#include "GLUtils.h"
+#include "OpenGLTimerQuery.h"
 
+#include <backend/platforms/OpenGLPlatform.h>
+#include <backend/DriverEnums.h>
+
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/Log.h>
+#include <utils/ostream.h>
+
+#include <functional>
+#include <string_view>
 #include <utility>
 
+#include <stddef.h>
+#include <stdio.h>
+#include <string.h>
+
 // change to true to display all GL extensions in the console on start-up
 #define DEBUG_PRINT_EXTENSIONS false
 
@@ -48,7 +63,8 @@ bool OpenGLContext::queryOpenGLVersion(GLint* major, GLint* minor) noexcept {
 #endif
 }
 
-OpenGLContext::OpenGLContext() noexcept {
+OpenGLContext::OpenGLContext(OpenGLPlatform& platform) noexcept
+        : mPlatform(platform) {
 
     state.vao.p = &mDefaultVAO;
 
@@ -231,6 +247,49 @@ OpenGLContext::OpenGLContext() noexcept {
         glDebugMessageCallback(cb, nullptr);
     }
 #endif
+
+    mTimerQueryFactory = TimerQueryFactory::init(platform, *this);
+}
+
+OpenGLContext::~OpenGLContext() noexcept {
+    delete mTimerQueryFactory;
+}
+
+void OpenGLContext::destroyWithContext(
+        size_t index, std::function<void(OpenGLContext&)> const& closure) noexcept {
+    if (index == 0) {
+        // Note: we only need to delay the destruction of objects on the unprotected context
+        // (index 0) because the protected context is always immediately destroyed and all its
+        // active objects and bindings are then automatically destroyed.
+        // TODO: this is only guaranteed for EGLPlatform, but that's the only one we care about.
+        mDestroyWithNormalContext.push_back(closure);
+    }
+}
+
+void OpenGLContext::unbindEverything() noexcept {
+    // TODO:  we're supposed to unbind everything here so that resources don't get
+    //        stuck in this context (contextIndex) when destroyed in the other context.
+    //        However, because EGLPlatform always immediately destroys the protected context (1),
+    //        the bindings will automatically be severed when we switch back to the default context.
+    //        Since bindings now only exist in one context, we don't have a ref-counting issue to
+    //        worry about.
+}
+
+void OpenGLContext::synchronizeStateAndCache(size_t index) noexcept {
+
+    // if we're just switching back to context 0, run all the pending destructors
+    if (index == 0) {
+        auto list = std::move(mDestroyWithNormalContext);
+        for (auto&& fn: list) {
+            fn(*this);
+        }
+    }
+
+    // the default FBO could be invalid
+    mDefaultFbo[index].reset();
+
+    contextIndex = index;
+    resetState();
 }
 
 void OpenGLContext::setDefaultState() noexcept {
@@ -599,6 +658,7 @@ void OpenGLContext::initExtensionsGLES(Extensions* ext, GLint major, GLint minor
     ext->EXT_disjoint_timer_query = exts.has("GL_EXT_disjoint_timer_query"sv);
     ext->EXT_multisampled_render_to_texture = exts.has("GL_EXT_multisampled_render_to_texture"sv);
     ext->EXT_multisampled_render_to_texture2 = exts.has("GL_EXT_multisampled_render_to_texture2"sv);
+    ext->EXT_protected_textures = exts.has("GL_EXT_protected_textures"sv);
 #endif
     ext->EXT_shader_framebuffer_fetch = exts.has("GL_EXT_shader_framebuffer_fetch"sv);
 #ifndef __EMSCRIPTEN__
@@ -709,6 +769,51 @@ void OpenGLContext::initExtensionsGL(Extensions* ext, GLint major, GLint minor)
 
 #endif // BACKEND_OPENGL_VERSION_GL
 
+
+GLuint OpenGLContext::bindFramebuffer(GLenum target, GLuint buffer) noexcept {
+    if (UTILS_UNLIKELY(buffer == 0)) {
+        // we're binding the default frame buffer, resolve its actual name
+        auto& defaultFboForThisContext = mDefaultFbo[contextIndex];
+        if (UTILS_UNLIKELY(!defaultFboForThisContext.has_value())) {
+            defaultFboForThisContext = GLuint(mPlatform.getDefaultFramebufferObject());
+        }
+        buffer = defaultFboForThisContext.value();
+    }
+    bindFramebufferResolved(target, buffer);
+    return buffer;
+}
+
+void OpenGLContext::unbindFramebuffer(GLenum target) noexcept {
+    bindFramebufferResolved(target, 0);
+}
+
+void OpenGLContext::bindFramebufferResolved(GLenum target, GLuint buffer) noexcept {
+    switch (target) {
+        case GL_FRAMEBUFFER:
+            if (state.draw_fbo != buffer || state.read_fbo != buffer) {
+                state.draw_fbo = state.read_fbo = buffer;
+                glBindFramebuffer(target, buffer);
+            }
+            break;
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+        case GL_DRAW_FRAMEBUFFER:
+            if (state.draw_fbo != buffer) {
+                state.draw_fbo = buffer;
+                glBindFramebuffer(target, buffer);
+            }
+            break;
+        case GL_READ_FRAMEBUFFER:
+            if (state.read_fbo != buffer) {
+                state.read_fbo = buffer;
+                glBindFramebuffer(target, buffer);
+            }
+            break;
+#endif
+        default:
+            break;
+    }
+}
+
 void OpenGLContext::bindBuffer(GLenum target, GLuint buffer) noexcept {
     if (target == GL_ELEMENT_ARRAY_BUFFER) {
         constexpr size_t targetIndex = getIndexForBufferTarget(GL_ELEMENT_ARRAY_BUFFER);
@@ -819,19 +924,22 @@ void OpenGLContext::deleteBuffers(GLsizei n, const GLuint* buffers, GLenum targe
 #endif
 }
 
-void OpenGLContext::deleteVertexArrays(GLsizei n, const GLuint* arrays) noexcept {
-    procs.deleteVertexArrays(n, arrays);
-    // if one of the destroyed VAO is bound, clear the binding.
-    for (GLsizei i = 0; i < n; ++i) {
-        if (state.vao.p->vao == arrays[i]) {
+void OpenGLContext::deleteVertexArray(GLuint vao) noexcept {
+    if (UTILS_LIKELY(vao)) {
+        procs.deleteVertexArrays(1, &vao);
+        // if the destroyed VAO is bound, clear the binding.
+        if (state.vao.p->vao[contextIndex] == vao) {
             bindVertexArray(nullptr);
-            break;
         }
     }
 }
 
 void OpenGLContext::resetState() noexcept {
     // Force GL state to match the Filament state
+
+    // increase the state version so other parts of the state know to reset
+    state.age++;
+
     if (state.major > 2) {
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
         glBindFramebuffer(GL_DRAW_FRAMEBUFFER, state.draw_fbo);
@@ -848,11 +956,8 @@ void OpenGLContext::resetState() noexcept {
     glUseProgram(state.program.use);
 
     // state.vao
-    if (state.vao.p) {
-        procs.bindVertexArray(state.vao.p->vao);
-    } else {
-        bindVertexArray(nullptr);
-    }
+    state.vao.p = nullptr;
+    bindVertexArray(nullptr);
 
     // state.raster
     glFrontFace(state.raster.frontFace);
@@ -1008,7 +1113,22 @@ void OpenGLContext::resetState() noexcept {
         state.window.viewport.w
     );
     glDepthRangef(state.window.depthRange.x, state.window.depthRange.y);
-    
+}
+
+void OpenGLContext::createTimerQuery(GLTimerQuery* query) {
+    mTimerQueryFactory->createTimerQuery(query);
+}
+
+void OpenGLContext::destroyTimerQuery(GLTimerQuery* query) {
+    mTimerQueryFactory->destroyTimerQuery(query);
+}
+
+void OpenGLContext::beginTimeElapsedQuery(GLTimerQuery* query) {
+    mTimerQueryFactory->beginTimeElapsedQuery(query);
+}
+
+void OpenGLContext::endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQuery* query) {
+    mTimerQueryFactory->endTimeElapsedQuery(driver, query);
 }
 
 } // namesapce filament

filament/backend/src/opengl/OpenGLContext.h

@@ -17,25 +17,37 @@
 #ifndef TNT_FILAMENT_BACKEND_OPENGLCONTEXT_H
 #define TNT_FILAMENT_BACKEND_OPENGLCONTEXT_H
 
-#include <math/vec4.h>
 
-#include <utils/CString.h>
-#include <utils/debug.h>
+#include "OpenGLTimerQuery.h"
 
+#include <backend/platforms/OpenGLPlatform.h>
+
+#include <backend/DriverEnums.h>
 #include <backend/Handle.h>
 
-#include "GLUtils.h"
+#include "gl_headers.h"
+
+#include <utils/compiler.h>
+#include <utils/bitset.h>
+#include <utils/debug.h>
+
+#include <math/vec2.h>
+#include <math/vec4.h>
 
 #include <array>
-#include <set>
+#include <functional>
+#include <optional>
 #include <tuple>
-#include <utility>
+#include <vector>
+
+#include <stddef.h>
+#include <stdint.h>
 
 namespace filament::backend {
 
 class OpenGLPlatform;
 
-class OpenGLContext {
+class OpenGLContext final : public TimerQueryFactoryInterface {
 public:
     static constexpr const size_t MAX_TEXTURE_UNIT_COUNT = MAX_SAMPLER_COUNT;
     static constexpr const size_t DUMMY_TEXTURE_BINDING = 7; // highest binding guaranteed to work with ES2
@@ -46,19 +58,29 @@ public:
     struct RenderPrimitive {
         static_assert(MAX_VERTEX_ATTRIBUTE_COUNT <= 16);
 
-        GLuint vao = 0;                                         // 4
+        GLuint vao[2] = {};                                     // 4
         GLuint elementArray = 0;                                // 4
-        utils::bitset<uint16_t> vertexAttribArray;              // 2
+        mutable utils::bitset<uint16_t> vertexAttribArray;      // 2
 
-        // If this version number does not match vertexBufferWithObjects->bufferObjectsVersion,
-        // then the VAO needs to be updated.
+        // if this differs from vertexBufferWithObjects->bufferObjectsVersion, this VAO needs to
+        // be updated (see OpenGLDriver::updateVertexArrayObject())
         uint8_t vertexBufferVersion = 0;                        // 1
+
+        // if this differs from OpenGLContext::state.age, this VAO needs to
+        // be updated (see OpenGLDriver::updateVertexArrayObject())
+        uint8_t stateVersion = 0;                               // 1
+
+        // If this differs from OpenGLContext::state.age, this VAO's name needs to be updated.
+        // See OpenGLContext::bindVertexArray()
+        uint8_t nameVersion = 0;                                // 1
+
+        // Size in bytes of indices in the index buffer
         uint8_t indicesSize = 0;                                // 1
 
         // The optional 32-bit handle to a GLVertexBuffer is necessary only if the referenced
         // VertexBuffer supports buffer objects. If this is zero, then the VBO handles array is
         // immutable.
-        Handle<HwVertexBuffer> vertexBufferWithObjects = {};    // 4
+        Handle<HwVertexBuffer> vertexBufferWithObjects;         // 4
 
         GLenum getIndicesType() const noexcept {
             return indicesSize == 4 ? GL_UNSIGNED_INT : GL_UNSIGNED_SHORT;
@@ -67,7 +89,18 @@ public:
 
     static bool queryOpenGLVersion(GLint* major, GLint* minor) noexcept;
 
-    OpenGLContext() noexcept;
+    explicit OpenGLContext(OpenGLPlatform& platform) noexcept;
+    ~OpenGLContext() noexcept final;
+
+    // TimerQueryInterface ------------------------------------------------------------------------
+
+    // note: OpenGLContext being final ensures (clang) these are not called through the vtable
+    void createTimerQuery(GLTimerQuery* query) override;
+    void destroyTimerQuery(GLTimerQuery* query) override;
+    void beginTimeElapsedQuery(GLTimerQuery* query) override;
+    void endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQuery* query) override;
+
+    // --------------------------------------------------------------------------------------------
 
     template<int MAJOR, int MINOR>
     inline bool isAtLeastGL() const noexcept {
@@ -123,10 +156,11 @@ public:
     inline void bindBufferRange(GLenum target, GLuint index, GLuint buffer,
             GLintptr offset, GLsizeiptr size) noexcept;
 
-    inline void bindFramebuffer(GLenum target, GLuint buffer) noexcept;
+    GLuint bindFramebuffer(GLenum target, GLuint buffer) noexcept;
+    void unbindFramebuffer(GLenum target) noexcept;
 
-    inline void enableVertexAttribArray(GLuint index) noexcept;
-    inline void disableVertexAttribArray(GLuint index) noexcept;
+    inline void enableVertexAttribArray(RenderPrimitive const* rp, GLuint index) noexcept;
+    inline void disableVertexAttribArray(RenderPrimitive const* rp, GLuint index) noexcept;
     inline void enable(GLenum cap) noexcept;
     inline void disable(GLenum cap) noexcept;
     inline void frontFace(GLenum mode) noexcept;
@@ -148,7 +182,9 @@ public:
     inline void depthRange(GLclampf near, GLclampf far) noexcept;
 
     void deleteBuffers(GLsizei n, const GLuint* buffers, GLenum target) noexcept;
-    void deleteVertexArrays(GLsizei n, const GLuint* arrays) noexcept;
+    void deleteVertexArray(GLuint vao) noexcept;
+
+    void destroyWithContext(size_t index, std::function<void(OpenGLContext&)> const& closure) noexcept;
 
     // glGet*() values
     struct Gets {
@@ -183,6 +219,7 @@ public:
         bool EXT_discard_framebuffer;
         bool EXT_multisampled_render_to_texture2;
         bool EXT_multisampled_render_to_texture;
+        bool EXT_protected_textures;
         bool EXT_shader_framebuffer_fetch;
         bool EXT_texture_compression_bptc;
         bool EXT_texture_compression_etc2;
@@ -291,8 +328,19 @@ public:
 
     FeatureLevel getFeatureLevel() const noexcept { return mFeatureLevel; }
 
+    // This is the index of the context in use. Must be 0 or 1. This is used to manange the
+    // OpenGL name of ContainerObjects within each context.
+    uint32_t contextIndex = 0;
+
     // Try to keep the State structure sorted by data-access patterns
     struct State {
+        State() noexcept = default;
+        // make sure we don't copy this state by accident
+        State(State const& rhs) = delete;
+        State(State&& rhs) noexcept = delete;
+        State& operator=(State const& rhs) = delete;
+        State& operator=(State&& rhs) noexcept = delete;
+
         GLint major = 0;
         GLint minor = 0;
 
@@ -397,6 +445,7 @@ public:
             vec4gli viewport { 0 };
             vec2glf depthRange { 0.0f, 1.0f };
         } window;
+        uint8_t age = 0;
     } state;
 
     struct Procs {
@@ -416,9 +465,19 @@ public:
         void (* maxShaderCompilerThreadsKHR)(GLuint count);
     } procs{};
 
+    void unbindEverything() noexcept;
+    void synchronizeStateAndCache(size_t index) noexcept;
+
 private:
+    OpenGLPlatform& mPlatform;
     ShaderModel mShaderModel = ShaderModel::MOBILE;
     FeatureLevel mFeatureLevel = FeatureLevel::FEATURE_LEVEL_1;
+    TimerQueryFactoryInterface* mTimerQueryFactory = nullptr;
+    std::vector<std::function<void(OpenGLContext&)>> mDestroyWithNormalContext;
+    RenderPrimitive mDefaultVAO;
+    std::optional<GLuint> mDefaultFbo[2];
+
+    void bindFramebufferResolved(GLenum target, GLuint buffer) noexcept;
 
     const std::array<std::tuple<bool const&, char const*, char const*>, sizeof(bugs)> mBugDatabase{{
             {   bugs.disable_glFlush,
@@ -471,8 +530,6 @@ private:
                     ""},
     }};
 
-    RenderPrimitive mDefaultVAO;
-
     // this is chosen to minimize code size
 #if defined(BACKEND_OPENGL_VERSION_GLES)
     static void initExtensionsGLES(Extensions* ext, GLint major, GLint minor) noexcept;
@@ -630,11 +687,26 @@ void OpenGLContext::depthRange(GLclampf near, GLclampf far) noexcept {
 void OpenGLContext::bindVertexArray(RenderPrimitive const* p) noexcept {
     RenderPrimitive* vao = p ? const_cast<RenderPrimitive *>(p) : &mDefaultVAO;
     update_state(state.vao.p, vao, [&]() {
-        procs.bindVertexArray(vao->vao);
+
+        // See if we need to create a name for this VAO on the fly, this would happen if:
+        // - we're not the default VAO, because its name is always 0
+        // - our name is 0, this could happen if this VAO was created in the "other" context
+        // - the nameVersion is out of date *and* we're on the protected context, in this case:
+        //      - the name must be stale from a previous use of this context because we always
+        //        destroy the protected context when we're done with it.
+        bool const recreateVaoName = p != &mDefaultVAO &&
+                ((vao->vao[contextIndex] == 0) ||
+                        (vao->nameVersion != state.age && contextIndex == 1));
+        if (UTILS_UNLIKELY(recreateVaoName)) {
+            vao->nameVersion = state.age;
+            procs.genVertexArrays(1, &vao->vao[contextIndex]);
+        }
+
+        procs.bindVertexArray(vao->vao[contextIndex]);
         // update GL_ELEMENT_ARRAY_BUFFER, which is updated by glBindVertexArray
         size_t const targetIndex = getIndexForBufferTarget(GL_ELEMENT_ARRAY_BUFFER);
         state.buffers.genericBinding[targetIndex] = vao->elementArray;
-        if (UTILS_UNLIKELY(bugs.vao_doesnt_store_element_array_buffer_binding)) {
+        if (UTILS_UNLIKELY(bugs.vao_doesnt_store_element_array_buffer_binding || recreateVaoName)) {
             // This shouldn't be needed, but it looks like some drivers don't do the implicit
             // glBindBuffer().
             glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vao->elementArray);
@@ -672,33 +744,6 @@ void OpenGLContext::bindBufferRange(GLenum target, GLuint index, GLuint buffer,
 #endif
 }
 
-void OpenGLContext::bindFramebuffer(GLenum target, GLuint buffer) noexcept {
-    switch (target) {
-        case GL_FRAMEBUFFER:
-            if (state.draw_fbo != buffer || state.read_fbo != buffer) {
-                state.draw_fbo = state.read_fbo = buffer;
-                glBindFramebuffer(target, buffer);
-            }
-            break;
-#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-        case GL_DRAW_FRAMEBUFFER:
-            if (state.draw_fbo != buffer) {
-                state.draw_fbo = buffer;
-                glBindFramebuffer(target, buffer);
-            }
-            break;
-        case GL_READ_FRAMEBUFFER:
-            if (state.read_fbo != buffer) {
-                state.read_fbo = buffer;
-                glBindFramebuffer(target, buffer);
-            }
-            break;
-#endif
-        default:
-            break;
-    }
-}
-
 void OpenGLContext::bindTexture(GLuint unit, GLuint target, GLuint texId, size_t targetIndex) noexcept {
     assert_invariant(targetIndex == getIndexForTextureTarget(target));
     assert_invariant(targetIndex < TEXTURE_TARGET_COUNT);
@@ -718,20 +763,22 @@ void OpenGLContext::useProgram(GLuint program) noexcept {
     });
 }
 
-void OpenGLContext::enableVertexAttribArray(GLuint index) noexcept {
-    assert_invariant(state.vao.p);
-    assert_invariant(index < state.vao.p->vertexAttribArray.size());
-    if (UTILS_UNLIKELY(!state.vao.p->vertexAttribArray[index])) {
-        state.vao.p->vertexAttribArray.set(index);
+void OpenGLContext::enableVertexAttribArray(RenderPrimitive const* rp, GLuint index) noexcept {
+    assert_invariant(rp);
+    assert_invariant(index < rp->vertexAttribArray.size());
+    bool const force = rp->stateVersion != state.age;
+    if (UTILS_UNLIKELY(force || !rp->vertexAttribArray[index])) {
+        rp->vertexAttribArray.set(index);
         glEnableVertexAttribArray(index);
     }
 }
 
-void OpenGLContext::disableVertexAttribArray(GLuint index) noexcept {
-    assert_invariant(state.vao.p);
-    assert_invariant(index < state.vao.p->vertexAttribArray.size());
-    if (UTILS_UNLIKELY(state.vao.p->vertexAttribArray[index])) {
-        state.vao.p->vertexAttribArray.unset(index);
+void OpenGLContext::disableVertexAttribArray(RenderPrimitive const* rp, GLuint index) noexcept {
+    assert_invariant(rp);
+    assert_invariant(index < rp->vertexAttribArray.size());
+    bool const force = rp->stateVersion != state.age;
+    if (UTILS_UNLIKELY(force || rp->vertexAttribArray[index])) {
+        rp->vertexAttribArray.unset(index);
         glDisableVertexAttribArray(index);
     }
 }

filament/backend/src/opengl/OpenGLDriver.cpp

@@ -16,21 +16,54 @@
 
 #include "OpenGLDriver.h"
 
-#include "private/backend/DriverApi.h"
-
 #include "CommandStreamDispatcher.h"
+#include "GLUtils.h"
 #include "OpenGLContext.h"
 #include "OpenGLDriverFactory.h"
 #include "OpenGLProgram.h"
 #include "OpenGLTimerQuery.h"
+#include "gl_headers.h"
 
 #include <backend/platforms/OpenGLPlatform.h>
-#include <backend/SamplerDescriptor.h>
 
-#include <utils/compiler.h>
+#include <backend/BufferDescriptor.h>
+#include <backend/CallbackHandler.h>
+#include <backend/DriverApiForward.h>
+#include <backend/DriverEnums.h>
+#include <backend/Handle.h>
+#include <backend/PipelineState.h>
+#include <backend/Platform.h>
+#include <backend/Program.h>
+#include <backend/SamplerDescriptor.h>
+#include <backend/TargetBufferInfo.h>
+
+#include "private/backend/Dispatcher.h"
+#include "private/backend/DriverApi.h"
+
+#include <utils/BitmaskEnum.h>
+#include <utils/CString.h>
 #include <utils/Log.h>
 #include <utils/Panic.h>
 #include <utils/Systrace.h>
+#include <utils/compiler.h>
+#include <utils/debug.h>
+#include <utils/ostream.h>
+
+#include <math/vec2.h>
+#include <math/vec3.h>
+
+#include <algorithm>
+#include <chrono>
+#include <functional>
+#include <limits>
+#include <memory>
+#include <mutex>
+#include <new>
+#include <utility>
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
 
 #if defined(__EMSCRIPTEN__)
 #include <emscripten.h>
@@ -94,10 +127,10 @@ Driver* OpenGLDriver::create(OpenGLPlatform* const platform,
     //    GLSamplerGroup            :  16       few
     //    GLSwapChain               :  16       few
     //    GLTimerQuery              :  16       few
-    //    GLRenderPrimitive         :  20       many
     //    GLFence                   :  24       few
-    // -- less than or equal 24 bytes
+    //    GLRenderPrimitive         :  32       many
     //    GLBufferObject            :  32       many
+    // -- less than or equal 32 bytes
     //    OpenGLProgram             :  56       moderate
     //    GLTexture                 :  64       moderate
     // -- less than or equal 64 bytes
@@ -170,9 +203,11 @@ OpenGLDriver::DebugMarker::~DebugMarker() noexcept {
 
 OpenGLDriver::OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfig& driverConfig) noexcept
         : mPlatform(*platform),
-          mContext(),
+          mContext(mPlatform),
           mShaderCompilerService(*this),
-          mHandleAllocator("Handles", driverConfig.handleArenaSize),
+          mHandleAllocator("Handles",
+                  driverConfig.handleArenaSize,
+                  driverConfig.disableHandleUseAfterFreeCheck),
           mSamplerMap(32),
           mDriverConfig(driverConfig) {
 
@@ -194,8 +229,6 @@ OpenGLDriver::OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfi
     assert_invariant(mContext.ext.EXT_disjoint_timer_query);
 #endif
 
-    mTimerQueryImpl = OpenGLTimerQueryFactory::init(mPlatform, *this);
-
     mShaderCompilerService.init();
 }
 
@@ -240,8 +273,6 @@ void OpenGLDriver::terminate() {
     }
 #endif
 
-    delete mTimerQueryImpl;
-
     mPlatform.terminate();
 }
 
@@ -522,13 +553,20 @@ void OpenGLDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph,
     GLIndexBuffer const* const ib = handle_cast<const GLIndexBuffer*>(ibh);
     assert_invariant(ib->elementSize == 2 || ib->elementSize == 4);
 
-    GLRenderPrimitive* rp = handle_cast<GLRenderPrimitive*>(rph);
+    GLVertexBuffer* const vb = handle_cast<GLVertexBuffer*>(vbh);
+    GLRenderPrimitive* const rp = handle_cast<GLRenderPrimitive*>(rph);
     rp->gl.indicesSize = (ib->elementSize == 4u) ? 4u : 2u;
     rp->gl.vertexBufferWithObjects = vbh;
     rp->type = pt;
+    rp->vbih = vb->vbih;
 
-    gl.procs.genVertexArrays(1, &rp->gl.vao);
+    // create a name for this VAO in the current context
+    gl.procs.genVertexArrays(1, &rp->gl.vao[gl.contextIndex]);
 
+    // this implies our name is up-to-date
+    rp->gl.nameVersion = gl.state.age;
+
+    // binding the VAO will actually create it
     gl.bindVertexArray(&rp->gl);
 
     // Note: we don't update the vertex buffer bindings in the VAO just yet, we will do that
@@ -580,13 +618,22 @@ void OpenGLDriver::createSamplerGroupR(Handle<HwSamplerGroup> sbh, uint32_t size
 
 UTILS_NOINLINE
 void OpenGLDriver::textureStorage(OpenGLDriver::GLTexture* t,
-        uint32_t width, uint32_t height, uint32_t depth) noexcept {
+        uint32_t width, uint32_t height, uint32_t depth, bool useProtectedMemory) noexcept {
 
     auto& gl = mContext;
 
     bindTexture(OpenGLContext::DUMMY_TEXTURE_BINDING, t);
     gl.activeTexture(OpenGLContext::DUMMY_TEXTURE_BINDING);
 
+#ifdef GL_EXT_protected_textures
+#ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
+    if (UTILS_UNLIKELY(useProtectedMemory)) {
+        assert_invariant(gl.ext.EXT_protected_textures);
+        glTexParameteri(t->gl.target, GL_TEXTURE_PROTECTED_EXT, 1);
+    }
+#endif
+#endif
+
     switch (t->gl.target) {
         case GL_TEXTURE_2D:
         case GL_TEXTURE_CUBE_MAP:
@@ -677,6 +724,12 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
     GLenum internalFormat = getInternalFormat(format);
     assert_invariant(internalFormat);
 
+    if (UTILS_UNLIKELY(usage & TextureUsage::PROTECTED)) {
+        // renderbuffers don't have a protected mode, so we need to use a texture
+        // because protected textures are only supported on GLES 3.2, MSAA will be available.
+        usage |= TextureUsage::SAMPLEABLE;
+    }
+
     auto& gl = mContext;
     samples = std::clamp(samples, uint8_t(1u), uint8_t(gl.gets.max_samples));
     GLTexture* t = construct<GLTexture>(th, target, levels, samples, w, h, depth, format, usage);
@@ -712,23 +765,23 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
                     // we can't be here -- doesn't matter what we do
                 case SamplerType::SAMPLER_2D:
                     t->gl.target = GL_TEXTURE_2D;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D);
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D);
                     break;
                 case SamplerType::SAMPLER_3D:
                     t->gl.target = GL_TEXTURE_3D;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_3D);
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_3D);
                     break;
                 case SamplerType::SAMPLER_2D_ARRAY:
                     t->gl.target = GL_TEXTURE_2D_ARRAY;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
                     break;
                 case SamplerType::SAMPLER_CUBEMAP:
                     t->gl.target = GL_TEXTURE_CUBE_MAP;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
                     break;
                 case SamplerType::SAMPLER_CUBEMAP_ARRAY:
                     t->gl.target = GL_TEXTURE_CUBE_MAP_ARRAY;
-                    t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
+                    t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
                     break;
             }
 
@@ -740,13 +793,14 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
                     // multi-sample texture on GL 3.2 / GLES 3.1 and above
                     t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
                     t->gl.targetIndex = (uint8_t)
-                            gl.getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
+                            OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
                 } else {
                     // Turn off multi-sampling for that texture. It's just not supported.
                 }
 #endif
             }
-            textureStorage(t, w, h, depth);
+
+            textureStorage(t, w, h, depth, bool(usage & TextureUsage::PROTECTED));
         }
     } else {
         assert_invariant(any(usage & (
@@ -755,6 +809,7 @@ void OpenGLDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
                 TextureUsage::STENCIL_ATTACHMENT)));
         assert_invariant(levels == 1);
         assert_invariant(target == SamplerType::SAMPLER_2D);
+        assert_invariant(none(usage & TextureUsage::PROTECTED));
         t->gl.internalFormat = internalFormat;
         t->gl.target = GL_RENDERBUFFER;
         glGenRenderbuffers(1, &t->gl.id);
@@ -809,27 +864,27 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
     switch (target) {
         case SamplerType::SAMPLER_EXTERNAL:
             t->gl.target = GL_TEXTURE_EXTERNAL_OES;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_EXTERNAL_OES);
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_EXTERNAL_OES);
             break;
         case SamplerType::SAMPLER_2D:
             t->gl.target = GL_TEXTURE_2D;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D);
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D);
             break;
         case SamplerType::SAMPLER_3D:
             t->gl.target = GL_TEXTURE_3D;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_3D);
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_3D);
             break;
         case SamplerType::SAMPLER_2D_ARRAY:
             t->gl.target = GL_TEXTURE_2D_ARRAY;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_ARRAY);
             break;
         case SamplerType::SAMPLER_CUBEMAP:
             t->gl.target = GL_TEXTURE_CUBE_MAP;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP);
             break;
         case SamplerType::SAMPLER_CUBEMAP_ARRAY:
             t->gl.target = GL_TEXTURE_CUBE_MAP_ARRAY;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_CUBE_MAP_ARRAY);
             break;
     }
 
