Android P 图像显示系统(三)Android HWUI 绘制流程

0.913字数 5619阅读 7752

Android中,绘图的API很多,比如2D的绘图skia;3D的绘图OpenGLES,Vulkan等。Android 开始,的View系统中,多数都是采用2D的模式的View Widget,比如绘制一张Bitmap图片,显示一个按钮等。随着Android系统的更新,和用户对视觉效果的追求,以前的这套2D View系统,不仅不能满足要求,而且渲染非常的慢。所以Android一方面完善对3D的API的支持,另一方面修改原来View Widget的渲染机制。

渲染机制的更新,Android提出了硬件加速的机制,其作用就是将2D的绘图操纵,转换为对应的3D的绘图操纵,这个转换的过程,我们把它叫做录制。需要显示的时候,再用OpenGLES通过GPU去渲染。界面创建时,第一次全部录制,后续的过程中,界面如果只有部分区域的widget更新,只需要重新录制更新的widget。录制好的绘图操纵,保存在一个显示列表DisplayList中,需要真正显示到界面的时候,直接显示DisplayList中的绘图 操纵。这样,一方面利用GPU去渲染,比Skia要快;另一方面,采用DisplayList,值重新录制,有更新区域,最大程度利用上一帧的数据,效率自然就快很多。这就是硬件加速的来源。
roundRectClipState
语言苍白,实践为先,我们结合测试示例,来看看硬件加速是怎么回事~

应用使用硬件(GPU)绘制实例

这个是Android原生的测试硬件绘制的应用:

* frameworks/base/tests/HwAccelerationTest/src/com/android/test/hwui/HardwareCanvasSurfaceViewActivity.java

    private static class RenderingThread extends Thread {
        private final SurfaceHolder mSurface;
        private volatile boolean mRunning = true;
        private int mWidth, mHeight;

        public RenderingThread(SurfaceHolder surface) {
            mSurface = surface;
        }

        void setSize(int width, int height) {
            mWidth = width;
            mHeight = height;
        }

        @Override
        public void run() {
            float x = 0.0f;
            float y = 0.0f;
            float speedX = 5.0f;
            float speedY = 3.0f;

            Paint paint = new Paint();
            paint.setColor(0xff00ff00);

            while (mRunning && !Thread.interrupted()) {
                final Canvas canvas = mSurface.lockHardwareCanvas();
                try {
                    canvas.drawColor(0x00000000, PorterDuff.Mode.CLEAR);
                    canvas.drawRect(x, y, x + 20.0f, y + 20.0f, paint);
                } finally {
                    mSurface.unlockCanvasAndPost(canvas);
                }

                ... ...

                try {
                    Thread.sleep(15);
                } catch (InterruptedException e) {
                    // Interrupted
                }
            }
        }

        void stopRendering() {
            interrupt();
            mRunning = false;
        }
    }

应用这里拿到一个Surface,然后lock一个HardwareCanvas,用lock的HardwareCanvas进行绘制,我们绘制的就可以使用硬件GPU进行绘制。这里每隔15秒循环一次,绘制一个小方块,在屏幕上不停的运动。而背景,被绘制成0x00000000,黑色。

硬件绘制Java层相关流程

通过前面的代码,关键的是在lockHardwareCanvas。

lockHardwareCanvas的代码如下:

* frameworks/base/core/java/android/view/SurfaceView.java

        @Override
        public Canvas lockHardwareCanvas() {
            return internalLockCanvas(null, true);
        }

        private Canvas internalLockCanvas(Rect dirty, boolean hardware) {
            mSurfaceLock.lock();

            if (DEBUG) Log.i(TAG, System.identityHashCode(this) + " " + "Locking canvas... stopped="
                    + mDrawingStopped + ", surfaceControl=" + mSurfaceControl);

            Canvas c = null;
            if (!mDrawingStopped && mSurfaceControl != null) {
                try {
                    if (hardware) {
                        c = mSurface.lockHardwareCanvas();
                    } else {
                        c = mSurface.lockCanvas(dirty);
                    }
                } catch (Exception e) {
                    Log.e(LOG_TAG, "Exception locking surface", e);
                }
            }

            if (DEBUG) Log.i(TAG, System.identityHashCode(this) + " " + "Returned canvas: " + c);
            if (c != null) {
                mLastLockTime = SystemClock.uptimeMillis();
                return c;
            }

            ... ...

            return null;
        }

这里Canvas是通过mSurface来申请的。

* frameworks/base/core/java/android/view/Surface.java

    public Canvas lockHardwareCanvas() {
        synchronized (mLock) {
            checkNotReleasedLocked();
            if (mHwuiContext == null) {
                mHwuiContext = new HwuiContext();
            }
            return mHwuiContext.lockCanvas(
                    nativeGetWidth(mNativeObject),
                    nativeGetHeight(mNativeObject));
        }
    }

Surface中封装了一个 HwuiContext ,其构造函数如下:

        HwuiContext() {
            mRenderNode = RenderNode.create("HwuiCanvas", null);
            mRenderNode.setClipToBounds(false);
            mHwuiRenderer = nHwuiCreate(mRenderNode.mNativeRenderNode, mNativeObject);
        }

在HwuiContext的构造函数中,创建了一个RenderNode,创建了一个HwuiRenderer。nHwuiCreate创建一个native的HwuiRender。

这里的HwuiContext,就是和HWUI打交道了。

HwuiContext的lockCanvas实现如下:

        Canvas lockCanvas(int width, int height) {
            if (mCanvas != null) {
                throw new IllegalStateException("Surface was already locked!");
            }
            mCanvas = mRenderNode.start(width, height);
            return mCanvas;
        }

RenderNode的start函数:

    public DisplayListCanvas start(int width, int height) {
        return DisplayListCanvas.obtain(this, width, height);
    }

    static DisplayListCanvas obtain(@NonNull RenderNode node, int width, int height) {
        if (node == null) throw new IllegalArgumentException("node cannot be null");
        DisplayListCanvas canvas = sPool.acquire();
        if (canvas == null) {
            canvas = new DisplayListCanvas(node, width, height);
        } else {
            nResetDisplayListCanvas(canvas.mNativeCanvasWrapper, node.mNativeRenderNode,
                    width, height);
        }
        canvas.mNode = node;
        canvas.mWidth = width;
        canvas.mHeight = height;
        return canvas;
    }

