Flutter渲染机制—UI线程

Posted by Gityuan on June 15, 2019

基于Flutter 1.5,从源码视角来深入剖析flutter渲染机制,相关源码目录见文末附录

一、UI线程渲染

Flutter是谷歌开源的移动UI框架,可以快速在Android和iOS上构建出高质量的原生用户界面,目前全世界越来越多的开发者加入到Flutter的队伍。 Flutter相比RN性能更好,由于Flutter自己实现了一套UI框架,丢弃了原生的UI框架,非常接近原生的体验。

为了揭秘Flutter高性能,本文从源码角度来看看Flutter的渲染绘制机制,跟渲染直接相关的两个线程是UI线程和GPU线程:

  • UI线程:运行着UI Task Runner,是Flutter Engine用于执行Dart root isolate代码,将其转换为layer tree视图结构;
  • GPU线程:该线程依然是在CPU上执行,运行着GPU Task Runner,处理layer tree,将其转换成为GPU命令并发送到GPU。

通过VSYNC信号使UI线程和GPU线程有条不紊的周期性的渲染界面,本文介绍VSYNC的产生过程、UI线程在引擎和框架的绘制工作,下一篇文章会介绍GPU线程的绘制工作。

1.1 UI渲染原理

1.1.1 UI渲染概览

通过VSYNC信号使UI线程和GPU线程有条不紊的周期性的渲染界面,如下图所示:

flutter_draw

  • 当需要渲染则会调用到Engine的ScheduleFrame()来注册VSYNC信号回调,一旦触发回调doFrame()执行完成后,便会移除回调方法,也就是说一次注册一次回调;
  • 当需要再次绘制则需要重新调用到ScheduleFrame()方法,该方法的唯一重要参数regenerate_layer_tree决定在帧绘制过程是否需要重新生成layer tree,还是直接复用上一次的layer tree;
  • UI线程的绘制过程,最核心的是执行WidgetsBinding的drawFrame()方法,然后会创建layer tree视图树
  • 再交由GPU Task Runner将layer tree提供的信息转化为平台可执行的GPU指令。

1.1.2 UI绘制核心工作

1)Vsync单注册模式:保证在一帧的时间窗口里UI线程只会生成一个layer tree发送给GPU线程,原理如下:

Animator中的信号量pending_frame_semaphore_用于控制不能连续频繁地调用Vsync请求,一次只能存在Vsync注册。 pending_frame_semaphore_初始值为1,在Animator::RequestFrame()消费信号会减1,当而后再次调用则会失败直接返回; Animator的BeginFrame()或者DrawLastLayerTree()方法会执行信号加1操作。

2)UI绘制最核心的方法是drawFrame(),包含以下几个过程:

  • Animate: 遍历_transientCallbacks,执行动画回调方法;
  • Build: 对于dirty的元素会执行build构造,没有dirty元素则不会执行,对应于buildScope()
  • Layout: 计算渲染对象的大小和位置,对应于flushLayout(),这个过程可能会嵌套再调用build操作;
  • Compositing bits: 更新具有脏合成位的任何渲染对象, 对应于flushCompositingBits();
  • Paint: 将绘制命令记录到Layer, 对应于flushPaint();
  • Compositing: 将Compositing bits发送给GPU, 对应于compositeFrame();
  • Semantics: 编译渲染对象的语义,并将语义发送给操作系统, 对应于flushSemantics()。

UI线程的耗时从doFrame(frameTimeNanos)中的frameTimeNanos为起点,以小节[4.10.6]Animator::Render()方法结束为终点, 并将结果保存到LayerTree的成员变量construction_time_,这便是UI线程的耗时时长。

1.1.3 Timeline说明

3)以上几个过程在Timeline中ui线程中都有体现,如下图所示

draw_ui

另外Timeline中还有两个比较常见的标签项

  • “Frame Request Pending”:从Animator::RequestFrame 到Animator::BeginFrame()结束;
  • ”PipelineProduce“: 从Animator::BeginFrame()到Animator::Render()结束。

1.2 UI线程渲染流程图

1.2.1 VSYNC注册流程

VSYNC注册流程图

Vysnc

当调用到引擎Engine的ScheduleFrame()方法过程则会注册VSYNC信号回调,一旦Vsync信号达到,则会调用到doFrame()方法。 对于调用ScheduleFrame()的场景有多种,比如动画的执行AnimationController.forward(),再比如比如surface创建的时候shell::SurfaceCreated()。

1.2.2 Engine层绘制

Engine层处理流程图

UIDraw_engine

doFrame()经过多层调用后通过PostTask将任务异步post到UI TaskRunner线程来执行,最后调用到Window的BeginFrame()方法。

1.2.3 Framework层绘制

Framework层处理流程图

UIDraw_fwk

其中window.cc中的一个BeginFrame()方法,会调用到window.dart中的onBeginFrame()和onDrawFrame()两个方法。

1.3 核心类图

类关系图

ClassEngine

为了让大家更容易理解源码,先看一张关于Shell、Engine、Animator等Flutter等Flutter引擎中核心类的类图。

  • Window类:是连接Flutter框架层(Dart)与引擎层(C++)的关键类,在框架层中window.dart文件里的一些方法在引擎层的window.cc文件有相对应的方法,比如scheduleFrame()方法。 在window.cc里面通过Window::RegisterNatives()注册了一些框架层与引擎层的方法对应关系;
  • RuntimeController类:可通过其成员root_isolate_找到Window类;
  • Shell类:同时继承了PlatformView::Delegate,Animator::Delegate,Engine::Delegate,所以在Engine,Animator,PlatformView中的成员变量delegate_都是指Shell对象, 从图中也能看出其中心地位,代理多项业务,该类是由AndroidShellHolder过程中初始化创建的;另外Shell类还继承了ServiceProtocol::Handler,图中省略而已。
  • PlatformViewAndroid类:在Android平台上PlatformView的实例采用的便是PlatformViewAndroid类。
  • Dart层与C层之间可以相互调用,从Window一路能调用到Shell类,也能从Shell类一路调用回Window。

接下来带着大家从源码角度来依次讲解Vsync注册以及UI线程的绘制处理流程,下一篇文章会介绍GPU线程的绘制工作。

二、 VSYNC注册流程

2.1 Engine::ScheduleFrame

[-> flutter/shell/common/engine.cc]

void Engine::ScheduleFrame(bool regenerate_layer_tree) {
    //[见小节2.2]
    animator_->RequestFrame(regenerate_layer_tree);
}

该方法说明:

  • animator_的赋值过程是在Engine对象初始化过程完成,而Engine初始化过程在Shell创建过程,此处animator_便是Animator对象;
  • ScheduleFrame的参数regenerate_layer_tree决定是否需要重新生成layer tree,还是直接复用上一次生成的layer tree;
  • 绝大多数情况下,调用RequestFrame()时将regenerate_layer_tree_设置为true或者用默认值true,执行完Animator::BeginFrame()则设置该变量为false;
    • 当无参数调用该方法时,regenerate_layer_tree为默认值为true。
    • 特别的例子就是Shell::OnPlatformViewMarkTextureFrameAvailable()过程,设置参数为false,那么计划绘制一帧的时候就不需要重绘layer tree;

2.2 Animator::RequestFrame

[-> flutter/shell/common/animator.cc]

void Animator::RequestFrame(bool regenerate_layer_tree) {
  if (regenerate_layer_tree) {
    // regenerate_layer_tree_决定Vsync信号到来时,是否执行BeginFrame
    regenerate_layer_tree_ = true;
  }

  //当调用Animator::Stop()则会停止动画绘制
  if (paused_ && !dimension_change_pending_) {
    return;
  }

  //调用sem_trywait来保证不会同时有多个vsync请求
  if (!pending_frame_semaphore_.TryWait()) {
    return;
  }

  task_runners_.GetUITaskRunner()->PostTask([self = weak_factory_.GetWeakPtr(),
                                             frame_number = frame_number_]() {
    if (!self.get()) {
      return;
    }
    TRACE_EVENT_ASYNC_BEGIN0("flutter", "Frame Request Pending", frame_number);
    self->AwaitVSync();  // [见小节2.3]
  });
  frame_scheduled_ = true;  //标注已经schedule绘画帧
}

过程说明:

  • pending_frame_semaphore_:非负信号量,初始值为1,第一次调用TryWait减1,而后再次调用则会失败直接返回。当消费了这次vsync回调,也就是调用了Animator的BeginFrame()或者DrawLastLayerTree()方法后,改信号量会加1[见小节3.6],可以再次执行vysnc的注册;
  • 通过Animator的Start()或者BeginFrame调用到的RequestFrame方法,则肯定需要重新生成layer tree;通过Engine的ScheduleFrame方法是否重建layer tree看小节2.1;
  • 此处通过post把Animator::AwaitVSync任务放入到UI Task Runner来执行。

2.3 Animator::AwaitVSync

[-> flutter/shell/common/animator.cc]

void Animator::AwaitVSync() {
  // [见小节2.4]
  waiter_->AsyncWaitForVsync(
      [self = weak_factory_.GetWeakPtr()](fml::TimePoint frame_start_time,
                                          fml::TimePoint frame_target_time) {
        if (self) {
          //是否能复用上次layer树,取决于regenerate_layer_tree_
          if (self->CanReuseLastLayerTree()) {
            //复用上次layer树,直接把任务post到gpu线程做栅格化操作
            self->DrawLastLayerTree();
          } else {
            self->BeginFrame(frame_start_time, frame_target_time);
          }
        }
      });

  delegate_.OnAnimatorNotifyIdle(dart_frame_deadline_);
}

waiter_的赋值是在Animator初始化过程,取值为VsyncWaiterAndroid对象,当调用了RequestFrame(),默认参数regenerate_layer_tree_为true,意味着需要重新生成layer树,故不能重复使用上一次的layer树,接着来看一下AsyncWaitForVsync()方法的实现。

2.4 VsyncWaiter::AsyncWaitForVsync

[-> flutter/shell/common/vsync_waiter.cc]

void VsyncWaiter::AsyncWaitForVsync(Callback callback) {
  {
    std::lock_guard<std::mutex> lock(callback_mutex_);
    //赋值callback_
    callback_ = std::move(callback);
  }
  TRACE_EVENT0("flutter", "AsyncWaitForVsync");
  AwaitVSync(); // [见小节2.5]
}

此次的callback_便是[小节2.3]方法中的参数,该方法根据regenerate_layer_tree_来决定执行流。

  • 当regenerate_layer_tree_=false,则执行DrawLastLayerTree();
  • 当regenerate_layer_tree_=false,则执行BeginFrame();

2.5 VsyncWaiterAndroid::AwaitVSync

[-> flutter/shell/platform/android/vsync_waiter_android.cc]

void VsyncWaiterAndroid::AwaitVSync() {
  std::weak_ptr<VsyncWaiter>* weak_this =
      new std::weak_ptr<VsyncWaiter>(shared_from_this());
  //获取VsyncWaiter的弱引用
  jlong java_baton = reinterpret_cast<jlong>(weak_this);

  JNIEnv* env = fml::jni::AttachCurrentThread();
  // 此次调用到Java层的asyncWaitForVsync方法,java_baton指向VsyncWaiter
  env->CallStaticVoidMethod(g_vsync_waiter_class->obj(),     //
                            g_async_wait_for_vsync_method_,  //
                            java_baton                       //
  );
}

此处g_vsync_waiter_class,g_async_wait_for_vsync_method_的赋值过程是由JNI_OnLoad完成,如下所示。

2.5.1 JNI_OnLoad

[-> flutter/shell/platform/android/library_loader.cc]

JNIEXPORT jint JNI_OnLoad(JavaVM* vm, void* reserved) {
  // 初始化Java虚拟机
  fml::jni::InitJavaVM(vm);

  JNIEnv* env = fml::jni::AttachCurrentThread();
  bool result = false;

  // 注册FlutterMain.
  result = shell::FlutterMain::Register(env);

  // 注册PlatformView [见小节2.5.2]
  result = shell::PlatformViewAndroid::Register(env);

  // 注册VSyncWaiter [见小节2.5.3]
  result = shell::VsyncWaiterAndroid::Register(env);

  return JNI_VERSION_1_4;
}

首次加载共享库时虚拟机会调用此方法。

2.5.2 Register

[-> flutter/shell/platform/android/platform_view_android_jni.cc]

bool PlatformViewAndroid::Register(JNIEnv* env) {
  //记录FlutterCallbackInformation类的全局引用
  g_flutter_callback_info_class = new fml::jni::ScopedJavaGlobalRef<jclass>(
      env, env->FindClass("io/flutter/view/FlutterCallbackInformation"));
  //记录FlutterCallbackInformation构造函数
  g_flutter_callback_info_constructor = env->GetMethodID(
      g_flutter_callback_info_class->obj(), "<init>",
      "(Ljava/lang/String;Ljava/lang/String;Ljava/lang/String;)V");
  //记录FlutterJNI类的全局引用
  g_flutter_jni_class = new fml::jni::ScopedJavaGlobalRef<jclass>(
      env, env->FindClass("io/flutter/embedding/engine/FlutterJNI"));
  //记录SurfaceTexture类的全局引用
  g_surface_texture_class = new fml::jni::ScopedJavaGlobalRef<jclass>(
      env, env->FindClass("android/graphics/SurfaceTexture"));

  static const JNINativeMethod callback_info_methods[] = {
      {
          .name = "nativeLookupCallbackInformation",
          .signature = "(J)Lio/flutter/view/FlutterCallbackInformation;",
          .fnPtr = reinterpret_cast<void*>(&shell::LookupCallbackInformation),
      },
  };
  //注册FlutterCallbackInformation的nativeLookupCallbackInformation()方法
  env->RegisterNatives(g_flutter_callback_info_class->obj(),
                           callback_info_methods,
                           arraysize(callback_info_methods)) != 0);

  g_is_released_method =
      env->GetMethodID(g_surface_texture_class->obj(), "isReleased", "()Z");

  fml::jni::ClearException(env);

  g_attach_to_gl_context_method = env->GetMethodID(
      g_surface_texture_class->obj(), "attachToGLContext", "(I)V");

  g_update_tex_image_method =
      env->GetMethodID(g_surface_texture_class->obj(), "updateTexImage", "()V");

  g_get_transform_matrix_method = env->GetMethodID(
      g_surface_texture_class->obj(), "getTransformMatrix", "([F)V");

  g_detach_from_gl_context_method = env->GetMethodID(
      g_surface_texture_class->obj(), "detachFromGLContext", "()V");

  return RegisterApi(env);
}

该方法的主要工作:

  • 记录和注册类FlutterCallbackInformation、FlutterJNI以及SurfaceTexture类的相关方法,用于Java和C++层方法的相互调用。

2.5.3 Register

[-> flutter/shell/platform/android/vsync_waiter_android.cc]

bool VsyncWaiterAndroid::Register(JNIEnv* env) {
  static const JNINativeMethod methods[] = ;

  jclass clazz = env->FindClass("io/flutter/view/VsyncWaiter");

  g_vsync_waiter_class = new fml::jni::ScopedJavaGlobalRef<jclass>(env, clazz);

  g_async_wait_for_vsync_method_ = env->GetStaticMethodID(
      g_vsync_waiter_class->obj(), "asyncWaitForVsync", "(J)V");

  return env->RegisterNatives(clazz, methods, arraysize(methods)) == 0;
}

该注册过程主要工作:

  • 将Java层的VsyncWaiter类的nativeOnVsync()方法,映射到C++层的OnNativeVsync()方法,用于该方法的Java调用C++的过程;
  • 将Java层的VsyncWaiter类的asyncWaitForVsync()方法,保存到C++层的g_async_wait_for_vsync_method_变量,用于该方法C++调用Java的过程。

可见,将调用VsyncWaiter类的asyncWaitForVsync()方法

2.6 asyncWaitForVsync[Java]

[-> flutter/shell/platform/android/io/flutter/view/VsyncWaiter.java]

public class VsyncWaiter {
    // FlutterView的刷新时间周期(16.7ms)
    public static long refreshPeriodNanos = 1000000000 / 60;

    private static HandlerThread handlerThread;
    private static Handler handler;

    static {
        handlerThread = new HandlerThread("FlutterVsyncThread");
        handlerThread.start();
    }

    public static void asyncWaitForVsync(final long cookie) {
        if (handler == null) {
            handler = new Handler(handlerThread.getLooper());
        }
        handler.post(new Runnable() {
            @Override
            public void run() {
                //注册帧回调方法,见小节[2.6.1]/[2.6.2]
                Choreographer.getInstance().postFrameCallback(new Choreographer.FrameCallback() {
                    @Override
                    public void doFrame(long frameTimeNanos) {
                        //frameTimeNanos是VYSNC触发的时间点,也就是计划绘制的时间点
                        nativeOnVsync(frameTimeNanos, frameTimeNanos + refreshPeriodNanos, cookie);
                    }
                });
            }
        });
    }
}