@@ -840,7 +895,7 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
         if (gl.features.multisample_texture) {
             // multi-sample texture on GL 3.2 / GLES 3.1 and above
             t->gl.target = GL_TEXTURE_2D_MULTISAMPLE;
-            t->gl.targetIndex = (uint8_t)gl.getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
+            t->gl.targetIndex = OpenGLContext::getIndexForTextureTarget(GL_TEXTURE_2D_MULTISAMPLE);
         } else {
             // Turn off multi-sampling for that texture. It's just not supported.
         }
@@ -851,22 +906,20 @@ void OpenGLDriver::importTextureR(Handle<HwTexture> th, intptr_t id,
 }
 
 void OpenGLDriver::updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer const* vb) {
+    // NOTE: this is called from draw() and must be as efficient as possible.
 
     auto& gl = mContext;
 
-    // NOTE: this is called from draw() and must be as efficient as possible.
-
     if (UTILS_LIKELY(gl.ext.OES_vertex_array_object)) {
         // The VAO for the given render primitive must already be bound.
-#ifndef NDEBUG
         GLint vaoBinding;
         glGetIntegerv(GL_VERTEX_ARRAY_BINDING, &vaoBinding);
-        assert_invariant(vaoBinding == (GLint)rp->gl.vao);
-#endif
-        rp->gl.vertexBufferVersion = vb->bufferObjectsVersion;
-    } else {
-        // if we don't have OES_vertex_array_object, we never update the buffer version so
-        // that it's always reset in draw
+        assert_invariant(vaoBinding == (GLint)rp->gl.vao[gl.contextIndex]);
+    }
+
+    if (UTILS_LIKELY(rp->gl.vertexBufferVersion == vb->bufferObjectsVersion &&
+                     rp->gl.stateVersion == gl.state.age)) {
+        return;
     }
 
     GLVertexBufferInfo const* const vbi = handle_cast<const GLVertexBufferInfo*>(vb->vbih);
@@ -883,7 +936,7 @@ void OpenGLDriver::updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer
 
             gl.bindBuffer(GL_ARRAY_BUFFER, vb->gl.buffers[bi]);
             GLuint const index = i;
-            GLint const size = getComponentCount(attribute.type);
+            GLint const size = (GLint)getComponentCount(attribute.type);
             GLenum const type = getComponentType(attribute.type);
             GLboolean const normalized = getNormalization(attribute.flags & Attribute::FLAG_NORMALIZED);
             GLsizei const stride = attribute.stride;
@@ -901,7 +954,7 @@ void OpenGLDriver::updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer
                 glVertexAttribPointer(index, size, type, normalized, stride, pointer);
             }
 
-            gl.enableVertexAttribArray(GLuint(i));
+            gl.enableVertexAttribArray(&rp->gl, GLuint(i));
         } else {
             // In some OpenGL implementations, we must supply a properly-typed placeholder for
             // every integer input that is declared in the vertex shader.
@@ -924,13 +977,21 @@ void OpenGLDriver::updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer
                 glVertexAttrib4f(GLuint(i), 0, 0, 0, 0);
             }
 
-            gl.disableVertexAttribArray(GLuint(i));
+            gl.disableVertexAttribArray(&rp->gl, GLuint(i));
         }
     }
+
+    rp->gl.stateVersion = gl.state.age;
+    if (UTILS_LIKELY(gl.ext.OES_vertex_array_object)) {
+        rp->gl.vertexBufferVersion = vb->bufferObjectsVersion;
+    } else {
+        // if we don't have OES_vertex_array_object, we never update the buffer version so
+        // that it's always reset in draw
+    }
 }
 
 void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
-        GLRenderTarget const* rt, GLenum attachment) noexcept {
+        GLRenderTarget const* rt, GLenum attachment, uint8_t layerCount) noexcept {
 
 #if !defined(NDEBUG)
     // Only used by assert_invariant() checks below
@@ -1068,9 +1129,20 @@ void OpenGLDriver::framebufferTexture(TargetBufferInfo const& binfo,
             case GL_TEXTURE_2D_ARRAY:
             case GL_TEXTURE_CUBE_MAP_ARRAY:
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
-                // GL_TEXTURE_2D_MULTISAMPLE_ARRAY is not supported in GLES
-                glFramebufferTextureLayer(GL_FRAMEBUFFER, attachment,
+
+                // TODO: support multiview for iOS and WebGL
+#if !defined(__EMSCRIPTEN__) && !defined(IOS)
+                if (layerCount > 1) {
+                    // if layerCount > 1, it means we use the multiview extension.
+                    glFramebufferTextureMultiviewOVR(GL_FRAMEBUFFER, attachment,
+                        t->gl.id, 0, binfo.baseViewIndex, layerCount);
+                } else
+#endif // !defined(__EMSCRIPTEN__) && !defined(IOS)
+                {
+                    // GL_TEXTURE_2D_MULTISAMPLE_ARRAY is not supported in GLES
+                    glFramebufferTextureLayer(GL_FRAMEBUFFER, attachment,
                         t->gl.id, binfo.level, binfo.layer);
+                }
 #endif
                 break;
             default:
@@ -1231,11 +1303,9 @@ void OpenGLDriver::createDefaultRenderTargetR(
 
     construct<GLRenderTarget>(rth, 0, 0);  // FIXME: we don't know the width/height
 
-    uint32_t const framebuffer = mPlatform.createDefaultRenderTarget();
-
     GLRenderTarget* rt = handle_cast<GLRenderTarget*>(rth);
     rt->gl.isDefault = true;
-    rt->gl.fbo = framebuffer;
+    rt->gl.fbo = 0; // the actual id is resolved at binding time
     rt->gl.samples = 1;
     // FIXME: these flags should reflect the actual attachments present
     rt->targets = TargetBufferFlags::COLOR0 | TargetBufferFlags::DEPTH;
@@ -1246,6 +1316,7 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         uint32_t width,
         uint32_t height,
         uint8_t samples,
+        uint8_t layerCount,
         MRT color,
         TargetBufferInfo depth,
         TargetBufferInfo stencil) {
@@ -1312,7 +1383,7 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
             if (any(targets & getTargetBufferFlagsAt(i))) {
                 assert_invariant(color[i].handle);
                 rt->gl.color[i] = handle_cast<GLTexture*>(color[i].handle);
-                framebufferTexture(color[i], rt, GL_COLOR_ATTACHMENT0 + i);
+                framebufferTexture(color[i], rt, GL_COLOR_ATTACHMENT0 + i, layerCount);
                 bufs[i] = GL_COLOR_ATTACHMENT0 + i;
                 checkDimensions(rt->gl.color[i], color[i].level);
             }
@@ -1334,7 +1405,7 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         // either we supplied only the depth handle or both depth/stencil are identical and not null
         if (depth.handle && (stencil.handle == depth.handle || !stencil.handle)) {
             rt->gl.depth = handle_cast<GLTexture*>(depth.handle);
-            framebufferTexture(depth, rt, GL_DEPTH_STENCIL_ATTACHMENT);
+            framebufferTexture(depth, rt, GL_DEPTH_STENCIL_ATTACHMENT, layerCount);
             specialCased = true;
             checkDimensions(rt->gl.depth, depth.level);
         }
@@ -1345,13 +1416,13 @@ void OpenGLDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         if (any(targets & TargetBufferFlags::DEPTH)) {
             assert_invariant(depth.handle);
             rt->gl.depth = handle_cast<GLTexture*>(depth.handle);
-            framebufferTexture(depth, rt, GL_DEPTH_ATTACHMENT);
+            framebufferTexture(depth, rt, GL_DEPTH_ATTACHMENT, layerCount);
             checkDimensions(rt->gl.depth, depth.level);
         }
         if (any(targets & TargetBufferFlags::STENCIL)) {
             assert_invariant(stencil.handle);
             rt->gl.stencil = handle_cast<GLTexture*>(stencil.handle);
-            framebufferTexture(stencil, rt, GL_STENCIL_ATTACHMENT);
+            framebufferTexture(stencil, rt, GL_STENCIL_ATTACHMENT, layerCount);
             checkDimensions(rt->gl.stencil, stencil.level);
         }
     }
@@ -1393,6 +1464,13 @@ void OpenGLDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
     GLSwapChain* sc = handle_cast<GLSwapChain*>(sch);
     sc->swapChain = mPlatform.createSwapChain(nativeWindow, flags);
 
+#if !defined(__EMSCRIPTEN__)
+    // note: in practice this should never happen on Android
+    ASSERT_POSTCONDITION(sc->swapChain,
+            "createSwapChain(%p, 0x%lx) failed. See logs for details.",
+            nativeWindow, flags);
+#endif
+
     // See if we need the emulated rec709 output conversion
     if (UTILS_UNLIKELY(mContext.isES2())) {
         sc->rec709 = (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE &&
@@ -1407,6 +1485,13 @@ void OpenGLDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch,
     GLSwapChain* sc = handle_cast<GLSwapChain*>(sch);
     sc->swapChain = mPlatform.createSwapChain(width, height, flags);
 
+#if !defined(__EMSCRIPTEN__)
+    // note: in practice this should never happen on Android
+    ASSERT_POSTCONDITION(sc->swapChain,
+            "createSwapChainHeadless(%u, %u, 0x%lx) failed. See logs for details.",
+            width, height, flags);
+#endif
+
     // See if we need the emulated rec709 output conversion
     if (UTILS_UNLIKELY(mContext.isES2())) {
         sc->rec709 = (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE &&
@@ -1417,7 +1502,7 @@ void OpenGLDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch,
 void OpenGLDriver::createTimerQueryR(Handle<HwTimerQuery> tqh, int) {
     DEBUG_MARKER()
     GLTimerQuery* tq = handle_cast<GLTimerQuery*>(tqh);
-    mTimerQueryImpl->createTimerQuery(tq);
+    mContext.createTimerQuery(tq);
 }
 
 // ------------------------------------------------------------------------------------------------
@@ -1471,7 +1556,20 @@ void OpenGLDriver::destroyRenderPrimitive(Handle<HwRenderPrimitive> rph) {
     if (rph) {
         auto& gl = mContext;
         GLRenderPrimitive const* rp = handle_cast<const GLRenderPrimitive*>(rph);
-        gl.deleteVertexArrays(1, &rp->gl.vao);
+        gl.deleteVertexArray(rp->gl.vao[gl.contextIndex]);
+
+        // If we have a name in the "other" context, we need to schedule the destroy for
+        // later, because it can't be done here. VAOs are "container objects" and are not
+        // shared between contexts.
+        size_t const otherContextIndex = 1 - gl.contextIndex;
+        GLuint const nameInOtherContext = rp->gl.vao[otherContextIndex];
+        if (UTILS_UNLIKELY(nameInOtherContext)) {
+            gl.destroyWithContext(otherContextIndex,
+                    [name = nameInOtherContext](OpenGLContext& gl) {
+                gl.deleteVertexArray(name);
+            });
+        }
+
         destruct(rph, rp);
     }
 }
@@ -1536,11 +1634,11 @@ void OpenGLDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
         GLRenderTarget* rt = handle_cast<GLRenderTarget*>(rth);
         if (rt->gl.fbo) {
             // first unbind this framebuffer if needed
-            gl.bindFramebuffer(GL_FRAMEBUFFER, 0);
+            gl.unbindFramebuffer(GL_FRAMEBUFFER);
         }
         if (rt->gl.fbo_read) {
             // first unbind this framebuffer if needed
-            gl.bindFramebuffer(GL_FRAMEBUFFER, 0);
+            gl.unbindFramebuffer(GL_FRAMEBUFFER);
         }
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
@@ -1612,7 +1710,7 @@ void OpenGLDriver::destroyTimerQuery(Handle<HwTimerQuery> tqh) {
 
     if (tqh) {
         GLTimerQuery* tq = handle_cast<GLTimerQuery*>(tqh);
-        mTimerQueryImpl->destroyTimerQuery(tq);
+        mContext.destroyTimerQuery(tq);
         destruct(tqh, tq);
     }
 }
@@ -1903,7 +2001,7 @@ bool OpenGLDriver::isFrameBufferFetchMultiSampleSupported() {
 }
 
 bool OpenGLDriver::isFrameTimeSupported() {
-    return OpenGLTimerQueryFactory::isGpuTimeSupported();
+    return TimerQueryFactory::isGpuTimeSupported();
 }
 
 bool OpenGLDriver::isAutoDepthResolveSupported() {
@@ -1921,6 +2019,10 @@ bool OpenGLDriver::isSRGBSwapChainSupported() {
     return mPlatform.isSRGBSwapChainSupported();
 }
 
+bool OpenGLDriver::isProtectedContentSupported() {
+    return mPlatform.isProtectedContextSupported();
+}
+
 bool OpenGLDriver::isStereoSupported(backend::StereoscopicType stereoscopicType) {
     // Instanced-stereo requires instancing and EXT_clip_cull_distance.
     // Multiview-stereo requires ES 3.0 and OVR_multiview2.
@@ -1945,6 +2047,10 @@ bool OpenGLDriver::isDepthStencilResolveSupported() {
     return true;
 }
 
+bool OpenGLDriver::isProtectedTexturesSupported() {
+    return getContext().ext.EXT_protected_textures;
+}
+
 bool OpenGLDriver::isWorkaroundNeeded(Workaround workaround) {
     switch (workaround) {
         case Workaround::SPLIT_EASU:
@@ -2012,7 +2118,18 @@ void OpenGLDriver::makeCurrent(Handle<HwSwapChain> schDraw, Handle<HwSwapChain>
 
     GLSwapChain* scDraw = handle_cast<GLSwapChain*>(schDraw);
     GLSwapChain* scRead = handle_cast<GLSwapChain*>(schRead);
-    mPlatform.makeCurrent(scDraw->swapChain, scRead->swapChain);
+
+    mPlatform.makeCurrent(scDraw->swapChain, scRead->swapChain,
+            [this]() {
+                // OpenGL context is about to change, unbind everything
+                mContext.unbindEverything();
+            },
+            [this](size_t index) {
+                // OpenGL context has changed, resynchronize the state with the cache
+                mContext.synchronizeStateAndCache(index);
+                slog.d << "*** OpenGL context change : " << (index ? "protected" : "default") << io::endl;
+            });
+
     mCurrentDrawSwapChain = scDraw;
 
     // From the GL spec for glViewport and glScissor:
@@ -2355,7 +2472,9 @@ void OpenGLDriver::setTextureData(GLTexture* t, uint32_t level,
     size_t const stride = p.stride ? p.stride : width;
     size_t const bpp = PBD::computeDataSize(p.format, p.type, 1, 1, 1);
     size_t const bpr = PBD::computeDataSize(p.format, p.type, stride, 1, p.alignment);
-    void const* const buffer = static_cast<char const*>(p.buffer) + p.left * bpp + bpr * p.top;
+    size_t const bpl = bpr * height; // TODO: PBD should have a "layer stride"
+    void const* const buffer = static_cast<char const*>(p.buffer)
+            + bpp* p.left + bpr * p.top + bpl * 0; // TODO: PBD should have a p.depth
 
     switch (t->target) {
         case SamplerType::SAMPLER_EXTERNAL:
@@ -2420,7 +2539,7 @@ void OpenGLDriver::setTextureData(GLTexture* t, uint32_t level,
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     if (!gl.isES2()) {
-        // update the base/max LOD, so we don't access undefined LOD. this allows the app to
+        // Update the base/max LOD, so we don't access undefined LOD. this allows the app to
         // specify levels as they become available.
         if (int8_t(level) < t->gl.baseLevel) {
             t->gl.baseLevel = int8_t(level);
@@ -2521,7 +2640,7 @@ void OpenGLDriver::setCompressedTextureData(GLTexture* t, uint32_t level,
 
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     if (!gl.isES2()) {
-        // update the base/max LOD, so we don't access undefined LOD. this allows the app to
+        // Update the base/max LOD, so we don't access undefined LOD. this allows the app to
         // specify levels as they become available.
         if (int8_t(level) < t->gl.baseLevel) {
             t->gl.baseLevel = int8_t(level);
@@ -2660,21 +2779,21 @@ void OpenGLDriver::replaceStream(GLTexture* texture, GLStream* newStream) noexce
 void OpenGLDriver::beginTimerQuery(Handle<HwTimerQuery> tqh) {
     DEBUG_MARKER()
     GLTimerQuery* tq = handle_cast<GLTimerQuery*>(tqh);
-    mTimerQueryImpl->beginTimeElapsedQuery(tq);
+    mContext.beginTimeElapsedQuery(tq);
 }
 
 void OpenGLDriver::endTimerQuery(Handle<HwTimerQuery> tqh) {
     DEBUG_MARKER()
     GLTimerQuery* tq = handle_cast<GLTimerQuery*>(tqh);
-    mTimerQueryImpl->endTimeElapsedQuery(tq);
+    mContext.endTimeElapsedQuery(*this, tq);
 }
 
-bool OpenGLDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapsedTime) {
+TimerQueryResult OpenGLDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapsedTime) {
     GLTimerQuery* tq = handle_cast<GLTimerQuery*>(tqh);
-    return OpenGLTimerQueryInterface::getTimerQueryValue(tq, elapsedTime);
+    return TimerQueryFactoryInterface::getTimerQueryValue(tq, elapsedTime);
 }
 
-void OpenGLDriver::compilePrograms(CompilerPriorityQueue priority,
+void OpenGLDriver::compilePrograms(CompilerPriorityQueue,
         CallbackHandler* handler, CallbackHandler::Callback callback, void* user) {
     if (callback) {
         getShaderCompilerService().notifyWhenAllProgramsAreReady(handler, callback, user);
@@ -2702,13 +2821,13 @@ void OpenGLDriver::beginRenderPass(Handle<HwRenderTarget> rth,
     const TargetBufferFlags clearFlags = params.flags.clear & rt->targets;
     TargetBufferFlags discardFlags = params.flags.discardStart & rt->targets;
 
-    gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
+    GLuint const fbo = gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
     CHECK_GL_FRAMEBUFFER_STATUS(utils::slog.e, GL_FRAMEBUFFER)
 
     if (gl.ext.EXT_discard_framebuffer
             && !gl.bugs.disable_invalidate_framebuffer) {
         AttachmentArray attachments; // NOLINT
-        GLsizei const attachmentCount = getAttachments(attachments, rt, discardFlags);
+        GLsizei const attachmentCount = getAttachments(attachments, discardFlags, !fbo);
         if (attachmentCount) {
             gl.procs.invalidateFramebuffer(GL_FRAMEBUFFER, attachmentCount, attachments.data());
         }
@@ -2796,9 +2915,9 @@ void OpenGLDriver::endRenderPass(int) {
         }
         if (!gl.bugs.disable_invalidate_framebuffer) {
             // we wouldn't have to bind the framebuffer if we had glInvalidateNamedFramebuffer()
-            gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
+            GLuint const fbo = gl.bindFramebuffer(GL_FRAMEBUFFER, rt->gl.fbo);
             AttachmentArray attachments; // NOLINT
-            GLsizei const attachmentCount = getAttachments(attachments, rt, effectiveDiscardFlags);
+            GLsizei const attachmentCount = getAttachments(attachments, effectiveDiscardFlags, !fbo);
             if (attachmentCount) {
                 gl.procs.invalidateFramebuffer(GL_FRAMEBUFFER, attachmentCount, attachments.data());
             }
@@ -2860,50 +2979,48 @@ void OpenGLDriver::resolvePass(ResolveAction action, GLRenderTarget const* rt,
 }
 
 GLsizei OpenGLDriver::getAttachments(AttachmentArray& attachments,
-        GLRenderTarget const* rt, TargetBufferFlags buffers) noexcept {
-    assert_invariant(buffers <= rt->targets);
-
+        TargetBufferFlags buffers, bool isDefaultFramebuffer) noexcept {
     GLsizei attachmentCount = 0;
     // the default framebuffer uses different constants!!!
-    const bool defaultFramebuffer = (rt->gl.fbo == 0);
+
     if (any(buffers & TargetBufferFlags::COLOR0)) {
-        attachments[attachmentCount++] = defaultFramebuffer ? GL_COLOR : GL_COLOR_ATTACHMENT0;
+        attachments[attachmentCount++] = isDefaultFramebuffer ? GL_COLOR : GL_COLOR_ATTACHMENT0;
     }
 #ifndef FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2
     if (any(buffers & TargetBufferFlags::COLOR1)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT1;
     }
     if (any(buffers & TargetBufferFlags::COLOR2)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT2;
     }
     if (any(buffers & TargetBufferFlags::COLOR3)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT3;
     }
     if (any(buffers & TargetBufferFlags::COLOR4)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT4;
     }
     if (any(buffers & TargetBufferFlags::COLOR5)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT5;
     }
     if (any(buffers & TargetBufferFlags::COLOR6)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT6;
     }
     if (any(buffers & TargetBufferFlags::COLOR7)) {
-        assert_invariant(!defaultFramebuffer);
+        assert_invariant(!isDefaultFramebuffer);
         attachments[attachmentCount++] = GL_COLOR_ATTACHMENT7;
     }
 #endif
     if (any(buffers & TargetBufferFlags::DEPTH)) {
-        attachments[attachmentCount++] = defaultFramebuffer ? GL_DEPTH : GL_DEPTH_ATTACHMENT;
+        attachments[attachmentCount++] = isDefaultFramebuffer ? GL_DEPTH : GL_DEPTH_ATTACHMENT;
     }
     if (any(buffers & TargetBufferFlags::STENCIL)) {
-        attachments[attachmentCount++] = defaultFramebuffer ? GL_STENCIL : GL_STENCIL_ATTACHMENT;
+        attachments[attachmentCount++] = isDefaultFramebuffer ? GL_STENCIL : GL_STENCIL_ATTACHMENT;
     }
     return attachmentCount;
 }
@@ -3623,8 +3740,8 @@ void OpenGLDriver::blit(
             mask, GL_NEAREST);
     CHECK_GL_ERROR(utils::slog.e)
 
-    gl.bindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
-    gl.bindFramebuffer(GL_READ_FRAMEBUFFER, 0);
+    gl.unbindFramebuffer(GL_DRAW_FRAMEBUFFER);
+    gl.unbindFramebuffer(GL_READ_FRAMEBUFFER);
     glDeleteFramebuffers(2, fbo);
 
     if (any(d->usage & TextureUsage::SAMPLEABLE)) {
@@ -3728,42 +3845,43 @@ void OpenGLDriver::updateTextureLodRange(GLTexture* texture, int8_t targetLevel)
 #endif
 }
 
-void OpenGLDriver::draw(PipelineState state, Handle<HwRenderPrimitive> rph,
-        uint32_t const indexOffset, uint32_t const indexCount, uint32_t const instanceCount) {
+void OpenGLDriver::bindPipeline(PipelineState state) {
     DEBUG_MARKER()
     auto& gl = mContext;
-
+    setRasterState(state.rasterState);
+    setStencilState(state.stencilState);
+    gl.polygonOffset(state.polygonOffset.slope, state.polygonOffset.constant);
     OpenGLProgram* const p = handle_cast<OpenGLProgram*>(state.program);
+    mValidProgram = useProgram(p);
+}
 
-    bool const success = useProgram(p);
-    if (UTILS_UNLIKELY(!success)) {
-        // Avoid fatal (or cascading) errors that can occur during the draw call when the program
-        // is invalid. The shader compile error has already been dumped to the console at this
-        // point, so it's fine to simply return early.
-        return;
-    }
+void OpenGLDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
+    DEBUG_MARKER()
+    auto& gl = mContext;
 
     GLRenderPrimitive* const rp = handle_cast<GLRenderPrimitive*>(rph);
 
     // Gracefully do nothing if the render primitive has not been set up.
     VertexBufferHandle vb = rp->gl.vertexBufferWithObjects;
     if (UTILS_UNLIKELY(!vb)) {
+        mBoundRenderPrimitive = nullptr;
         return;
     }
 
-    gl.bindVertexArray(&rp->gl);
-
     // If necessary, mutate the bindings in the VAO.
+    gl.bindVertexArray(&rp->gl);
     GLVertexBuffer const* const glvb = handle_cast<GLVertexBuffer*>(vb);
-    if (UTILS_UNLIKELY(rp->gl.vertexBufferVersion != glvb->bufferObjectsVersion)) {
-        updateVertexArrayObject(rp, glvb);
+    updateVertexArrayObject(rp, glvb);
+
+    mBoundRenderPrimitive = rp;
+}
+
+void OpenGLDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
+    GLRenderPrimitive const* const rp = mBoundRenderPrimitive;
+    if (UTILS_UNLIKELY(!rp || !mValidProgram)) {
+        return;
     }
 
-    setRasterState(state.rasterState);
-    setStencilState(state.stencilState);
-
-    gl.polygonOffset(state.polygonOffset.slope, state.polygonOffset.constant);
-
     if (UTILS_LIKELY(instanceCount <= 1)) {
         glDrawElements(GLenum(rp->type), (GLsizei)indexCount, rp->gl.getIndicesType(),
                 reinterpret_cast<const void*>(indexOffset * rp->gl.indicesSize));
@@ -3788,6 +3906,17 @@ void OpenGLDriver::scissor(Viewport scissor) {
     setScissor(scissor);
 }
 
+void OpenGLDriver::draw(PipelineState state, Handle<HwRenderPrimitive> rph,
+        uint32_t const indexOffset, uint32_t const indexCount, uint32_t const instanceCount) {
+    DEBUG_MARKER()
+    GLRenderPrimitive* const rp = handle_cast<GLRenderPrimitive*>(rph);
+    state.primitiveType = rp->type;
+    state.vertexBufferInfo = rp->vbih;
+    bindPipeline(state);
+    bindRenderPrimitive(rph);
+    draw2(indexOffset, indexCount, instanceCount);
+}
+
 void OpenGLDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGroupCount) {
     getShaderCompilerService().tick();
 

filament/backend/src/opengl/OpenGLDriver.h

@@ -18,27 +18,42 @@
 #define TNT_FILAMENT_BACKEND_OPENGL_OPENGLDRIVER_H
 
 #include "DriverBase.h"
-#include "GLUtils.h"
 #include "OpenGLContext.h"
+#include "OpenGLTimerQuery.h"
 #include "ShaderCompilerService.h"
 
-#include "private/backend/Driver.h"
-#include "private/backend/HandleAllocator.h"
-
 #include <backend/platforms/OpenGLPlatform.h>
 
 #include <backend/AcquiredImage.h>
+#include <backend/DriverEnums.h>
+#include <backend/Handle.h>
+#include <backend/Platform.h>
 #include <backend/Program.h>
 #include <backend/TargetBufferInfo.h>
 
+#include "private/backend/Driver.h"
+#include "private/backend/HandleAllocator.h"
+
+#include <utils/FixedCapacityVector.h>
 #include <utils/compiler.h>
-#include <utils/Allocator.h>
+#include <utils/debug.h>
 
 #include <math/vec4.h>
 
 #include <tsl/robin_map.h>
 
-#include <set>
+#include <array>
+#include <condition_variable>
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include <stddef.h>
+#include <stdint.h>
 
 #ifndef FILAMENT_OPENGL_HANDLE_ARENA_SIZE_IN_MB
 #    define FILAMENT_OPENGL_HANDLE_ARENA_SIZE_IN_MB 4
@@ -51,15 +66,17 @@ class PixelBufferDescriptor;
 struct TargetBufferInfo;
 
 class OpenGLProgram;
-class OpenGLTimerQueryInterface;
+class TimerQueryFactoryInterface;
 
 class OpenGLDriver final : public DriverBase {
-    inline explicit OpenGLDriver(OpenGLPlatform* platform, const Platform::DriverConfig& driverConfig) noexcept;
+    inline explicit OpenGLDriver(OpenGLPlatform* platform,
+            const Platform::DriverConfig& driverConfig) noexcept;
     ~OpenGLDriver() noexcept final;
     Dispatcher getDispatcher() const noexcept final;
 
 public:
-    static Driver* create(OpenGLPlatform* platform, void* sharedGLContext, const Platform::DriverConfig& driverConfig) noexcept;
+    static Driver* create(OpenGLPlatform* platform, void* sharedGLContext,
+            const Platform::DriverConfig& driverConfig) noexcept;
 
     class DebugMarker {
         OpenGLDriver& driver;
@@ -137,6 +154,7 @@ public:
     struct GLRenderPrimitive : public HwRenderPrimitive {
         using HwRenderPrimitive::HwRenderPrimitive;
         OpenGLContext::RenderPrimitive gl;
+        Handle<HwVertexBufferInfo> vbih;
     };
 
     struct GLTexture : public HwTexture {
@@ -161,15 +179,7 @@ public:
         OpenGLPlatform::ExternalTexture* externalTexture = nullptr;
     };
 
-    struct GLTimerQuery : public HwTimerQuery {
-        struct State {
-            struct {
-                GLuint query;
-            } gl;
-            std::atomic<int64_t> elapsed{};
-        };
-        std::shared_ptr<State> state;
-    };
+    using GLTimerQuery = filament::backend::GLTimerQuery;
 
     struct GLStream : public HwStream {
         using HwStream::HwStream;
@@ -223,8 +233,8 @@ private:
     OpenGLContext mContext;
     ShaderCompilerService mShaderCompilerService;
 
-    friend class OpenGLTimerQueryFactory;
-    friend class TimerQueryNative;
+    friend class TimerQueryFactory;
+    friend class TimerQueryNativeFactory;
     OpenGLContext& getContext() noexcept { return mContext; }
 
     ShaderCompilerService& getShaderCompilerService() noexcept {
@@ -304,7 +314,7 @@ private:
     void updateVertexArrayObject(GLRenderPrimitive* rp, GLVertexBuffer const* vb);
 
     void framebufferTexture(TargetBufferInfo const& binfo,
-            GLRenderTarget const* rt, GLenum attachment) noexcept;
+            GLRenderTarget const* rt, GLenum attachment, uint8_t layerCount) noexcept;
 
     void setRasterState(RasterState rs) noexcept;
 
@@ -325,8 +335,8 @@ private:
     void renderBufferStorage(GLuint rbo, GLenum internalformat, uint32_t width,
             uint32_t height, uint8_t samples) const noexcept;
 
-    void textureStorage(GLTexture* t,
-            uint32_t width, uint32_t height, uint32_t depth) noexcept;
+    void textureStorage(OpenGLDriver::GLTexture* t, uint32_t width, uint32_t height,
+            uint32_t depth, bool useProtectedMemory) noexcept;
 