RenderNode,start时,将创建一个DisplayListCanvas。DisplayListCanvas是显示列表的Canvas。DisplayListCanvas 构建时,将通过nCreateDisplayListCanvas创建一个native的DisplayListCanvas。

    private DisplayListCanvas(@NonNull RenderNode node, int width, int height) {
        super(nCreateDisplayListCanvas(node.mNativeRenderNode, width, height));
        mDensity = 0; // disable bitmap density scaling
    }

DisplayListCanvas和RecordingCanvas的构造函数都比较简单,但是留意一下Canvas的构造函数:

    public Canvas(long nativeCanvas) {
        if (nativeCanvas == 0) {
            throw new IllegalStateException();
        }
        mNativeCanvasWrapper = nativeCanvas;
        mFinalizer = NoImagePreloadHolder.sRegistry.registerNativeAllocation(
                this, mNativeCanvasWrapper);
        mDensity = Bitmap.getDefaultDensity();
    }

这里的mNativeCanvasWrapper,就是nCreateDisplayListCanvas时,创建的native对应的Canvas。后续,JNI中都是通过mNativeCanvasWrapper去找到对应的nativ的Canvas的。

我们先来看这些相关的类之间的关系~


Hwui Context相关类图

其中,RenderNode,DisplayListCanvas,HwuiRenderer构成了硬件绘制的重要元素。

再回到我们的测试代码,我们这里有两个绘制操纵:

  • drawColor
  • drawRect

drawColor是在DisplayListCanvas的父类RecordingCanvas中实现的:

    public final void drawColor(@ColorInt int color, @NonNull PorterDuff.Mode mode) {
        nDrawColor(mNativeCanvasWrapper, color, mode.nativeInt);
    }

这里调用native的nDrawColor方法。

drawRect也是在DisplayListCanvas的父类RecordingCanvas中实现的:

    @Override
    public final void drawRect(float left, float top, float right, float bottom,
            @NonNull Paint paint) {
        nDrawRect(mNativeCanvasWrapper, left, top, right, bottom, paint.getNativeInstance());
    }

调用native的nDrawRect方法。

native处理流程

native的Canvas创建

DisplayListCanvas的JNI实现如下:

* frameworks/base/core/jni/android_view_DisplayListCanvas.cpp

const char* const kClassPathName = "android/view/DisplayListCanvas";

static JNINativeMethod gMethods[] = {

    // ------------ @FastNative ------------------

    { "nCallDrawGLFunction", "(JJLjava/lang/Runnable;)V",
            (void*) android_view_DisplayListCanvas_callDrawGLFunction },

    // ------------ @CriticalNative --------------
    { "nCreateDisplayListCanvas", "(JII)J",     (void*) android_view_DisplayListCanvas_createDisplayListCanvas },
    { "nResetDisplayListCanvas",  "(JJII)V",    (void*) android_view_DisplayListCanvas_resetDisplayListCanvas },
    { "nGetMaximumTextureWidth",  "()I",        (void*) android_view_DisplayListCanvas_getMaxTextureWidth },
    { "nGetMaximumTextureHeight", "()I",        (void*) android_view_DisplayListCanvas_getMaxTextureHeight },
    { "nInsertReorderBarrier",    "(JZ)V",      (void*) android_view_DisplayListCanvas_insertReorderBarrier },
    { "nFinishRecording",         "(J)J",       (void*) android_view_DisplayListCanvas_finishRecording },
    { "nDrawRenderNode",          "(JJ)V",      (void*) android_view_DisplayListCanvas_drawRenderNode },
    { "nDrawLayer",               "(JJ)V",      (void*) android_view_DisplayListCanvas_drawLayer },
    { "nDrawCircle",              "(JJJJJ)V",   (void*) android_view_DisplayListCanvas_drawCircleProps },
    { "nDrawRoundRect",           "(JJJJJJJJ)V",(void*) android_view_DisplayListCanvas_drawRoundRectProps },
};

nCreateDisplayListCanvas对应的实现为android_view_DisplayListCanvas_createDisplayListCanvas。

static jlong android_view_DisplayListCanvas_createDisplayListCanvas(jlong renderNodePtr,
        jint width, jint height) {
    RenderNode* renderNode = reinterpret_cast<RenderNode*>(renderNodePtr);
    return reinterpret_cast<jlong>(Canvas::create_recording_canvas(width, height, renderNode));
}

注意我们这里的renderNodePtr。这个是RenderNode在native层的对象(地址)。

Canvas的create_recording_canvas函数如下:

Canvas* Canvas::create_recording_canvas(int width, int height, uirenderer::RenderNode* renderNode) {
    if (uirenderer::Properties::isSkiaEnabled()) {
        return new uirenderer::skiapipeline::SkiaRecordingCanvas(renderNode, width, height);
    }
    return new uirenderer::RecordingCanvas(width, height);
}

isSkiaEnabled没有被enable的,所以创建的是native的RecordingCanvas。Android 8.0开始,对HWUI进行了重构,增加了RenderPipeline的概念。目前有三种类型的pipeline,分别对应不同的渲染。

enum class RenderPipelineType {
    OpenGL = 0,
    SkiaGL,
    SkiaVulkan,
    NotInitialized = 128
};

默认还是OpenGL类型。

native的RecordingCanvas如下:

* frameworks/base/libs/hwui/RecordingCanvas.cpp

RecordingCanvas::RecordingCanvas(size_t width, size_t height)
        : mState(*this), mResourceCache(ResourceCache::getInstance()) {
    resetRecording(width, height);
}

RecordingCanvas创建时,创建了对应的CanvasState,和ResourceCache。CanvasState是Canvas的状态,管理Snapshot的栈,实现matrix,save/restore,clipping等Renderer的接口。ResourceCache主要是做资源cache,cache为点九类型。

在resetRecording函数中,又做了很多初始化。

void RecordingCanvas::resetRecording(int width, int height, RenderNode* node) {
    LOG_ALWAYS_FATAL_IF(mDisplayList, "prepareDirty called a second time during a recording!");
    mDisplayList = new DisplayList();

    mState.initializeRecordingSaveStack(width, height);

    mDeferredBarrierType = DeferredBarrierType::InOrder;
}
  • 创建了显示列表mDisplayList,这个很重要,稍后我们再介绍。它主要用来保存显示列表的绘制命令。
  • 初始化CanvasState