通过Handler将工作post到FlutterVsyncThread线程,具体的工作是通过Choreographer来注册回调方法doFrame()以监听系统VSYNC信号。

2.6.1 Choreographer.getInstance

[-> Choreographer.java]

public static Choreographer getInstance() {
    return sThreadInstance.get(); //单例模式
}

private static final ThreadLocal<Choreographer> sThreadInstance =
    new ThreadLocal<Choreographer>() {

    protected Choreographer initialValue() {
        //获取当前线程FlutterVsyncThread的Looper
        Looper looper = Looper.myLooper();
        // 初始化Choreographer对象
        return new Choreographer(looper);
    }
};

private Choreographer(Looper looper) {
    mLooper = looper;
    //创建Handler对象
    mHandler = new FrameHandler(looper);
    //创建用于接收VSync信号的对象
    mDisplayEventReceiver = USE_VSYNC ? new FrameDisplayEventReceiver(looper) : null;
    mLastFrameTimeNanos = Long.MIN_VALUE;  //上一次帧绘制时间点
    mFrameIntervalNanos = (long)(1000000000 / getRefreshRate());
    mCallbackQueues = new CallbackQueue[CALLBACK_LAST + 1];  
    for (int i = 0; i <= CALLBACK_LAST; i++) {
        mCallbackQueues[i] = new CallbackQueue();
    }
}

此处Choreographer的mLooper和mHandler都运行在FlutterVsyncThread线程。

2.6.2 postFrameCallback

[-> Choreographer.java]

public void postFrameCallback(FrameCallback callback) {
    postFrameCallbackDelayed(callback, 0);
}

public void postFrameCallbackDelayed(FrameCallback callback, long delayMillis) {
    postCallbackDelayedInternal(CALLBACK_ANIMATION,
            callback, FRAME_CALLBACK_TOKEN, delayMillis);
}

private void postCallbackDelayedInternal(int callbackType, Object action, Object token, long delayMillis) {
    synchronized (mLock) {
        final long now = SystemClock.uptimeMillis();
        final long dueTime = now + delayMillis;
        //添加到mCallbackQueues队列
        mCallbackQueues[callbackType].addCallbackLocked(dueTime, action, token);
        if (dueTime <= now) {
          scheduleFrameLocked(now);
        } else {
          ...
        }
    }
}

将FrameCallback方法加入到mCallbackQueues[CALLBACK_ANIMATION]回调队列中。

三、Engine层绘制

3.1 doFrame

[-> Choreographer.java]

public void doFrame(long frameTimeNanos) {
    //Android FW每次当vsync信号触发,则会调用该方法 [见下方]
    nativeOnVsync(frameTimeNanos, frameTimeNanos + refreshPeriodNanos, cookie);
}

Vsync注册过程见[小节2.6] Choreographer.FrameCallback()。注册了Vysnc信号后,一旦底层Vsync信号触发,经过层层调用回到FrameDisplayEventReceiver的过程,然后会有一个通过handler的方式post到线程”FlutterVsyncThread”来执行操作, 具体流程见Choreographer原理。紧接着再处理所有注册的doCallbacks方法,则会执行Choreographer.FrameCallback中的doFrame()方法,如下所示。

new Choreographer.FrameCallback() {
    @Override
    public void doFrame(long frameTimeNanos) {
        //frameTimeNanos是VYSNC触发的时间点,也就是计划绘制的时间点 [见小节3.2]
        nativeOnVsync(frameTimeNanos, frameTimeNanos + refreshPeriodNanos, cookie);
    }
}

3.2 OnNativeVsync

[-> flutter/shell/platform/android/io/flutter/view/VsyncWaiter.java]

public class VsyncWaiter {
    ...
    // [见小节3.2.1]
    private static native void nativeOnVsync(long frameTimeNanos,
                                             long frameTargetTimeNanos,
                                             long cookie);
    ...
}

由[小节2.5.3]可知,VsyncWaiter.java中的nativeOnVsync对应于vsync_waiter_android.cc的OnNativeVsync(),具体过程在jni加载过程初始化,如下所示。

3.2.1 OnNativeVsync[C++]

[-> flutter/shell/platform/android/vsync_waiter_android.cc]

static void OnNativeVsync(JNIEnv* env, jclass jcaller,
                          jlong frameTimeNanos,
                          jlong frameTargetTimeNanos,
                          jlong java_baton) {
  auto frame_time = fml::TimePoint::FromEpochDelta(
      fml::TimeDelta::FromNanoseconds(frameTimeNanos));
  auto target_time = fml::TimePoint::FromEpochDelta(
      fml::TimeDelta::FromNanoseconds(frameTargetTimeNanos));
  //消费pending回调[见小节3.3]
  ConsumePendingCallback(java_baton, frame_time, target_time);
}

3.3 ConsumePendingCallback

[-> flutter/shell/platform/android/vsync_waiter_android.cc]

static void ConsumePendingCallback(jlong java_baton,
                                   fml::TimePoint frame_start_time,
                                   fml::TimePoint frame_target_time) {
  auto* weak_this = reinterpret_cast<std::weak_ptr<VsyncWaiter>*>(java_baton);
  auto shared_this = weak_this->lock();
  delete weak_this;

  if (shared_this) {
    //shared_this指向VsyncWaiter的弱引用 [见小节3.4]
    shared_this->FireCallback(frame_start_time, frame_target_time);
  }
}

3.4 VsyncWaiter::FireCallback

[-> flutter/shell/common/vsync_waiter.cc]

void VsyncWaiter::FireCallback(fml::TimePoint frame_start_time,
                               fml::TimePoint frame_target_time) {
  Callback callback;
  {
    std::lock_guard<std::mutex> lock(callback_mutex_);
    callback = std::move(callback_);
  }
  if (!callback) {
    TRACE_EVENT_INSTANT0("flutter", "MismatchedFrameCallback");
    return;
  }

  TRACE_EVENT0("flutter", "VsyncFireCallback");
  //将任务放入task队列[见小节3.4.1]
  task_runners_.GetUITaskRunner()->PostTaskForTime(
    [callback, flow_identifier, frame_start_time, frame_target_time]() {
      FML_TRACE_EVENT("flutter", kVsyncTraceName, "StartTime",
                      frame_start_time, "TargetTime", frame_target_time);
      fml::tracing::TraceEventAsyncComplete(
          "flutter", "VsyncSchedulingOverhead", fml::TimePoint::Now(),
          frame_start_time);
      //开始执行vync [见小节3.5]
      callback(frame_start_time, frame_target_time);
      TRACE_FLOW_END("flutter", kVsyncFlowName, flow_identifier);
    },
    frame_start_time);
}

将任务闭包放入task队列,消息Loop一旦接受到消息则会读取出来。

3.4.1 MessageLoopImpl::RunExpiredTasks

[-> flutter/fml/message_loop_impl.cc]

void MessageLoopImpl::RunExpiredTasks() {
  TRACE_EVENT0("fml", "MessageLoop::RunExpiredTasks");
  std::vector<fml::closure> invocations;

  {
    std::lock_guard<std::mutex> lock(delayed_tasks_mutex_);
    //当没有待处理的task则直接返回
    if (delayed_tasks_.empty()) {
      return;
    }

    auto now = fml::TimePoint::Now();
    while (!delayed_tasks_.empty()) {
      const auto& top = delayed_tasks_.top();
      if (top.target_time > now) {
        break;
      }
      invocations.emplace_back(std::move(top.task));
      delayed_tasks_.pop();
    }
    WakeUp(delayed_tasks_.empty() ? fml::TimePoint::Max()
                                  : delayed_tasks_.top().target_time);
  }

  for (const auto& invocation : invocations) {
    invocation();  // [见小节3.5]
    for (const auto& observer : task_observers_) {
      observer.second();
    }
  }
}

对于ui线程处于消息loop状态,一旦有时间到达的任务则开始执行,否则处于空闲等等状态。前面[小节3.4] VsyncWaiter::FireCallback过程已经向该ui线程postTask。 对于不可复用layer tree的情况则调用Animator::BeginFrame()方法。