     /* State tracking GL wrappers... */
 
@@ -359,8 +369,8 @@ private:
     }
 
     using AttachmentArray = std::array<GLenum, MRT::MAX_SUPPORTED_RENDER_TARGET_COUNT + 2>;
-    static GLsizei getAttachments(AttachmentArray& attachments,
-            GLRenderTarget const* rt, TargetBufferFlags buffers) noexcept;
+    static GLsizei getAttachments(AttachmentArray& attachments, TargetBufferFlags buffers,
+            bool isDefaultFramebuffer) noexcept;
 
     // state required to represent the current render pass
     Handle<HwRenderTarget> mRenderPassTarget;
@@ -369,6 +379,10 @@ private:
     GLboolean mRenderPassDepthWrite{};
     GLboolean mRenderPassStencilWrite{};
 
+    GLRenderPrimitive const* mBoundRenderPrimitive = nullptr;
+    bool mValidProgram = false;
+
+
     void clearWithRasterPipe(TargetBufferFlags clearFlags,
             math::float4 const& linearColor, GLfloat depth, GLint stencil) noexcept;
 
@@ -411,9 +425,6 @@ private:
     void executeEveryNowAndThenOps() noexcept;
     std::vector<std::function<bool()>> mEveryNowAndThenOps;
 
-    // timer query implementation
-    OpenGLTimerQueryInterface* mTimerQueryImpl = nullptr;
-
     const Platform::DriverConfig mDriverConfig;
     Platform::DriverConfig const& getDriverConfig() const noexcept { return mDriverConfig; }
 

filament/backend/src/opengl/OpenGLPlatform.cpp

@@ -24,7 +24,10 @@
 
 #include <utils/compiler.h>
 
+#include <utils/Invocable.h>
+#include <stddef.h>
 #include <stdint.h>
+#include <stdlib.h>
 
 namespace filament::backend {
 
@@ -35,14 +38,27 @@ Driver* OpenGLPlatform::createDefaultDriver(OpenGLPlatform* platform,
 
 OpenGLPlatform::~OpenGLPlatform() noexcept = default;
 
+void OpenGLPlatform::makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain,
+        utils::Invocable<void()>, utils::Invocable<void(size_t)>) noexcept {
+    makeCurrent(getCurrentContextType(), drawSwapChain, readSwapChain);
+}
+
+bool OpenGLPlatform::isProtectedContextSupported() const noexcept {
+    return false;
+}
+
 bool OpenGLPlatform::isSRGBSwapChainSupported() const noexcept {
     return false;
 }
 
-uint32_t OpenGLPlatform::createDefaultRenderTarget() noexcept {
+uint32_t OpenGLPlatform::getDefaultFramebufferObject() noexcept {
     return 0;
 }
 
+OpenGLPlatform::ContextType OpenGLPlatform::getCurrentContextType() const noexcept {
+    return ContextType::UNPROTECTED;
+}
+
 void OpenGLPlatform::setPresentationTime(
         UTILS_UNUSED int64_t presentationTimeInNanosecond) noexcept {
 }
@@ -113,10 +129,14 @@ AcquiredImage OpenGLPlatform::transformAcquiredImage(AcquiredImage source) noexc
     return source;
 }
 
-TargetBufferFlags OpenGLPlatform::getPreservedFlags(UTILS_UNUSED SwapChain* swapChain) noexcept {
+TargetBufferFlags OpenGLPlatform::getPreservedFlags(UTILS_UNUSED SwapChain*) noexcept {
     return TargetBufferFlags::NONE;
 }
 
+bool OpenGLPlatform::isSwapChainProtected(UTILS_UNUSED SwapChain*) noexcept {
+    return false;
+}
+
 bool OpenGLPlatform::isExtraContextSupported() const noexcept {
     return false;
 }

filament/backend/src/opengl/OpenGLProgram.cpp

@@ -16,16 +16,25 @@
 
 #include "OpenGLProgram.h"
 
-#include "BlobCacheKey.h"
+#include "GLUtils.h"
 #include "OpenGLDriver.h"
 #include "ShaderCompilerService.h"
 
+#include <backend/Program.h>
+
+#include <private/backend/BackendUtils.h>
+
 #include <utils/debug.h>
 #include <utils/compiler.h>
 #include <utils/Log.h>
 #include <utils/Systrace.h>
 
-#include <private/backend/BackendUtils.h>
+#include <array>
+#include <string_view>
+#include <utility>
+#include <new>
+
+#include <stddef.h>
 
 namespace filament::backend {
 

filament/backend/src/opengl/OpenGLProgram.h

@@ -28,6 +28,9 @@
 #include <utils/compiler.h>
 #include <utils/FixedCapacityVector.h>
 
+#include <array>
+#include <limits>
+
 #include <stddef.h>
 #include <stdint.h>
 

filament/backend/src/opengl/OpenGLTimerQuery.cpp

@@ -16,110 +16,132 @@
 
 #include "OpenGLTimerQuery.h"
 
+#include "GLUtils.h"
+#include "OpenGLDriver.h"
+
+#include <backend/Platform.h>
 #include <backend/platforms/OpenGLPlatform.h>
+#include <backend/DriverEnums.h>
 
 #include <utils/compiler.h>
+#include <utils/debug.h>
 #include <utils/JobSystem.h>
 #include <utils/Log.h>
+#include <utils/Mutex.h>
 #include <utils/Systrace.h>
-#include <utils/debug.h>
+
+#include <atomic>
+#include <memory>
+#include <mutex>
+#include <new>
+#include <utility>
+
+#include <stdint.h>
 
 namespace filament::backend {
 
 using namespace backend;
 using namespace GLUtils;
 
+class OpenGLDriver;
+
 // ------------------------------------------------------------------------------------------------
 
-bool OpenGLTimerQueryFactory::mGpuTimeSupported = false;
+bool TimerQueryFactory::mGpuTimeSupported = false;
 
-OpenGLTimerQueryInterface* OpenGLTimerQueryFactory::init(
-        OpenGLPlatform& platform, OpenGLDriver& driver) noexcept {
-    (void)driver;
+TimerQueryFactoryInterface* TimerQueryFactory::init(
+        OpenGLPlatform& platform, OpenGLContext& context) noexcept {
+    (void)context;
 
-    OpenGLTimerQueryInterface* impl;
+    TimerQueryFactoryInterface* impl = nullptr;
 
 #if defined(BACKEND_OPENGL_VERSION_GL) || defined(GL_EXT_disjoint_timer_query)
-    auto& context = driver.getContext();
     if (context.ext.EXT_disjoint_timer_query) {
         // timer queries are available
         if (context.bugs.dont_use_timer_query && platform.canCreateFence()) {
             // however, they don't work well, revert to using fences if we can.
-            impl = new(std::nothrow) OpenGLTimerQueryFence(platform);
+            impl = new(std::nothrow) TimerQueryFenceFactory(platform);
         } else {
-            impl = new(std::nothrow) TimerQueryNative(driver);
+            impl = new(std::nothrow) TimerQueryNativeFactory(context);
         }
         mGpuTimeSupported = true;
     } else
 #endif
     if (platform.canCreateFence()) {
         // no timer queries, but we can use fences
-        impl = new(std::nothrow) OpenGLTimerQueryFence(platform);
+        impl = new(std::nothrow) TimerQueryFenceFactory(platform);
         mGpuTimeSupported = true;
     } else {
         // no queries, no fences -- that's a problem
-        impl = new(std::nothrow) TimerQueryFallback();
+        impl = new(std::nothrow) TimerQueryFallbackFactory();
         mGpuTimeSupported = false;
     }
+    assert_invariant(impl);
     return impl;
 }
 
 // ------------------------------------------------------------------------------------------------
 
-OpenGLTimerQueryInterface::~OpenGLTimerQueryInterface() = default;
+TimerQueryFactoryInterface::~TimerQueryFactoryInterface() = default;
 
 // This is a backend synchronous call
-bool OpenGLTimerQueryInterface::getTimerQueryValue(GLTimerQuery* tq, uint64_t* elapsedTime) noexcept {
+TimerQueryResult TimerQueryFactoryInterface::getTimerQueryValue(
+        GLTimerQuery* tq, uint64_t* elapsedTime) noexcept {
     if (UTILS_LIKELY(tq->state)) {
         int64_t const elapsed = tq->state->elapsed.load(std::memory_order_relaxed);
-        bool const available = elapsed > 0;
-        if (available) {
+        if (elapsed > 0) {
             *elapsedTime = elapsed;
+            return TimerQueryResult::AVAILABLE;
         }
-        return available;
+        return TimerQueryResult(elapsed);
     }
-    return false;
+    return TimerQueryResult::ERROR;
 }
 
 // ------------------------------------------------------------------------------------------------
 
 #if defined(BACKEND_OPENGL_VERSION_GL) || defined(GL_EXT_disjoint_timer_query)
 
-TimerQueryNative::TimerQueryNative(OpenGLDriver& driver)
-        : mDriver(driver) {
+TimerQueryNativeFactory::TimerQueryNativeFactory(OpenGLContext& context)
+        : mContext(context) {
 }
 
-TimerQueryNative::~TimerQueryNative() = default;
+TimerQueryNativeFactory::~TimerQueryNativeFactory() = default;
 
-void TimerQueryNative::createTimerQuery(GLTimerQuery* tq) {
-    if (UTILS_UNLIKELY(!tq->state)) {
-        tq->state = std::make_shared<GLTimerQuery::State>();
-    }
-    mDriver.getContext().procs.genQueries(1u, &tq->state->gl.query);
+void TimerQueryNativeFactory::createTimerQuery(GLTimerQuery* tq) {
+    assert_invariant(!tq->state);
+
+    tq->state = std::make_shared<GLTimerQuery::State>();
+    mContext.procs.genQueries(1u, &tq->state->gl.query);
     CHECK_GL_ERROR(utils::slog.e)
 }
 
-void TimerQueryNative::destroyTimerQuery(GLTimerQuery* tq) {
+void TimerQueryNativeFactory::destroyTimerQuery(GLTimerQuery* tq) {
     assert_invariant(tq->state);
-    mDriver.getContext().procs.deleteQueries(1u, &tq->state->gl.query);
+
+    mContext.procs.deleteQueries(1u, &tq->state->gl.query);
+    CHECK_GL_ERROR(utils::slog.e)
+
+    tq->state.reset();
+}
+
+void TimerQueryNativeFactory::beginTimeElapsedQuery(GLTimerQuery* tq) {
+    assert_invariant(tq->state);
+
+    tq->state->elapsed.store(int64_t(TimerQueryResult::NOT_READY), std::memory_order_relaxed);
+    mContext.procs.beginQuery(GL_TIME_ELAPSED, tq->state->gl.query);
     CHECK_GL_ERROR(utils::slog.e)
 }
 
-void TimerQueryNative::beginTimeElapsedQuery(GLTimerQuery* tq) {
+void TimerQueryNativeFactory::endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQuery* tq) {
     assert_invariant(tq->state);
-    tq->state->elapsed.store(0);
-    mDriver.getContext().procs.beginQuery(GL_TIME_ELAPSED, tq->state->gl.query);
-    CHECK_GL_ERROR(utils::slog.e)
-}
 
-void TimerQueryNative::endTimeElapsedQuery(GLTimerQuery* tq) {
-    assert_invariant(tq->state);
-    mDriver.getContext().procs.endQuery(GL_TIME_ELAPSED);
+    mContext.procs.endQuery(GL_TIME_ELAPSED);
     CHECK_GL_ERROR(utils::slog.e)
 
     std::weak_ptr<GLTimerQuery::State> const weak = tq->state;
 
-    mDriver.runEveryNowAndThen([context = mDriver.getContext(), weak]() -> bool {
+    driver.runEveryNowAndThen([&context = mContext, weak]() -> bool {
         auto state = weak.lock();
         if (state) {
             GLuint available = 0;
@@ -133,6 +155,8 @@ void TimerQueryNative::endTimeElapsedQuery(GLTimerQuery* tq) {
             // we won't end-up here if we're on ES and don't have GL_EXT_disjoint_timer_query
             context.procs.getQueryObjectui64v(state->gl.query, GL_QUERY_RESULT, &elapsedTime);
             state->elapsed.store((int64_t)elapsedTime, std::memory_order_relaxed);
+        } else {
+            state->elapsed.store(int64_t(TimerQueryResult::ERROR), std::memory_order_relaxed);
         }
         return true;
     });
@@ -142,7 +166,7 @@ void TimerQueryNative::endTimeElapsedQuery(GLTimerQuery* tq) {
 
 // ------------------------------------------------------------------------------------------------
 
-OpenGLTimerQueryFence::OpenGLTimerQueryFence(OpenGLPlatform& platform)
+TimerQueryFenceFactory::TimerQueryFenceFactory(OpenGLPlatform& platform)
         : mPlatform(platform) {
     mQueue.reserve(2);
     mThread = std::thread([this]() {
@@ -166,7 +190,8 @@ OpenGLTimerQueryFence::OpenGLTimerQueryFence(OpenGLPlatform& platform)
     });
 }
 
-OpenGLTimerQueryFence::~OpenGLTimerQueryFence() {
+TimerQueryFenceFactory::~TimerQueryFenceFactory() {
+    assert_invariant(mQueue.empty());
     if (mThread.joinable()) {
         std::unique_lock<utils::Mutex> lock(mLock);
         mExitRequested = true;
@@ -178,27 +203,26 @@ OpenGLTimerQueryFence::~OpenGLTimerQueryFence() {
     }
 }
 
-void OpenGLTimerQueryFence::enqueue(OpenGLTimerQueryFence::Job&& job) {
+void TimerQueryFenceFactory::push(TimerQueryFenceFactory::Job&& job) {
     std::unique_lock<utils::Mutex> const lock(mLock);
-    mQueue.push_back(std::forward<Job>(job));
+    mQueue.push_back(std::move(job));
     mCondition.notify_one();
 }
 
-void OpenGLTimerQueryFence::createTimerQuery(GLTimerQuery* tq) {
-    if (UTILS_UNLIKELY(!tq->state)) {
-        tq->state = std::make_shared<GLTimerQuery::State>();
-    }
+void TimerQueryFenceFactory::createTimerQuery(GLTimerQuery* tq) {
+    assert_invariant(!tq->state);
+    tq->state = std::make_shared<GLTimerQuery::State>();
 }
 
-void OpenGLTimerQueryFence::destroyTimerQuery(GLTimerQuery* tq) {
+void TimerQueryFenceFactory::destroyTimerQuery(GLTimerQuery* tq) {
     assert_invariant(tq->state);
+    tq->state.reset();
 }
 
-void OpenGLTimerQueryFence::beginTimeElapsedQuery(GLTimerQuery* tq) {
+void TimerQueryFenceFactory::beginTimeElapsedQuery(GLTimerQuery* tq) {
     assert_invariant(tq->state);
-    tq->state->elapsed.store(0);
+    tq->state->elapsed.store(int64_t(TimerQueryResult::NOT_READY), std::memory_order_relaxed);
 
-    Platform::Fence* fence = mPlatform.createFence();
     std::weak_ptr<GLTimerQuery::State> const weak = tq->state;
 
     // FIXME: this implementation of beginTimeElapsedQuery is usually wrong; it ends up
@@ -207,12 +231,11 @@ void OpenGLTimerQueryFence::beginTimeElapsedQuery(GLTimerQuery* tq) {
     //    on a dummy target for instance, or somehow latch the begin time at the next renderpass
     //    start.
 
-    push([&platform = mPlatform, fence, weak]() {
+    push([&platform = mPlatform, fence = mPlatform.createFence(), weak]() {
         auto state = weak.lock();
         if (state) {
             platform.waitFence(fence, FENCE_WAIT_FOR_EVER);
-            int64_t const then = clock::now().time_since_epoch().count();
-            state->elapsed.store(-then, std::memory_order_relaxed);
+            state->then = clock::now().time_since_epoch().count();
             SYSTRACE_CONTEXT();
             SYSTRACE_ASYNC_BEGIN("OpenGLTimerQueryFence", intptr_t(state.get()));
         }
@@ -220,19 +243,16 @@ void OpenGLTimerQueryFence::beginTimeElapsedQuery(GLTimerQuery* tq) {
     });
 }
 
-void OpenGLTimerQueryFence::endTimeElapsedQuery(GLTimerQuery* tq) {
+void TimerQueryFenceFactory::endTimeElapsedQuery(OpenGLDriver&, GLTimerQuery* tq) {
     assert_invariant(tq->state);
-    Platform::Fence* fence = mPlatform.createFence();
     std::weak_ptr<GLTimerQuery::State> const weak = tq->state;
 
-    push([&platform = mPlatform, fence, weak]() {
+    push([&platform = mPlatform, fence = mPlatform.createFence(), weak]() {
         auto state = weak.lock();
         if (state) {
             platform.waitFence(fence, FENCE_WAIT_FOR_EVER);
             int64_t const now = clock::now().time_since_epoch().count();
-            int64_t const then = state->elapsed.load(std::memory_order_relaxed);
-            assert_invariant(then < 0);
-            state->elapsed.store(now + then, std::memory_order_relaxed);
+            state->elapsed.store(now - state->then, std::memory_order_relaxed);
             SYSTRACE_CONTEXT();
             SYSTRACE_ASYNC_END("OpenGLTimerQueryFence", intptr_t(state.get()));
         }
@@ -242,34 +262,32 @@ void OpenGLTimerQueryFence::endTimeElapsedQuery(GLTimerQuery* tq) {
 
 // ------------------------------------------------------------------------------------------------
 
-TimerQueryFallback::TimerQueryFallback() = default;
+TimerQueryFallbackFactory::TimerQueryFallbackFactory() = default;
 
-TimerQueryFallback::~TimerQueryFallback() = default;
+TimerQueryFallbackFactory::~TimerQueryFallbackFactory() = default;
 
-void TimerQueryFallback::createTimerQuery(GLTimerQuery* tq) {
-    if (UTILS_UNLIKELY(!tq->state)) {
-        tq->state = std::make_shared<GLTimerQuery::State>();
-    }
+void TimerQueryFallbackFactory::createTimerQuery(GLTimerQuery* tq) {
+    assert_invariant(!tq->state);
+    tq->state = std::make_shared<GLTimerQuery::State>();
 }
 
-void TimerQueryFallback::destroyTimerQuery(GLTimerQuery* tq) {
+void TimerQueryFallbackFactory::destroyTimerQuery(GLTimerQuery* tq) {
     assert_invariant(tq->state);
+    tq->state.reset();
 }
 
-void TimerQueryFallback::beginTimeElapsedQuery(OpenGLTimerQueryInterface::GLTimerQuery* tq) {
+void TimerQueryFallbackFactory::beginTimeElapsedQuery(GLTimerQuery* tq) {
     assert_invariant(tq->state);
     // this implementation measures the CPU time, but we have no h/w support
-    int64_t const then = clock::now().time_since_epoch().count();
-    tq->state->elapsed.store(-then, std::memory_order_relaxed);
+    tq->state->then = clock::now().time_since_epoch().count();
+    tq->state->elapsed.store(int64_t(TimerQueryResult::NOT_READY), std::memory_order_relaxed);
 }
 
-void TimerQueryFallback::endTimeElapsedQuery(OpenGLTimerQueryInterface::GLTimerQuery* tq) {
+void TimerQueryFallbackFactory::endTimeElapsedQuery(OpenGLDriver&, GLTimerQuery* tq) {
     assert_invariant(tq->state);
     // this implementation measures the CPU time, but we have no h/w support
     int64_t const now = clock::now().time_since_epoch().count();
-    int64_t const then = tq->state->elapsed.load(std::memory_order_relaxed);
-    assert_invariant(then < 0);
-    tq->state->elapsed.store(now + then, std::memory_order_relaxed);
+    tq->state->elapsed.store(now - tq->state->then, std::memory_order_relaxed);
 }
 
 } // namespace filament::backend

filament/backend/src/opengl/OpenGLTimerQuery.h

@@ -17,18 +17,41 @@
 #ifndef TNT_FILAMENT_BACKEND_OPENGL_TIMERQUERY_H
 #define TNT_FILAMENT_BACKEND_OPENGL_TIMERQUERY_H
 
-#include "OpenGLDriver.h"
+#include <backend/DriverEnums.h>
+
+#include "DriverBase.h"
 
 #include <utils/Condition.h>
 #include <utils/Mutex.h>
 
+#include "gl_headers.h"
+
+#include <atomic>
+#include <chrono>
+#include <functional>
+#include <memory>
 #include <thread>
 #include <vector>
 
+#include <stdint.h>
+
 namespace filament::backend {
 
 class OpenGLPlatform;
-class OpenGLTimerQueryInterface;
+class OpenGLContext;
+class OpenGLDriver;
+class TimerQueryFactoryInterface;
+
+struct GLTimerQuery : public HwTimerQuery {
+    struct State {
+        struct {
+            GLuint query;
+        } gl;
+        int64_t then{};
+        std::atomic<int64_t> elapsed{};
+    };
+    std::shared_ptr<State> state;
+};
 
 /*
  * We need two implementation of timer queries (only elapsed time), because
@@ -38,83 +61,80 @@ class OpenGLTimerQueryInterface;
  * These classes implement the various strategies...
  */
 
-
-class OpenGLTimerQueryFactory {
+class TimerQueryFactory {
     static bool mGpuTimeSupported;
 public:
-    static OpenGLTimerQueryInterface* init(
-            OpenGLPlatform& platform, OpenGLDriver& driver) noexcept;
+    static TimerQueryFactoryInterface* init(
+            OpenGLPlatform& platform, OpenGLContext& context) noexcept;
 
     static bool isGpuTimeSupported() noexcept {
         return mGpuTimeSupported;
     }
 };
 
-class OpenGLTimerQueryInterface {
+class TimerQueryFactoryInterface {
 protected:
-    using GLTimerQuery = OpenGLDriver::GLTimerQuery;
+    using GLTimerQuery = filament::backend::GLTimerQuery;
     using clock = std::chrono::steady_clock;
 
 public:
-    virtual ~OpenGLTimerQueryInterface();
+    virtual ~TimerQueryFactoryInterface();
     virtual void createTimerQuery(GLTimerQuery* query) = 0;
     virtual void destroyTimerQuery(GLTimerQuery* query) = 0;
     virtual void beginTimeElapsedQuery(GLTimerQuery* query) = 0;
-    virtual void endTimeElapsedQuery(GLTimerQuery* query) = 0;
+    virtual void endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQuery* query) = 0;
 
-    static bool getTimerQueryValue(GLTimerQuery* tq, uint64_t* elapsedTime) noexcept;
+    static TimerQueryResult getTimerQueryValue(GLTimerQuery* tq, uint64_t* elapsedTime) noexcept;
 };
 
 #if defined(BACKEND_OPENGL_VERSION_GL) || defined(GL_EXT_disjoint_timer_query)
 
-class TimerQueryNative : public OpenGLTimerQueryInterface {
+class TimerQueryNativeFactory final : public TimerQueryFactoryInterface {
 public:
-    explicit TimerQueryNative(OpenGLDriver& driver);
-    ~TimerQueryNative() override;
+    explicit TimerQueryNativeFactory(OpenGLContext& context);
+    ~TimerQueryNativeFactory() override;
 private:
     void createTimerQuery(GLTimerQuery* query) override;
     void destroyTimerQuery(GLTimerQuery* query) override;
     void beginTimeElapsedQuery(GLTimerQuery* query) override;
-    void endTimeElapsedQuery(GLTimerQuery* query) override;
-    OpenGLDriver& mDriver;
+    void endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQuery* query) override;
+    OpenGLContext& mContext;
 };
 
 #endif
 
-class OpenGLTimerQueryFence : public OpenGLTimerQueryInterface {
+class TimerQueryFenceFactory final : public TimerQueryFactoryInterface {
 public:
-    explicit OpenGLTimerQueryFence(OpenGLPlatform& platform);
-    ~OpenGLTimerQueryFence() override;
+    explicit TimerQueryFenceFactory(OpenGLPlatform& platform);
+    ~TimerQueryFenceFactory() override;
 private:
     using Job = std::function<void()>;
+    using Container = std::vector<Job>;
+
     void createTimerQuery(GLTimerQuery* query) override;
     void destroyTimerQuery(GLTimerQuery* query) override;
     void beginTimeElapsedQuery(GLTimerQuery* tq) override;
-    void endTimeElapsedQuery(GLTimerQuery* tq) override;
+    void endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQuery* tq) override;
 
-    void enqueue(Job&& job);
-    template<typename CALLABLE, typename ... ARGS>
-    void push(CALLABLE&& func, ARGS&& ... args) {
-        enqueue(Job(std::bind(std::forward<CALLABLE>(func), std::forward<ARGS>(args)...)));
-    }
+    void push(Job&& job);
 
     OpenGLPlatform& mPlatform;
     std::thread mThread;
     mutable utils::Mutex mLock;
     mutable utils::Condition mCondition;
-    std::vector<Job> mQueue;
+    Container mQueue;
     bool mExitRequested = false;
 };
 
-class TimerQueryFallback : public OpenGLTimerQueryInterface {
+class TimerQueryFallbackFactory final : public TimerQueryFactoryInterface {
 public:
-    explicit TimerQueryFallback();
-    ~TimerQueryFallback() override;
+    explicit TimerQueryFallbackFactory();
+    ~TimerQueryFallbackFactory() override;
 private:
     void createTimerQuery(GLTimerQuery* query) override;
     void destroyTimerQuery(GLTimerQuery* query) override;
     void beginTimeElapsedQuery(GLTimerQuery* query) override;
-    void endTimeElapsedQuery(GLTimerQuery* query) override;
+    void endTimeElapsedQuery(OpenGLDriver& driver, GLTimerQuery* query) override;
 };
 
 } // namespace filament::backend

filament/backend/src/opengl/ShaderCompilerService.cpp

@@ -30,6 +30,7 @@
 #include <utils/Log.h>
 #include <utils/Systrace.h>
 
+#include <cctype>
 #include <chrono>
 #include <string>
 #include <string_view>
@@ -264,6 +265,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
                     compileShaders(gl,
                             std::move(program.getShadersSource()),
                             program.getSpecializationConstants(),
+                            program.isMultiview(),
                             shaders,
                             token->shaderSourceCode);
 
@@ -299,6 +301,7 @@ ShaderCompilerService::program_token_t ShaderCompilerService::createProgram(
         compileShaders(gl,
                 std::move(program.getShadersSource()),
                 program.getSpecializationConstants(),
+                program.isMultiview(),
                 token->gl.shaders,
                 token->shaderSourceCode);
 
@@ -501,6 +504,7 @@ GLuint ShaderCompilerService::initialize(program_token_t& token) noexcept {
 void ShaderCompilerService::compileShaders(OpenGLContext& context,
         Program::ShaderSource shadersSource,
         utils::FixedCapacityVector<Program::SpecializationConstant> const& specializationConstants,
+        bool multiview,
         std::array<GLuint, Program::SHADER_TYPE_COUNT>& outShaders,
         UTILS_UNUSED_IN_RELEASE std::array<CString, Program::SHADER_TYPE_COUNT>& outShaderSourceCode) noexcept {
 
@@ -514,8 +518,16 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
     };
 
     std::string specializationConstantString;
+    int32_t numViews = 2;
     for (auto const& sc : specializationConstants) {
         appendSpecConstantString(specializationConstantString, sc);
+        if (sc.id == 8) {
+            // This constant must match
+            // ReservedSpecializationConstants::CONFIG_STEREO_EYE_COUNT
+            // which we can't use here because it's defined in EngineEnums.h.
+            // (we're breaking layering here, but it's for the good cause).
+            numViews = std::get<int32_t>(sc.value);
+        }
     }
     if (!specializationConstantString.empty()) {
         specializationConstantString += '\n';
@@ -544,17 +556,23 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
 
         if (UTILS_LIKELY(!shadersSource[i].empty())) {
             Program::ShaderBlob& shader = shadersSource[i];
+            char* shader_src = reinterpret_cast<char*>(shader.data());
+            size_t shader_len = shader.size();
 
             // remove GOOGLE_cpp_style_line_directive
-            std::string_view const source = process_GOOGLE_cpp_style_line_directive(context,
-                    reinterpret_cast<char*>(shader.data()), shader.size());
+            process_GOOGLE_cpp_style_line_directive(context, shader_src, shader_len);
+
+            // replace the value of layout(num_views = X) for multiview extension
+            if (multiview && stage == ShaderStage::VERTEX) {
+                process_OVR_multiview2(context, numViews, shader_src, shader_len);
+            }
 
             // add support for ARB_shading_language_packing if needed
             auto const packingFunctions = process_ARB_shading_language_packing(context);
 
             // split shader source, so we can insert the specialization constants and the packing
             // functions
-            auto const [prolog, body] = splitShaderSource(source);
+            auto const [prolog, body] = splitShaderSource({ shader_src, shader_len });
 
             const std::array<const char*, 4> sources = {
                     prolog.data(),
@@ -577,7 +595,7 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
 #ifndef NDEBUG
             // for debugging we return the original shader source (without the modifications we
             // made here), otherwise the line numbers wouldn't match.
-            outShaderSourceCode[i] = { source.data(), source.length() };
+            outShaderSourceCode[i] = { shader_src, shader_len };
 #endif
 
             outShaders[i] = shaderId;
@@ -586,15 +604,59 @@ void ShaderCompilerService::compileShaders(OpenGLContext& context,
 }
 