到此,native的Canvas创建完成。

Draw操纵的录制

测试代码中,一共两个绘制操纵,我们以这两个绘制操纵为例,来说明绘制的操纵的录制。
nDrawColor nDrawRect

* frameworks/base/core/jni/android_graphics_Canvas.cpp

static const JNINativeMethod gDrawMethods[] = {
    {"nDrawColor","(JII)V", (void*) CanvasJNI::drawColor},
    {"nDrawPaint","(JJ)V", (void*) CanvasJNI::drawPaint},
    {"nDrawPoint", "(JFFJ)V", (void*) CanvasJNI::drawPoint},
    {"nDrawPoints", "(J[FIIJ)V", (void*) CanvasJNI::drawPoints},
    {"nDrawLine", "(JFFFFJ)V", (void*) CanvasJNI::drawLine},
    {"nDrawLines", "(J[FIIJ)V", (void*) CanvasJNI::drawLines},
    {"nDrawRect","(JFFFFJ)V", (void*) CanvasJNI::drawRect},

drawColor函数

static void drawColor(JNIEnv* env, jobject, jlong canvasHandle, jint color, jint modeHandle) {
    SkBlendMode mode = static_cast<SkBlendMode>(modeHandle);
    get_canvas(canvasHandle)->drawColor(color, mode);
}

canvasHandle为native RecordingCanvas的handle,所以get_canvas获取到的是RecordingCanvas。

RecordingCanvas的drawColor函数如下:

* frameworks/base/libs/hwui/RecordingCanvas.cpp

void RecordingCanvas::drawColor(int color, SkBlendMode mode) {
    addOp(alloc().create_trivial<ColorOp>(getRecordedClip(), color, mode));
}
  • alloc()获取到的是DisplayList的allocator
  • create_trivial是一个模板函数
    template <class T, typename... Params>
    T* create_trivial(Params&&... params) {
        static_assert(std::is_trivially_destructible<T>::value,
                      "Error, called create_trivial on a non-trivial type");
        return new (allocImpl(sizeof(T))) T(std::forward<Params>(params)...);
    }

类型 T为ColorOp,参数params为(getRecordedClip(), color, mode),其作用就是构造已给ColorOp。

  • allocImpl,分配内存空间

ColorOp的定义在头文件中:

frameworks/base/libs/hwui/RecordedOp.h

struct ColorOp : RecordedOp {
    // Note: unbounded op that will fillclip, so no bounds/matrix needed
    ColorOp(const ClipBase* localClip, int color, SkBlendMode mode)
            : RecordedOp(RecordedOpId::ColorOp, Rect(), Matrix4::identity(), localClip, nullptr)
            , color(color)
            , mode(mode) {}
    const int color;
    const SkBlendMode mode;
};

RecordedOp.h中定义了所以的绘图操纵。

如nDrawRect对应的操纵为RectOp:

void RecordingCanvas::drawRect(float left, float top, float right, float bottom,
                               const SkPaint& paint) {
    if (CC_UNLIKELY(paint.nothingToDraw())) return;

    addOp(alloc().create_trivial<RectOp>(Rect(left, top, right, bottom),
                                         *(mState.currentSnapshot()->transform), getRecordedClip(),
                                         refPaint(&paint)));
}

struct RectOp : RecordedOp {
    RectOp(BASE_PARAMS) : SUPER(RectOp) {}
};

所有的绘图操作都继承RecordedOp。

RecordedOp定义如下:

struct RecordedOp {
    /* ID from RecordedOpId - generally used for jumping into function tables */
    const int opId;

    /* bounds in *local* space, without accounting for DisplayList transformation, or stroke */
    const Rect unmappedBounds;

    /* transform in recording space (vs DisplayList origin) */
    const Matrix4 localMatrix;

    /* clip in recording space - nullptr if not clipped */
    const ClipBase* localClip;

    /* optional paint, stored in base object to simplify merging logic */
    const SkPaint* paint;

protected:
    RecordedOp(unsigned int opId, BASE_PARAMS)
            : opId(opId)
            , unmappedBounds(unmappedBounds)
            , localMatrix(localMatrix)
            , localClip(localClip)
            , paint(paint) {}
};
  • opId,RecordedOpId中的ID,用以调转到对应的函数
  • unmappedBounds,绘制区域的大小
  • localMatrix,transform
  • ClipBase,截取
  • paint,画笔

绘图操纵创建后,通过addOp方法,添加到DisplayList中:

int RecordingCanvas::addOp(RecordedOp* op) {
    // skip op with empty clip
    if (op->localClip && op->localClip->rect.isEmpty()) {
        // NOTE: this rejection happens after op construction/content ref-ing, so content ref'd
        // and held by renderthread isn't affected by clip rejection.
        // Could rewind alloc here if desired, but callers would have to not touch op afterwards.
        return -1;
    }

    int insertIndex = mDisplayList->ops.size();
    mDisplayList->ops.push_back(op);
    if (mDeferredBarrierType != DeferredBarrierType::None) {
        // op is first in new chunk
        mDisplayList->chunks.emplace_back();
        DisplayList::Chunk& newChunk = mDisplayList->chunks.back();
        newChunk.beginOpIndex = insertIndex;
        newChunk.endOpIndex = insertIndex + 1;
        newChunk.reorderChildren = (mDeferredBarrierType == DeferredBarrierType::OutOfOrder);
        newChunk.reorderClip = mDeferredBarrierClip;

        int nextChildIndex = mDisplayList->children.size();
        newChunk.beginChildIndex = newChunk.endChildIndex = nextChildIndex;
        mDeferredBarrierType = DeferredBarrierType::None;
    } else {
        // standard case - append to existing chunk
        mDisplayList->chunks.back().endOpIndex = insertIndex + 1;
    }
    return insertIndex;
}