3.5 callback

[-> flutter/shell/common/animator.cc]

[self = weak_factory_.GetWeakPtr()](fml::TimePoint frame_start_time,
                                    fml::TimePoint frame_target_time) {
  if (self) {
    if (self->CanReuseLastLayerTree()) {
      self->DrawLastLayerTree();
    } else {
      //根据默认参数regenerate_layer_tree_为true,则执行该分支 [见小节3.6]
      self->BeginFrame(frame_start_time, frame_target_time);
    }
  }
}

此次的callback赋值过程位于[小节2.3]Animator::AwaitVSync()方法的闭包参数,相关说明:

  • frame_start_time:计划开始绘制时间点,来源于doFrame()方法中的参数;
  • frame_target_time:从frame_start_time加上一帧时间(16.7ms)的时间,作为本次绘制的deadline。

3.6 Animator::BeginFrame

[-> flutter/shell/common/animator.cc]

void Animator::BeginFrame(fml::TimePoint frame_start_time,
                          fml::TimePoint frame_target_time) {
  TRACE_EVENT_ASYNC_END0("flutter", "Frame Request Pending", frame_number_++);
  TRACE_EVENT0("flutter", "Animator::BeginFrame");

  frame_scheduled_ = false;
  notify_idle_task_id_++;
  regenerate_layer_tree_ = false;
  //信号量加1,可以注册新的vsync信号,也就是能执行Animator::RequestFrame()
  pending_frame_semaphore_.Signal();

  if (!producer_continuation_) {
    //[小节3.6.1]/[小节3.6.2]
    producer_continuation_ = layer_tree_pipeline_->Produce();
    //pipeline已满,说明GPU线程繁忙,则结束本次UI绘制,重新注册Vsync
    if (!producer_continuation_) {
      RequestFrame();
      return;
    }
  }

  //从pipeline中获取有效的continuation,并准备为可能的frame服务
  last_begin_frame_time_ = frame_start_time;
  //获取当前帧绘制截止时间,用于告知可GC的空闲时长
  dart_frame_deadline_ = FxlToDartOrEarlier(frame_target_time);
  {
    TRACE_EVENT2("flutter", "Framework Workload", "mode", "basic", "frame",
                 FrameParity());
    //此处delegate_为Shell [小节3.7]
    delegate_.OnAnimatorBeginFrame(last_begin_frame_time_);
  }

  if (!frame_scheduled_) {
    task_runners_.GetUITaskRunner()->PostDelayedTask(
        [self = weak_factory_.GetWeakPtr(),
         notify_idle_task_id = notify_idle_task_id_]() {
          if (!self.get()) {
            return;
          }
          // 该任务id和当前任务id一致,则不再需要审查frame,可以通知引擎当前处于空闲状态,100ms
          if (notify_idle_task_id == self->notify_idle_task_id_) {
            self->delegate_.OnAnimatorNotifyIdle(Dart_TimelineGetMicros() +
                                                 100000);
          }
        },
        kNotifyIdleTaskWaitTime); //延迟51ms再通知引擎空闲状态
  }
}

该方法主要功能说明:

  • layer_tree_pipeline_是在Animator对象初始化的过程中创建的LayerTreePipeline,其类型为Pipeline
  • 此处kNotifyIdleTaskWaitTime等于51ms,等于3帧的时间+1ms,之所以这样设计是由于在某些工作负载下(比如父视图调整大小,通过viewport metrics事件传达给子视图)实际上还没有schedule帧,尽管在下一个vsync会生成一帧(将在收到viewport事件后schedule),因此推迟调用OnAnimatorNotifyIdle一点点,从而避免可能垃圾回收在不希望的时间触发。

3.6.1 LayerTreePipeline初始化

[-> flutter/shell/common/animator.cc]

Animator::Animator(Delegate& delegate,
                   TaskRunners task_runners,
                   std::unique_ptr<VsyncWaiter> waiter)
    : delegate_(delegate),
      task_runners_(std::move(task_runners)),
      waiter_(std::move(waiter)),
      last_begin_frame_time_(),
      dart_frame_deadline_(0),
      layer_tree_pipeline_(fml::MakeRefCounted<LayerTreePipeline>(2)),
      ... {}

此处LayerTreePipeline的初始化过程如下:

using LayerTreePipeline = Pipeline<flutter::LayerTree>;

在pipeline.h的过程会初始化Pipeline,可见初始值empty_ = 2,available_ = 0;

Pipeline(uint32_t depth) : empty_(depth), available_(0) {}

3.6.2 Pipeline::Produce

[-> flutter/synchronization/pipeline.h]

ProducerContinuation Produce() {
  //当管道不为空,则不允许再次向管道加入数据
  if (!empty_.TryWait()) {
    return {};
  }

  //[见小节3.6.3]
  return ProducerContinuation{
      std::bind(&Pipeline::ProducerCommit, this, std::placeholders::_1,
                std::placeholders::_2),  // continuation
      GetNextPipelineTraceID()};  
}

通过信号量empty_的初始值为depth(默认等于2),来保证同一个管道的任务最多不超过depth个,每次UI线程执行Produce()会减1,当GPU线程执行完成Consume()方法后才会执行加1操作。

3.6.3 ProducerContinuation初始化

[-> flutter/synchronization/pipeline.h]

ProducerContinuation(Continuation continuation, size_t trace_id)
    : continuation_(continuation), trace_id_(trace_id) {
  TRACE_FLOW_BEGIN("flutter", "PipelineItem", trace_id_);
  TRACE_EVENT_ASYNC_BEGIN0("flutter", "PipelineProduce", trace_id_);
}

3.6.3 Pipeline.ProducerCommit

[-> flutter/synchronization/pipeline.h]

void ProducerCommit(ResourcePtr resource, size_t trace_id) {
  {
    std::lock_guard<std::mutex> lock(queue_mutex_);
    queue_.emplace(std::move(resource), trace_id);
  }

  available_.Signal();
}

3.7 Shell::OnAnimatorBeginFrame

[-> flutter/shell/common/shell.cc]

void Shell::OnAnimatorBeginFrame(fml::TimePoint frame_time) {
  if (engine_) {
    engine_->BeginFrame(frame_time);  // [小节3.8]
  }
}

3.8 Engine::BeginFrame

[-> flutter/shell/common/engine.cc]

void Engine::BeginFrame(fml::TimePoint frame_time) {
  TRACE_EVENT0("flutter", "Engine::BeginFrame");
  runtime_controller_->BeginFrame(frame_time);  // [小节3.9]
}

3.9 RuntimeController::BeginFrame

[-> flutter/runtime/runtime_controller.cc]

bool RuntimeController::BeginFrame(fml::TimePoint frame_time) {
  if (auto* window = GetWindowIfAvailable()) {
    window->BeginFrame(frame_time);  // [小节3.10]
    return true;
  }
  return false;
}

3.10 Window::BeginFrame

[-> flutter/lib/ui/window/window.cc]

void Window::BeginFrame(fml::TimePoint frameTime) {
  std::shared_ptr<tonic::DartState> dart_state = library_.dart_state().lock();
  if (!dart_state)
    return;
  tonic::DartState::Scope scope(dart_state);
  //注意此处的frameTime便是前面小节3.1中doFrame方法中的参数frameTimeNanos
  int64_t microseconds = (frameTime - fml::TimePoint()).ToMicroseconds();

  // [见小节4.2]
  DartInvokeField(library_.value(), "_beginFrame",
                  {Dart_NewInteger(microseconds)});

  //执行MicroTask
  UIDartState::Current()->FlushMicrotasksNow();

  // [见小节4.4]
  DartInvokeField(library_.value(), "_drawFrame", {});
}

Window::BeginFrame()过程主要工作:

  • 执行_beginFrame
  • 执行FlushMicrotasksNow
  • 执行_drawFrame

可见,Microtask位于beginFrame和drawFrame之间,那么Microtask的耗时会影响ui绘制过程。

DartInvokeField()通过dart虚拟机调用了window.onBeginFrame()和onDrawFrame方法,见hooks.dart文件中如下过程:

@pragma('vm:entry-point')
void _beginFrame(int microseconds) {
  _invoke1<Duration>(window.onBeginFrame, window._onBeginFrameZone, new Duration(microseconds: microseconds));
}