 // If usages of the Google-style line directive are present, remove them, as some
-// drivers don't allow the quotation marks. This happens in-place.
-std::string_view ShaderCompilerService::process_GOOGLE_cpp_style_line_directive(OpenGLContext& context,
+// drivers don't allow the quotation marks. This source modification happens in-place.
+void ShaderCompilerService::process_GOOGLE_cpp_style_line_directive(OpenGLContext& context,
         char* source, size_t len) noexcept {
     if (!context.ext.GOOGLE_cpp_style_line_directive) {
         if (UTILS_UNLIKELY(requestsGoogleLineDirectivesExtension({ source, len }))) {
             removeGoogleLineDirectives(source, len); // length is unaffected
         }
     }
-    return { source, len };
+}
+
+// Look up the `source` to replace the number of eyes for multiview with the given number. This is
+// necessary for OpenGL because OpenGL relies on the number specified in shader files to determine
+// the number of views, which is assumed as a single digit, for multiview.
+// This source modification happens in-place.
+void ShaderCompilerService::process_OVR_multiview2(OpenGLContext& context,
+        int32_t eyeCount, char* source, size_t len) noexcept {
+    // We don't use regular expression in favor of performance.
+    if (context.ext.OVR_multiview2) {
+        const std::string_view shader{ source, len };
+        const std::string_view layout = "layout";
+        const std::string_view num_views = "num_views";
+        size_t found = 0;
+        while (true) {
+            found = shader.find(layout, found);
+            if (found == std::string_view::npos) {
+                break;
+            }
+            found = shader.find_first_not_of(' ', found + layout.size());
+            if (found == std::string_view::npos || shader[found] != '(') {
+                continue;
+            }
+            found = shader.find_first_not_of(' ', found + 1);
+            if (found == std::string_view::npos) {
+                continue;
+            }
+            if (shader.compare(found, num_views.size(), num_views) != 0) {
+                continue;
+            }
+            found = shader.find_first_not_of(' ', found + num_views.size());
+            if (found == std::string_view::npos || shader[found] != '=') {
+                continue;
+            }
+            found = shader.find_first_not_of(' ', found + 1);
+            if (found == std::string_view::npos) {
+                continue;
+            }
+            // We assume the value should be one-digit number.
+            assert_invariant(eyeCount < 10);
+            assert_invariant(!::isdigit(source[found + 1]));
+            source[found] = '0' + eyeCount;
+            break;
+        }
+    }
 }
 
 // Tragically, OpenGL 4.1 doesn't support unpackHalf2x16 (appeared in 4.2) and

filament/backend/src/opengl/ShaderCompilerService.h

@@ -134,10 +134,14 @@ private:
             OpenGLContext& context,
             Program::ShaderSource shadersSource,
             utils::FixedCapacityVector<Program::SpecializationConstant> const& specializationConstants,
+            bool multiview,
             std::array<GLuint, Program::SHADER_TYPE_COUNT>& outShaders,
             std::array<utils::CString, Program::SHADER_TYPE_COUNT>& outShaderSourceCode) noexcept;
 
-    static std::string_view process_GOOGLE_cpp_style_line_directive(OpenGLContext& context,
+    static void process_GOOGLE_cpp_style_line_directive(OpenGLContext& context,
+            char* source, size_t len) noexcept;
+
+    static void process_OVR_multiview2(OpenGLContext& context, int32_t eyeCount,
             char* source, size_t len) noexcept;
 
     static std::string_view process_ARB_shading_language_packing(OpenGLContext& context) noexcept;

filament/backend/src/opengl/gl_headers.cpp

@@ -67,6 +67,9 @@ PFNGLDISCARDFRAMEBUFFEREXTPROC glDiscardFramebufferEXT;
 #ifdef GL_KHR_parallel_shader_compile
 PFNGLMAXSHADERCOMPILERTHREADSKHRPROC glMaxShaderCompilerThreadsKHR;
 #endif
+#ifdef GL_OVR_multiview
+PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC glFramebufferTextureMultiviewOVR;
+#endif
 
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
 // On Android, If we want to support a build system less than ANDROID_API 21, we need to
@@ -117,6 +120,9 @@ void importGLESExtensionsEntryPoints() {
 #ifdef GL_KHR_parallel_shader_compile
         getProcAddress(glMaxShaderCompilerThreadsKHR, "glMaxShaderCompilerThreadsKHR");
 #endif
+#ifdef GL_OVR_multiview
+        getProcAddress(glFramebufferTextureMultiviewOVR, "glFramebufferTextureMultiviewOVR");
+#endif
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
         getProcAddress(glDispatchCompute, "glDispatchCompute");
 #endif

filament/backend/src/opengl/gl_headers.h

@@ -151,6 +151,9 @@ extern PFNGLDISCARDFRAMEBUFFEREXTPROC glDiscardFramebufferEXT;
 #ifdef GL_KHR_parallel_shader_compile
 extern PFNGLMAXSHADERCOMPILERTHREADSKHRPROC glMaxShaderCompilerThreadsKHR;
 #endif
+#ifdef GL_OVR_multiview
+extern PFNGLFRAMEBUFFERTEXTUREMULTIVIEWOVRPROC glFramebufferTextureMultiviewOVR;
+#endif
 #if defined(__ANDROID__) && !defined(FILAMENT_SILENCE_NOT_SUPPORTED_BY_ES2)
 extern PFNGLDISPATCHCOMPUTEPROC glDispatchCompute;
 #endif

filament/backend/src/opengl/platforms/PlatformCocoaGL.mm

@@ -247,8 +247,8 @@ void PlatformCocoaGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept
     delete cocoaSwapChain;
 }
 
-void PlatformCocoaGL::makeCurrent(Platform::SwapChain* drawSwapChain,
-        Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformCocoaGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
             "ContextManagerCocoa does not support using distinct draw/read swap chains.");
     CocoaGLSwapChain* swapChain = (CocoaGLSwapChain*)drawSwapChain;
@@ -287,6 +287,7 @@ void PlatformCocoaGL::makeCurrent(Platform::SwapChain* drawSwapChain,
 
     swapChain->previousBounds = currentBounds;
     swapChain->previousWindowFrame = currentWindowFrame;
+    return true;
 }
 
 void PlatformCocoaGL::commit(Platform::SwapChain* swapChain) noexcept {

filament/backend/src/opengl/platforms/PlatformCocoaTouchGL.mm

@@ -143,11 +143,12 @@ void PlatformCocoaTouchGL::destroySwapChain(Platform::SwapChain* swapChain) noex
     }
 }
 
-uint32_t PlatformCocoaTouchGL::createDefaultRenderTarget() noexcept {
+uint32_t PlatformCocoaTouchGL::getDefaultFramebufferObject() noexcept {
     return pImpl->mDefaultFramebuffer;
 }
 
-void PlatformCocoaTouchGL::makeCurrent(SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept {
+bool PlatformCocoaTouchGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
             "PlatformCocoaTouchGL does not support using distinct draw/read swap chains.");
     CAEAGLLayer* const glLayer = (__bridge CAEAGLLayer*) drawSwapChain;
@@ -182,6 +183,7 @@ void PlatformCocoaTouchGL::makeCurrent(SwapChain* drawSwapChain, SwapChain* read
         ASSERT_POSTCONDITION(status == GL_FRAMEBUFFER_COMPLETE, "Incomplete framebuffer.");
         glBindFramebuffer(GL_FRAMEBUFFER, oldFramebuffer);
     }
+    return true;
 }
 
 void PlatformCocoaTouchGL::commit(Platform::SwapChain* swapChain) noexcept {

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

@@ -25,13 +25,15 @@
 
 #include <EGL/egl.h>
 #include <EGL/eglext.h>
-#if defined(__ANDROID__)
+#include <EGL/eglplatform.h>
 
+#if defined(__ANDROID__)
 #include <sys/system_properties.h>
 #endif
 #include <utils/compiler.h>
 
 #include <utils/debug.h>
+#include <utils/Invocable.h>
 #include <utils/Log.h>
 #include <utils/ostream.h>
 
@@ -152,6 +154,7 @@ Driver* PlatformEGL::createDriver(void* sharedContext, const Platform::DriverCon
     ext.egl.KHR_create_context = extensions.has("EGL_KHR_create_context");
     ext.egl.KHR_no_config_context = extensions.has("EGL_KHR_no_config_context");
     ext.egl.KHR_surfaceless_context = extensions.has("EGL_KHR_surfaceless_context");
+    ext.egl.EXT_protected_content = extensions.has("EGL_EXT_protected_content");
     if (ext.egl.KHR_create_context) {
         // KHR_create_context implies KHR_surfaceless_context for ES3.x contexts
         ext.egl.KHR_surfaceless_context = true;
@@ -280,12 +283,14 @@ Driver* PlatformEGL::createDriver(void* sharedContext, const Platform::DriverCon
         }
     }
 
-    if (UTILS_UNLIKELY(!makeCurrent(mEGLDummySurface, mEGLDummySurface))) {
+    if (UTILS_UNLIKELY(
+            egl.makeCurrent(mEGLContext, mEGLDummySurface, mEGLDummySurface) == EGL_FALSE)) {
         // eglMakeCurrent failed
         logEglError("eglMakeCurrent");
         goto error;
     }
 
+    mCurrentContextType = ContextType::UNPROTECTED;
     mContextAttribs = std::move(contextAttribs);
 
     initializeGlExtensions();
@@ -304,9 +309,13 @@ error:
     if (mEGLContext) {
         eglDestroyContext(mEGLDisplay, mEGLContext);
     }
+    if (mEGLContextProtected) {
+        eglDestroyContext(mEGLDisplay, mEGLContextProtected);
+    }
 
     mEGLDummySurface = EGL_NO_SURFACE;
     mEGLContext = EGL_NO_CONTEXT;
+    mEGLContextProtected = EGL_NO_CONTEXT;
 
     eglTerminate(mEGLDisplay);
     eglReleaseThread();
@@ -318,6 +327,10 @@ bool PlatformEGL::isExtraContextSupported() const noexcept {
     return ext.egl.KHR_surfaceless_context;
 }
 
+bool PlatformEGL::isProtectedContextSupported() const noexcept {
+    return ext.egl.EXT_protected_content;
+}
+
 void PlatformEGL::createContext(bool shared) {
     EGLConfig config = ext.egl.KHR_no_config_context ? EGL_NO_CONFIG_KHR : mEGLConfig;
 
@@ -352,15 +365,6 @@ void PlatformEGL::releaseContext() noexcept {
     eglReleaseThread();
 }
 
-EGLBoolean PlatformEGL::makeCurrent(EGLSurface drawSurface, EGLSurface readSurface) noexcept {
-    if (UTILS_UNLIKELY((drawSurface != mCurrentDrawSurface || readSurface != mCurrentReadSurface))) {
-        mCurrentDrawSurface = drawSurface;
-        mCurrentReadSurface = readSurface;
-        return eglMakeCurrent(mEGLDisplay, drawSurface, readSurface, mEGLContext);
-    }
-    return EGL_TRUE;
-}
-
 void PlatformEGL::terminate() noexcept {
     // it's always allowed to use EGL_NO_SURFACE, EGL_NO_CONTEXT
     eglMakeCurrent(mEGLDisplay, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
@@ -368,6 +372,9 @@ void PlatformEGL::terminate() noexcept {
         eglDestroySurface(mEGLDisplay, mEGLDummySurface);
     }
     eglDestroyContext(mEGLDisplay, mEGLContext);
+    if (mEGLContextProtected != EGL_NO_CONTEXT) {
+        eglDestroyContext(mEGLDisplay, mEGLContextProtected);
+    }
     for (auto context : mAdditionalContexts) {
         eglDestroyContext(mEGLDisplay, context);
     }
@@ -438,6 +445,8 @@ EGLConfig PlatformEGL::findSwapChainConfig(uint64_t flags, bool window, bool pbu
     return config;
 }
 
+// -----------------------------------------------------------------------------------------------
+
 bool PlatformEGL::isSRGBSwapChainSupported() const noexcept {
     return ext.egl.KHR_gl_colorspace;
 }
@@ -445,6 +454,23 @@ bool PlatformEGL::isSRGBSwapChainSupported() const noexcept {
 Platform::SwapChain* PlatformEGL::createSwapChain(
         void* nativeWindow, uint64_t flags) noexcept {
 
+    Config attribs;
+    if (ext.egl.KHR_gl_colorspace) {
+        if (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) {
+            attribs[EGL_GL_COLORSPACE_KHR] = EGL_GL_COLORSPACE_SRGB_KHR;
+        }
+    } else {
+        flags &= ~SWAP_CHAIN_CONFIG_SRGB_COLORSPACE;
+    }
+
+    if (ext.egl.EXT_protected_content) {
+        if (flags & SWAP_CHAIN_CONFIG_PROTECTED_CONTENT) {
+            attribs[EGL_PROTECTED_CONTENT_EXT] = EGL_TRUE;
+        }
+    } else {
+        flags &= ~SWAP_CHAIN_CONFIG_PROTECTED_CONTENT;
+    }
+
     EGLConfig config = EGL_NO_CONFIG_KHR;
     if (UTILS_LIKELY(ext.egl.KHR_no_config_context)) {
         config = findSwapChainConfig(flags, true, false);
@@ -452,34 +478,27 @@ Platform::SwapChain* PlatformEGL::createSwapChain(
         config = mEGLConfig;
     }
 
-    if (UTILS_UNLIKELY(config == EGL_NO_CONFIG_KHR)) {
-        return nullptr;
-    }
+    EGLSurface sur = EGL_NO_SURFACE;
+    if (UTILS_LIKELY(config != EGL_NO_CONFIG_KHR)) {
+        sur = eglCreateWindowSurface(mEGLDisplay, config,
+                (EGLNativeWindowType)nativeWindow, attribs.data());
 
-    Config attribs;
-
-    if (ext.egl.KHR_gl_colorspace) {
-        if (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) {
-            attribs[EGL_GL_COLORSPACE_KHR] = EGL_GL_COLORSPACE_SRGB_KHR;
+        if (UTILS_LIKELY(sur != EGL_NO_SURFACE)) {
+            // this is not fatal
+            eglSurfaceAttrib(mEGLDisplay, sur, EGL_SWAP_BEHAVIOR, EGL_BUFFER_DESTROYED);
+        } else {
+            logEglError("PlatformEGL::createSwapChain: eglCreateWindowSurface");
         }
+    } else {
+        // error already logged
     }
 
-    EGLSurface sur = eglCreateWindowSurface(mEGLDisplay, config,
-            (EGLNativeWindowType)nativeWindow, attribs.data());
-
-    if (UTILS_UNLIKELY(sur == EGL_NO_SURFACE)) {
-        logEglError("eglCreateWindowSurface");
-        return nullptr;
-    }
-
-    // this is not fatal
-    eglSurfaceAttrib(mEGLDisplay, sur, EGL_SWAP_BEHAVIOR, EGL_BUFFER_DESTROYED);
-
     SwapChainEGL* const sc = new(std::nothrow) SwapChainEGL({
         .sur = sur,
         .attribs = std::move(attribs),
         .nativeWindow = (EGLNativeWindowType)nativeWindow,
-        .config = config
+        .config = config,
+        .flags = flags
     });
     return sc;
 }
@@ -487,17 +506,6 @@ Platform::SwapChain* PlatformEGL::createSwapChain(
 Platform::SwapChain* PlatformEGL::createSwapChain(
         uint32_t width, uint32_t height, uint64_t flags) noexcept {
 
-    EGLConfig config = EGL_NO_CONFIG_KHR;
-    if (UTILS_LIKELY(ext.egl.KHR_no_config_context)) {
-        config = findSwapChainConfig(flags, false, true);
-    } else {
-        config = mEGLConfig;
-    }
-
-    if (UTILS_UNLIKELY(config == EGL_NO_CONFIG_KHR)) {
-        return nullptr;
-    }
-
     Config attribs = {
             { EGL_WIDTH,  EGLint(width) },
             { EGL_HEIGHT, EGLint(height) },
@@ -507,19 +515,40 @@ Platform::SwapChain* PlatformEGL::createSwapChain(
         if (flags & SWAP_CHAIN_CONFIG_SRGB_COLORSPACE) {
             attribs[EGL_GL_COLORSPACE_KHR] = EGL_GL_COLORSPACE_SRGB_KHR;
         }
+    } else {
+        flags &= ~SWAP_CHAIN_CONFIG_SRGB_COLORSPACE;
     }
 
-    EGLSurface sur = eglCreatePbufferSurface(mEGLDisplay, config, attribs.data());
+    if (ext.egl.EXT_protected_content) {
+        if (flags & SWAP_CHAIN_CONFIG_PROTECTED_CONTENT) {
+            attribs[EGL_PROTECTED_CONTENT_EXT] = EGL_TRUE;
+        }
+    } else {
+        flags &= ~SWAP_CHAIN_CONFIG_PROTECTED_CONTENT;
+    }
 
-    if (UTILS_UNLIKELY(sur == EGL_NO_SURFACE)) {
-        logEglError("eglCreatePbufferSurface");
-        return nullptr;
+    EGLConfig config = EGL_NO_CONFIG_KHR;
+    if (UTILS_LIKELY(ext.egl.KHR_no_config_context)) {
+        config = findSwapChainConfig(flags, true, false);
+    } else {
+        config = mEGLConfig;
+    }
+
+    EGLSurface sur = EGL_NO_SURFACE;
+    if (UTILS_LIKELY(config != EGL_NO_CONFIG_KHR)) {
+        sur = eglCreatePbufferSurface(mEGLDisplay, config, attribs.data());
+        if (UTILS_UNLIKELY(sur == EGL_NO_SURFACE)) {
+            logEglError("PlatformEGL::createSwapChain: eglCreatePbufferSurface");
+        }
+    } else {
+        // error already logged
     }
 
     SwapChainEGL* const sc = new(std::nothrow) SwapChainEGL({
             .sur = sur,
             .attribs = std::move(attribs),
-            .config = config
+            .config = config,
+            .flags = flags
     });
     return sc;
 }
@@ -528,21 +557,94 @@ void PlatformEGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
     if (swapChain) {
         SwapChainEGL const* const sc = static_cast<SwapChainEGL const*>(swapChain);
         if (sc->sur != EGL_NO_SURFACE) {
-            makeCurrent(mEGLDummySurface, mEGLDummySurface);
+            egl.makeCurrent(mEGLDummySurface, mEGLDummySurface);
             eglDestroySurface(mEGLDisplay, sc->sur);
             delete sc;
         }
     }
 }
 
-void PlatformEGL::makeCurrent(Platform::SwapChain* drawSwapChain,
-                              Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformEGL::isSwapChainProtected(Platform::SwapChain* swapChain) noexcept {
+    if (swapChain) {
+        SwapChainEGL const* const sc = static_cast<SwapChainEGL const*>(swapChain);
+        return bool(sc->flags & SWAP_CHAIN_CONFIG_PROTECTED_CONTENT);
+    }
+    return false;
+}
+
+OpenGLPlatform::ContextType PlatformEGL::getCurrentContextType() const noexcept {
+    return mCurrentContextType;
+}
+
+bool PlatformEGL::makeCurrent(ContextType type,
+        SwapChain* drawSwapChain, SwapChain* readSwapChain) noexcept {
     SwapChainEGL const* const dsc = static_cast<SwapChainEGL const*>(drawSwapChain);
     SwapChainEGL const* const rsc = static_cast<SwapChainEGL const*>(readSwapChain);
-    if (UTILS_UNLIKELY(dsc->sur == EGL_NO_SURFACE && rsc->sur == EGL_NO_SURFACE)) {
-        return;
+    EGLContext context = getContextForType(type);
+    EGLBoolean const success = egl.makeCurrent(context, dsc->sur, rsc->sur);
+    return success == EGL_TRUE ? true : false;
+}
+
+void PlatformEGL::makeCurrent(Platform::SwapChain* drawSwapChain,
+        Platform::SwapChain* readSwapChain,
+        utils::Invocable<void()> preContextChange,
+        utils::Invocable<void(size_t index)> postContextChange) noexcept {
+
+    assert_invariant(drawSwapChain);
+    assert_invariant(readSwapChain);
+
+    ContextType type = ContextType::UNPROTECTED;
+    if (ext.egl.EXT_protected_content) {
+        bool const swapChainProtected = isSwapChainProtected(drawSwapChain);
+        if (UTILS_UNLIKELY(swapChainProtected)) {
+            // we need a protected context
+            if (UTILS_UNLIKELY(mEGLContextProtected == EGL_NO_CONTEXT)) {
+                // we don't have one, create it!
+                EGLConfig config = ext.egl.KHR_no_config_context ? EGL_NO_CONFIG_KHR : mEGLConfig;
+                Config protectedContextAttribs{ mContextAttribs };
+                protectedContextAttribs[EGL_PROTECTED_CONTENT_EXT] = EGL_TRUE;
+                mEGLContextProtected = eglCreateContext(mEGLDisplay, config, mEGLContext,
+                        protectedContextAttribs.data());
+                if (UTILS_UNLIKELY(mEGLContextProtected == EGL_NO_CONTEXT)) {
+                    // couldn't create the protected context
+                    logEglError("eglCreateContext[EGL_PROTECTED_CONTENT_EXT]");
+                    ext.egl.EXT_protected_content = false;
+                    goto error;
+                }
+            }
+            type = ContextType::PROTECTED;
+            error: ;
+        }
+
+        bool const contextChange = type != mCurrentContextType;
+        mCurrentContextType = type;
+
+        if (UTILS_UNLIKELY(contextChange)) {
+            preContextChange();
+            bool const success = makeCurrent(mCurrentContextType, drawSwapChain, readSwapChain);
+            if (UTILS_UNLIKELY(!success)) {
+                logEglError("PlatformEGL::makeCurrent");
+                if (mEGLContextProtected != EGL_NO_CONTEXT) {
+                    eglDestroyContext(mEGLDisplay, mEGLContextProtected);
+                    mEGLContextProtected = EGL_NO_CONTEXT;
+                }
+                mCurrentContextType = ContextType::UNPROTECTED;
+            }
+            if (UTILS_LIKELY(!swapChainProtected && mEGLContextProtected != EGL_NO_CONTEXT)) {
+                // We don't need the protected context anymore, unbind and destroy right away.
+                eglDestroyContext(mEGLDisplay, mEGLContextProtected);
+                mEGLContextProtected = EGL_NO_CONTEXT;
+            }
+            size_t const contextIndex = (mCurrentContextType == ContextType::PROTECTED) ? 1 : 0;
+            postContextChange(contextIndex);
+            return;
+        }
+    }
+
+    bool const success = makeCurrent(mCurrentContextType, drawSwapChain, readSwapChain);
+    if (UTILS_UNLIKELY(!success)) {
+        logEglError("PlatformEGL::makeCurrent");
     }
-    makeCurrent(dsc->sur, rsc->sur);
 }
 
 void PlatformEGL::commit(Platform::SwapChain* swapChain) noexcept {
@@ -554,6 +656,8 @@ void PlatformEGL::commit(Platform::SwapChain* swapChain) noexcept {
     }
 }
 
+// -----------------------------------------------------------------------------------------------
+
 bool PlatformEGL::canCreateFence() noexcept {
     return true;
 }
@@ -592,6 +696,8 @@ FenceStatus PlatformEGL::waitFence(
     return FenceStatus::ERROR;
 }
 
+// -----------------------------------------------------------------------------------------------
+
 OpenGLPlatform::ExternalTexture* PlatformEGL::createExternalImageTexture() noexcept {
     ExternalTexture* outTexture = new(std::nothrow) ExternalTexture{};
     glGenTextures(1, &outTexture->id);
@@ -622,6 +728,8 @@ bool PlatformEGL::setExternalImage(void* externalImage,
     return true;
 }
 
+// -----------------------------------------------------------------------------------------------
+
 void PlatformEGL::initializeGlExtensions() noexcept {
     // We're guaranteed to be on an ES platform, since we're using EGL
     GLUtils::unordered_string_set glExtensions;
@@ -630,6 +738,17 @@ void PlatformEGL::initializeGlExtensions() noexcept {
     ext.gl.OES_EGL_image_external_essl3 = glExtensions.has("GL_OES_EGL_image_external_essl3");
 }
 
+EGLContext PlatformEGL::getContextForType(OpenGLPlatform::ContextType type) const noexcept {
+    switch (type) {
+        case ContextType::NONE:
+            return EGL_NO_CONTEXT;
+        case ContextType::UNPROTECTED:
+            return mEGLContext;
+        case ContextType::PROTECTED:
+            return mEGLContextProtected;
+    }
+}
+
 // ---------------------------------------------------------------------------------------------
 
 PlatformEGL::Config::Config() = default;
@@ -666,6 +785,23 @@ void PlatformEGL::Config::erase(EGLint name) noexcept {
     }
 }
 
-} // namespace filament::backend
+// ------------------------------------------------------------------------------------------------
 
-// ---------------------------------------------------------------------------------------------
+EGLBoolean PlatformEGL::EGL::makeCurrent(EGLContext context, EGLSurface drawSurface,
+        EGLSurface readSurface) noexcept {
+    if (UTILS_UNLIKELY((
+            mCurrentContext != context ||
+            drawSurface != mCurrentDrawSurface || readSurface != mCurrentReadSurface))) {
+        EGLBoolean const success = eglMakeCurrent(
+                mEGLDisplay, drawSurface, readSurface, context);
+        if (success) {
+            mCurrentDrawSurface = drawSurface;
+            mCurrentReadSurface = readSurface;
+            mCurrentContext = context;
+        }
+        return success;
+    }
+    return EGL_TRUE;
+}
+
+} // namespace filament::backend

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

@@ -14,21 +14,34 @@
  * limitations under the License.
  */
 
+#include <backend/AcquiredImage.h>
+#include <backend/Platform.h>
+#include <backend/platforms/PlatformEGL.h>
 #include <backend/platforms/PlatformEGLAndroid.h>
 
 #include "opengl/GLUtils.h"
 #include "ExternalStreamManagerAndroid.h"
 
 #include <android/api-level.h>
+#include <android/hardware_buffer.h>
+
+#include <utils/compiler.h>
+#include <utils/ostream.h>
+#include <utils/Log.h>
 
 #include <EGL/egl.h>
 #include <EGL/eglext.h>
 
-#include <utils/compiler.h>
-#include <utils/Log.h>
-
 #include <sys/system_properties.h>
 
+#include <jni.h>
+
+#include <new>
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdlib.h>
+
 // We require filament to be built with an API 19 toolchain, before that, OpenGLES 3.0 didn't exist
 // Actually, OpenGL ES 3.0 was added to API 18, but API 19 is the better target and
 // the minimum for Jetpack at the time of this comment.
@@ -129,11 +142,12 @@ Driver* PlatformEGLAndroid::createDriver(void* sharedContext,
 }
 
 void PlatformEGLAndroid::setPresentationTime(int64_t presentationTimeInNanosecond) noexcept {
-    if (mCurrentDrawSurface != EGL_NO_SURFACE) {
+    EGLSurface currentDrawSurface = eglGetCurrentSurface(EGL_DRAW);
+    if (currentDrawSurface != EGL_NO_SURFACE) {
         if (eglPresentationTimeANDROID) {
             eglPresentationTimeANDROID(
                     mEGLDisplay,
-                    mCurrentDrawSurface,
+                    currentDrawSurface,
                     presentationTimeInNanosecond);
         }
     }
@@ -165,14 +179,28 @@ int PlatformEGLAndroid::getOSVersion() const noexcept {
 
 AcquiredImage PlatformEGLAndroid::transformAcquiredImage(AcquiredImage source) noexcept {
     // Convert the AHardwareBuffer to EGLImage.
-    EGLClientBuffer clientBuffer = eglGetNativeClientBufferANDROID((const AHardwareBuffer*)source.image);
+    AHardwareBuffer const* const pHardwareBuffer = (const AHardwareBuffer*)source.image;
+
+    EGLClientBuffer clientBuffer = eglGetNativeClientBufferANDROID(pHardwareBuffer);
     if (!clientBuffer) {
         slog.e << "Unable to get EGLClientBuffer from AHardwareBuffer." << io::endl;
         return {};
     }
-    // Note that this cannot be used to stream protected video (for now) because we do not set EGL_PROTECTED_CONTENT_EXT.
-    EGLint attrs[] = { EGL_NONE, EGL_NONE };
-    EGLImageKHR eglImage = eglCreateImageKHR(mEGLDisplay, EGL_NO_CONTEXT, EGL_NATIVE_BUFFER_ANDROID, clientBuffer, attrs);
+
+    PlatformEGL::Config attributes;
+
+    if (__builtin_available(android 26, *)) {
+        AHardwareBuffer_Desc desc;
+        AHardwareBuffer_describe(pHardwareBuffer, &desc);
+        bool const isProtectedContent =
+                desc.usage & AHardwareBuffer_UsageFlags::AHARDWAREBUFFER_USAGE_PROTECTED_CONTENT;
+        if (isProtectedContent) {
+            attributes[EGL_PROTECTED_CONTENT_EXT] = EGL_TRUE;
+        }
+    }
+
+    EGLImageKHR eglImage = eglCreateImageKHR(mEGLDisplay,
+            EGL_NO_CONTEXT, EGL_NATIVE_BUFFER_ANDROID, clientBuffer, attributes.data());
     if (eglImage == EGL_NO_IMAGE_KHR) {
         slog.e << "eglCreateImageKHR returned no image." << io::endl;
         return {};

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

@@ -265,10 +265,11 @@ void PlatformGLX::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
     }
 }
 
-void PlatformGLX::makeCurrent(
-        Platform::SwapChain* drawSwapChain, Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformGLX::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
     g_glx.setCurrentContext(mGLXDisplay,
             (GLXDrawable)drawSwapChain, (GLXDrawable)readSwapChain, mGLXContext);
+    return true;
 }
 
 void PlatformGLX::commit(Platform::SwapChain* swapChain) noexcept {

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

@@ -255,8 +255,8 @@ void PlatformWGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
     wglMakeCurrent(mWhdc, mContext);
 }
 
-void PlatformWGL::makeCurrent(Platform::SwapChain* drawSwapChain,
-                              Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformWGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
     ASSERT_PRECONDITION_NON_FATAL(drawSwapChain == readSwapChain,
                                   "PlatformWGL does not support distinct draw/read swap chains.");
 
@@ -269,6 +269,7 @@ void PlatformWGL::makeCurrent(Platform::SwapChain* drawSwapChain,
             wglMakeCurrent(0, NULL);
         }
     }
+    return true;
 }
 
 void PlatformWGL::commit(Platform::SwapChain* swapChain) noexcept {

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

@@ -46,8 +46,9 @@ Platform::SwapChain* PlatformWebGL::createSwapChain(
 void PlatformWebGL::destroySwapChain(Platform::SwapChain* swapChain) noexcept {
 }
 