不得不说,这里有点复杂,但是很巧妙。

  • 所有的绘图操纵,我们把它叫做Ops,都保存在ops中。ops就好比一个公司,而Ops就是一个员工。而每个Ops都有一个序号insertIndex,按照加入的先后顺序,相当与工号。
  • chunk中还没有元素时,mDeferredBarrierType为DeferredBarrierType::InOrder,这个时候就会增加一个Chunk。除非重新插入Barrier,即insertReorderBarrier,要不然,后续添加的Ops都是在同一个Chunk中的。Chunk就好比公司里面的部门,部门说,工号从多少号到多少号的归属于这个部门。beginOpIndex是开始的序号,endOpIndex是结束的序号,这之间的,都是属于同一个Chunk,每加入一个Ops,endOpIndex就会加1。
  • 怎么来理解children呢?按照前面的类比,可以理解为一个部门里面的小组。beginChildIndex和endChildIndex之间的Ops都属于同一个Children。

其实,这的Ops,chunk,children就是对Android View系统的抽象化。Chunk对应RootView,而children对应ViewGroup,Ops再对应,绘制Color,Rect等操纵。就是这么神奇~

我们来看一下DisplayList和Ops之间的关系


DisplayList显示列表

绘制操纵完成后,所有绘制操纵极其参数都保存在DisplayList中了。那么这些绘制操纵什么时候显示出来呢?我们继续看。

创建RenderNode

RenderNode用以录制绘图操纵的批处理,当绘制的时候,可以store和apply。
java层的代码如下:其实RenderNode就对应前面我们所说的ViewGroup,有一个RootView,同样也有一个RootNode。

我们先来看RenderNode是怎么创建的

    public static RenderNode create(String name, @Nullable View owningView) {
        return new RenderNode(name, owningView);
    }

    private RenderNode(String name, View owningView) {
        mNativeRenderNode = nCreate(name);
        NoImagePreloadHolder.sRegistry.registerNativeAllocation(this, mNativeRenderNode);
        mOwningView = owningView;
    }

nCreate是JNI方法。

RenderNode的JNI实现如下:

const char* const kClassPathName = "android/view/RenderNode";

static const JNINativeMethod gMethods[] = {
// ----------------------------------------------------------------------------
// Regular JNI
// ----------------------------------------------------------------------------
    { "nCreate",               "(Ljava/lang/String;)J", (void*) android_view_RenderNode_create },
    { "nGetNativeFinalizer",   "()J",    (void*) android_view_RenderNode_getNativeFinalizer },
    { "nOutput",               "(J)V",    (void*) android_view_RenderNode_output },
    { "nGetDebugSize",         "(J)I",    (void*) android_view_RenderNode_getDebugSize },
    { "nAddAnimator",              "(JJ)V", (void*) android_view_RenderNode_addAnimator },
    { "nEndAllAnimators",          "(J)V", (void*) android_view_RenderNode_endAllAnimators },
    { "nRequestPositionUpdates",   "(JLandroid/view/SurfaceView;)V", (void*) android_view_RenderNode_requestPositionUpdates },
    { "nSetDisplayList",       "(JJ)V",   (void*) android_view_RenderNode_setDisplayList },

nCreate函数实现为android_view_RenderNode_create

static jlong android_view_RenderNode_create(JNIEnv* env, jobject, jstring name) {
    RenderNode* renderNode = new RenderNode();
    renderNode->incStrong(0);
    if (name != NULL) {
        const char* textArray = env->GetStringUTFChars(name, NULL);
        renderNode->setName(textArray);
        env->ReleaseStringUTFChars(name, textArray);
    }
    return reinterpret_cast<jlong>(renderNode);
}

在JNI中就创建了一个native的RenderNode

* frameworks/base/libs/hwui/RenderNode.cpp

RenderNode::RenderNode()
        : mDirtyPropertyFields(0)
        , mNeedsDisplayListSync(false)
        , mDisplayList(nullptr)
        , mStagingDisplayList(nullptr)
        , mAnimatorManager(*this)
        , mParentCount(0) {}

创建完成的RenderNode,是给到DisplayListCanvas的。

HwuiContext和HwuiRenderer

nHwuiCreate创建HwuiRenderer

* frameworks/base/core/jni/android_view_Surface.cpp
static const JNINativeMethod gSurfaceMethods[] = {
    ... ...

    // HWUI context
    {"nHwuiCreate", "(JJ)J", (void*) hwui::create },
    {"nHwuiSetSurface", "(JJ)V", (void*) hwui::setSurface },
    {"nHwuiDraw", "(J)V", (void*) hwui::draw },
    {"nHwuiDestroy", "(J)V", (void*) hwui::destroy },
};

nHwuiCreate函数实现如下:

static jlong create(JNIEnv* env, jclass clazz, jlong rootNodePtr, jlong surfacePtr) {
    RenderNode* rootNode = reinterpret_cast<RenderNode*>(rootNodePtr);
    sp<Surface> surface(reinterpret_cast<Surface*>(surfacePtr));
    ContextFactory factory;
    RenderProxy* proxy = new RenderProxy(false, rootNode, &factory);
    proxy->loadSystemProperties();
    proxy->setSwapBehavior(SwapBehavior::kSwap_discardBuffer);
    proxy->initialize(surface);
    // Shadows can't be used via this interface, so just set the light source
    // to all 0s.
    proxy->setup(0, 0, 0);
    proxy->setLightCenter((Vector3){0, 0, 0});
    return (jlong) proxy;
}

创建了一个RenderProxy,nHwuiCreate返回的是一个RenderProxy实例。

RenderProxy的构造函数如下:

* frameworks/base/libs/hwui/renderthread/RenderProxy.cpp

RenderProxy::RenderProxy(bool translucent, RenderNode* rootRenderNode,
                         IContextFactory* contextFactory)
        : mRenderThread(RenderThread::getInstance()), mContext(nullptr) {
    mContext = mRenderThread.queue().runSync([&]() -> CanvasContext* {
        return CanvasContext::create(mRenderThread, translucent, rootRenderNode, contextFactory);
    });
    mDrawFrameTask.setContext(&mRenderThread, mContext, rootRenderNode);
}

这里诞生了很多东西:

  • RenderProxy是一个代理者,严格的单线程。所有的方法都必须在自己的线程中调用。
  • RenderThread,渲染线程,是一个单例,也就是说,一个进程中只有一个,所有的绘制操纵都必须在这个线程中完成。应用端很多操纵,都以RenderTask的形式post到RenderThread线程中完成。
  • CanvasContext,上下文,由于OpenGL是单线程的,所以,我们给到GPU的绘图命令都封装在各自的上下文中。这个和上层的HwuiRenderer是对应的。
  • DrawFrameTask,比较特殊的一个RenderTask。可重复使用的绘制Task。