@pragma('vm:entry-point')
void _drawFrame() {
  _invoke(window.onDrawFrame, window._onDrawFrameZone);
}

四、Framework层绘制

在引擎层的处理过程会调用到window.onBeginFrame()和onDrawFrame,回到framework层从这个两个方法开始说起。

4.1 SchedulerBinding.initInstances

[-> lib/src/scheduler/binding.dart:: SchedulerBinding]

mixin SchedulerBinding on BindingBase, ServicesBinding {
  @override
  void initInstances() {
    super.initInstances();
    _instance = this;
    ui.window.onBeginFrame = _handleBeginFrame; //[见小节4.2]
    ui.window.onDrawFrame = _handleDrawFrame;  //[见小节4.4]
    SystemChannels.lifecycle.setMessageHandler(_handleLifecycleMessage);
  }
}

可见,引擎层中的Window::BeginFrame()调用的两个方法,进入到dart层则分别是_handleBeginFrame()和_handleDrawFrame()方法

4.1.1 Window初始化

[-> flutter/lib/ui/window.dart]

class Window {
    Window._()

    FrameCallback get onBeginFrame => _onBeginFrame;
    FrameCallback _onBeginFrame;

    VoidCallback get onDrawFrame => _onDrawFrame;
    VoidCallback _onDrawFrame;
    ...
}

Window初始化过程,可以知道onBeginFrame和onDrawFrame分别保存_onBeginFrame和_onDrawFrame方法。

4.2 _handleBeginFrame

[-> lib/src/scheduler/binding.dart:: SchedulerBinding]

void _handleBeginFrame(Duration rawTimeStamp) {
  if (_warmUpFrame) {
    _ignoreNextEngineDrawFrame = true;
    return;
  }
  handleBeginFrame(rawTimeStamp);  //[见小节4.3]
}

4.3 handleBeginFrame

[-> lib/src/scheduler/binding.dart:: SchedulerBinding]

void handleBeginFrame(Duration rawTimeStamp) {
  Timeline.startSync('Frame', arguments: timelineWhitelistArguments);
  _firstRawTimeStampInEpoch ??= rawTimeStamp;
  _currentFrameTimeStamp = _adjustForEpoch(rawTimeStamp ?? _lastRawTimeStamp);
  if (rawTimeStamp != null)
    _lastRawTimeStamp = rawTimeStamp;

  profile(() {
    _profileFrameNumber += 1;
    _profileFrameStopwatch.reset();
    _profileFrameStopwatch.start();
  });

  //此时阶段等于SchedulerPhase.idle;
  _hasScheduledFrame = false;
  try {
    Timeline.startSync('Animate', arguments: timelineWhitelistArguments);
    _schedulerPhase = SchedulerPhase.transientCallbacks;
    //执行动画的回调方法
    final Map<int, _FrameCallbackEntry> callbacks = _transientCallbacks;
    _transientCallbacks = <int, _FrameCallbackEntry>{};
    callbacks.forEach((int id, _FrameCallbackEntry callbackEntry) {
      if (!_removedIds.contains(id))
        _invokeFrameCallback(callbackEntry.callback, _currentFrameTimeStamp, callbackEntry.debugStack);
    });
    _removedIds.clear();
  } finally {
    _schedulerPhase = SchedulerPhase.midFrameMicrotasks;
  }
}

该方法主要功能是遍历_transientCallbacks,执行相应的Animate操作,可通过scheduleFrameCallback()/cancelFrameCallbackWithId()来完成添加和删除成员,再来简单看看这两个方法。

4.3.1 scheduleFrameCallback

[-> lib/src/scheduler/binding.dart:: SchedulerBinding]

int scheduleFrameCallback(FrameCallback callback, { bool rescheduling = false }) {
  scheduleFrame();  //触发帧绘制的调度
  _nextFrameCallbackId += 1;
  _transientCallbacks[_nextFrameCallbackId] = _FrameCallbackEntry(callback, rescheduling: rescheduling);
  return _nextFrameCallbackId;
}

callback保存在_FrameCallbackEntry对象里面

4.3.2 cancelFrameCallbackWithId

[-> lib/src/scheduler/binding.dart:: SchedulerBinding]

void cancelFrameCallbackWithId(int id) {
  assert(id > 0);
  _transientCallbacks.remove(id);
  _removedIds.add(id);
}

4.4 _handleDrawFrame

[-> lib/src/scheduler/binding.dart:: SchedulerBinding]

void _handleDrawFrame() {
  if (_ignoreNextEngineDrawFrame) {
    _ignoreNextEngineDrawFrame = false;
    return;
  }
  handleDrawFrame();  //[见小节4.5]
}

4.5 handleDrawFrame

[-> lib/src/scheduler/binding.dart:: SchedulerBinding]

void handleDrawFrame() {
  assert(_schedulerPhase == SchedulerPhase.midFrameMicrotasks);
  Timeline.finishSync(); // 标识结束"Animate"阶段
  try {
    _schedulerPhase = SchedulerPhase.persistentCallbacks;
    //执行PERSISTENT FRAME回调
    for (FrameCallback callback in _persistentCallbacks)
      _invokeFrameCallback(callback, _currentFrameTimeStamp); //[见小节4.5.1]

    _schedulerPhase = SchedulerPhase.postFrameCallbacks;
    // 执行POST-FRAME回调
    final List<FrameCallback> localPostFrameCallbacks = List<FrameCallback>.from(_postFrameCallbacks);
    _postFrameCallbacks.clear();
    for (FrameCallback callback in localPostFrameCallbacks)
      _invokeFrameCallback(callback, _currentFrameTimeStamp);
  } finally {
    _schedulerPhase = SchedulerPhase.idle;
    Timeline.finishSync(); //标识结束”Frame“阶段
    profile(() {
      _profileFrameStopwatch.stop();
      _profileFramePostEvent();
    });
    _currentFrameTimeStamp = null;
  }
}

该方法主要功能:

  • 遍历_persistentCallbacks,执行相应的回调方法,可通过addPersistentFrameCallback()注册,一旦注册后不可移除,后续每一次frame回调都会执行;
  • 遍历_postFrameCallbacks,执行相应的回调方法,可通过addPostFrameCallback()注册,handleDrawFrame()执行完成后会清空_postFrameCallbacks内容。

4.5.1 _invokeFrameCallback

[-> lib/src/scheduler/binding.dart:: SchedulerBinding]

void _invokeFrameCallback(FrameCallback callback, Duration timeStamp, [ StackTrace callbackStack ]) {
  try {
    callback(timeStamp);  //[见小节4.5.2]
  } catch (exception, exceptionStack) {
    FlutterError.reportError(FlutterErrorDetails(...));
  }
}

这里的callback是_persistentCallbacks列表中的成员,再来看看其成员是如何添加进去的。

4.5.2 WidgetsBinding.initInstances

[-> lib/src/widgets/binding.dart]

mixin WidgetsBinding on BindingBase, SchedulerBinding, GestureBinding, RendererBinding, SemanticsBinding {
  @override
  void initInstances() {
    super.initInstances();  //[见小节4.5.3]
    _instance = this;
    buildOwner.onBuildScheduled = _handleBuildScheduled;
    ui.window.onLocaleChanged = handleLocaleChanged;
    ui.window.onAccessibilityFeaturesChanged = handleAccessibilityFeaturesChanged;
    SystemChannels.navigation.setMethodCallHandler(_handleNavigationInvocation);
    SystemChannels.system.setMessageHandler(_handleSystemMessage);
  }
}

在flutter app启动过程,也就是执行runApp过程会有WidgetsFlutterBinding初始化过程,WidgetsBinding的initInstances(),根据mixin的顺序,可知此处的super.initInstances() 便是RendererBinding类。

4.5.3 RendererBinding.initInstances

[-> lib/src/rendering/binding.dart]

mixin RendererBinding on BindingBase, ServicesBinding, SchedulerBinding, SemanticsBinding, HitTestable {

  void initInstances() {
    super.initInstances();
    _instance = this;
    _pipelineOwner = PipelineOwner(
      onNeedVisualUpdate: ensureVisualUpdate,
      onSemanticsOwnerCreated: _handleSemanticsOwnerCreated,
      onSemanticsOwnerDisposed: _handleSemanticsOwnerDisposed,
    );
    ui.window
      ..onMetricsChanged = handleMetricsChanged
      ..onTextScaleFactorChanged = handleTextScaleFactorChanged
      ..onSemanticsEnabledChanged = _handleSemanticsEnabledChanged
      ..onSemanticsAction = _handleSemanticsAction;
    initRenderView();
    _handleSemanticsEnabledChanged();
    addPersistentFrameCallback(_handlePersistentFrameCallback); //[见小节4.5.4]
  }

  void _handlePersistentFrameCallback(Duration timeStamp) {
    drawFrame();  //[见小节4.6]
  }
}