-void PlatformWebGL::makeCurrent(Platform::SwapChain* drawSwapChain,
-        Platform::SwapChain* readSwapChain) noexcept {
+bool PlatformWebGL::makeCurrent(ContextType type, SwapChain* drawSwapChain,
+        SwapChain* readSwapChain) noexcept {
+    return true;
 }
 
 void PlatformWebGL::commit(Platform::SwapChain* swapChain) noexcept {

filament/backend/src/vulkan/VulkanBlitter.cpp

@@ -154,15 +154,9 @@ struct BlitterUniforms {
 
 }// anonymous namespace
 
-VulkanBlitter::VulkanBlitter() noexcept = default;
-
-void VulkanBlitter::initialize(VkPhysicalDevice physicalDevice, VkDevice device,
-        VmaAllocator allocator, VulkanCommands* commands) noexcept {
-    mPhysicalDevice = physicalDevice;
-    mDevice = device;
-    mAllocator = allocator;
-    mCommands = commands;
-}
+VulkanBlitter::VulkanBlitter(VkPhysicalDevice physicalDevice, VulkanCommands* commands) noexcept
+    : mPhysicalDevice(physicalDevice),
+      mCommands(commands) {}
 
 void VulkanBlitter::resolve(VulkanAttachment dst, VulkanAttachment src) {
 

filament/backend/src/vulkan/VulkanBlitter.h

@@ -32,10 +32,7 @@ struct VulkanProgram;
 
 class VulkanBlitter {
 public:
-    VulkanBlitter() noexcept;
-
-    void initialize(VkPhysicalDevice physicalDevice, VkDevice device,
-            VmaAllocator allocator, VulkanCommands* commands) noexcept;
+    VulkanBlitter(VkPhysicalDevice physicalDevice, VulkanCommands* commands) noexcept;
 
     void blit(VkFilter filter,
             VulkanAttachment dst, const VkOffset3D* dstRectPair,
@@ -47,8 +44,6 @@ public:
 
 private:
     UTILS_UNUSED VkPhysicalDevice mPhysicalDevice;
-    VkDevice mDevice;
-    VmaAllocator mAllocator;
     VulkanCommands* mCommands;
 };
 

filament/backend/src/vulkan/VulkanCommands.cpp

@@ -155,7 +155,7 @@ VulkanCommands::VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFam
 #endif
 }
 
-VulkanCommands::~VulkanCommands() {
+void VulkanCommands::terminate() {
     wait();
     gc();
     vkDestroyCommandPool(mDevice, mPool, VKALLOC);
@@ -435,8 +435,8 @@ void VulkanCommands::popGroupMarker() {
         auto const [marker, startTime] = mGroupMarkers->pop();
         auto const endTime = std::chrono::high_resolution_clock::now();
         std::chrono::duration<double> diff = endTime - startTime;
-        utils::slog.d << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms\n"
-                      << utils::io::flush;
+        utils::slog.d << "<---- " << marker << " elapsed: " << (diff.count() * 1000) << " ms"
+                      << utils::io::endl;
 #else
         mGroupMarkers->pop();
 #endif

filament/backend/src/vulkan/VulkanCommands.h

@@ -139,71 +139,72 @@ public:
 //    - We do this because vkGetFenceStatus must be called from the rendering thread.
 //
 class VulkanCommands {
-    public:
-        VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFamilyIndex,
-                VulkanContext* context, VulkanResourceAllocator* allocator);
-        ~VulkanCommands();
+public:
+    VulkanCommands(VkDevice device, VkQueue queue, uint32_t queueFamilyIndex,
+            VulkanContext* context, VulkanResourceAllocator* allocator);
 
-        // Creates a "current" command buffer if none exists, otherwise returns the current one.
-        VulkanCommandBuffer& get();
+    void terminate();
 
-        // Submits the current command buffer if it exists, then sets "current" to null.
-        // If there are no outstanding commands then nothing happens and this returns false.
-        bool flush();
+    // Creates a "current" command buffer if none exists, otherwise returns the current one.
+    VulkanCommandBuffer& get();
 
-        // Returns the "rendering finished" semaphore for the most recent flush and removes
-        // it from the existing dependency chain. This is especially useful for setting up
-        // vkQueuePresentKHR.
-        VkSemaphore acquireFinishedSignal();
+    // Submits the current command buffer if it exists, then sets "current" to null.
+    // If there are no outstanding commands then nothing happens and this returns false.
+    bool flush();
 
-        // Takes a semaphore that signals when the next flush can occur. Only one injected
-        // semaphore is allowed per flush. Useful after calling vkAcquireNextImageKHR.
-        void injectDependency(VkSemaphore next);
+    // Returns the "rendering finished" semaphore for the most recent flush and removes
+    // it from the existing dependency chain. This is especially useful for setting up
+    // vkQueuePresentKHR.
+    VkSemaphore acquireFinishedSignal();
 
-        // Destroys all command buffers that are no longer in use.
-        void gc();
+    // Takes a semaphore that signals when the next flush can occur. Only one injected
+    // semaphore is allowed per flush. Useful after calling vkAcquireNextImageKHR.
+    void injectDependency(VkSemaphore next);
 
-        // Waits for all outstanding command buffers to finish.
-        void wait();
+    // Destroys all command buffers that are no longer in use.
+    void gc();
 
-        // Updates the atomic "status" variable in every extant fence.
-        void updateFences();
+    // Waits for all outstanding command buffers to finish.
+    void wait();
 
-        // Sets an observer who is notified every time a new command buffer has been made "current".
-        // The observer's event handler can only be called during get().
-        void setObserver(CommandBufferObserver* observer) { mObserver = observer; }
+    // Updates the atomic "status" variable in every extant fence.
+    void updateFences();
+
+    // Sets an observer who is notified every time a new command buffer has been made "current".
+    // The observer's event handler can only be called during get().
+    void setObserver(CommandBufferObserver* observer) { mObserver = observer; }
 
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-        void pushGroupMarker(char const* str, VulkanGroupMarkers::Timestamp timestamp = {});
+    void pushGroupMarker(char const* str, VulkanGroupMarkers::Timestamp timestamp = {});
 
-        void popGroupMarker();
+    void popGroupMarker();
 
-        void insertEventMarker(char const* string, uint32_t len);
+    void insertEventMarker(char const* string, uint32_t len);
 
-        std::string getTopGroupMarker() const;
+    std::string getTopGroupMarker() const;
 #endif
 
-    private:
-        static constexpr int CAPACITY = FVK_MAX_COMMAND_BUFFERS;
-        VkDevice const mDevice;
-        VkQueue const mQueue;
-        VkCommandPool const mPool;
-        VulkanContext const* mContext;
+private:
+    static constexpr int CAPACITY = FVK_MAX_COMMAND_BUFFERS;
+    VkDevice const mDevice;
+    VkQueue const mQueue;
+    VkCommandPool const mPool;
+    VulkanContext const* mContext;
 
-        // int8 only goes up to 127, therefore capacity must be less than that.
-        static_assert(CAPACITY < 128);
-        int8_t mCurrentCommandBufferIndex = -1;
-        VkSemaphore mSubmissionSignal = {};
-        VkSemaphore mInjectedSignal = {};
-        utils::FixedCapacityVector<std::unique_ptr<VulkanCommandBuffer>> mStorage;
-        VkFence mFences[CAPACITY] = {};
-        VkSemaphore mSubmissionSignals[CAPACITY] = {};
-        uint8_t mAvailableBufferCount = CAPACITY;
-        CommandBufferObserver* mObserver = nullptr;
+    // int8 only goes up to 127, therefore capacity must be less than that.
+    static_assert(CAPACITY < 128);
+    int8_t mCurrentCommandBufferIndex = -1;
+    VkSemaphore mSubmissionSignal = {};
+    VkSemaphore mInjectedSignal = {};
+    utils::FixedCapacityVector<std::unique_ptr<VulkanCommandBuffer>> mStorage;
+    VkFence mFences[CAPACITY] = {};
+    VkSemaphore mSubmissionSignals[CAPACITY] = {};
+    uint8_t mAvailableBufferCount = CAPACITY;
+    CommandBufferObserver* mObserver = nullptr;
 
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-        std::unique_ptr<VulkanGroupMarkers> mGroupMarkers;
-        std::unique_ptr<VulkanGroupMarkers> mCarriedOverMarkers;
+    std::unique_ptr<VulkanGroupMarkers> mGroupMarkers;
+    std::unique_ptr<VulkanGroupMarkers> mCarriedOverMarkers;
 #endif
 };
 

filament/backend/src/vulkan/VulkanConstants.h

@@ -43,9 +43,12 @@
 // Enables Android systrace
 #define FVK_DEBUG_SYSTRACE                0x00000001
 
-// Group markers are used to denote collection of GPU commands.  It is typically at the granualarity
-// of a renderpass.
-#define FVK_DEBUG_GROUP_MARKERS           0x00000002
+// Group markers are used to denote collections of GPU commands.  It is typically at the
+// granualarity of a renderpass. You can enable this along with FVK_DEBUG_DEBUG_UTILS to take
+// advantage of vkCmdBegin/EndDebugUtilsLabelEXT. You can also just enable this with
+// FVK_DEBUG_PRINT_GROUP_MARKERS to print the current marker to stdout.
+#define FVK_DEBUG_GROUP_MARKERS 0x00000002
+
 #define FVK_DEBUG_TEXTURE                 0x00000004
 #define FVK_DEBUG_LAYOUT_TRANSITION       0x00000008
 #define FVK_DEBUG_COMMAND_BUFFER          0x00000010
@@ -63,11 +66,14 @@
 // Enable the debug utils extension if it is available.
 #define FVK_DEBUG_DEBUG_UTILS             0x00008000
 
-// Usefaul default combinations
+// Use this to debug potential Handle/Resource leakage. It will print out reference counts for all
+// the currently active resources.
+#define FVK_DEBUG_RESOURCE_LEAK           0x00010000
+
+// Useful default combinations
 #define FVK_DEBUG_EVERYTHING              0xFFFFFFFF
 #define FVK_DEBUG_PERFORMANCE     \
-    FVK_DEBUG_SYSTRACE |          \
-    FVK_DEBUG_GROUP_MARKERS
+    FVK_DEBUG_SYSTRACE
 
 #define FVK_DEBUG_CORRECTNESS     \
     FVK_DEBUG_VALIDATION |        \
@@ -80,14 +86,12 @@
     FVK_DEBUG_PRINT_GROUP_MARKERS
 
 #ifndef NDEBUG
-#define FVK_DEBUG_FLAGS (FVK_DEBUG_PERFORMANCE | FVK_DEBUG_DEBUG_UTILS | FVK_DEBUG_VALIDATION)
+#define FVK_DEBUG_FLAGS (FVK_DEBUG_PERFORMANCE)
 #else
 #define FVK_DEBUG_FLAGS 0
 #endif
 
-#define FVK_ENABLED(flags) ((FVK_DEBUG_FLAGS) & (flags))
-#define FVK_ENABLED_BOOL(flags) ((bool) FVK_ENABLED(flags))
-
+#define FVK_ENABLED(flags) (((FVK_DEBUG_FLAGS) & (flags)) == (flags))
 
 // Group marker only works only if validation or debug utils is enabled since it uses
 // vkCmd(Begin/End)DebugUtilsLabelEXT or vkCmdDebugMarker(Begin/End)EXT
@@ -107,6 +111,15 @@ static_assert(FVK_ENABLED(FVK_DEBUG_VALIDATION));
 
 // end dependcy checks
 
+// Shorthand for combination of enabled debug flags
+#if FVK_ENABLED(FVK_DEBUG_DEBUG_UTILS) || FVK_ENABLED(FVK_DEBUG_TEXTURE)
+#define FVK_ENABLED_DEBUG_SAMPLER_NAME 1
+#else
+#define FVK_ENABLED_DEBUG_SAMPLER_NAME 0
+#endif
+
+// end shorthands
+
 #if FVK_ENABLED(FVK_DEBUG_SYSTRACE)
 
 #include <utils/Systrace.h>

filament/backend/src/vulkan/VulkanDriver.cpp

@@ -209,10 +209,16 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
       mAllocator(createAllocator(mPlatform->getInstance(), mPlatform->getPhysicalDevice(),
               mPlatform->getDevice())),
       mContext(context),
-      mResourceAllocator(driverConfig.handleArenaSize),
+      mResourceAllocator(driverConfig.handleArenaSize, driverConfig.disableHandleUseAfterFreeCheck),
       mResourceManager(&mResourceAllocator),
       mThreadSafeResourceManager(&mResourceAllocator),
+      mCommands(mPlatform->getDevice(), mPlatform->getGraphicsQueue(),
+              mPlatform->getGraphicsQueueFamilyIndex(), &mContext, &mResourceAllocator),
       mPipelineCache(&mResourceAllocator),
+      mStagePool(mAllocator, &mCommands),
+      mFramebufferCache(mPlatform->getDevice()),
+      mSamplerCache(mPlatform->getDevice()),
+      mBlitter(mPlatform->getPhysicalDevice(), &mCommands),
       mReadPixels(mPlatform->getDevice()),
       mIsSRGBSwapChainSupported(mPlatform->getCustomization().isSRGBSwapChainSupported) {
 
@@ -237,23 +243,13 @@ VulkanDriver::VulkanDriver(VulkanPlatform* platform, VulkanContext const& contex
 #endif
 
     mTimestamps = std::make_unique<VulkanTimestamps>(mPlatform->getDevice());
-    mCommands = std::make_unique<VulkanCommands>(mPlatform->getDevice(),
-            mPlatform->getGraphicsQueue(), mPlatform->getGraphicsQueueFamilyIndex(), &mContext,
-            &mResourceAllocator);
-    mCommands->setObserver(&mPipelineCache);
+    mCommands.setObserver(&mPipelineCache);
     mPipelineCache.setDevice(mPlatform->getDevice(), mAllocator);
 
-    // TOOD: move them all to be initialized by constructor
-    mStagePool.initialize(mAllocator, mCommands.get());
-    mFramebufferCache.initialize(mPlatform->getDevice());
-    mSamplerCache.initialize(mPlatform->getDevice());
-
-    mEmptyTexture.reset(createEmptyTexture(mPlatform->getDevice(), mPlatform->getPhysicalDevice(),
-            mContext, mAllocator, mCommands.get(), mStagePool));
+    mEmptyTexture = createEmptyTexture(mPlatform->getDevice(), mPlatform->getPhysicalDevice(),
+            mContext, mAllocator, &mCommands, mStagePool);
 
     mPipelineCache.setDummyTexture(mEmptyTexture->getPrimaryImageView());
-    mBlitter.initialize(mPlatform->getPhysicalDevice(), mPlatform->getDevice(), mAllocator,
-            mCommands.get());
 }
 
 VulkanDriver::~VulkanDriver() noexcept = default;
@@ -264,26 +260,28 @@ Driver* VulkanDriver::create(VulkanPlatform* platform, VulkanContext const& cont
 #if 0
     // this is useful for development, but too verbose even for debug builds
     // For reference on a 64-bits machine in Release mode:
-    //    VulkanSamplerGroup            :  24       few
+    //    VulkanSamplerGroup            :  16       few
     //    HwStream                      :  24       few
-    //    VulkanFence                   :  40       few
-    //    VulkanProgram                 :  40       moderate
-    //    VulkanIndexBuffer             :  72       moderate
-    //    VulkanBufferObject            :  72       many
-    // -- less than or equal 80 bytes
-    //    VulkanRenderPrimitive         : 104       many
-    //    VulkanSwapChain               : 112       few
-    //    VulkanTimerQuery              : 168       few
-    // -- less than or equal 176 bytes
-    //    VulkanTexture                 : 232       moderate
-    //    VulkanVertexBuffer            : 312       moderate
-    //    VulkanRenderTarget            : 320       few
-    // -- less than or equal to 320 bytes
+    //    VulkanFence                   :  32       few
+    //    VulkanProgram                 :  32       moderate
+    //    VulkanIndexBuffer             :  64       moderate
+    //    VulkanBufferObject            :  64       many
+    // -- less than or equal 64 bytes
+    //    VulkanRenderPrimitive         :  72       many
+    //    VulkanVertexBufferInfo        :  96       moderate
+    //    VulkanSwapChain               : 104       few
+    //    VulkanTimerQuery              : 160       few
+    // -- less than or equal 160 bytes
+    //    VulkanVertexBuffer            : 192       moderate
+    //    VulkanTexture                 : 224       moderate
+    //    VulkanRenderTarget            : 312       few
+    // -- less than or equal to 312 bytes
 
     utils::slog.d
            << "\nVulkanSwapChain: " << sizeof(VulkanSwapChain)
            << "\nVulkanBufferObject: " << sizeof(VulkanBufferObject)
            << "\nVulkanVertexBuffer: " << sizeof(VulkanVertexBuffer)
+           << "\nVulkanVertexBufferInfo: " << sizeof(VulkanVertexBufferInfo)
            << "\nVulkanIndexBuffer: " << sizeof(VulkanIndexBuffer)
            << "\nVulkanSamplerGroup: " << sizeof(VulkanSamplerGroup)
            << "\nVulkanRenderPrimitive: " << sizeof(VulkanRenderPrimitive)
@@ -314,8 +312,11 @@ ShaderModel VulkanDriver::getShaderModel() const noexcept {
 void VulkanDriver::terminate() {
     // Command buffers should come first since it might have commands depending on resources that
     // are about to be destroyed.
-    mCommands.reset();
-    mEmptyTexture.reset();
+    mCommands.terminate();
+
+    delete mEmptyTexture;
+    mResourceManager.clear();
+
     mTimestamps.reset();
 
     mBlitter.terminate();
@@ -344,7 +345,7 @@ void VulkanDriver::terminate() {
 }
 
 void VulkanDriver::tick(int) {
-    mCommands->updateFences();
+    mCommands.updateFences();
 }
 
 // Garbage collection should not occur too frequently, only about once per frame. Internally, the
@@ -356,9 +357,14 @@ void VulkanDriver::collectGarbage() {
     FVK_SYSTRACE_START("gc");
     // Command buffers need to be submitted and completed before other resources can be gc'd. And
     // its gc() function carrys out the *wait*.
-    mCommands->gc();
+    mCommands.gc();
     mStagePool.gc();
     mFramebufferCache.gc();
+
+#if FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
+    mResourceAllocator.print();
+#endif
+
     FVK_SYSTRACE_END();
 }
 void VulkanDriver::beginFrame(int64_t monotonic_clock_ns, uint32_t frameId) {
@@ -379,7 +385,7 @@ void VulkanDriver::setPresentationTime(int64_t monotonic_clock_ns) {
 void VulkanDriver::endFrame(uint32_t frameId) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("endframe");
-    mCommands->flush();
+    mCommands.flush();
     collectGarbage();
     FVK_SYSTRACE_END();
 }
@@ -387,7 +393,7 @@ void VulkanDriver::endFrame(uint32_t frameId) {
 void VulkanDriver::flush(int) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("flush");
-    mCommands->flush();
+    mCommands.flush();
     FVK_SYSTRACE_END();
 }
 
@@ -395,8 +401,8 @@ void VulkanDriver::finish(int dummy) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("finish");
 
-    mCommands->flush();
-    mCommands->wait();
+    mCommands.flush();
+    mCommands.wait();
 
     mReadPixels.runUntilComplete();
     FVK_SYSTRACE_END();
@@ -411,9 +417,9 @@ void VulkanDriver::createSamplerGroupR(Handle<HwSamplerGroup> sbh, uint32_t coun
 void VulkanDriver::createRenderPrimitiveR(Handle<HwRenderPrimitive> rph,
         Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh,
         PrimitiveType pt) {
-    auto rp = mResourceAllocator.construct<VulkanRenderPrimitive>(rph, &mResourceAllocator);
+    auto rp = mResourceAllocator.construct<VulkanRenderPrimitive>(rph,
+            &mResourceAllocator, pt, vbh, ibh);
     mResourceManager.acquire(rp);
-    VulkanDriver::setRenderPrimitiveBuffer(rph, pt, vbh, ibh);
 }
 
 void VulkanDriver::destroyRenderPrimitive(Handle<HwRenderPrimitive> rph) {
@@ -493,7 +499,7 @@ void VulkanDriver::createTextureR(Handle<HwTexture> th, SamplerType target, uint
         TextureFormat format, uint8_t samples, uint32_t w, uint32_t h, uint32_t depth,
         TextureUsage usage) {
     auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, mCommands.get(), target, levels,
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, target, levels,
             format, samples, w, h, depth, usage, mStagePool);
     mResourceManager.acquire(vktexture);
 }
@@ -505,7 +511,7 @@ void VulkanDriver::createTextureSwizzledR(Handle<HwTexture> th, SamplerType targ
     TextureSwizzle swizzleArray[] = {r, g, b, a};
     const VkComponentMapping swizzleMap = getSwizzleMap(swizzleArray);
     auto vktexture = mResourceAllocator.construct<VulkanTexture>(th, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, mCommands.get(), target, levels,
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, target, levels,
             format, samples, w, h, depth, usage, mStagePool, false /*heap allocated */, swizzleMap);
     mResourceManager.acquire(vktexture);
 }
@@ -548,7 +554,7 @@ void VulkanDriver::createDefaultRenderTargetR(Handle<HwRenderTarget> rth, int) {
 
 void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
         TargetBufferFlags targets, uint32_t width, uint32_t height, uint8_t samples,
-        MRT color, TargetBufferInfo depth, TargetBufferInfo stencil) {
+        uint8_t layerCount, MRT color, TargetBufferInfo depth, TargetBufferInfo stencil) {
     UTILS_UNUSED_IN_RELEASE math::vec2<uint32_t> tmin = {std::numeric_limits<uint32_t>::max()};
     UTILS_UNUSED_IN_RELEASE math::vec2<uint32_t> tmax = {0};
     UTILS_UNUSED_IN_RELEASE size_t attachmentCount = 0;
@@ -600,7 +606,7 @@ void VulkanDriver::createRenderTargetR(Handle<HwRenderTarget> rth,
     assert_invariant(tmin.x >= width && tmin.y >= height);
 
     auto renderTarget = mResourceAllocator.construct<VulkanRenderTarget>(rth, mPlatform->getDevice(),
-            mPlatform->getPhysicalDevice(), mContext, mAllocator, mCommands.get(), width, height,
+            mPlatform->getPhysicalDevice(), mContext, mAllocator, &mCommands, width, height,
             samples, colorTargets, depthStencil, mStagePool);
     mResourceManager.acquire(renderTarget);
 }
@@ -618,7 +624,7 @@ void VulkanDriver::destroyRenderTarget(Handle<HwRenderTarget> rth) {
 }
 
 void VulkanDriver::createFenceR(Handle<HwFence> fh, int) {
-    VulkanCommandBuffer const& commandBuffer = mCommands->get();
+    VulkanCommandBuffer const& commandBuffer = mCommands.get();
     mResourceAllocator.construct<VulkanFence>(fh, commandBuffer.fence);
 }
 
@@ -629,7 +635,7 @@ void VulkanDriver::createSwapChainR(Handle<HwSwapChain> sch, void* nativeWindow,
         flags = flags | ~(backend::SWAP_CHAIN_CONFIG_SRGB_COLORSPACE);
     }
     auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, mCommands.get(), mStagePool, nativeWindow, flags);
+            mAllocator, &mCommands, mStagePool, nativeWindow, flags);
     mResourceManager.acquire(swapChain);
 }
 
@@ -642,7 +648,7 @@ void VulkanDriver::createSwapChainHeadlessR(Handle<HwSwapChain> sch, uint32_t wi
     }
     assert_invariant(width > 0 && height > 0 && "Vulkan requires non-zero swap chain dimensions.");
     auto swapChain = mResourceAllocator.construct<VulkanSwapChain>(sch, mPlatform, mContext,
-            mAllocator, mCommands.get(), mStagePool, nullptr, flags, VkExtent2D{width, height});
+            mAllocator, &mCommands, mStagePool, nullptr, flags, VkExtent2D{width, height});
     mResourceManager.acquire(swapChain);
 }
 
@@ -868,6 +874,11 @@ bool VulkanDriver::isSRGBSwapChainSupported() {
     return mIsSRGBSwapChainSupported;
 }
 
+bool VulkanDriver::isProtectedContentSupported() {
+    // the SWAP_CHAIN_CONFIG_PROTECTED_CONTENT flag is not supported
+    return false;
+}
+
 bool VulkanDriver::isStereoSupported(backend::StereoscopicType stereoscopicType) {
     switch (stereoscopicType) {
     case backend::StereoscopicType::INSTANCED:
@@ -887,6 +898,10 @@ bool VulkanDriver::isDepthStencilResolveSupported() {
     return false;
 }
 
+bool VulkanDriver::isProtectedTexturesSupported() {
+    return false;
+}
+
 bool VulkanDriver::isWorkaroundNeeded(Workaround workaround) {
     switch (workaround) {
         case Workaround::SPLIT_EASU: {
@@ -973,13 +988,12 @@ void VulkanDriver::setVertexBufferObject(Handle<HwVertexBuffer> vbh, uint32_t in
     auto vb = mResourceAllocator.handle_cast<VulkanVertexBuffer*>(vbh);
     auto bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
     assert_invariant(bo->bindingType == BufferObjectBinding::VERTEX);
-
     vb->setBuffer(mResourceAllocator, bo, index);
 }
 
 void VulkanDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor&& p,
         uint32_t byteOffset) {
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
     auto ib = mResourceAllocator.handle_cast<VulkanIndexBuffer*>(ibh);
     commands.acquire(ib);
     ib->buffer.loadFromCpu(commands.buffer(), p.buffer, byteOffset, p.size);
@@ -989,7 +1003,7 @@ void VulkanDriver::updateIndexBuffer(Handle<HwIndexBuffer> ibh, BufferDescriptor
 
 void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescriptor&& bd,
         uint32_t byteOffset) {
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
 
     auto bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
     commands.acquire(bo);
@@ -1000,7 +1014,7 @@ void VulkanDriver::updateBufferObject(Handle<HwBufferObject> boh, BufferDescript
 
 void VulkanDriver::updateBufferObjectUnsynchronized(Handle<HwBufferObject> boh,
         BufferDescriptor&& bd, uint32_t byteOffset) {
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
     auto bo = mResourceAllocator.handle_cast<VulkanBufferObject*>(boh);
     commands.acquire(bo);
     // TODO: implement unsynchronized version
@@ -1033,10 +1047,10 @@ void VulkanDriver::update3DImage(Handle<HwTexture> th, uint32_t level, uint32_t
 void VulkanDriver::setupExternalImage(void* image) {
 }
 
-bool VulkanDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapsedTime) {
+TimerQueryResult VulkanDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapsedTime) {
     VulkanTimerQuery* vtq = mResourceAllocator.handle_cast<VulkanTimerQuery*>(tqh);
     if (!vtq->isCompleted()) {
-        return false;
+        return TimerQueryResult::NOT_READY;
     }
 
     auto results = mTimestamps->getResult(vtq);
@@ -1046,7 +1060,7 @@ bool VulkanDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapse
     uint64_t available1 = results[3];
 
     if (available0 == 0 || available1 == 0) {
-        return false;
+        return TimerQueryResult::NOT_READY;
     }
 
     ASSERT_POSTCONDITION(timestamp1 >= timestamp0, "Timestamps are not monotonically increasing.");
@@ -1058,7 +1072,7 @@ bool VulkanDriver::getTimerQueryValue(Handle<HwTimerQuery> tqh, uint64_t* elapse
     float const period = mContext.getPhysicalDeviceLimits().timestampPeriod;
     uint64_t delta = uint64_t(float(timestamp1 - timestamp0) * period);
     *elapsedTime = delta;
-    return true;
+    return TimerQueryResult::AVAILABLE;
 }
 
 void VulkanDriver::setExternalImage(Handle<HwTexture> th, void* image) {
@@ -1165,7 +1179,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
     // If that's the case, we need to change the layout of the texture to DEPTH_SAMPLER, which is a
     // more general layout. Otherwise, we prefer the DEPTH_ATTACHMENT layout, which is optimal for
     // the non-sampling case.
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer const cmdbuffer = commands.buffer();
 
     UTILS_NOUNROLL
@@ -1292,7 +1306,7 @@ void VulkanDriver::beginRenderPass(Handle<HwRenderTarget> rth, const RenderPassP
 
 // Assign a label to the framebuffer for debugging purposes.
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS | FVK_DEBUG_DEBUG_UTILS)
-    auto const topMarker = mCommands->getTopGroupMarker();
+    auto const topMarker = mCommands.getTopGroupMarker();
     if (!topMarker.empty()) {
         DebugUtils::setName(VK_OBJECT_TYPE_FRAMEBUFFER, reinterpret_cast<uint64_t>(vkfb),
                 topMarker.c_str());
@@ -1381,7 +1395,7 @@ void VulkanDriver::endRenderPass(int) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("endRenderPass");
 
-    VulkanCommandBuffer& commands = mCommands->get();
+    VulkanCommandBuffer& commands = mCommands.get();
     VkCommandBuffer cmdbuffer = commands.buffer();
     vkCmdEndRenderPass(cmdbuffer);
 
@@ -1434,7 +1448,7 @@ void VulkanDriver::nextSubpass(int) {
     assert_invariant(renderTarget);
     assert_invariant(mCurrentRenderPass.params.subpassMask);
 
-    vkCmdNextSubpass(mCommands->get().buffer(), VK_SUBPASS_CONTENTS_INLINE);
+    vkCmdNextSubpass(mCommands.get().buffer(), VK_SUBPASS_CONTENTS_INLINE);
 
     mPipelineCache.bindRenderPass(mCurrentRenderPass.renderPass,
             ++mCurrentRenderPass.currentSubpass);
@@ -1451,14 +1465,6 @@ void VulkanDriver::nextSubpass(int) {
     }
 }
 