我们先来理解这个HWUI的Thread。

RenderThread

hwui中很多C++的新特性,代码比较难理解。

* frameworks/base/libs/hwui/renderthread/RenderThread.h

class RenderThread : private ThreadBase {
    PREVENT_COPY_AND_ASSIGN(RenderThread);
    
  • PREVENT_COPY_AND_ASSIG阻止拷贝构造函数和=重载
  • 继承ThreadBase,ThreadBase继承Android的基本类Thread

在构造RenderThread时,就启动了RenderThread线程。

RenderThread::RenderThread()
        : ThreadBase()
        , mDisplayEventReceiver(nullptr)
        , mVsyncRequested(false)
        , mFrameCallbackTaskPending(false)
        , mRenderState(nullptr)
        , mEglManager(nullptr)
        , mVkManager(nullptr) {
    Properties::load();
    start("RenderThread");
}

ThreadBase的构造函数值得一看:

    ThreadBase()
            : Thread(false)
            , mLooper(new Looper(false))
            , mQueue([this]() { mLooper->wake(); }, mLock) {}

mQueue的实例化,C++的新特性。其实就是构造一个Queue,第一个参数是一个函数。函数体为:

{ mLooper->wake(); }

这个函数执行的时候,就唤醒mLooper,线程开始工作。

WorkQueue的构造函数如下:

    WorkQueue(std::function<void()>&& wakeFunc, std::mutex& lock)
            : mWakeFunc(move(wakeFunc)), mLock(lock) {}

我们再来看RenderThread是怎么工作的。RenderThread起来后,就会执行RenderThread的threadLoop。

threadLoop如下:

bool RenderThread::threadLoop() {
    setpriority(PRIO_PROCESS, 0, PRIORITY_DISPLAY);
    if (gOnStartHook) {
        gOnStartHook();
    }
    initThreadLocals();

    while (true) {
        waitForWork();
        processQueue();

        if (mPendingRegistrationFrameCallbacks.size() && !mFrameCallbackTaskPending) {
            drainDisplayEventQueue();
            mFrameCallbacks.insert(mPendingRegistrationFrameCallbacks.begin(),
                                   mPendingRegistrationFrameCallbacks.end());
            mPendingRegistrationFrameCallbacks.clear();
            requestVsync();
        }

        if (!mFrameCallbackTaskPending && !mVsyncRequested && mFrameCallbacks.size()) {
            // TODO: Clean this up. This is working around an issue where a combination
            // of bad timing and slow drawing can result in dropping a stale vsync
            // on the floor (correct!) but fails to schedule to listen for the
            // next vsync (oops), so none of the callbacks are run.
            requestVsync();
        }
    }

    return false;
}
  • initThreadLocals初始化Thread的本地变量
  • threadLoop中while循环,不停处理请求。如果没有任务时,等在waitForWork

前面是创建完RenderProxy后,还会设置一些参数

    RenderProxy* proxy = new RenderProxy(false, rootNode, &factory);
    proxy->loadSystemProperties();
    proxy->setSwapBehavior(SwapBehavior::kSwap_discardBuffer);
    proxy->initialize(surface);
    // Shadows can't be used via this interface, so just set the light source
    // to all 0s.
    proxy->setup(0, 0, 0);
    proxy->setLightCenter((Vector3){0, 0, 0});

我们以initialize为例。

void RenderProxy::initialize(const sp<Surface>& surface) {
    mRenderThread.queue().post(
            [ this, surf = surface ]() mutable { mContext->setSurface(std::move(surf)); });
}

initialize时,将给mRenderThread的队列中post一个东西,Oops...现在还不知道它是什么。下面我们将来看它是什么。

post是一个模板函数:

* frameworks/base/libs/hwui/thread/WorkQueue.h

    template <class F>
    void post(F&& func) {
        postAt(0, std::forward<F>(func));
    }
    
    template <class F>
    void postAt(nsecs_t time, F&& func) {
        enqueue(WorkItem{time, std::function<void()>(std::forward<F>(func))});
    }

post的时候,将根据传进来的参数,创建一个WorkItem,enqueue到消息队列mWorkQueue中。

    void enqueue(WorkItem&& item) {
        bool needsWakeup;
        {
            std::unique_lock _lock{mLock};
            auto insertAt = std::find_if(
                    std::begin(mWorkQueue), std::end(mWorkQueue),
                    [time = item.runAt](WorkItem & item) { return item.runAt > time; });
            needsWakeup = std::begin(mWorkQueue) == insertAt;
            mWorkQueue.emplace(insertAt, std::move(item));
        }
        if (needsWakeup) {
            mWakeFunc();
        }
    }

mWakeFunc如果需要唤醒,就通过mWakeFunc函数,唤醒mLooper。还记得吗?mWakeFunc是ThreadBase中构建WorkQueue时,传下来的无名函数。

WorkItem定义如下。

    struct WorkItem {
        WorkItem() = delete;
        WorkItem(const WorkItem& other) = delete;
        WorkItem& operator=(const WorkItem& other) = delete;
        WorkItem(WorkItem&& other) = default;
        WorkItem& operator=(WorkItem&& other) = default;

        WorkItem(nsecs_t runAt, std::function<void()>&& work)
                : runAt(runAt), work(std::move(work)) {}

        nsecs_t runAt;
        std::function<void()> work;
    };

对于我们的initialize函数而言,这里的WorkItem中的work是不是mContext->setSurface?答案是肯定的。

再来看RenderThread,收到新消息后怎么处理。

首先用processQueue处理Queue。

void processQueue() { mQueue.process(); }

最终还是 回到WorkQueue 中。

    void process() {
        auto now = clock::now();
        std::vector<WorkItem> toProcess;
        {
            std::unique_lock _lock{mLock};
            if (mWorkQueue.empty()) return;
            toProcess = std::move(mWorkQueue);
            auto moveBack = find_if(std::begin(toProcess), std::end(toProcess),
                                    [&now](WorkItem& item) { return item.runAt > now; });
            if (moveBack != std::end(toProcess)) {
                mWorkQueue.reserve(std::distance(moveBack, std::end(toProcess)) + 5);
                std::move(moveBack, std::end(toProcess), std::back_inserter(mWorkQueue));
                toProcess.erase(moveBack, std::end(toProcess));
            }
        }
        for (auto& item : toProcess) {
            item.work();
        }
    }