4.5.4 SchedulerBinding.addPersistentFrameCallback

[-> lib/src/scheduler/binding.dart]

mixin SchedulerBinding on BindingBase, ServicesBinding {

  void addPersistentFrameCallback(FrameCallback callback) {
    _persistentCallbacks.add(callback);
  }
}

4.6 WidgetsBinding.drawFrame

[-> lib/src/widgets/binding.dart]

void drawFrame() {
  try {
    if (renderViewElement != null)
      buildOwner.buildScope(renderViewElement);   //[见小节4.6.1]
    super.drawFrame();   //[见小节4.6.4]
    buildOwner.finalizeTree();  //[见小节4.12]
  } finally {
  }
}

4.6.1 BuildOwner.buildScope

[-> lib/src/widgets/framework.dart]

void buildScope(Element context, [VoidCallback callback]) {
  if (callback == null && _dirtyElements.isEmpty)
    return;
  Timeline.startSync('Build', arguments: timelineWhitelistArguments);
  try {
    _scheduledFlushDirtyElements = true;
    if (callback != null) {
      _dirtyElementsNeedsResorting = false;
      callback();  //执行回调方法
    }
    _dirtyElements.sort(Element._sort); //排序
    _dirtyElementsNeedsResorting = false;
    int dirtyCount = _dirtyElements.length;
    int index = 0;
    while (index < dirtyCount) {
      try {
        //具体Element子类执行重建操作 [见小节4.6.2]
        _dirtyElements[index].rebuild();
      } catch (e, stack) {
      }
      index += 1;
      if (dirtyCount < _dirtyElements.length || _dirtyElementsNeedsResorting) {
        _dirtyElements.sort(Element._sort);
        _dirtyElementsNeedsResorting = false;
        dirtyCount = _dirtyElements.length;
        while (index > 0 && _dirtyElements[index - 1].dirty) {
          index -= 1;
        }
      }
    }
  } finally {
    for (Element element in _dirtyElements) {
      element._inDirtyList = false;
    }
    _dirtyElements.clear();
    _scheduledFlushDirtyElements = false;
    _dirtyElementsNeedsResorting = null;
    Timeline.finishSync();
  }
}

4.6.2 Element.rebuild

[-> lib/src/widgets/framework.dart]

void rebuild() {
  if (!_active || !_dirty)
    return;
  performRebuild();
}

performRebuild具体执行方法,取决于相应的Element子类,这里以ComponentElement为例

4.6.3 ComponentElement.performRebuild

[-> lib/src/widgets/framework.dart]

void performRebuild() {
  Widget built;
  try {
    built = build();  //执行build方法
  } catch (e, stack) {
    built = ErrorWidget.builder(_debugReportException('building $this', e, stack));
  } finally {
    _dirty = false;
  }
  try {
    _child = updateChild(_child, built, slot); //更新子元素
  } catch (e, stack) {
    built = ErrorWidget.builder(_debugReportException('building $this', e, stack));
    _child = updateChild(null, built, slot);
  }
}

4.6.4 RendererBinding.drawFrame

[-> lib/src/rendering/binding.dart]

void drawFrame() {
  pipelineOwner.flushLayout();  //[见小节4.7]
  pipelineOwner.flushCompositingBits();  //[见小节4.8]
  pipelineOwner.flushPaint(); //[见小节4.9]
  renderView.compositeFrame();  //[见小节4.10]
  pipelineOwner.flushSemantics(); //[见小节4.11]
}

RendererBinding的initInstances()过程注册了一个Persistent的帧回调方法_handlePersistentFrameCallback(),故handleDrawFrame()过程会调用该方法。pipelineOwner管理渲染管道,提供了一个用于驱动渲染管道的接口,并存储了哪些渲染对象请求访问状态,要刷新管道,需要按顺序运行如下5个阶段:

  1. [flushLayout]:更新需要计算其布局的渲染对象,在此阶段计算每个渲染对象的大小和位置,渲染对象可能会弄脏其绘画或者合成状态,这个过程可能还会调用到build过程。
    • 耗时对应timeline的‘Layout’过程
  2. [flushCompositingBits]:更新具有脏合成位的任何渲染对象,在此阶段每个渲染对象都会了解其子项是否需要合成。在绘制阶段使用此信息选择如何实现裁剪等视觉效果。如果渲染对象有一个自己合成的子项,它需要使用布局信息来创建裁剪,以便将裁剪应用于已合成的子项
    • 耗时对应timeline的‘Compositing bits’过程
  3. [flushPaint]:访问需要绘制的任何渲染对象,在此阶段,渲染对象有机会将绘制命令记录到[PictureLayer],并构建其他合成的[Layer];
    • 耗时对应timeline的‘Paint’过程
  4. [compositeFrame]:将Compositing bits发送给GPU;
    • 耗时对应timeline的‘Compositing’过程
  5. [flushSemantics]:编译渲染对象的语义,并将语义发送给操作系统;
    • 耗时对应timeline的‘Semantics’过程

packages/flutter/lib/src/rendering/debug.dart,这里面记录着关于render过程相关的调试开关,可以逐一实践。

4.7 PipelineOwner.flushLayout

[-> lib/src/rendering/object.dart]

void flushLayout() {
  profile(() {
    Timeline.startSync('Layout', arguments: timelineWhitelistArguments);
  });
  try {
    //遍历所有的渲染对象
    while (_nodesNeedingLayout.isNotEmpty) {
      final List<RenderObject> dirtyNodes = _nodesNeedingLayout;
      _nodesNeedingLayout = <RenderObject>[];
      for (RenderObject node in dirtyNodes..sort((RenderObject a, RenderObject b) => a.depth - b.depth)) {
        //如果渲染对象需要重新布局,则执行布局操作 [见小节4.7.1]
        if (node._needsLayout && node.owner == this)
          node._layoutWithoutResize();
      }
    }
  } finally {
    profile(() {
      Timeline.finishSync();
    });
  }
}

4.7.1 _layoutWithoutResize

[-> lib/src/rendering/object.dart]

void _layoutWithoutResize() {
  try {
    performLayout();  //执行布局操作[]
    markNeedsSemanticsUpdate();  //[见小节4.7.2]
  } catch (e, stack) {
    _debugReportException('performLayout', e, stack);
  }
  _needsLayout = false; //完成layout操作
  markNeedsPaint(); // [见小节4.7.3]
}

该方法主要工作:

  • performLayout操作:参数sizedByParent为false需要同时改变渲染对象和指导子项的布局,性能更慢;
  • markNeedsSemanticsUpdate:标记需要更新语义;
  • markNeedsPaint:标记需要绘制;
SchedulerBinding.scheduleWarmUpFrame
  RenderView.performLayout
    RenderObject.layout
      _RenderLayoutBuilder.performLayout
        _LayoutBuilderElement._layout
          BuildOwner.buildScope

4.7.2 markNeedsSemanticsUpdate

[-> lib/src/rendering/object.dart]

void markNeedsSemanticsUpdate() {
  if (!attached || owner._semanticsOwner == null) {
    _cachedSemanticsConfiguration = null;
    return;
  }

  final bool wasSemanticsBoundary = _semantics != null && _cachedSemanticsConfiguration?.isSemanticBoundary == true;
  _cachedSemanticsConfiguration = null;
  bool isEffectiveSemanticsBoundary = _semanticsConfiguration.isSemanticBoundary && wasSemanticsBoundary;
  RenderObject node = this;

  while (!isEffectiveSemanticsBoundary && node.parent is RenderObject) {
    if (node != this && node._needsSemanticsUpdate)
      break;
    node._needsSemanticsUpdate = true;

    node = node.parent;
    isEffectiveSemanticsBoundary = node._semanticsConfiguration.isSemanticBoundary;
    if (isEffectiveSemanticsBoundary && node._semantics == null) {
      return;
    }
  }
  if (node != this && _semantics != null && _needsSemanticsUpdate) {
    owner._nodesNeedingSemantics.remove(this);
  }
  if (!node._needsSemanticsUpdate) {
    node._needsSemanticsUpdate = true;
    if (owner != null) {
      //记录需要更新语义的渲染对象
      owner._nodesNeedingSemantics.add(node);
      owner.requestVisualUpdate();
    }
  }
}