-void VulkanDriver::setRenderPrimitiveBuffer(Handle<HwRenderPrimitive> rph, PrimitiveType pt,
-        Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh) {
-    auto primitive = mResourceAllocator.handle_cast<VulkanRenderPrimitive*>(rph);
-    primitive->setBuffers(mResourceAllocator.handle_cast<VulkanVertexBuffer*>(vbh),
-            mResourceAllocator.handle_cast<VulkanIndexBuffer*>(ibh));
-    primitive->setPrimitiveType(pt);
-}
-
 void VulkanDriver::makeCurrent(Handle<HwSwapChain> drawSch, Handle<HwSwapChain> readSch) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("makeCurrent");
@@ -1523,14 +1529,14 @@ void VulkanDriver::bindSamplers(uint32_t index, Handle<HwSamplerGroup> sbh) {
 
 void VulkanDriver::insertEventMarker(char const* string, uint32_t len) {
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    mCommands->insertEventMarker(string, len);
+    mCommands.insertEventMarker(string, len);
 #endif
 }
 
 void VulkanDriver::pushGroupMarker(char const* string, uint32_t) {
     // Turns out all the markers are 0-terminated, so we can just pass it without len.
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    mCommands->pushGroupMarker(string);
+    mCommands.pushGroupMarker(string);
 #endif
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START(string);
@@ -1538,7 +1544,7 @@ void VulkanDriver::pushGroupMarker(char const* string, uint32_t) {
 
 void VulkanDriver::popGroupMarker(int) {
 #if FVK_ENABLED(FVK_DEBUG_GROUP_MARKERS)
-    mCommands->popGroupMarker();
+    mCommands.popGroupMarker();
 #endif
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_END();
@@ -1551,7 +1557,7 @@ void VulkanDriver::stopCapture(int) {}
 void VulkanDriver::readPixels(Handle<HwRenderTarget> src, uint32_t x, uint32_t y, uint32_t width,
         uint32_t height, PixelBufferDescriptor&& pbd) {
     VulkanRenderTarget* srcTarget = mResourceAllocator.handle_cast<VulkanRenderTarget*>(src);
-    mCommands->flush();
+    mCommands.flush();
     mReadPixels.run(
             srcTarget, x, y, width, height, mPlatform->getGraphicsQueueFamilyIndex(),
             std::move(pbd),
@@ -1704,14 +1710,13 @@ void VulkanDriver::blitDEPRECATED(TargetBufferFlags buffers,
     FVK_SYSTRACE_END();
 }
 
-void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> rph,
-        uint32_t const indexOffset, uint32_t const indexCount, uint32_t const instanceCount) {
+void VulkanDriver::bindPipeline(PipelineState pipelineState) {
     FVK_SYSTRACE_CONTEXT();
     FVK_SYSTRACE_START("draw");
 
-    VulkanCommandBuffer* commands = &mCommands->get();
-    VkCommandBuffer cmdbuffer = commands->buffer();
-    const VulkanRenderPrimitive& prim = *mResourceAllocator.handle_cast<VulkanRenderPrimitive*>(rph);
+    VulkanCommandBuffer* commands = &mCommands.get();
+    const VulkanVertexBufferInfo& vbi =
+            *mResourceAllocator.handle_cast<VulkanVertexBufferInfo*>(pipelineState.vertexBufferInfo);
 
     Handle<HwProgram> programHandle = pipelineState.program;
     RasterState rasterState = pipelineState.rasterState;
@@ -1719,15 +1724,6 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
 
     auto* program = mResourceAllocator.handle_cast<VulkanProgram*>(programHandle);
     commands->acquire(program);
-    commands->acquire(prim.indexBuffer);
-    commands->acquire(prim.vertexBuffer);
-
-    // If this is a debug build, validate the current shader.
-#if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
-    if (program->bundle.vertex == VK_NULL_HANDLE || program->bundle.fragment == VK_NULL_HANDLE) {
-        utils::slog.e << "Binding missing shader: " << program->name.c_str() << utils::io::endl;
-    }
-#endif
 
     // Update the VK raster state.
     const VulkanRenderTarget* rt = mCurrentRenderPass.renderTarget;
@@ -1753,20 +1749,19 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
         .depthBiasSlopeFactor = depthOffset.slope
     };
 
+    // unfortunately in Vulkan the topology is per pipeline
+    VkPrimitiveTopology const topology =
+            VulkanPipelineCache::getPrimitiveTopology(pipelineState.primitiveType);
+
     // Declare fixed-size arrays that get passed to the pipeCache and to vkCmdBindVertexBuffers.
-    VulkanVertexBufferInfo const* const vbi =
-            mResourceAllocator.handle_cast<VulkanVertexBufferInfo*>(prim.vertexBuffer->vbih);
-    uint32_t const bufferCount = vbi->attributes.size();
-    VkVertexInputAttributeDescription const* attribDesc = prim.vertexBuffer->getAttribDescriptions();
-    VkVertexInputBindingDescription const* bufferDesc =  prim.vertexBuffer->getBufferDescriptions();
-    VkBuffer const* buffers = prim.vertexBuffer->getVkBuffers();
-    VkDeviceSize const* offsets = prim.vertexBuffer->getOffsets();
+    VkVertexInputAttributeDescription const* attribDesc = vbi.getAttribDescriptions();
+    VkVertexInputBindingDescription const* bufferDesc =  vbi.getBufferDescriptions();
 
     // Push state changes to the VulkanPipelineCache instance. This is fast and does not make VK calls.
     mPipelineCache.bindProgram(program);
     mPipelineCache.bindRasterState(vulkanRasterState);
-    mPipelineCache.bindPrimitiveTopology(prim.primitiveTopology);
-    mPipelineCache.bindVertexArray(attribDesc, bufferDesc, bufferCount);
+    mPipelineCache.bindPrimitiveTopology(topology);
+    mPipelineCache.bindVertexArray(attribDesc, bufferDesc, vbi.getAttributeCount());
 
     // Query the program for the mapping from (SamplerGroupBinding,Offset) to (SamplerBinding),
     // where "SamplerBinding" is the integer in the GLSL, and SamplerGroupBinding is the abstract
@@ -1776,7 +1771,11 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
     VulkanTexture* samplerTextures[VulkanPipelineCache::SAMPLER_BINDING_COUNT] = {nullptr};
 
     auto const& bindingToSamplerIndex = program->getBindingToSamplerIndex();
-    VulkanPipelineCache::UsageFlags  usage = program->getUsage();
+    VulkanPipelineCache::UsageFlags usage = program->getUsage();
+
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+    auto const& bindingToName = program->getBindingToName();
+#endif
 
     UTILS_NOUNROLL
     for (uint8_t binding = 0; binding < VulkanPipelineCache::SAMPLER_BINDING_COUNT; binding++) {
@@ -1810,15 +1809,21 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
         // matching characteristics. (e.g. if the missing texture is a 3D texture)
         if (UTILS_UNLIKELY(texture->getPrimaryImageLayout() == VulkanLayout::UNDEFINED)) {
 #if FVK_ENABLED(FVK_DEBUG_TEXTURE)
-            utils::slog.w << "Uninitialized texture bound to '" << sampler.name.c_str() << "'";
+            utils::slog.w << "Uninitialized texture bound to '" << bindingToName[binding] << "'";
             utils::slog.w << " in material '" << program->name.c_str() << "'";
-            utils::slog.w << " at binding point " << +sampler.binding << utils::io::endl;
+            utils::slog.w << " at binding point " << +binding << utils::io::endl;
 #endif
-            texture = mEmptyTexture.get();
+            texture = mEmptyTexture;
         }
 
         SamplerParams const& samplerParams = boundSampler->s;
         VkSampler const vksampler = mSamplerCache.getSampler(samplerParams);
+
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+        VulkanDriver::DebugUtils::setName(VK_OBJECT_TYPE_SAMPLER,
+                reinterpret_cast<uint64_t>(vksampler), bindingToName[binding].c_str());
+#endif
+
         VkImageView imageView = VK_NULL_HANDLE;
         VkImageSubresourceRange const range = texture->getPrimaryViewRange();
         if (any(texture->usage & TextureUsage::DEPTH_ATTACHMENT) &&
@@ -1842,19 +1847,33 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
 
     mPipelineCache.bindSamplers(samplerInfo, samplerTextures, usage);
 
-    // Bind new descriptor sets if they need to change.
-    // If descriptor set allocation failed, skip the draw call and bail. No need to emit an error
-    // message since the validation layers already do so.
-    if (!mPipelineCache.bindDescriptors(cmdbuffer)) {
-        return;
-    }
-
     // Bind a new pipeline if the pipeline state changed.
     // If allocation failed, skip the draw call and bail. We do not emit an error since the
     // validation layer will already do so.
     if (!mPipelineCache.bindPipeline(commands)) {
         return;
     }
+    FVK_SYSTRACE_END();
+}
+
+void VulkanDriver::bindRenderPrimitive(Handle<HwRenderPrimitive> rph) {
+    FVK_SYSTRACE_CONTEXT();
+    FVK_SYSTRACE_START("bindRenderPrimitive");
+
+    VulkanCommandBuffer* commands = &mCommands.get();
+    VkCommandBuffer cmdbuffer = commands->buffer();
+    const VulkanRenderPrimitive& prim = *mResourceAllocator.handle_cast<VulkanRenderPrimitive*>(rph);
+    commands->acquire(prim.indexBuffer);
+    commands->acquire(prim.vertexBuffer);
+
+    // This *must* match the VulkanVertexBufferInfo that was bound in bindPipeline(). But we want
+    // to allow to call this before bindPipeline(), so the validation can only happen in draw()
+    VulkanVertexBufferInfo const* const vbi =
+            mResourceAllocator.handle_cast<VulkanVertexBufferInfo*>(prim.vertexBuffer->vbih);
+
+    uint32_t const bufferCount = vbi->getAttributeCount();
+    VkDeviceSize const* offsets = vbi->getOffsets();
+    VkBuffer const* buffers = prim.vertexBuffer->getVkBuffers();
 
     // Next bind the vertex buffers and index buffer. One potential performance improvement is to
     // avoid rebinding these if they are already bound, but since we do not (yet) support subranges
@@ -1863,22 +1882,50 @@ void VulkanDriver::draw(PipelineState pipelineState, Handle<HwRenderPrimitive> r
     vkCmdBindIndexBuffer(cmdbuffer, prim.indexBuffer->buffer.getGpuBuffer(), 0,
             prim.indexBuffer->indexType);
 
+    FVK_SYSTRACE_END();
+}
+
+void VulkanDriver::draw2(uint32_t indexOffset, uint32_t indexCount, uint32_t instanceCount) {
+    FVK_SYSTRACE_CONTEXT();
+    FVK_SYSTRACE_START("draw2");
+
+    VulkanCommandBuffer& commands = mCommands.get();
+    VkCommandBuffer cmdbuffer = commands.buffer();
+
+    // Bind new descriptor sets if they need to change.
+    // If descriptor set allocation failed, skip the draw call and bail. No need to emit an error
+    // message since the validation layers already do so.
+    if (!mPipelineCache.bindDescriptors(cmdbuffer)) {
+        return;
+    }
+
     // Finally, make the actual draw call. TODO: support subranges
     const uint32_t firstIndex = indexOffset;
     const int32_t vertexOffset = 0;
     const uint32_t firstInstId = 0;
 
     vkCmdDrawIndexed(cmdbuffer, indexCount, instanceCount, firstIndex, vertexOffset, firstInstId);
+
     FVK_SYSTRACE_END();
 }
 
+void VulkanDriver::draw(PipelineState state, Handle<HwRenderPrimitive> rph,
+        uint32_t const indexOffset, uint32_t const indexCount, uint32_t const instanceCount) {
+    VulkanRenderPrimitive* const rp = mResourceAllocator.handle_cast<VulkanRenderPrimitive*>(rph);
+    state.primitiveType = rp->type;
+    state.vertexBufferInfo = rp->vertexBuffer->vbih;
+    bindPipeline(state);
+    bindRenderPrimitive(rph);
+    draw2(indexOffset, indexCount, instanceCount);
+}
+
 void VulkanDriver::dispatchCompute(Handle<HwProgram> program, math::uint3 workGroupCount) {
     // FIXME: implement me
 }
 
 void VulkanDriver::scissor(Viewport scissorBox) {
-    VulkanCommandBuffer* commands = &mCommands->get();
-    VkCommandBuffer cmdbuffer = commands->buffer();
+    VulkanCommandBuffer& commands = mCommands.get();
+    VkCommandBuffer cmdbuffer = commands.buffer();
 
     // Set scissoring.
     // clamp left-bottom to 0,0 and avoid overflows
@@ -1905,12 +1952,12 @@ void VulkanDriver::scissor(Viewport scissorBox) {
 
 void VulkanDriver::beginTimerQuery(Handle<HwTimerQuery> tqh) {
     VulkanTimerQuery* vtq = mResourceAllocator.handle_cast<VulkanTimerQuery*>(tqh);
-    mTimestamps->beginQuery(&(mCommands->get()), vtq);
+    mTimestamps->beginQuery(&(mCommands.get()), vtq);
 }
 
 void VulkanDriver::endTimerQuery(Handle<HwTimerQuery> tqh) {
     VulkanTimerQuery* vtq = mResourceAllocator.handle_cast<VulkanTimerQuery*>(tqh);
-    mTimestamps->endQuery(&(mCommands->get()), vtq);
+    mTimestamps->endQuery(&(mCommands.get()), vtq);
 }
 
 void VulkanDriver::debugCommandBegin(CommandStream* cmds, bool synchronous, const char* methodName) noexcept {

filament/backend/src/vulkan/VulkanDriver.h

@@ -101,16 +101,12 @@ private:
     VulkanDriver& operator=(VulkanDriver const&) = delete;
 
 private:
-    inline void setRenderPrimitiveBuffer(Handle<HwRenderPrimitive> rph, PrimitiveType pt,
-            Handle<HwVertexBuffer> vbh,
-            Handle<HwIndexBuffer> ibh);
-
     void collectGarbage();
 
     VulkanPlatform* mPlatform = nullptr;
-    std::unique_ptr<VulkanCommands> mCommands;
     std::unique_ptr<VulkanTimestamps> mTimestamps;
-    std::unique_ptr<VulkanTexture> mEmptyTexture;
+
+    VulkanTexture* mEmptyTexture;
 
     VulkanSwapChain* mCurrentSwapChain = nullptr;
     VulkanRenderTarget* mDefaultRenderTarget = nullptr;
@@ -126,6 +122,7 @@ private:
     // thread.
     VulkanThreadSafeResourceManager mThreadSafeResourceManager;
 
+    VulkanCommands mCommands;
     VulkanPipelineCache mPipelineCache;
     VulkanStagePool mStagePool;
     VulkanFboCache mFramebufferCache;

filament/backend/src/vulkan/VulkanFboCache.cpp

@@ -64,7 +64,8 @@ bool VulkanFboCache::FboKeyEqualFn::operator()(const FboKey& k1, const FboKey& k
     return true;
 }
 
-void VulkanFboCache::initialize(VkDevice device) noexcept { mDevice = device; }
+VulkanFboCache::VulkanFboCache(VkDevice device)
+    : mDevice(device) {}
 
 VulkanFboCache::~VulkanFboCache() {
     ASSERT_POSTCONDITION(mFramebufferCache.empty() && mRenderPassCache.empty(),
@@ -238,7 +239,7 @@ VkRenderPass VulkanFboCache::getRenderPass(RenderPassKey config) noexcept {
             .format = config.colorFormat[i],
             .samples = (VkSampleCountFlagBits) config.samples,
             .loadOp = clear ? kClear : (discard ? kDontCare : kKeep),
-            .storeOp = config.samples == 1 ? kEnableStore : kDisableStore,
+            .storeOp = kEnableStore,
             .stencilLoadOp = kDontCare,
             .stencilStoreOp = kDisableStore,
             .initialLayout = ((!discard && config.initialColorLayoutMask & (1 << i)) || clear)

filament/backend/src/vulkan/VulkanFboCache.h

@@ -104,10 +104,9 @@ public:
         bool operator()(const FboKey& k1, const FboKey& k2) const;
     };
 
+    explicit VulkanFboCache(VkDevice device);
     ~VulkanFboCache();
 
-    void initialize(VkDevice device) noexcept;
-
     // Retrieves or creates a VkFramebuffer handle.
     VkFramebuffer getFramebuffer(FboKey config) noexcept;
 

filament/backend/src/vulkan/VulkanHandles.cpp

@@ -99,6 +99,10 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
 #endif
     }
 
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+    auto& bindingToName = mInfo->bindingToName;
+#endif
+
     auto& groupInfo = builder.getSamplerGroupInfo();
     auto& bindingToSamplerIndex = mInfo->bindingToSamplerIndex;
     auto& usage = mInfo->usage;
@@ -109,12 +113,16 @@ VulkanProgram::VulkanProgram(VkDevice device, Program const& builder) noexcept
             uint32_t const binding = samplers[i].binding;
             bindingToSamplerIndex[binding] = (groupInd << 8) | (0xff & i);
             usage = VulkanPipelineCache::getUsageFlags(binding, group.stageFlags, usage);
+
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+            bindingToName[binding] = samplers[i].name.c_str();
+#endif
         }
     }
 
 #if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
-    utils::slog.d << "Created VulkanProgram " << builder << ", shaders = (" << bundle.vertex
-                  << ", " << bundle.fragment << ")" << utils::io::endl;
+    utils::slog.d << "Created VulkanProgram " << builder << ", shaders = (" << modules[0]
+                  << ", " << modules[1] << ")" << utils::io::endl;
 #endif
 }
 
@@ -281,32 +289,21 @@ uint8_t VulkanRenderTarget::getColorTargetCount(const VulkanRenderPass& pass) co
     return count;
 }
 
-VulkanVertexBufferInfo::VulkanVertexBufferInfo(uint8_t bufferCount, uint8_t attributeCount,
-        AttributeArray const& attributes)
-        : HwVertexBufferInfo(bufferCount, attributeCount),
-          VulkanResource(VulkanResourceType::VERTEX_BUFFER_INFO),
-          attributes(attributes) {
-}
+VulkanVertexBufferInfo::VulkanVertexBufferInfo(
+        uint8_t bufferCount, uint8_t attributeCount, AttributeArray const& attributes)
+    : HwVertexBufferInfo(bufferCount, attributeCount),
+      VulkanResource(VulkanResourceType::VERTEX_BUFFER_INFO),
+      mInfo(attributes.size()) {
 
-VulkanVertexBuffer::VulkanVertexBuffer(VulkanContext& context, VulkanStagePool& stagePool,
-        VulkanResourceAllocator* allocator, uint32_t vertexCount, Handle<HwVertexBufferInfo> vbih)
-    : HwVertexBuffer(vertexCount),
-      VulkanResource(VulkanResourceType::VERTEX_BUFFER),
-      vbih(vbih),
-      mResources(allocator) {
+    auto attribDesc = mInfo.mSoa.data<PipelineInfo::ATTRIBUTE_DESCRIPTION>();
+    auto bufferDesc = mInfo.mSoa.data<PipelineInfo::BUFFER_DESCRIPTION>();
+    auto offsets = mInfo.mSoa.data<PipelineInfo::OFFSETS>();
+    auto attribToBufferIndex = mInfo.mSoa.data<PipelineInfo::ATTRIBUTE_TO_BUFFER_INDEX>();
+    std::fill(mInfo.mSoa.begin<PipelineInfo::ATTRIBUTE_TO_BUFFER_INDEX>(),
+            mInfo.mSoa.end<PipelineInfo::ATTRIBUTE_TO_BUFFER_INDEX>(), -1);
 
-    VulkanVertexBufferInfo const* const vbi = allocator->handle_cast<VulkanVertexBufferInfo*>(vbih);
-    mInfo = new PipelineInfo(vbi->attributes.size());
-
-    auto attribDesc = mInfo->mSoa.data<PipelineInfo::ATTRIBUTE_DESCRIPTION>();
-    auto bufferDesc = mInfo->mSoa.data<PipelineInfo::BUFFER_DESCRIPTION>();
-    auto offsets = mInfo->mSoa.data<PipelineInfo::OFFSETS>();
-    auto attribToBufferIndex = mInfo->mSoa.data<PipelineInfo::ATTRIBUTE_TO_BUFFER_INDEX>();
-    std::fill(mInfo->mSoa.begin<PipelineInfo::ATTRIBUTE_TO_BUFFER_INDEX>(),
-            mInfo->mSoa.end<PipelineInfo::ATTRIBUTE_TO_BUFFER_INDEX>(), -1);
-
-    for (uint32_t attribIndex = 0; attribIndex < vbi->attributes.size(); attribIndex++) {
-        Attribute attrib = vbi->attributes[attribIndex];
+    for (uint32_t attribIndex = 0; attribIndex < attributes.size(); attribIndex++) {
+        Attribute attrib = attributes[attribIndex];
         bool const isInteger = attrib.flags & Attribute::FLAG_INTEGER_TARGET;
         bool const isNormalized = attrib.flags & Attribute::FLAG_NORMALIZED;
         VkFormat vkformat = getVkFormat(attrib.type, isNormalized, isInteger);
@@ -318,7 +315,7 @@ VulkanVertexBuffer::VulkanVertexBuffer(VulkanContext& context, VulkanStagePool&
         // expects to receive floats or ints.
         if (attrib.buffer == Attribute::BUFFER_UNUSED) {
             vkformat = isInteger ? VK_FORMAT_R8G8B8A8_UINT : VK_FORMAT_R8G8B8A8_SNORM;
-            attrib = vbi->attributes[0];
+            attrib = attributes[0];
         }
         offsets[attribIndex] = attrib.offset;
         attribDesc[attribIndex] = {
@@ -334,18 +331,23 @@ VulkanVertexBuffer::VulkanVertexBuffer(VulkanContext& context, VulkanStagePool&
     }
 }
 
-VulkanVertexBuffer::~VulkanVertexBuffer() {
-    delete mInfo;
+VulkanVertexBuffer::VulkanVertexBuffer(VulkanContext& context, VulkanStagePool& stagePool,
+        VulkanResourceAllocator* allocator,
+        uint32_t vertexCount, Handle<HwVertexBufferInfo> vbih)
+    : HwVertexBuffer(vertexCount),
+      VulkanResource(VulkanResourceType::VERTEX_BUFFER),
+      vbih(vbih),
+      mBuffers(MAX_VERTEX_BUFFER_COUNT), // TODO: can we do better here?
+      mResources(allocator) {
 }
 
 void VulkanVertexBuffer::setBuffer(VulkanResourceAllocator const& allocator,
         VulkanBufferObject* bufferObject, uint32_t index) {
-
     VulkanVertexBufferInfo const* const vbi =
             const_cast<VulkanResourceAllocator&>(allocator).handle_cast<VulkanVertexBufferInfo*>(vbih);
-    size_t count = vbi->attributes.size();
-    auto vkbuffers = mInfo->mSoa.data<PipelineInfo::VK_BUFFER>();
-    auto attribToBuffer = mInfo->mSoa.data<PipelineInfo::ATTRIBUTE_TO_BUFFER_INDEX>();
+    size_t const count = vbi->getAttributeCount();
+    VkBuffer* const vkbuffers = getVkBuffers();
+    int8_t const* const attribToBuffer = vbi->getAttributeToBuffer();
     for (uint8_t attribIndex = 0; attribIndex < count; attribIndex++) {
         if (attribToBuffer[attribIndex] == static_cast<int8_t>(index)) {
             vkbuffers[attribIndex] = bufferObject->buffer.getGpuBuffer();
@@ -361,35 +363,6 @@ VulkanBufferObject::VulkanBufferObject(VmaAllocator allocator, VulkanStagePool&
       buffer(allocator, stagePool, getBufferObjectUsage(bindingType), byteCount),
       bindingType(bindingType) {}
 
-void VulkanRenderPrimitive::setPrimitiveType(PrimitiveType pt) {
-    this->type = pt;
-    switch (pt) {
-        case PrimitiveType::POINTS:
-            primitiveTopology = VK_PRIMITIVE_TOPOLOGY_POINT_LIST;
-            break;
-        case PrimitiveType::LINES:
-            primitiveTopology = VK_PRIMITIVE_TOPOLOGY_LINE_LIST;
-            break;
-        case PrimitiveType::LINE_STRIP:
-            primitiveTopology = VK_PRIMITIVE_TOPOLOGY_LINE_STRIP;
-            break;
-        case PrimitiveType::TRIANGLES:
-            primitiveTopology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
-            break;
-        case PrimitiveType::TRIANGLE_STRIP:
-            primitiveTopology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP;
-            break;
-    }
-}
-
-void VulkanRenderPrimitive::setBuffers(VulkanVertexBuffer* vertexBuffer,
-        VulkanIndexBuffer* indexBuffer) {
-    this->vertexBuffer = vertexBuffer;
-    this->indexBuffer = indexBuffer;
-    mResources.acquire(vertexBuffer);
-    mResources.acquire(indexBuffer);
-}
-
 VulkanTimerQuery::VulkanTimerQuery(std::tuple<uint32_t, uint32_t> indices)
     : VulkanThreadSafeResource(VulkanResourceType::TIMER_QUERY),
       mStartingQueryIndex(std::get<0>(indices)),
@@ -421,4 +394,15 @@ bool VulkanTimerQuery::isCompleted() noexcept {
 
 VulkanTimerQuery::~VulkanTimerQuery() = default;
 
+VulkanRenderPrimitive::VulkanRenderPrimitive(VulkanResourceAllocator* resourceAllocator,
+        PrimitiveType pt, Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh)
+        : VulkanResource(VulkanResourceType::RENDER_PRIMITIVE),
+          mResources(resourceAllocator) {
+    type = pt;
+    vertexBuffer = resourceAllocator->handle_cast<VulkanVertexBuffer*>(vbh);
+    indexBuffer = resourceAllocator->handle_cast<VulkanIndexBuffer*>(ibh);
+    mResources.acquire(vertexBuffer);
+    mResources.acquire(indexBuffer);
+}
+
 } // namespace filament::backend

filament/backend/src/vulkan/VulkanHandles.h

@@ -64,6 +64,12 @@ struct VulkanProgram : public HwProgram, VulkanResource {
         return mInfo->bindingToSamplerIndex;
     }
 
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+    inline utils::FixedCapacityVector<std::string> const& getBindingToName() const {
+        return mInfo->bindingToName;
+    }
+#endif
+
 private:
     // TODO: handle compute shaders.
     // The expected order of shaders - from frontend to backend - is vertex, fragment, compute.
@@ -71,7 +77,11 @@ private:
 
     struct PipelineInfo {
         PipelineInfo()
-            : bindingToSamplerIndex(MAX_SAMPLER_COUNT, 0xffff) {}
+            : bindingToSamplerIndex(MAX_SAMPLER_COUNT, 0xffff)
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+            , bindingToName(MAX_SAMPLER_COUNT, "")
+#endif              
+            {}
 
         // This bitset maps to each of the sampler in the sampler groups associated with this
         // program, and whether each sampler is used in which shader (i.e. vert, frag, compute).
@@ -80,6 +90,12 @@ private:
         // We store the samplerGroupIndex as the top 8-bit and the index within each group as the lower 8-bit.
         utils::FixedCapacityVector<uint16_t> bindingToSamplerIndex;
         VkShaderModule shaders[MAX_SHADER_MODULES] = { VK_NULL_HANDLE };
+
+#if FVK_ENABLED_DEBUG_SAMPLER_NAME
+        // We store the sampler name mapped from binding index (only for debug purposes).
+        utils::FixedCapacityVector<std::string> bindingToName;
+#endif
+
     };
 
     PipelineInfo* mInfo;
@@ -132,7 +148,48 @@ struct VulkanBufferObject;
 struct VulkanVertexBufferInfo : public HwVertexBufferInfo, VulkanResource {
     VulkanVertexBufferInfo(uint8_t bufferCount, uint8_t attributeCount,
             AttributeArray const& attributes);
-    AttributeArray attributes;
+
+    inline VkVertexInputAttributeDescription const* getAttribDescriptions() const {
+        return mInfo.mSoa.data<PipelineInfo::ATTRIBUTE_DESCRIPTION>();
+    }
+
+    inline VkVertexInputBindingDescription const* getBufferDescriptions() const {
+        return mInfo.mSoa.data<PipelineInfo::BUFFER_DESCRIPTION>();
+    }
+
+    inline int8_t const* getAttributeToBuffer() const {
+        return mInfo.mSoa.data<PipelineInfo::ATTRIBUTE_TO_BUFFER_INDEX>();
+    }
+
+    inline VkDeviceSize const* getOffsets() const {
+        return mInfo.mSoa.data<PipelineInfo::OFFSETS>();
+    }
+
+    size_t getAttributeCount() const noexcept {
+        return mInfo.mSoa.size();
+    }
+
+private:
+    struct PipelineInfo {
+        PipelineInfo(size_t size) : mSoa(size /* capacity */) {
+            mSoa.resize(size);
+        }
+
+        // These correspond to the index of the element in the SoA
+        static constexpr uint8_t ATTRIBUTE_DESCRIPTION = 0;
+        static constexpr uint8_t BUFFER_DESCRIPTION = 1;
+        static constexpr uint8_t OFFSETS = 2;
+        static constexpr uint8_t ATTRIBUTE_TO_BUFFER_INDEX = 3;
+
+        utils::StructureOfArrays<
+            VkVertexInputAttributeDescription,
+            VkVertexInputBindingDescription,
+            VkDeviceSize,
+            int8_t
+        > mSoa;
+    };
+
+    PipelineInfo mInfo;
 };
 
 struct VulkanVertexBuffer : public HwVertexBuffer, VulkanResource {
@@ -140,53 +197,20 @@ struct VulkanVertexBuffer : public HwVertexBuffer, VulkanResource {
             VulkanResourceAllocator* allocator,
             uint32_t vertexCount, Handle<HwVertexBufferInfo> vbih);
 
-    ~VulkanVertexBuffer();
-
     void setBuffer(VulkanResourceAllocator const& allocator,
             VulkanBufferObject* bufferObject, uint32_t index);
 
-    inline VkVertexInputAttributeDescription const* getAttribDescriptions() {
-        return mInfo->mSoa.data<PipelineInfo::ATTRIBUTE_DESCRIPTION>();
-    }
-
-    inline VkVertexInputBindingDescription const* getBufferDescriptions() {
-        return mInfo->mSoa.data<PipelineInfo::BUFFER_DESCRIPTION>();
-    }
-
     inline VkBuffer const* getVkBuffers() const {
-        return mInfo->mSoa.data<PipelineInfo::VK_BUFFER>();
+        return mBuffers.data();
     }
 
-    inline VkDeviceSize const* getOffsets() const {
-        return mInfo->mSoa.data<PipelineInfo::OFFSETS>();
+    inline VkBuffer* getVkBuffers() {
+        return mBuffers.data();
     }
 