这里将mWorkQueue中未处理的WorkItem找处理,放到toProcess中。再调用每个Item的work方法。

对于我们的initialize函数而言,这里是不是就是mContext->setSurface?也就是CanvasContext的setSurface方法:

void CanvasContext::setSurface(sp<Surface>&& surface) {
    ATRACE_CALL();

    mNativeSurface = std::move(surface);

    ColorMode colorMode = mWideColorGamut ? ColorMode::WideColorGamut : ColorMode::Srgb;
    bool hasSurface = mRenderPipeline->setSurface(mNativeSurface.get(), mSwapBehavior, colorMode);

    mFrameNumber = -1;

    if (hasSurface) {
        mHaveNewSurface = true;
        mSwapHistory.clear();
    } else {
        mRenderThread.removeFrameCallback(this);
    }
}

神奇吧~

很多RenderProxy中的操作,都是通过这种方式post到CanvasContext中,且运行在RenderThread线程中。

我们再来看一个特殊的Task DrawFrameTask。

RenderProxy创建时,创建的DrawFrameTask

* frameworks/base/libs/hwui/renderthread/DrawFrameTask.cpp

DrawFrameTask::DrawFrameTask()
        : mRenderThread(nullptr)
        , mContext(nullptr)
        , mContentDrawBounds(0, 0, 0, 0)
        , mSyncResult(SyncResult::OK) {}

DrawFrameTask::~DrawFrameTask() {}

void DrawFrameTask::setContext(RenderThread* thread, CanvasContext* context,
                               RenderNode* targetNode) {
    mRenderThread = thread;
    mContext = context;
    mTargetNode = targetNode;
}

到目前位置,DisplayList有了,RenderThread有了,但是绘制在哪儿呢?我们这里直接解密吧,具体的流程就不介绍了,我们单看hwui这部分的逻辑。

显示时,上层会调syncAndDrawFrame

int RenderProxy::syncAndDrawFrame() {
    return mDrawFrameTask.drawFrame();
}
int DrawFrameTask::drawFrame() {
    LOG_ALWAYS_FATAL_IF(!mContext, "Cannot drawFrame with no CanvasContext!");

    mSyncResult = SyncResult::OK;
    mSyncQueued = systemTime(CLOCK_MONOTONIC);
    postAndWait();

    return mSyncResult;
}

void DrawFrameTask::postAndWait() {
    AutoMutex _lock(mLock);
    mRenderThread->queue().post([this]() { run(); });
    mSignal.wait(mLock);
}

这类,drawFrame,也就通过RenderThread,post一个WorkItem到RenderThread的队列里面,在RenderThread线程中执行的。

RenderThread处理Queue时,执行的确是这里的run函数。

void DrawFrameTask::run() {
    ATRACE_NAME("DrawFrame");

    bool canUnblockUiThread;
    bool canDrawThisFrame;
    {
        TreeInfo info(TreeInfo::MODE_FULL, *mContext);
        canUnblockUiThread = syncFrameState(info);
        canDrawThisFrame = info.out.canDrawThisFrame;
    }

    // Grab a copy of everything we need
    CanvasContext* context = mContext;

    // From this point on anything in "this" is *UNSAFE TO ACCESS*
    if (canUnblockUiThread) {
        unblockUiThread();
    }

    if (CC_LIKELY(canDrawThisFrame)) {
        context->draw();
    } else {
        // wait on fences so tasks don't overlap next frame
        context->waitOnFences();
    }

    if (!canUnblockUiThread) {
        unblockUiThread();
    }
}
  • 先调用syncFrameState,同步一下Frame的状态
  • 再通过CanvasContext的draw方法去绘制

OK,现在,主要的流程就到CanvasContext,我们看看CanvasContext

CanvasContext

渲染的上下文。

* frameworks/base/libs/hwui/renderthread/CanvasContext.cpp

CanvasContext* CanvasContext::create(RenderThread& thread, bool translucent,
                                     RenderNode* rootRenderNode, IContextFactory* contextFactory) {
    auto renderType = Properties::getRenderPipelineType();

    switch (renderType) {
        case RenderPipelineType::OpenGL:
            return new CanvasContext(thread, translucent, rootRenderNode, contextFactory,
                                     std::make_unique<OpenGLPipeline>(thread));
        case RenderPipelineType::SkiaGL:
            return new CanvasContext(thread, translucent, rootRenderNode, contextFactory,
                                     std::make_unique<skiapipeline::SkiaOpenGLPipeline>(thread));
        case RenderPipelineType::SkiaVulkan:
            return new CanvasContext(thread, translucent, rootRenderNode, contextFactory,
                                     std::make_unique<skiapipeline::SkiaVulkanPipeline>(thread));
        default:
            LOG_ALWAYS_FATAL("canvas context type %d not supported", (int32_t)renderType);
            break;
    }
    return nullptr;
}

前面我们已经说过,渲染Pipeline有几种类型,Pipeline由IRenderPipeline描述。创建CanvasContext时,会根据pipeline的类型,创建对应的Pipeline,他们的关系如下:


HWUI处理管线Pipeline

IRenderPipeline是统一的接口。默认的类型是OpenGLPipeline,用的是OpenGL实现。这可以可通过属性debug.hwui.renderer来设置。对应地逻辑如下:

* frameworks/base/libs/hwui/Properties.cpp
#define PROPERTY_RENDERER "debug.hwui.renderer"

RenderPipelineType Properties::getRenderPipelineType() {
    if (sRenderPipelineType != RenderPipelineType::NotInitialized) {
        return sRenderPipelineType;
    }
    char prop[PROPERTY_VALUE_MAX];
    property_get(PROPERTY_RENDERER, prop, "skiagl");
    if (!strcmp(prop, "skiagl")) {
        ALOGD("Skia GL Pipeline");
        sRenderPipelineType = RenderPipelineType::SkiaGL;
    } else if (!strcmp(prop, "skiavk")) {
        ALOGD("Skia Vulkan Pipeline");
        sRenderPipelineType = RenderPipelineType::SkiaVulkan;
    } else {  //"opengl"
        ALOGD("HWUI GL Pipeline");
        sRenderPipelineType = RenderPipelineType::OpenGL;
    }
    return sRenderPipelineType;
}