4.7.3 markNeedsPaint

[-> lib/src/rendering/object.dart]

void markNeedsPaint() {
  if (_needsPaint)
    return;
  _needsPaint = true;
  if (isRepaintBoundary) {
    if (owner != null) {
      //记录需要重新绘制的渲染对象
      owner._nodesNeedingPaint.add(this);
      owner.requestVisualUpdate();
    }
  } else if (parent is RenderObject) {
    final RenderObject parent = this.parent;
    parent.markNeedsPaint();
  } else {
    if (owner != null)
      owner.requestVisualUpdate();
  }
}

4.8 PipelineOwner.flushCompositingBits

[-> lib/src/rendering/object.dart]

void flushCompositingBits() {
  profile(() { Timeline.startSync('Compositing bits'); });
  _nodesNeedingCompositingBitsUpdate.sort((RenderObject a, RenderObject b) => a.depth - b.depth);
  for (RenderObject node in _nodesNeedingCompositingBitsUpdate) {
    //根据需要来决定是否更新位合成
    if (node._needsCompositingBitsUpdate && node.owner == this)
      node._updateCompositingBits(); // [见小节4.8.1]
  }
  _nodesNeedingCompositingBitsUpdate.clear();  //清空需要位合成的渲染对象
  profile(() { Timeline.finishSync(); });
}

4.8.1 _updateCompositingBits

[-> lib/src/rendering/object.dart]

void _updateCompositingBits() {
  if (!_needsCompositingBitsUpdate)
    return;
  final bool oldNeedsCompositing = _needsCompositing;
  _needsCompositing = false;
  visitChildren((RenderObject child) {
    //遍历所有子项来更新位合成
    child._updateCompositingBits();
    if (child.needsCompositing)
      _needsCompositing = true;
  });
  if (isRepaintBoundary || alwaysNeedsCompositing)
    _needsCompositing = true;
  if (oldNeedsCompositing != _needsCompositing)
    markNeedsPaint();
  _needsCompositingBitsUpdate = false;
}

4.9 PipelineOwner.flushPaint

[-> lib/src/rendering/object.dart]

void flushPaint() {
  profile(() { Timeline.startSync('Paint', arguments: timelineWhitelistArguments); });
  try {
    final List<RenderObject> dirtyNodes = _nodesNeedingPaint;
    _nodesNeedingPaint = <RenderObject>[];
    //排序脏节点,深度最大的节点排第一位
    for (RenderObject node in dirtyNodes..sort((RenderObject a, RenderObject b) => b.depth - a.depth)) {
      if (node._needsPaint && node.owner == this) {
        //此节点是否连接到树中,如果连接则重绘,否则跳过
        if (node._layer.attached) {
          PaintingContext.repaintCompositedChild(node);  //[小节4.9.1]
        } else {
          node._skippedPaintingOnLayer();
        }
      }
    }
  } finally {
    profile(() { Timeline.finishSync(); });
  }
}

4.9.1 repaintCompositedChild

[-> lib/src/rendering/object.dart]

static void repaintCompositedChild(RenderObject child, { bool debugAlsoPaintedParent = false }) {
  _repaintCompositedChild(
    child,
    debugAlsoPaintedParent: debugAlsoPaintedParent,
  );
}

static void _repaintCompositedChild(
  RenderObject child, {
  bool debugAlsoPaintedParent = false,
  PaintingContext childContext,
}) {
  if (child._layer == null) {
    child._layer = OffsetLayer();
  } else {
    child._layer.removeAllChildren();
  }
  childContext ??= PaintingContext(child._layer, child.paintBounds);
  child._paintWithContext(childContext, Offset.zero);
  childContext.stopRecordingIfNeeded();
}

4.10 RenderView.compositeFrame

[-> lib/src/rendering/view.dart]

void compositeFrame() {
  Timeline.startSync('Compositing', arguments: timelineWhitelistArguments);
  try {
    //创建SceneBuilder [见小节4.10.1]
    final ui.SceneBuilder builder = ui.SceneBuilder();
    //创建Scene [见小节4.10.2]
    final ui.Scene scene = layer.buildScene(builder);
    if (automaticSystemUiAdjustment)
      _updateSystemChrome();
    ui.window.render(scene); // [见小节4.10.3]
    scene.dispose();
  } finally {
    Timeline.finishSync();
  }
}

该方法主要工作:

  • 分别创建Flutter框架(dart)和引擎层(C++)的两个SceneBuilder;
  • 分别创建Flutter框架(dart)和引擎层(C++)的两个Scene;
  • 执行render()将layer树发送给GPU线程;

4.10.1 SceneBuilder初始化

[-> lib/ui/compositing.dart]

class SceneBuilder extends NativeFieldWrapperClass2 {
  @pragma('vm:entry-point')
  SceneBuilder() { _constructor(); }
  void _constructor() native 'SceneBuilder_constructor';
  ...
}

SceneBuilder_constructor这是native方法,最终调用到引擎中的lib/ui/compositing/scene_builder.h中的SceneBuilder::create()方法, 创建C++的SceneBuilder对象。

4.10.2 OffsetLayer.buildScene

[-> lib/src/rendering/layer.dart]

ui.Scene buildScene(ui.SceneBuilder builder) {
  updateSubtreeNeedsAddToScene();  //遍历layer树,将需要子树加入到scene
  addToScene(builder); //将layer添加到SceneBuilder
  return builder.build(); //调用C++层的build来构建Scene对象。
}

遍历layer树,将需要更新的全部都加入到SceneBuilder。再调用build(),同样也是native方法,执行SceneBuilder::build()来构建Scene对象。

4.10.3 Window::Render

[-> flutter/lib/ui/window/window.cc]

void Render(Dart_NativeArguments args) {
  Dart_Handle exception = nullptr;
  Scene* scene = tonic::DartConverter<Scene*>::FromArguments(args, 1, exception);
  if (exception) {
    Dart_ThrowException(exception);
    return;
  }
  UIDartState::Current()->window()->client()->Render(scene);  // [4.10.4]
}

ui.window.render()位于window.dart文件,这是一个native方法,会调用到window.cc的Render()方法。

4.10.4 RuntimeController::Render

[-> flutter/runtime/runtime_controller.cc]

void RuntimeController::Render(Scene* scene) {
  //从scene中取出layer树 [见小节4.10.5]
  client_.Render(scene->takeLayerTree());
}

4.10.5 Engine::Render

[-> flutter/shell/common/engine.cc]

void Engine::Render(std::unique_ptr<flow::LayerTree> layer_tree) {
  if (!layer_tree)
    return;

  SkISize frame_size = SkISize::Make(viewport_metrics_.physical_width,
                                     viewport_metrics_.physical_height);
  if (frame_size.isEmpty())
    return;

  layer_tree->set_frame_size(frame_size);
  animator_->Render(std::move(layer_tree));  // [4.10.6]
}

4.10.6 Animator::Render

[-> flutter/shell/common/animator.cc]

void Animator::Render(std::unique_ptr<flow::LayerTree> layer_tree) {
  if (dimension_change_pending_ &&
      layer_tree->frame_size() != last_layer_tree_size_) {
    dimension_change_pending_ = false;
  }
  last_layer_tree_size_ = layer_tree->frame_size();

  if (layer_tree) {
    layer_tree->set_construction_time(fml::TimePoint::Now() -
                                      last_begin_frame_time_);
  }

  //提交待处理的continuation,本次PipelineProduce完成 //[见小节4.10.7]
  producer_continuation_.Complete(std::move(layer_tree));

  delegate_.OnAnimatorDraw(layer_tree_pipeline_); //[见小节4.10.8]
}

UI线程的耗时从doFrame(frameTimeNanos)中的frameTimeNanos为起点,以Animator::Render()方法结束为终点, 并将结果保存到LayerTree的成员变量construction_time_,这便是UI线程的耗时时长。

4.10.7 ProducerContinuation.Complete

[-> flutter/synchronization/pipeline.h]

class ProducerContinuation {
  void Complete(ResourcePtr resource) {
    if (continuation_) {
      continuation_(std::move(resource), trace_id_);
      continuation_ = nullptr;
      TRACE_EVENT_ASYNC_END0("flutter", "PipelineProduce", trace_id_);
      TRACE_FLOW_STEP("flutter", "PipelineItem", trace_id_);
    }
  }