     Handle<HwVertexBufferInfo> vbih;
-
 private:
-    struct PipelineInfo {
-        PipelineInfo(size_t size)
-            : mSoa(size /* capacity */) {
-            mSoa.resize(size);
-        }
-
-        // These correspond to the index of the element in the SoA
-        static constexpr uint8_t ATTRIBUTE_DESCRIPTION = 0;
-        static constexpr uint8_t BUFFER_DESCRIPTION = 1;
-        static constexpr uint8_t VK_BUFFER = 2;
-        static constexpr uint8_t OFFSETS = 3;
-        static constexpr uint8_t ATTRIBUTE_TO_BUFFER_INDEX = 4;
-
-        utils::StructureOfArrays<
-            VkVertexInputAttributeDescription,
-            VkVertexInputBindingDescription,
-            VkBuffer,
-            VkDeviceSize,
-            int8_t
-        > mSoa;
-    };
-
-    PipelineInfo* mInfo;
+    utils::FixedCapacityVector<VkBuffer> mBuffers;
     FixedSizeVulkanResourceManager<MAX_VERTEX_BUFFER_COUNT> mResources;
 };
 
@@ -219,19 +243,15 @@ struct VulkanSamplerGroup : public HwSamplerGroup, VulkanResource {
 };
 
 struct VulkanRenderPrimitive : public HwRenderPrimitive, VulkanResource {
-    VulkanRenderPrimitive(VulkanResourceAllocator* allocator)
-        : VulkanResource(VulkanResourceType::RENDER_PRIMITIVE),
-          mResources(allocator) {}
+    VulkanRenderPrimitive(VulkanResourceAllocator* resourceAllocator,
+            PrimitiveType pt, Handle<HwVertexBuffer> vbh, Handle<HwIndexBuffer> ibh);
 
     ~VulkanRenderPrimitive() {
         mResources.clear();
     }
 
-    void setPrimitiveType(PrimitiveType pt);
-    void setBuffers(VulkanVertexBuffer* vertexBuffer, VulkanIndexBuffer* indexBuffer);
     VulkanVertexBuffer* vertexBuffer = nullptr;
     VulkanIndexBuffer* indexBuffer = nullptr;
-    VkPrimitiveTopology primitiveTopology;
 
 private:
     // Keep references to the vertex buffer and the index buffer.

filament/backend/src/vulkan/VulkanPipelineCache.cpp

@@ -563,6 +563,14 @@ VulkanPipelineCache::PipelineLayoutCacheEntry* VulkanPipelineCache::getOrCreateP
 void VulkanPipelineCache::bindProgram(VulkanProgram* program) noexcept {
     mPipelineRequirements.shaders[0] = program->getVertexShader();
     mPipelineRequirements.shaders[1] = program->getFragmentShader();
+
+    // If this is a debug build, validate the current shader.
+#if FVK_ENABLED(FVK_DEBUG_SHADER_MODULE)
+    if (mPipelineRequirements.shaders[0] == VK_NULL_HANDLE ||
+            mPipelineRequirements.shaders[1] == VK_NULL_HANDLE) {
+        utils::slog.e << "Binding missing shader: " << program->name.c_str() << utils::io::endl;
+    }
+#endif
 }
 
 void VulkanPipelineCache::bindRasterState(const RasterState& rasterState) noexcept {
@@ -628,12 +636,8 @@ void VulkanPipelineCache::unbindImageView(VkImageView imageView) noexcept {
 
 void VulkanPipelineCache::bindUniformBufferObject(uint32_t bindingIndex,
         VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept {
-    bindUniformBuffer(bindingIndex, bufferObject->buffer.getGpuBuffer(), offset, size);
-    mPipelineBoundResources.acquire(bufferObject);
-}
+    VkBuffer buffer = bufferObject->buffer.getGpuBuffer();
 
-void VulkanPipelineCache::bindUniformBuffer(uint32_t bindingIndex, VkBuffer buffer,
-        VkDeviceSize offset, VkDeviceSize size) noexcept {
     ASSERT_POSTCONDITION(bindingIndex < UBUFFER_BINDING_COUNT,
             "Uniform bindings overflow: index = %d, capacity = %d.", bindingIndex,
             UBUFFER_BINDING_COUNT);
@@ -649,6 +653,8 @@ void VulkanPipelineCache::bindUniformBuffer(uint32_t bindingIndex, VkBuffer buff
 
     key.uniformBufferOffsets[bindingIndex] = offset;
     key.uniformBufferSizes[bindingIndex] = size;
+
+    mPipelineBoundResources.acquire(bufferObject);
 }
 
 void VulkanPipelineCache::bindSamplers(VkDescriptorImageInfo samplers[SAMPLER_BINDING_COUNT],

filament/backend/src/vulkan/VulkanPipelineCache.h

@@ -157,8 +157,6 @@ public:
     void bindPrimitiveTopology(VkPrimitiveTopology topology) noexcept;
     void bindUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
             VkDeviceSize offset = 0, VkDeviceSize size = VK_WHOLE_SIZE) noexcept;
-    void bindUniformBuffer(uint32_t bindingIndex, VkBuffer buffer,
-            VkDeviceSize offset = 0, VkDeviceSize size = VK_WHOLE_SIZE) noexcept;
     void bindSamplers(VkDescriptorImageInfo samplers[SAMPLER_BINDING_COUNT],
             VulkanTexture* textures[SAMPLER_BINDING_COUNT], UsageFlags flags) noexcept;
     void bindInputAttachment(uint32_t bindingIndex, VkDescriptorImageInfo imageInfo) noexcept;
@@ -193,10 +191,19 @@ public:
         mDummyTargetInfo.imageView = imageView;
     }
 
-    // Acquires a resource to be bound to the current pipeline. The ownership of the resource
-    // will be transferred to the corresponding pipeline when pipeline is bound.
-    void acquireResource(VulkanResource* resource) {
-        mPipelineBoundResources.acquire(resource);
+    static VkPrimitiveTopology getPrimitiveTopology(PrimitiveType pt) noexcept {
+        switch (pt) {
+            case PrimitiveType::POINTS:
+                return VK_PRIMITIVE_TOPOLOGY_POINT_LIST;
+            case PrimitiveType::LINES:
+                return VK_PRIMITIVE_TOPOLOGY_LINE_LIST;
+            case PrimitiveType::LINE_STRIP:
+                return VK_PRIMITIVE_TOPOLOGY_LINE_STRIP;
+            case PrimitiveType::TRIANGLES:
+                return VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
+            case PrimitiveType::TRIANGLE_STRIP:
+                return VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP;
+        }
     }
 
 private:

filament/backend/src/vulkan/VulkanResourceAllocator.h

@@ -17,6 +17,7 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANRESOURCEALLOCATOR_H
 #define TNT_FILAMENT_BACKEND_VULKANRESOURCEALLOCATOR_H
 
+#include "VulkanConstants.h"
 #include "VulkanHandles.h"
 
 #include <private/backend/HandleAllocator.h>
@@ -29,18 +30,17 @@
 
 namespace filament::backend {
 
-// RESOURCE_TYPE_COUNT matches the count of enum VulkanResourceType.
-#define RESOURCE_TYPE_COUNT 12
-#define DEBUG_RESOURCE_LEAKS 0
+#define RESOURCE_TYPE_COUNT (static_cast<int>(VulkanResourceType::END_TYPE))
+#define DEBUG_RESOURCE_LEAKS FVK_ENABLED(FVK_DEBUG_RESOURCE_LEAK)
 
 #if DEBUG_RESOURCE_LEAKS
-    #define TRACK_INCREMENT()                                       \
-    if (!IS_MANAGED_TYPE(obj->mType)) {                             \
-        mDebugOnlyResourceCount[static_cast<size_t>(obj->mType)]++; \
+    #define TRACK_INCREMENT()                                           \
+    if (!IS_HEAP_ALLOC_TYPE(obj->getType())) {                          \
+        mDebugOnlyResourceCount[static_cast<size_t>(obj->getType())]++; \
     }
-    #define TRACK_DECREMENT()                                       \
-    if (!IS_MANAGED_TYPE(obj->mType)) {                             \
-        mDebugOnlyResourceCount[static_cast<size_t>(obj->mType)]--; \
+    #define TRACK_DECREMENT()                                           \
+    if (!IS_HEAP_ALLOC_TYPE(obj->getType())) {                          \
+        mDebugOnlyResourceCount[static_cast<size_t>(obj->getType())]--; \
     }
 #else
     // No-op
@@ -51,8 +51,8 @@ namespace filament::backend {
 class VulkanResourceAllocator {
 
 public:
-    VulkanResourceAllocator(size_t arenaSize)
-        : mHandleAllocatorImpl("Handles", arenaSize)
+    VulkanResourceAllocator(size_t arenaSize, bool disableUseAfterFreeCheck)
+        : mHandleAllocatorImpl("Handles", arenaSize, disableUseAfterFreeCheck)
 #if DEBUG_RESOURCE_LEAKS
         , mDebugOnlyResourceCount(RESOURCE_TYPE_COUNT) {
         std::memset(mDebugOnlyResourceCount.data(), 0, sizeof(size_t) * RESOURCE_TYPE_COUNT);

filament/backend/src/vulkan/VulkanResources.cpp

@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
-
 #include "VulkanResources.h"
 #include "VulkanHandles.h"
 #include "VulkanResourceAllocator.h"
+#include "caching/VulkanDescriptorSet.h"
 
 namespace filament::backend {
 
@@ -62,10 +62,15 @@ void deallocateResource(VulkanResourceAllocator* allocator, VulkanResourceType t
         case VulkanResourceType::RENDER_PRIMITIVE:
             allocator->destruct<VulkanRenderPrimitive>(Handle<VulkanRenderPrimitive>(id));
             break;
+        case VulkanResourceType::DESCRIPTOR_SET:
+            allocator->destruct<VulkanDescriptorSet>(Handle<VulkanDescriptorSet>(id));
+            break;
+
         // If the resource is heap allocated, then the resource manager just skip refcounted
         // destruction.
         case VulkanResourceType::FENCE:
         case VulkanResourceType::HEAP_ALLOCATED:
+        case VulkanResourceType::END_TYPE:
             break;
     }
 }

filament/backend/src/vulkan/VulkanResources.h

@@ -17,6 +17,8 @@
 #ifndef TNT_FILAMENT_BACKEND_VULKANRESOURCES_H
 #define TNT_FILAMENT_BACKEND_VULKANRESOURCES_H
 
+#include "VulkanUtility.h"
+
 #include <backend/Handle.h>
 
 #include <tsl/robin_set.h>
@@ -44,10 +46,13 @@ enum class VulkanResourceType : uint8_t {
     TIMER_QUERY,
     VERTEX_BUFFER,
     VERTEX_BUFFER_INFO,
+    DESCRIPTOR_SET,
 
     // Below are resources that are managed manually (i.e. not ref counted).
     FENCE,
     HEAP_ALLOCATED,
+
+    END_TYPE,  // A placeholder
 };
 
 #define IS_HEAP_ALLOC_TYPE(f)                                                                      \
@@ -162,62 +167,7 @@ namespace {
 // When the size of the resource set is known to be small, (for example for VulkanRenderPrimitive),
 // we just use a std::array to back the set.
 template<std::size_t SIZE>
-class FixedCapacityResourceSet {
-private:
-    using FixedSizeArray = std::array<VulkanResource*, SIZE>;
-
-public:
-    using const_iterator = typename FixedSizeArray::const_iterator;
-
-    inline ~FixedCapacityResourceSet() {
-        clear();
-    }
-
-    inline const_iterator begin() {
-        if (mInd == 0) {
-            return mArray.cend();
-        }
-        return mArray.cbegin();
-    }
-
-    inline const_iterator end() {
-        if (mInd == 0) {
-            return mArray.cend();
-        }
-        if (mInd < SIZE) {
-            return mArray.begin() + mInd;
-        }
-        return mArray.cend();
-    }
-
-    inline const_iterator find(VulkanResource* resource) {
-        return std::find(begin(), end(), resource);
-    }
-
-    inline void insert(VulkanResource* resource) {
-        assert_invariant(mInd < SIZE);
-        mArray[mInd++] = resource;
-    }
-
-    inline void erase(VulkanResource* resource) {
-        assert_invariant(false && "FixedCapacityResourceSet::erase should not be called");
-    }
-
-    inline void clear() {
-        if (mInd == 0) {
-            return;
-        }
-        mInd = 0;
-    }
-
-    inline size_t size() {
-        return mInd;
-    }
-
-private:
-    FixedSizeArray mArray{nullptr};
-    size_t mInd = 0;
-};
+using FixedCapacityResourceSet = CappedArray<VulkanResource*, SIZE>;
 
 // robin_set/map are useful for sets that are acquire only and the set will be iterated when the set
 // is cleared.

filament/backend/src/vulkan/VulkanSamplerCache.cpp

@@ -95,7 +95,8 @@ constexpr inline VkBool32 getCompareEnable(SamplerCompareMode mode) noexcept {
     return mode == SamplerCompareMode::NONE ? VK_FALSE : VK_TRUE;
 }
 
-void VulkanSamplerCache::initialize(VkDevice device) { mDevice = device; }
+VulkanSamplerCache::VulkanSamplerCache(VkDevice device)
+    : mDevice(device) {}
 
 VkSampler VulkanSamplerCache::getSampler(SamplerParams params) noexcept {
     auto iter = mCache.find(params);

filament/backend/src/vulkan/VulkanSamplerCache.h

@@ -27,7 +27,7 @@ namespace filament::backend {
 // Simple manager for VkSampler objects.
 class VulkanSamplerCache {
 public:
-    void initialize(VkDevice device);
+    explicit VulkanSamplerCache(VkDevice device);
     VkSampler getSampler(SamplerParams params) noexcept;
     void terminate() noexcept;
 private:

filament/backend/src/vulkan/VulkanStagePool.cpp

@@ -26,10 +26,9 @@ static constexpr uint32_t TIME_BEFORE_EVICTION = FVK_MAX_COMMAND_BUFFERS;
 
 namespace filament::backend {
 
-void VulkanStagePool::initialize(VmaAllocator allocator, VulkanCommands* commands) noexcept {
-    mAllocator = allocator;
-    mCommands = commands;
-}
+VulkanStagePool::VulkanStagePool(VmaAllocator allocator, VulkanCommands* commands)
+    : mAllocator(allocator),
+      mCommands(commands) {}
 
 VulkanStage const* VulkanStagePool::acquireStage(uint32_t numBytes) {
     // First check if a stage exists whose capacity is greater than or equal to the requested size.

filament/backend/src/vulkan/VulkanStagePool.h

@@ -45,7 +45,7 @@ struct VulkanStageImage {
 // This class manages two types of host-mappable staging areas: buffer stages and image stages.
 class VulkanStagePool {
 public:
-    void initialize(VmaAllocator allocator, VulkanCommands* commands) noexcept;
+    VulkanStagePool(VmaAllocator allocator, VulkanCommands* commands);
 
     // Finds or creates a stage whose capacity is at least the given number of bytes.
     // The stage is automatically released back to the pool after TIME_BEFORE_EVICTION frames.

filament/backend/src/vulkan/VulkanSwapChain.cpp

@@ -35,7 +35,7 @@ VulkanSwapChain::VulkanSwapChain(VulkanPlatform* platform, VulkanContext const&
       mStagePool(stagePool),
       mHeadless(extent.width != 0 && extent.height != 0 && !nativeWindow),
       mFlushAndWaitOnResize(platform->getCustomization().flushAndWaitOnWindowResize),
-      mImageReady(VK_NULL_HANDLE),
+      mCurrentImageReadyIndex(0),
       mAcquired(false),
       mIsFirstRenderPass(true) {
     swapChain = mPlatform->createSwapChain(nativeWindow, flags, extent);
@@ -46,10 +46,15 @@ VulkanSwapChain::VulkanSwapChain(VulkanPlatform* platform, VulkanContext const&
     };
 
     // No need to wait on this semaphore before drawing when in Headless mode.
-    if (!mHeadless) {
-        VkResult result =
-                vkCreateSemaphore(mPlatform->getDevice(), &createInfo, nullptr, &mImageReady);
-        ASSERT_POSTCONDITION(result == VK_SUCCESS, "Failed to create semaphore");
+    if (mHeadless) {
+		// Set all sempahores to VK_NULL_HANDLE
+		memset(mImageReady, 0, sizeof(mImageReady[0]) * IMAGE_READY_SEMAPHORE_COUNT);
+	} else {
+		for (uint32_t i = 0; i < IMAGE_READY_SEMAPHORE_COUNT; ++i) {
+			VkResult result =
+					vkCreateSemaphore(mPlatform->getDevice(), &createInfo, nullptr, mImageReady + i);
+			ASSERT_POSTCONDITION(result == VK_SUCCESS, "Failed to create semaphore");
+		}
     }
 
     update();
@@ -62,9 +67,11 @@ VulkanSwapChain::~VulkanSwapChain() {
     mCommands->wait();
 
     mPlatform->destroy(swapChain);
-    if (mImageReady != VK_NULL_HANDLE) {
-        vkDestroySemaphore(mPlatform->getDevice(), mImageReady, VKALLOC);
-    }
+	for (uint32_t i = 0; i < IMAGE_READY_SEMAPHORE_COUNT; ++i) {
+		if (mImageReady[i] != VK_NULL_HANDLE) {
+			vkDestroySemaphore(mPlatform->getDevice(), mImageReady[i], VKALLOC);
+		}
+	}
 }
 
 void VulkanSwapChain::update() {
@@ -131,11 +138,13 @@ void VulkanSwapChain::acquire(bool& resized) {
         update();
     }
 
-    VkResult const result = mPlatform->acquire(swapChain, mImageReady, &mCurrentSwapIndex);
+	mCurrentImageReadyIndex = (mCurrentImageReadyIndex + 1) % IMAGE_READY_SEMAPHORE_COUNT;
+	const VkSemaphore imageReady = mImageReady[mCurrentImageReadyIndex];
+    VkResult const result = mPlatform->acquire(swapChain, imageReady, &mCurrentSwapIndex);
     ASSERT_POSTCONDITION(result == VK_SUCCESS || result == VK_SUBOPTIMAL_KHR,
             "Cannot acquire in swapchain.");
-    if (mImageReady != VK_NULL_HANDLE) {
-        mCommands->injectDependency(mImageReady);
+    if (imageReady != VK_NULL_HANDLE) {
+        mCommands->injectDependency(imageReady);
     }
     mAcquired = true;
 }

filament/backend/src/vulkan/VulkanSwapChain.h

@@ -69,6 +69,8 @@ struct VulkanSwapChain : public HwSwapChain, VulkanResource {
     }
 
 private:
+	static constexpr int IMAGE_READY_SEMAPHORE_COUNT = FVK_MAX_COMMAND_BUFFERS;
+
     void update();
 
     VulkanPlatform* mPlatform;
@@ -83,7 +85,8 @@ private:
     utils::FixedCapacityVector<std::unique_ptr<VulkanTexture>> mColors;
     std::unique_ptr<VulkanTexture> mDepth;
     VkExtent2D mExtent;
-    VkSemaphore mImageReady;
+    VkSemaphore mImageReady[IMAGE_READY_SEMAPHORE_COUNT];
+	uint32_t mCurrentImageReadyIndex;
     uint32_t mCurrentSwapIndex;
     bool mAcquired;
     bool mIsFirstRenderPass;

filament/backend/src/vulkan/VulkanTexture.cpp

@@ -18,10 +18,10 @@
 #include "VulkanTexture.h"
 #include "VulkanUtility.h"
 
+#include <DataReshaper.h>
+#include <backend/DriverEnums.h>
 #include <private/backend/BackendUtils.h>
 
-#include "DataReshaper.h"
-
 #include <utils/Panic.h>
 
 using namespace bluevk;
@@ -100,9 +100,13 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
 
     // Filament expects blit() to work with any texture, so we almost always set these usage flags.
     // TODO: investigate performance implications of setting these flags.
-    const VkImageUsageFlags blittable = VK_IMAGE_USAGE_TRANSFER_DST_BIT |
+    constexpr VkImageUsageFlags blittable = VK_IMAGE_USAGE_TRANSFER_DST_BIT |
             VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
 
+    if (any(usage & (TextureUsage::BLIT_DST | TextureUsage::BLIT_SRC))) {
+        imageInfo.usage |= blittable;
+    }
+
     if (any(usage & TextureUsage::SAMPLEABLE)) {
 
 #if FVK_ENABLED(FVK_DEBUG_TEXTURE)
@@ -159,7 +163,7 @@ VulkanTexture::VulkanTexture(VkDevice device, VkPhysicalDevice physicalDevice,
     imageInfo.samples = (VkSampleCountFlagBits) samples;
 
     VkResult error = vkCreateImage(mDevice, &imageInfo, VKALLOC, &mTextureImage);
-    if (error || FVK_ENABLED_BOOL(FVK_DEBUG_TEXTURE)) {
+    if (error || FVK_ENABLED(FVK_DEBUG_TEXTURE)) {
         utils::slog.d << "vkCreateImage: "
             << "image = " << mTextureImage << ", "
             << "result = " << error << ", "

filament/backend/src/vulkan/VulkanUtility.h

@@ -23,6 +23,8 @@
 
 #include <bluevk/BlueVK.h>
 
+#include <utility>
+
 namespace filament::backend {
 
 VkFormat getVkFormat(ElementType type, bool normalized, bool integer);
@@ -88,10 +90,100 @@ utils::FixedCapacityVector<OutType> enumerate(
 #undef EXPAND_ENUM_NO_ARGS
 #undef EXPAND_ENUM_ARGS
 
-
 // Useful shorthands
 using VkFormatList = utils::FixedCapacityVector<VkFormat>;
 
+// An Array that will be fixed capacity, but the "size" (as in user added elements) is variable.
+// Note that this class is movable.
+template<typename T, uint16_t CAPACITY>
+class CappedArray {
+private:
+    using FixedSizeArray = std::array<T, CAPACITY>;
+public:
+    using const_iterator = typename FixedSizeArray::const_iterator;
+    using iterator = typename FixedSizeArray::iterator;
+
+    CappedArray() = default;
+
+    // Delete copy constructor/assignment.
+    CappedArray(CappedArray const& rhs) = delete;
+    CappedArray& operator=(CappedArray& rhs) = delete;
+
+    CappedArray(CappedArray&& rhs) noexcept {
+        this->swap(rhs);
+    }
+
+    CappedArray& operator=(CappedArray&& rhs) noexcept {
+        this->swap(rhs);
+        return *this;
+    }
+
+    inline ~CappedArray() {
+        clear();
+    }
+
+    inline const_iterator begin() {
+        if (mInd == 0) {
+            return mArray.cend();
+        }
+        return mArray.cbegin();
+    }
+
+    inline const_iterator end() {
+        if (mInd > 0 && mInd < CAPACITY) {
+            return mArray.begin() + mInd;
+        }
+        return mArray.cend();
+    }
+
+    inline T back() {
+        assert_invariant(mInd > 0);
+        return *(mArray.begin() + mInd);
+    }
+
+    inline void pop_back() {
+        assert_invariant(mInd > 0);
+        mInd--;
+    }
+
+    inline const_iterator find(T item) {
+        return std::find(begin(), end(), item);
+    }
+
+    inline void insert(T item) {
+        assert_invariant(mInd < CAPACITY);
+        mArray[mInd++] = item;
+    }
+
+    inline void erase(T item) {
+        PANIC_PRECONDITION("CappedArray::erase should not be called");
+    }
+
+    inline void clear() {
+        if (mInd == 0) {
+            return;
+        }
+        mInd = 0;
+    }
+
+    inline T& operator[](uint16_t ind) {
+        return mArray[ind];
+    }
+
+    inline size_t size() {
+        return mInd;
+    }
+
+private:
+    void swap(CappedArray& rhs) {
+        std::swap(mArray, rhs.mArray);
+        std::swap(mInd, rhs.mInd);
+    }
+
+    FixedSizeArray mArray;
+    size_t mInd = 0;
+};
+
 } // namespace filament::backend
 