SkiaOpenGLPipeline和SkiaVulkanPipeline,两者都用到skia进行Ops的渲染,也就是说,Ops的录制是用skia来完成的。后面的显示才用到OpenGL或Vulkan。

我们再来看一下CanvasContext的构造函数:

CanvasContext::CanvasContext(RenderThread& thread, bool translucent, RenderNode* rootRenderNode,
                             IContextFactory* contextFactory,
                             std::unique_ptr<IRenderPipeline> renderPipeline)
        : mRenderThread(thread)
        , mOpaque(!translucent)
        , mAnimationContext(contextFactory->createAnimationContext(mRenderThread.timeLord()))
        , mJankTracker(&thread.globalProfileData(), thread.mainDisplayInfo())
        , mProfiler(mJankTracker.frames())
        , mContentDrawBounds(0, 0, 0, 0)
        , mRenderPipeline(std::move(renderPipeline)) {
    rootRenderNode->makeRoot();
    mRenderNodes.emplace_back(rootRenderNode);
    mRenderThread.renderState().registerCanvasContext(this);
    mProfiler.setDensity(mRenderThread.mainDisplayInfo().density);
}
  • contextFactory
    contextFactory是在Surface的JNI中创建RenderProxy时,传入的。主要是用来创建AnimationContext,AnimationContext主要用来处理动画Animation。
* frameworks/base/core/jni/android_view_Surface.cpp

class ContextFactory : public IContextFactory {
public:
    virtual AnimationContext* createAnimationContext(renderthread::TimeLord& clock) {
        return new AnimationContext(clock);
    }
};
  • rootRenderNod,rootRenderNode前面在做Ops录制时的RenderNode。这里通过makeRoot,将其设置为Root的RenderNode。它是mRenderNodes中的第一个RenderNode。

  • CanvasContext实现了IFrameCallback接口,所以,CanvasContext能接收编舞者Choreographer的callback,处理实时动画。

我们再回过头看DrawFrameTask的run。首先是syncFrameState处理,同步Frame的State:

bool DrawFrameTask::syncFrameState(TreeInfo& info) {
    ATRACE_CALL();
    int64_t vsync = mFrameInfo[static_cast<int>(FrameInfoIndex::Vsync)];
    mRenderThread->timeLord().vsyncReceived(vsync);
    bool canDraw = mContext->makeCurrent();
    mContext->unpinImages();

    for (size_t i = 0; i < mLayers.size(); i++) {
        mLayers[i]->apply();
    }
    mLayers.clear();
    mContext->setContentDrawBounds(mContentDrawBounds);
    mContext->prepareTree(info, mFrameInfo, mSyncQueued, mTargetNode);

    // This is after the prepareTree so that any pending operations
    // (RenderNode tree state, prefetched layers, etc...) will be flushed.
    if (CC_UNLIKELY(!mContext->hasSurface() || !canDraw)) {
        if (!mContext->hasSurface()) {
            mSyncResult |= SyncResult::LostSurfaceRewardIfFound;
        } else {
            // If we have a surface but can't draw we must be stopped
            mSyncResult |= SyncResult::ContextIsStopped;
        }
        info.out.canDrawThisFrame = false;
    }

    if (info.out.hasAnimations) {
        if (info.out.requiresUiRedraw) {
            mSyncResult |= SyncResult::UIRedrawRequired;
        }
    }
    // If prepareTextures is false, we ran out of texture cache space
    return info.prepareTextures;
}
  • makeCurrent,这个从早期的版本就有,早期只有Opengl pipeline时,Opengl只支持单线程。我们首先要通过makeCurrent,告诉GPU处理当前的上下文(context)。
  • unpinImages,hwui为了提高速度,对各种object都做了cache,这里的unpin,就是让cache去做unpin,以前的都不要了。
  • setContentDrawBounds,设置绘制的区域大小
  • prepareTree,前面我们也说过,Android View是树型结构的,这就是在绘制之前,去准备这些Tree节点的绘图操作Ops。这个过程也是非常的复杂。

回到run函数,syncFrameState后,如果,可以绘制,也就是存在更新。直接让CanvasContext去绘制了。

CanvasContext的draw是在RenderPipeline中完成的。而Ops的渲染则是通过BakedOpRenderer完成。默认用的是OpenGLPipeline,简单的来看,这段流程。


DrawFrame时序图

其中就两个主要的流程:PrepareTree和Draw。在流程图上,只是标记了一下,没有仔细的画。下面的我们来看看,这里都做了什么,我们的界面是怎么画出来的。

Node Tree的准备

离开我们的测试应用代码很久了,回来测试的代码。此时,RenderThread,DrawFrameTask,CanvasContext等已经就绪,绘制操纵已经被添加到了DisplayList中。

那么DisplayList,是怎么到CanvasContext中进行绘制的呢?

我们接着来看测试代码,接下来,就是Surface的unlock和post操纵。

mSurface.unlockCanvasAndPost(canvas);

SurfaceHolder直接调的Surface的unlockCanvasAndPost。

        @Override
        public void unlockCanvasAndPost(Canvas canvas) {
            mSurface.unlockCanvasAndPost(canvas);
            mSurfaceLock.unlock();
        }

由于我们采用的hardware Context,走的HwuiContext的分支。

    public void unlockCanvasAndPost(Canvas canvas) {
        synchronized (mLock) {
            checkNotReleasedLocked();

            if (mHwuiContext != null) {
                mHwuiContext.unlockAndPost(canvas);
            } else {
                unlockSwCanvasAndPost(canvas);
            }
        }
    }

HwuiContext的unlockAndPost函数如下:

        void unlockAndPost(Canvas canvas) {
            if (canvas != mCanvas) {
                throw new IllegalArgumentException("canvas object must be the same instance that "
                        + "was previously returned by lockCanvas");
            }
            mRenderNode.end(mCanvas);
            mCanvas = null;
            nHwuiDraw(mHwuiRenderer);
        }