4.10.8 Shell::OnAnimatorDraw

[-> flutter/shell/common/shell.cc]

void Shell::OnAnimatorDraw(
    fml::RefPtr<flutter::Pipeline<flow::LayerTree>> pipeline) {

  //向GPU线程提交绘制任务
  task_runners_.GetGPUTaskRunner()->PostTask(
      [rasterizer = rasterizer_->GetWeakPtr(),
       pipeline = std::move(pipeline)]() {
        if (rasterizer) {
          //由GPU线程来负责栅格化操作
          rasterizer->Draw(pipeline);
        }
      });
}

这个方法主要是向GPU线程提交绘制任务。

4.11 PipelineOwner.flushSemantics

[-> lib/src/rendering/view.dart]

void flushSemantics() {
  if (_semanticsOwner == null)
    return;
  profile(() { Timeline.startSync('Semantics'); });
  try {
    final List<RenderObject> nodesToProcess = _nodesNeedingSemantics.toList()
      ..sort((RenderObject a, RenderObject b) => a.depth - b.depth);
    _nodesNeedingSemantics.clear();
    //遍历_nodesNeedingSemantics,更新需要更新语义的渲染对象
    for (RenderObject node in nodesToProcess) {
      if (node._needsSemanticsUpdate && node.owner == this)
        node._updateSemantics(); // [见小节4.11.1]
    }
    _semanticsOwner.sendSemanticsUpdate(); // 发送语义更新[见小节4.11.2]
  } finally {
    profile(() { Timeline.finishSync(); });
  }
}

4.11.1 _updateSemantics

[-> lib/src/rendering/object.dart]

void _updateSemantics() {
  if (_needsLayout) {
    //此子树中没有足够的信息来计算语义,子树可能被视图窗口保持活着但没有布局
    return;
  }
  final _SemanticsFragment fragment = _getSemanticsForParent(
    mergeIntoParent: _semantics?.parent?.isPartOfNodeMerging ?? false,
  );
  final _InterestingSemanticsFragment interestingFragment = fragment;
  final SemanticsNode node = interestingFragment.compileChildren(
    parentSemanticsClipRect: _semantics?.parentSemanticsClipRect,
    parentPaintClipRect: _semantics?.parentPaintClipRect,
  ).single;
}

4.11.2 sendSemanticsUpdate

[-> lib/src/semantics/semantics.dart]

void sendSemanticsUpdate() {
  if (_dirtyNodes.isEmpty)
    return;
  final Set<int> customSemanticsActionIds = Set<int>();
  final List<SemanticsNode> visitedNodes = <SemanticsNode>[];
  while (_dirtyNodes.isNotEmpty) {
    final List<SemanticsNode> localDirtyNodes = _dirtyNodes.where((SemanticsNode node) => !_detachedNodes.contains(node)).toList();
    _dirtyNodes.clear();
    _detachedNodes.clear();
    localDirtyNodes.sort((SemanticsNode a, SemanticsNode b) => a.depth - b.depth);
    visitedNodes.addAll(localDirtyNodes);
    for (SemanticsNode node in localDirtyNodes) {
      if (node.isPartOfNodeMerging) {
        //如果合并到父节点,确保父节点已被添加到脏列表
        if (node.parent != null && node.parent.isPartOfNodeMerging)
          node.parent._markDirty(); //将节点添加到脏列表
      }
    }
  }
  visitedNodes.sort((SemanticsNode a, SemanticsNode b) => a.depth - b.depth);
  final ui.SemanticsUpdateBuilder builder = ui.SemanticsUpdateBuilder();
  for (SemanticsNode node in visitedNodes) {
    if (node._dirty && node.attached)
      node._addToUpdate(builder, customSemanticsActionIds);
  }
  _dirtyNodes.clear();
  for (int actionId in customSemanticsActionIds) {
    final CustomSemanticsAction action = CustomSemanticsAction.getAction(actionId);
    builder.updateCustomAction(id: actionId, label: action.label, hint: action.hint, overrideId: action.action?.index ?? -1);
  }
  ui.window.updateSemantics(builder.build());  // [见小节4.11.3]
  notifyListeners(); //通知已注册的监听器
}

可以看看监听器的数据,是否影响性能。

updateSemantics这是window.dart中的一个native方法,调用到如下方法。

4.11.3 Window::updateSemantics

[-> flutter/lib/ui/window/window.cc]

void UpdateSemantics(Dart_NativeArguments args) {
  Dart_Handle exception = nullptr;
  SemanticsUpdate* update =
      tonic::DartConverter<SemanticsUpdate*>::FromArguments(args, 1, exception);
  if (exception) {
    Dart_ThrowException(exception);
    return;
  }
  UIDartState::Current()->window()->client()->UpdateSemantics(update); // [见小节4.11.4]
}

4.11.4 RuntimeController::UpdateSemantics

[-> flutter/runtime/runtime_controller.cc]

void RuntimeController::UpdateSemantics(SemanticsUpdate* update) {
  if (window_data_.semantics_enabled) {
    client_.UpdateSemantics(update->takeNodes(), update->takeActions()); // [见小节4.11.5]
  }
}

4.11.5 Engine::UpdateSemantics

[-> flutter/shell/common/engine.cc]

void Engine::UpdateSemantics(blink::SemanticsNodeUpdates update,
                             blink::CustomAccessibilityActionUpdates actions) {
  delegate_.OnEngineUpdateSemantics(std::move(update), std::move(actions)); // [见小节4.11.6]
}

4.11.6 Shell::OnAnimatorDraw

[-> flutter/shell/common/shell.cc]

void Shell::OnEngineUpdateSemantics(
    blink::SemanticsNodeUpdates update,
    blink::CustomAccessibilityActionUpdates actions) {

  task_runners_.GetPlatformTaskRunner()->PostTask(
      [view = platform_view_->GetWeakPtr(), update = std::move(update),
       actions = std::move(actions)] {
        if (view) {
          view->UpdateSemantics(std::move(update), std::move(actions));
        }
      });
}

这个方法主要是向平台线程提交Semantic任务。

再回到小节4.6,可知接下来再执行finalizeTree()操作;

4.12 BuildOwner.finalizeTree

[-> lib/src/widgets/framework.dart]

void finalizeTree() {
  Timeline.startSync('Finalize tree', arguments: timelineWhitelistArguments);
  try {
    lockState(() {
      //遍历所有的Element,执行unmount()动作,且取消GlobalKeys的注册
      _inactiveElements._unmountAll();
    });
  } catch (e, stack) {
    _debugReportException('while finalizing the widget tree', e, stack);
  } finally {
    Timeline.finishSync();
  }
}

遍历所有的Element,执行相应具体Element子类的unmount()操作,下面以常见的StatefulElement为例来说明。

4.12.1 StatefulElement.unmount

[-> lib/src/widgets/framework.dart]

void unmount() {
  super.unmount(); //[见小节4.12.2]
  _state.dispose(); //执行State的dispose()方法
  _state._element = null;
  _state = null;
}

4.12.2 Element.unmount

[-> lib/src/widgets/framework.dart]

void unmount() {
  if (widget.key is GlobalKey) {
    final GlobalKey key = widget.key;
    key._unregister(this);  //取消GlobalKey的注册
  }
}

附录

本文涉及到相关源码文件

flutter/shell/common/
    - vsync_waiter.cc
    - engine.cc
    - animator.cc
    - shell.cc
    - rasterizer.cc

flutter/shell/platform/android/
    - vsync_waiter_android.cc
    - platform_view_android_jni.cc
    - library_loader.cc
    - io/flutter/view/VsyncWaiter.java

flutter/runtime/runtime_controller.cc
flutter/synchronization/pipeline.h
flutter/fml/message_loop_impl.cc
flutter/lib/ui/window/window.cc
flutter/lib/ui/window.dart
flutter/lib/ui/hooks.dart

lib/src/widgets/framework.dart
lib/src/widgets/binding.dart
lib/src/scheduler/binding.dart
lib/src/semantics/semantics.dart

lib/src/rendering/
    - binding.dart
    - object.dart
    - view.dart

微信公众号 Gityuan | 微博 weibo.com/gityuan | 博客 留言区交流