 #endif // TNT_FILAMENT_BACKEND_VULKANUTILITY_H

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

@@ -0,0 +1,647 @@
+/*
+ * Copyright (C) 2024 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 "VulkanDescriptorSet.h"
+
+#include <vulkan/VulkanConstants.h>
+#include <vulkan/VulkanResources.h>
+
+#include <type_traits>
+
+namespace filament::backend {
+
+namespace {
+
+template<typename MaskType>
+using LayoutMap = tsl::robin_map<MaskType, VkDescriptorSetLayout>;
+using LayoutArray = VulkanDescriptorSetManager::LayoutArray;
+using UniformBufferBitmask = VulkanDescriptorSetManager::UniformBufferBitmask;
+using SamplerBitmask = VulkanDescriptorSetManager::SamplerBitmask;
+
+constexpr uint8_t EXPECTED_IN_FLIGHT_FRAMES = 3; // Asssume triple buffering
+
+#define PORT_CONSTANT(NameSpace, K) constexpr decltype(NameSpace::K) K = NameSpace::K;
+
+PORT_CONSTANT(VulkanDescriptorSetManager, MAX_SUPPORTED_SHADER_STAGE)
+PORT_CONSTANT(VulkanDescriptorSetManager, VERTEX_STAGE)
+PORT_CONSTANT(VulkanDescriptorSetManager, FRAGMENT_STAGE)
+
+PORT_CONSTANT(Program, UNIFORM_BINDING_COUNT)
+PORT_CONSTANT(Program, SAMPLER_BINDING_COUNT)
+
+#undef PORT_CONSTANT
+
+// Use constexpr to statically generate a bit count table for 8-bit numbers.
+struct BitCountHelper {
+    constexpr BitCountHelper() : data{} {
+        for (uint16_t i = 0; i < 256; ++i) {
+            data[i] = 0;
+            for (auto j = i; j > 0; j /= 2) {
+                if (j & 1) {
+                    data[i]++;
+                }
+            }
+        }
+    }
+    uint8_t data[256];
+} BitCountImpl;
+
+inline uint8_t countBits(uint32_t num) {
+    static constexpr uint8_t const* BIT_COUNT = BitCountImpl.data;
+    return BIT_COUNT[num & 0xFF] + BIT_COUNT[(num >> 8) & 0xFF] + BIT_COUNT[(num >> 16) & 0xFF] +
+           BIT_COUNT[(num >> 24) & 0xFF];
+}
+
+struct SamplerSet {
+    using Mask = UniformBufferBitmask;
+    struct Key {
+        Mask mask;
+        static_assert(sizeof(Mask) == 4);
+        uint32_t padding;// We need a padding to ensure 8-byte alignement.
+        VkDescriptorSetLayout layout;
+        VkSampler sampler[SAMPLER_BINDING_COUNT];
+        VkImageView imageView[SAMPLER_BINDING_COUNT];
+        VkImageLayout imageLayout[SAMPLER_BINDING_COUNT];
+    };
+
+    struct Equal {
+        bool operator()(Key const& k1, Key const& k2) const {
+            if (k1.mask != k2.mask) return false;
+            if (k1.layout != k2.layout) return false;
+
+            for (uint8_t i = 0, bitCount = countBits(k1.mask); i < bitCount; ++i) {
+                if (k1.sampler[i] != k2.sampler[i] || k1.imageView[i] != k2.imageView[i] ||
+                        k1.imageLayout[i] != k2.imageLayout[i]) {
+                    return false;
+                }
+            }
+            return true;
+        }
+    };
+    using HashFn = utils::hash::MurmurHashFn<Key>;
+    using Cache = std::unordered_map<Key, VulkanDescriptorSet*, HashFn, Equal>;
+};
+
+struct UBOSet {
+    using Mask = SamplerBitmask;
+
+    struct Key {
+        Mask mask;
+        static_assert(sizeof(Mask) == 4);
+        uint32_t padding;// We need a padding to ensure 8-byte alignement.
+        VkDescriptorSetLayout layout;
+        VkBuffer buffers[UNIFORM_BINDING_COUNT];    //   80     0
+        VkDeviceSize offsets[UNIFORM_BINDING_COUNT];//   40  1592
+        VkDeviceSize sizes[UNIFORM_BINDING_COUNT];  //   40  1632
+    };
+
+    struct Equal {
+        bool operator()(Key const& k1, Key const& k2) const {
+            if (k1.mask != k2.mask) return false;
+            if (k1.layout != k2.layout) return false;
+
+            for (uint8_t i = 0, bitCount = countBits(k1.mask); i < bitCount; ++i) {
+                if (k1.buffers[i] != k2.buffers[i] || k1.offsets[i] != k2.offsets[i] ||
+                        k1.sizes[i] != k2.sizes[i]) {
+                    return false;
+                }
+            }
+            return true;
+        }
+    };
+    using HashFn = utils::hash::MurmurHashFn<Key>;
+    using Cache = std::unordered_map<Key, VulkanDescriptorSet*, HashFn, Equal>;
+};
+
+constexpr uint8_t MAX_BINDINGS = UNIFORM_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE;
+static_assert(MAX_BINDINGS >= SAMPLER_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE);
+
+template<typename TYPE>
+class DescriptorPool {
+private:
+    static constexpr uint32_t COUNT_MULTIPLIER =
+            MAX_SUPPORTED_SHADER_STAGE * EXPECTED_IN_FLIGHT_FRAMES;
+    static constexpr uint32_t CAPACITY =
+            std::is_same_v<TYPE, UBOSet> ? Program::UNIFORM_BINDING_COUNT * COUNT_MULTIPLIER
+                                         : Program::SAMPLER_BINDING_COUNT * COUNT_MULTIPLIER;
+
+public:
+    DescriptorPool()
+        : mDevice(VK_NULL_HANDLE),
+          mResourceAllocator(nullptr) {}
+
+    DescriptorPool(VkDevice device, VulkanResourceAllocator* allocator)
+        : mDevice(device),
+          mResourceAllocator(allocator) {
+        VkDescriptorPoolSize sizes[1] = {
+            {
+                .type = std::is_same_v<TYPE, UBOSet> ? VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
+                                                     : VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+                .descriptorCount = CAPACITY,
+            },
+        };
+        VkDescriptorPoolCreateInfo info{
+            .flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT,
+            .maxSets = EXPECTED_IN_FLIGHT_FRAMES,
+            .poolSizeCount = 1,
+            .pPoolSizes = sizes,
+        };
+        vkCreateDescriptorPool(mDevice, &info, VKALLOC, &mPool);
+    }
+
+    DescriptorPool(DescriptorPool const&) = delete;
+    DescriptorPool& operator=(DescriptorPool const&) = delete;
+
+    ~DescriptorPool() {
+        vkDestroyDescriptorPool(mDevice, mPool, VKALLOC);
+    }
+
+    VulkanDescriptorSet* obtainSet(VkDescriptorSetLayout layout) {
+        if (mCount == CAPACITY) {
+            return nullptr;
+        }
+
+        if (auto unused = mUnused.find(layout);
+                unused != mUnused.end() && !unused->second.empty()) {
+            auto& unusedList = unused->second;
+            auto set = unusedList.back();
+            unusedList.pop_back();
+            mCount++;
+            return set;
+        }
+
+        // Creating a new set
+        VkDescriptorSetLayout layouts[1] = {layout};
+        VkDescriptorSetAllocateInfo allocInfo = {
+            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
+            .descriptorPool = mPool,
+            .descriptorSetCount = 1,
+            .pSetLayouts = layouts,
+        };
+        VkDescriptorSet vkSet;
+        UTILS_UNUSED VkResult result = vkAllocateDescriptorSets(mDevice, &allocInfo, &vkSet);
+        ASSERT_POSTCONDITION(result == VK_SUCCESS, "Cannot allocate descriptor set. error=%d",
+                (int) result);
+
+        mCount++;
+        return createSet(layout, vkSet);
+    }
+
+private:
+    inline VulkanDescriptorSet* createSet(VkDescriptorSetLayout layout, VkDescriptorSet vkSet) {
+        return VulkanDescriptorSet::create(mResourceAllocator, vkSet, layout,
+                [this](VulkanDescriptorSet* set) {
+                    auto const layout = set->layout;
+                    auto const vkSet = set->vkSet;
+                    auto listIter = mUnused.find(layout);
+                    assert_invariant(listIter != mUnused.end());
+                    auto& unusedList = listIter->second;
+
+                    // At this point, we know that the pointer will be stale, and we need to remove
+                    // any non-ref-counted references to it.
+                    // TODO: this will be inefficient
+                    auto iter = std::find(unusedList.begin(), unusedList.end(), set);
+                    unusedList.erase(iter);
+
+                    // We are recycling - release the vk resource back into the pool. Note that the
+                    // vk handle has not changed, but we need to change the backend handle to allow
+                    // for proper refcounting of resources referenced in this set.
+                    unusedList.push_back(createSet(layout, vkSet));
+                    mCount--;
+                });
+    }
+
+    VkDevice mDevice;
+    VulkanResourceAllocator* mResourceAllocator;
+    VkDescriptorPool mPool;
+
+    uint32_t mCount = 0;
+    std::unordered_map<VkDescriptorSetLayout, std::vector<VulkanDescriptorSet*>> mUnused;
+};
+
+template<typename TYPE>
+class DescriptorInfinitePool {
+public:
+    DescriptorInfinitePool(VkDevice device, VulkanResourceAllocator* allocator)
+        : mDevice(device),
+          mResourceAllocator(allocator) {}
+
+    ~DescriptorInfinitePool() {
+        for (auto pool: mPools) {
+            delete pool;
+        }
+    }
+
+    VulkanDescriptorSet* obtainSet(VkDescriptorSetLayout layout) {
+        for (auto pool: mPools) {
+            auto set = pool->obtainSet(layout);
+            if (set) {
+                return set;
+            }
+        }
+        // We need to increase the number of pools
+        mPools.push_back(new DescriptorPool<TYPE>{mDevice, mResourceAllocator});
+        auto pool = mPools.back();
+        return pool->obtainSet(layout);
+    }
+
+private:
+    VkDevice mDevice;
+    VulkanResourceAllocator* mResourceAllocator;
+    std::vector<DescriptorPool<TYPE>*> mPools;
+};
+
+// This holds a cache of descriptor sets of a TYPE (UBO or Sampler). The Key is defined as the
+// layout and the content (for example specific samplers).
+template<typename TYPE, typename Cache, typename Key>
+class DescriptorSetCache {
+public:
+    DescriptorSetCache(VkDevice device, VulkanResourceAllocator* allocator)
+        : mPool(device, allocator),
+          mResourceManager(allocator) {}
+
+    std::pair<VulkanDescriptorSet*, bool> obtainSet(Key const& key) {
+        mAge++;
+        if (auto result = mCache.find(key); result != mCache.end()) {
+            auto set = result->second;
+            auto const oldAge = mReverseHistory[set];
+            mHistory.erase(oldAge);
+            mHistory[mAge] = set;
+            mReverseHistory[set] = mAge;
+            return {set, true};
+        }
+
+        VulkanDescriptorSet* set = mPool.obtainSet(key.layout);
+        mCache[key] = set;
+        mReverseCache[set] = key;
+        mHistory[mAge] = set;
+        mReverseHistory[set] = mAge;
+        mResourceManager.acquire(set);
+        return {set, false};
+    }
+
+    // We need to periodically purge the descriptor sets so that we're not holding on to resources
+    // unnecessarily. The strategy for purging needs to be examined more.
+    void gc() noexcept {
+        constexpr uint32_t ALLOWED_ENTRIES = EXPECTED_IN_FLIGHT_FRAMES * 3;
+        int32_t const toCutCount = mHistory.size() - ALLOWED_ENTRIES;
+
+        if (toCutCount <= 0) {
+            return;
+        }
+
+        std::vector<uint64_t> remove;
+        for (auto entry = mHistory.begin(); entry != mHistory.end(); entry++) {
+            auto const& set = entry->second;
+            Key const& key = mReverseCache[set];
+            mCache.erase(key);
+            mReverseCache.erase(set);
+            mReverseHistory.erase(set);
+
+            remove.push_back(entry->first);
+            mResourceManager.release(set);
+        }
+        for (auto removeAge: remove) {
+            mHistory.erase(removeAge);
+        }
+    }
+
+private:
+    DescriptorInfinitePool<TYPE> mPool;
+
+    // TODO: combine some of these data structures
+    Cache mCache;
+    std::unordered_map<VulkanDescriptorSet*, Key> mReverseCache;
+
+    // Use the ordering for purging if needed;
+    std::map<uint64_t, VulkanDescriptorSet*> mHistory;
+    std::unordered_map<VulkanDescriptorSet*, uint64_t> mReverseHistory;
+    uint64_t mAge;
+
+    // Note that instead of owning the resources (i.e. descriptor set) in the pools, we keep them
+    // here since all the allocated sets have to pass through this class, and this class has better
+    // knowledge to make decisions about gc.
+    VulkanResourceManager mResourceManager;
+};
+
+using UBOSetCache = DescriptorSetCache<UBOSet, UBOSet::Cache, UBOSet::Key>;
+using SamplerSetCache = DescriptorSetCache<SamplerSet, SamplerSet::Cache, SamplerSet::Key>;
+
+template<typename MaskType>
+class LayoutCache {
+public:
+    explicit LayoutCache(VkDevice device)
+        : mDevice(device) {}
+
+    ~LayoutCache() {
+        for (auto [mask, layout]: mLayouts) {
+            vkDestroyDescriptorSetLayout(mDevice, layout, VKALLOC);
+        }
+        mLayouts.clear();
+    }
+
+    VkDescriptorSetLayout getLayout(MaskType stageMask) noexcept {
+        if (stageMask == 0) {
+            return VK_NULL_HANDLE;
+        }
+        if (auto iter = mLayouts.find(stageMask); iter != mLayouts.end()) {
+            return iter->second;
+        }
+        VkDescriptorType descriptorType;
+        if constexpr (std::is_same_v<MaskType, UBOSet::Mask>) {
+            descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
+        } else if constexpr (std::is_same_v<MaskType, SamplerSet::Mask>) {
+            descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
+        }
+
+        VkDescriptorSetLayoutBinding toBind[MAX_BINDINGS];
+        uint32_t count = 0;
+        for (uint8_t i = 0, maxBindings = sizeof(stageMask) * 4; i < maxBindings; i++) {
+            VkShaderStageFlags stages = 0;
+            if ((stageMask & (VERTEX_STAGE << i)) != 0) {
+                stages |= VK_SHADER_STAGE_VERTEX_BIT;
+            }
+            if ((stageMask & (FRAGMENT_STAGE << i)) == 0) {
+                stages |= VK_SHADER_STAGE_FRAGMENT_BIT;
+            }
+            if (stages == 0) {
+                continue;
+            }
+
+            auto& bindInfo = toBind[count++];
+            bindInfo = {
+                .binding = i,
+                .descriptorType = descriptorType,
+                .descriptorCount = 1,
+                .stageFlags = stages,
+            };
+        }
+        VkDescriptorSetLayoutCreateInfo dlinfo = {
+            .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+            .bindingCount = count,
+            .pBindings = toBind,
+        };
+
+        VkDescriptorSetLayout handle;
+        vkCreateDescriptorSetLayout(mDevice, &dlinfo, VKALLOC, &handle);
+        mLayouts[stageMask] = handle;
+
+        return handle;
+    }
+
+private:
+    VkDevice mDevice;
+    LayoutMap<MaskType> mLayouts;
+};
+
+using UBOSetLayoutCache = LayoutCache<UBOSet::Mask>;
+using SamplerSetLayoutCache = LayoutCache<SamplerSet::Mask>;
+
+} // anonymous namespace
+
+VulkanDescriptorSet::VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
+        VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn)
+    : VulkanResource(VulkanResourceType::DESCRIPTOR_SET),
+      resources(allocator),
+      vkSet(rawSet),
+      layout(layout),
+      mOnRecycleFn(std::move(onRecycleFn)) {}
+
+VulkanDescriptorSet::~VulkanDescriptorSet() {
+    if (mOnRecycleFn) {
+        mOnRecycleFn(this);
+    }
+}
+
+VulkanDescriptorSet* VulkanDescriptorSet::create(VulkanResourceAllocator* allocator,
+        VkDescriptorSet rawSet, VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn) {
+    auto handle = allocator->allocHandle<VulkanDescriptorSet>();
+    auto set = allocator->construct<VulkanDescriptorSet>(handle, allocator, rawSet, layout,
+            std::move(onRecycleFn));
+    return set;
+}
+
+class VulkanDescriptorSetManager::Impl {
+public:
+    Impl(VkDevice device, VulkanResourceAllocator* allocator)
+        : mDevice(device),
+          mUBOLayoutCache(device),
+          mUBOCache(device, allocator),
+          mSamplerLayoutCache(device),
+          mSamplerCache(device, allocator),
+          mResources(allocator) {}
+
+    void setUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
+            VkDeviceSize offset, VkDeviceSize size) noexcept {
+        mUbos.push_back({bindingIndex, bufferObject, offset, size});
+
+        // Between "set" and "commit", we need to ref the buffer object to avoid it being collected.
+        mResources.acquire(bufferObject);
+    }
+
+    void setSamplers(SamplerArray&& samplers) {
+        mSamplers = std::move(samplers);
+        for (auto const& sampler: mSamplers) {
+            mResources.acquire(sampler.texture);
+        }
+    }
+
+    void gc() noexcept {
+        mUBOCache.gc();
+        mSamplerCache.gc();
+    }
+
+    // This will write/update all of the descriptor set (and create a set if a one of the same
+    // layout is not available).
+    void bind(VulkanCommandBuffer* commands, GetPipelineLayoutFunction& getPipelineLayoutFn) {
+        LayoutArray layouts = {VK_NULL_HANDLE, VK_NULL_HANDLE};
+        std::array<VkDescriptorSet, 2> descSets = {VK_NULL_HANDLE, VK_NULL_HANDLE};
+        uint8_t descSetCount = 0;
+
+        VkWriteDescriptorSet descriptorWrites[UNIFORM_BINDING_COUNT + SAMPLER_BINDING_COUNT];
+        VkDescriptorBufferInfo uboWrite[UNIFORM_BINDING_COUNT];
+        VkDescriptorImageInfo samplerWrite[SAMPLER_BINDING_COUNT];
+        uint32_t nwrites = 0;
+
+        UBOSet::Key uboKey = {};
+        auto& uboMask = uboKey.mask;
+        uint8_t uboCount = 0;
+        for (auto const& ubo: mUbos) {
+            auto const& binding = std::get<0>(ubo);
+
+            uboKey.buffers[uboCount] = std::get<1>(ubo)->buffer.getGpuBuffer();
+            uboKey.offsets[uboCount] = std::get<2>(ubo);
+            uboKey.sizes[uboCount] = std::get<3>(ubo);
+            uboCount++;
+
+            // Currently we let ubo be visible in all stages.
+            uboMask |= (VERTEX_STAGE << binding);
+            uboMask |= (FRAGMENT_STAGE << binding);
+        }
+
+        if (uboMask) {
+            layouts[UBO_SET_INDEX] = uboKey.layout = mUBOLayoutCache.getLayout(uboMask);
+
+            auto [descriptorSet, cached] = mUBOCache.obtainSet(uboKey);
+            descSets[descSetCount++] = descriptorSet->vkSet;
+
+            // We need to write to the descriptor set since it wasn't cached.
+            if (!cached) {
+                // If it wasn't cached before, we need to ref the resources that it touches.
+                for (auto const& ubo: mUbos) {
+                    auto const buffer = std::get<1>(ubo);
+                    descriptorSet->resources.acquire(buffer);
+                }
+
+                for (uint8_t i = 0; i < uboCount; ++i) {
+                    auto const& binding = std::get<0>(mUbos[i]);
+                    VkWriteDescriptorSet& descriptorWrite = descriptorWrites[nwrites++];
+                    auto& writeInfo = uboWrite[i];
+                    writeInfo = {
+                        .buffer = uboKey.buffers[i],
+                        .offset = uboKey.offsets[i],
+                        .range = uboKey.sizes[i],
+                    };
+                    descriptorWrite = {
+                        .dstSet = descriptorSet->vkSet,
+                        .dstBinding = binding,
+                        .descriptorCount = 1,
+                        .descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
+                        .pBufferInfo = &writeInfo,
+                    };
+                }
+            }
+            commands->acquire(descriptorSet);
+        }
+
+        SamplerSet::Key samplerKey = {};
+        auto& samplerMask = samplerKey.mask;
+        uint8_t samplerCount = 0;
+        for (auto const sampler: mSamplers) {
+            samplerMask |= (sampler.stage << sampler.binding);
+
+            samplerKey.sampler[samplerCount] = sampler.info.sampler;
+            samplerKey.imageView[samplerCount] = sampler.info.imageView;
+            samplerKey.imageLayout[samplerCount] = sampler.info.imageLayout;
+            samplerCount++;
+        }
+
+        if (samplerMask) {
+            layouts[SAMPLER_SET_INDEX] = samplerKey.layout =
+                    mSamplerLayoutCache.getLayout(samplerMask);
+
+            auto [descriptorSet, cached] = mSamplerCache.obtainSet(samplerKey);
+            descSets[descSetCount++] = descriptorSet->vkSet;
+
+            if (!cached) {
+                for (auto const& sampler: mSamplers) {
+                    descriptorSet->resources.acquire(sampler.texture);
+                }
+
+                for (uint8_t i = 0; i < samplerCount; ++i) {
+                    auto const& binding = mSamplers[i].binding;
+                    VkWriteDescriptorSet& descriptorWrite = descriptorWrites[nwrites++];
+                    auto& writeInfo = samplerWrite[i];
+                    writeInfo = mSamplers[i].info;
+                    descriptorWrite = {
+                        .dstSet = descriptorSet->vkSet,
+                        .dstBinding = binding,
+                        .descriptorCount = 1,
+                        .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+                        .pImageInfo = &writeInfo,
+                    };
+                }
+            }
+            commands->acquire(descriptorSet);
+        }
+        if (nwrites) {
+            vkUpdateDescriptorSets(mDevice, nwrites, descriptorWrites, 0, nullptr);
+        }
+
+        VkPipelineLayout const pipelineLayout = getPipelineLayoutFn(layouts);
+        VkCommandBuffer const cmdbuffer = commands->buffer();
+
+        BoundState state {
+            .cmdbuf = cmdbuffer,
+            .pipelineLayout = pipelineLayout,
+        };
+
+        if (state == mPreviousBoundState) {
+            return;
+        }
+
+        vkCmdBindDescriptorSets(cmdbuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0,
+                descSetCount, descSets.data(), 0, nullptr);
+
+        mPreviousBoundState = state;
+
+        // Once bound, the resources are now ref'd in the descriptor set and the references in this
+        // class can be released and the descriptor set is ref'd by the command buffer.
+        mResources.clear();
+    }
+
+private:
+    struct BoundState {
+        VkCommandBuffer cmdbuf;
+        VkPipelineLayout pipelineLayout;
+
+        bool operator==(BoundState const& b) const {
+            return cmdbuf == b.cmdbuf && pipelineLayout == b.pipelineLayout;
+        }
+    };
+
+    VkDevice mDevice;
+
+    UBOSetLayoutCache mUBOLayoutCache;
+    UBOSetCache mUBOCache;
+
+    SamplerSetLayoutCache mSamplerLayoutCache;
+    SamplerSetCache mSamplerCache;
+
+    std::vector<std::tuple<uint32_t, VulkanBufferObject*, VkDeviceSize, VkDeviceSize>> mUbos;
+    SamplerArray mSamplers;
+    VulkanAcquireOnlyResourceManager mResources;
+
+    BoundState mPreviousBoundState;
+};
+
+VulkanDescriptorSetManager::VulkanDescriptorSetManager(VkDevice device,
+        VulkanResourceAllocator* resourceAllocator)
+    : mImpl(new Impl(device, resourceAllocator)) {}
+
+void VulkanDescriptorSetManager::terminate() noexcept {
+    assert_invariant(mImpl);
+    delete mImpl;
+    mImpl = nullptr;
+}
+
+void VulkanDescriptorSetManager::gc() noexcept { mImpl->gc(); }
+
+void VulkanDescriptorSetManager::bind(VulkanCommandBuffer* commands,
+        GetPipelineLayoutFunction& getPipelineLayoutFn) {
+    mImpl->bind(commands, getPipelineLayoutFn);
+}
+
+void VulkanDescriptorSetManager::setUniformBufferObject(uint32_t bindingIndex,
+        VulkanBufferObject* bufferObject, VkDeviceSize offset, VkDeviceSize size) noexcept {
+    mImpl->setUniformBufferObject(bindingIndex, bufferObject, offset, size);
+}
+
+void VulkanDescriptorSetManager::setSamplers(SamplerArray&& samplers) {
+    mImpl->setSamplers(std::move(samplers));
+}
+
+} // namespace filament::backend

filament/backend/src/vulkan/caching/VulkanDescriptorSet.h

@@ -0,0 +1,111 @@
+/*
+ * Copyright (C) 2024 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H
+#define TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H
+
+#include <vulkan/VulkanResourceAllocator.h>
+#include <vulkan/VulkanTexture.h>
+#include <vulkan/VulkanUtility.h>
+
+#include <backend/DriverEnums.h>
+#include <backend/Program.h>
+#include <backend/TargetBufferInfo.h>
+
+#include <utils/bitset.h>
+
+#include <bluevk/BlueVK.h>
+#include <tsl/robin_map.h>
+
+namespace filament::backend {
+
+// We need to make this class public to enable allocation on the HandleAllocator.
+struct VulkanDescriptorSet : public VulkanResourceBase {
+public:
+    // Because we need to recycle descriptor set not used, we allow for a callback that the "Pool"
+    // can use to repackage the vk handle.
+    using OnRecycle = std::function<void(VulkanDescriptorSet*)>;
+
+    VulkanDescriptorSet(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
+            VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn);
+
+    ~VulkanDescriptorSet();
+
+    static VulkanDescriptorSet* create(VulkanResourceAllocator* allocator, VkDescriptorSet rawSet,
+            VkDescriptorSetLayout layout, OnRecycle&& onRecycleFn);
+
+    // TODO: maybe change to fixed size for performance.
+    VulkanAcquireOnlyResourceManager resources;
+
+    VkDescriptorSet const vkSet;
+    VkDescriptorSetLayout const layout;
+
+private:
+    OnRecycle mOnRecycleFn;
+};
+
+// Abstraction over the pool and the layout cache.
+class VulkanDescriptorSetManager {
+public:
+    using StageBitMask = uint32_t;
+    using UniformBufferBitmask = StageBitMask;
+    using SamplerBitmask = StageBitMask;
+
+    static constexpr uint8_t CONCURRENT_DESCRIPTOR_SET_COUNT = 2;// UBO and samplers
+    static constexpr uint8_t MAX_SUPPORTED_SHADER_STAGE = 2;     // Vertex and fragment.
+
+    static_assert(sizeof(UniformBufferBitmask) * 8 >=
+                  (Program::UNIFORM_BINDING_COUNT) *MAX_SUPPORTED_SHADER_STAGE);
+    static_assert(sizeof(SamplerBitmask) * 8 >=
+                  Program::SAMPLER_BINDING_COUNT * MAX_SUPPORTED_SHADER_STAGE);
+
+    static constexpr StageBitMask VERTEX_STAGE = 0x1;
+    static constexpr StageBitMask FRAGMENT_STAGE = (0x1 << (sizeof(StageBitMask) / 4));
+    static constexpr uint8_t UBO_SET_INDEX = 0;
+    static constexpr uint8_t SAMPLER_SET_INDEX = 1;
+
+    using LayoutArray = std::array<VkDescriptorSetLayout, CONCURRENT_DESCRIPTOR_SET_COUNT>;
+    struct SamplerBundle {
+        VkDescriptorImageInfo info = {};
+        VulkanTexture* texture = nullptr;
+        uint8_t binding = 0;
+        SamplerBitmask stage = 0;
+    };
+    using SamplerArray = CappedArray<SamplerBundle, Program::SAMPLER_BINDING_COUNT>;
+    using GetPipelineLayoutFunction = std::function<VkPipelineLayout(LayoutArray)>;
+
+    VulkanDescriptorSetManager(VkDevice device, VulkanResourceAllocator* resourceAllocator);
+
+    void terminate() noexcept;
+
+    void gc() noexcept;
+
+    // This will write/update all of the descriptor set.
+    void bind(VulkanCommandBuffer* commands, GetPipelineLayoutFunction& getPipelineFn);
+
+    void setUniformBufferObject(uint32_t bindingIndex, VulkanBufferObject* bufferObject,
+            VkDeviceSize offset, VkDeviceSize size) noexcept;
+
+    void setSamplers(SamplerArray&& samplers);
+
+private:
+    class Impl;
+    Impl* mImpl;
+};
+
+} // namespace filament::backend
+
+#endif// TNT_FILAMENT_BACKEND_CACHING_VULKANDESCRIPTORSET_H

filament/backend/test/BackendTest.cpp

@@ -51,7 +51,7 @@ void BackendTest::init(Backend backend, bool isMobilePlatform) {
 }
 
 BackendTest::BackendTest() : commandBufferQueue(CONFIG_MIN_COMMAND_BUFFERS_SIZE,
-            CONFIG_COMMAND_BUFFERS_SIZE) {
+        CONFIG_COMMAND_BUFFERS_SIZE, /*mPaused=*/false) {
     initializeDriver();
 }
 

filament/backend/test/ComputeTest.cpp

@@ -46,7 +46,8 @@ void ComputeTest::init(Backend backend) {
 }
 
 ComputeTest::ComputeTest()
-        : commandBufferQueue(CONFIG_MIN_COMMAND_BUFFERS_SIZE, CONFIG_COMMAND_BUFFERS_SIZE) {
+    : commandBufferQueue(CONFIG_MIN_COMMAND_BUFFERS_SIZE, CONFIG_COMMAND_BUFFERS_SIZE,
+            /*paused=*/false) {
 }
 
 ComputeTest::~ComputeTest() = default;

filament/backend/test/test_Blit.cpp

@@ -226,9 +226,9 @@ TEST_F(BackendTest, ColorMagnify) {
     Handle<HwRenderTarget> dstRenderTargets[kNumLevels];
     for (uint8_t level = 0; level < kNumLevels; level++) {
         srcRenderTargets[level] = api.createRenderTarget( TargetBufferFlags::COLOR,
-                kSrcTexWidth >> level, kSrcTexHeight >> level, 1, { srcTexture, level, 0 }, {}, {});
+                kSrcTexWidth >> level, kSrcTexHeight >> level, 1, 0, { srcTexture, level, 0 }, {}, {});
         dstRenderTargets[level] = api.createRenderTarget( TargetBufferFlags::COLOR,
-                kDstTexWidth >> level, kDstTexHeight >> level, 1, { dstTexture, level, 0 }, {}, {});
+                kDstTexWidth >> level, kDstTexHeight >> level, 1, 0, { dstTexture, level, 0 }, {}, {});
     }
 
     // Do a "magnify" blit from level 1 of the source RT to the level 0 of the destination RT.
@@ -301,9 +301,9 @@ TEST_F(BackendTest, ColorMinify) {
     Handle<HwRenderTarget> dstRenderTargets[kNumLevels];
     for (uint8_t level = 0; level < kNumLevels; level++) {
         srcRenderTargets[level] = api.createRenderTarget( TargetBufferFlags::COLOR,
-                kSrcTexWidth >> level, kSrcTexHeight >> level, 1, { srcTexture, level, 0 }, {}, {});
+                kSrcTexWidth >> level, kSrcTexHeight >> level, 1, 0, { srcTexture, level, 0 }, {}, {});
         dstRenderTargets[level] = api.createRenderTarget( TargetBufferFlags::COLOR,
-                kDstTexWidth >> level, kDstTexHeight >> level, 1, { dstTexture, level, 0 }, {}, {});
+                kDstTexWidth >> level, kDstTexHeight >> level, 1, 0, { dstTexture, level, 0 }, {}, {});
     }
 
     // Do a "minify" blit from level 1 of the source RT to the level 0 of the destination RT.
@@ -369,12 +369,12 @@ TEST_F(BackendTest, ColorResolve) {
 
     // Create a 4-sample render target with the 4-sample texture.
     Handle<HwRenderTarget> const srcRenderTarget = api.createRenderTarget(
-            TargetBufferFlags::COLOR, kSrcTexWidth, kSrcTexHeight, kSampleCount,
+            TargetBufferFlags::COLOR, kSrcTexWidth, kSrcTexHeight, kSampleCount, 0,
             {{ srcColorTexture }}, {}, {});
 
     // Create a 1-sample render target with the 1-sample texture.
     Handle<HwRenderTarget> const dstRenderTarget = api.createRenderTarget(
-            TargetBufferFlags::COLOR, kDstTexWidth, kDstTexHeight, 1,
+            TargetBufferFlags::COLOR, kDstTexWidth, kDstTexHeight, 1, 0,
             {{ dstColorTexture }}, {}, {});
 
     // Prep for rendering.
@@ -473,9 +473,9 @@ TEST_F(BackendTest, Blit2DTextureArray) {
     // Create two RenderTargets.
     const int level = 0;
     Handle<HwRenderTarget> srcRenderTarget = api.createRenderTarget( TargetBufferFlags::COLOR,
-            kSrcTexWidth >> level, kSrcTexHeight >> level, 1, { srcTexture, level, kSrcTexLayer }, {}, {});
+            kSrcTexWidth >> level, kSrcTexHeight >> level, 1, 0, { srcTexture, level, kSrcTexLayer }, {}, {});
     Handle<HwRenderTarget> dstRenderTarget = api.createRenderTarget( TargetBufferFlags::COLOR,
-            kDstTexWidth >> level, kDstTexHeight >> level, 1, { dstTexture, level, kDstTexLayer }, {}, {});
+            kDstTexWidth >> level, kDstTexHeight >> level, 1, 0, { dstTexture, level, kDstTexLayer }, {}, {});
 
     // Do a blit from kSrcTexLayer of the source RT to kDstTexLayer of the destination RT.
     const int srcLevel = 0;
@@ -565,10 +565,10 @@ TEST_F(BackendTest, BlitRegion) {
     // that this case is handled correctly.
     Handle<HwRenderTarget> srcRenderTarget =
             api.createRenderTarget(TargetBufferFlags::COLOR, srcRect.width,
-                    srcRect.height, 1, {srcTexture, kSrcLevel, 0}, {}, {});
+                    srcRect.height, 1, 0, {srcTexture, kSrcLevel, 0}, {}, {});
     Handle<HwRenderTarget> dstRenderTarget =
             api.createRenderTarget(TargetBufferFlags::COLOR, kDstTexWidth >> kDstLevel,
-                    kDstTexHeight >> kDstLevel, 1, {dstTexture, kDstLevel, 0}, {}, {});
+                    kDstTexHeight >> kDstLevel, 1, 0, {dstTexture, kDstLevel, 0}, {}, {});
 
     api.blitDEPRECATED(TargetBufferFlags::COLOR0, dstRenderTarget, dstRect, srcRenderTarget, srcRect,
             SamplerMagFilter::LINEAR);
@@ -637,7 +637,7 @@ TEST_F(BackendTest, BlitRegionToSwapChain) {
     Handle<HwRenderTarget> srcRenderTargets[kNumLevels];
     for (uint8_t level = 0; level < kNumLevels; level++) {
         srcRenderTargets[level] = api.createRenderTarget( TargetBufferFlags::COLOR,
-                kSrcTexWidth >> level, kSrcTexHeight >> level, 1, { srcTexture, level, 0 }, {}, {});
+                kSrcTexWidth >> level, kSrcTexHeight >> level, 1, 0, { srcTexture, level, 0 }, {}, {});
     }
 
     // Blit one-quarter of src level 1 to dst level 0.

filament/backend/test/test_BufferUpdates.cpp

@@ -216,7 +216,7 @@ TEST_F(BackendTest, BufferObjectUpdateWithOffset) {
     auto colorTexture = getDriverApi().createTexture(SamplerType::SAMPLER_2D, 1,
             TextureFormat::RGBA8, 1, 512, 512, 1, TextureUsage::COLOR_ATTACHMENT);
     auto renderTarget = getDriverApi().createRenderTarget(
-            TargetBufferFlags::COLOR0, 512, 512, 1, {{colorTexture}}, {}, {});
+            TargetBufferFlags::COLOR0, 512, 512, 1, 0, {{colorTexture}}, {}, {});
 
     // Upload uniforms for the first triangle.
     {

filament/backend/test/test_FeedbackLoops.cpp

@@ -160,7 +160,7 @@ TEST_F(BackendTest, FeedbackLoops) {
             slog.i << "Level " << int(level) << ": " <<
                     (kTexWidth >> level) << "x" << (kTexHeight >> level) << io::endl;
             renderTargets[level] = api.createRenderTarget( TargetBufferFlags::COLOR,
-                    kTexWidth >> level, kTexHeight >> level, 1, { texture, level, 0 }, {}, {});
+                    kTexWidth >> level, kTexHeight >> level, 1, 0, { texture, level, 0 }, {}, {});
         }
 
         // Fill the base level of the texture with interesting colors.