我们在lockCanvas时,mRenderNode.start,unlock时,调的mRenderNode.end。

Node结束时,先结束Canvas的录制,然后将录制的List,给到RenderNode。

    public void end(DisplayListCanvas canvas) {
        long displayList = canvas.finishRecording();
        nSetDisplayList(mNativeRenderNode, displayList);
        canvas.recycle();
    }

记住,Canvas录制的List,给到了RenderNode。这很重要。

finishRecording,我们直接看最后native的实现。

DisplayList* RecordingCanvas::finishRecording() {
    restoreToCount(1);
    mPaintMap.clear();
    mRegionMap.clear();
    mPathMap.clear();
    DisplayList* displayList = mDisplayList;
    mDisplayList = nullptr;
    mSkiaCanvasProxy.reset(nullptr);
    return displayList;
}

返回的就是前面我们已经录制好的mDisplayList。

录制好的DisplayList,最后给到哪儿呢?

nSetDisplayListJNI实现如下:

static void android_view_RenderNode_setDisplayList(JNIEnv* env,
        jobject clazz, jlong renderNodePtr, jlong displayListPtr) {
    RenderNode* renderNode = reinterpret_cast<RenderNode*>(renderNodePtr);
    DisplayList* newData = reinterpret_cast<DisplayList*>(displayListPtr);
    renderNode->setStagingDisplayList(newData);
}

JNI再通过setStagingDisplayList,给到RenderNode的mStagingDisplayList

void RenderNode::setStagingDisplayList(DisplayList* displayList) {
    mValid = (displayList != nullptr);
    mNeedsDisplayListSync = true;
    delete mStagingDisplayList;
    mStagingDisplayList = displayList;
}

到此,录制的Ops,是不是都给到RenderNode的mStagingDisplayList了。

现在,我们可以来看CanvasContext的PrepareTree了。

* frameworks/base/libs/hwui/renderthread/CanvasContext.cpp

void CanvasContext::prepareTree(TreeInfo& info, int64_t* uiFrameInfo, int64_t syncQueued,
                                RenderNode* target) {
    mRenderThread.removeFrameCallback(this);

    ... ... //处理frame信息

    info.damageAccumulator = &mDamageAccumulator;
    info.layerUpdateQueue = &mLayerUpdateQueue;

    mAnimationContext->startFrame(info.mode);
    mRenderPipeline->onPrepareTree();
    for (const sp<RenderNode>& node : mRenderNodes) {
        // 只有Primary的node是 FULL,其他都是实时
        info.mode = (node.get() == target ? TreeInfo::MODE_FULL : TreeInfo::MODE_RT_ONLY);
        node->prepareTree(info);
        GL_CHECKPOINT(MODERATE);
    }
    mAnimationContext->runRemainingAnimations(info);
    GL_CHECKPOINT(MODERATE);

    freePrefetchedLayers();
    GL_CHECKPOINT(MODERATE);

    mIsDirty = true;

    // 如果,窗口已经没有Native Surface,这一帧就丢掉。
    if (CC_UNLIKELY(!mNativeSurface.get())) {
        mCurrentFrameInfo->addFlag(FrameInfoFlags::SkippedFrame);
        info.out.canDrawThisFrame = false;
        return;
    }

    ... ...
}

第一个问题,info是什么,从哪儿来的?从DrawFrameTask中来的。

void DrawFrameTask::run() {
    ATRACE_NAME("DrawFrame");

    bool canUnblockUiThread;
    bool canDrawThisFrame;
    {
        TreeInfo info(TreeInfo::MODE_FULL, *mContext);
        canUnblockUiThread = syncFrameState(info);
        canDrawThisFrame = info.out.canDrawThisFrame;
    }

TreeInfo顾名思义,描述Viewtree的,也就是RenderNode tree。

    TreeInfo(TraversalMode mode, renderthread::CanvasContext& canvasContext)
            : mode(mode), prepareTextures(mode == MODE_FULL), canvasContext(canvasContext) {}

注意这里的mode为TreeInfo::MODE_FULL。只有Primary的node是 FULL,其他都是实时。

Context可能会有多个Node,每个Node都进行Prepare。

* frameworks/base/libs/hwui/RenderNode.cpp

void RenderNode::prepareTree(TreeInfo& info) {
    ATRACE_CALL();
    LOG_ALWAYS_FATAL_IF(!info.damageAccumulator, "DamageAccumulator missing");
    MarkAndSweepRemoved observer(&info);

    // The OpenGL renderer reserves the stencil buffer for overdraw debugging.  Functors
    // will need to be drawn in a layer.
    bool functorsNeedLayer = Properties::debugOverdraw && !Properties::isSkiaEnabled();

    prepareTreeImpl(observer, info, functorsNeedLayer);
}

在RenderNode进行Prepare时,先对TreeInfo进行封,MarkAndSweepRemoved,主要是对可能的Node进行标记,删除。MarkAndSweepRemoved的代码如下:

class MarkAndSweepRemoved : public TreeObserver {
    PREVENT_COPY_AND_ASSIGN(MarkAndSweepRemoved);

public:
    explicit MarkAndSweepRemoved(TreeInfo* info) : mTreeInfo(info) {}

    void onMaybeRemovedFromTree(RenderNode* node) override { mMarked.emplace_back(node); }

    ~MarkAndSweepRemoved() {
        for (auto& node : mMarked) {
            if (!node->hasParents()) {
                node->onRemovedFromTree(mTreeInfo);
            }
        }
    }

private:
    FatVector<sp<RenderNode>, 10> mMarked;
    TreeInfo* mTreeInfo;
};

能从tree上删除的就添加到mMarked中,在析构函数中,再对mMarked的mode进行删除。

prepareTreeImpl是RenderNode真正进行Prepare的地方。

void RenderNode::prepareTreeImpl(TreeObserver& observer, TreeInfo& info, bool functorsNeedLayer) {
    info.damageAccumulator->pushTransform(this);

    if (info.mode == TreeInfo::MODE_FULL) {
        pushStagingPropertiesChanges(info);
    }
    uint32_t animatorDirtyMask = 0;
    if (CC_LIKELY(info.runAnimations)) {
        animatorDirtyMask = mAnimatorManager.animate(info);
    }

    bool willHaveFunctor = false;
    if (info.mode == TreeInfo::MODE_FULL && mStagingDisplayList) {
        willHaveFunctor = mStagingDisplayList->hasFunctor();
    } else if (mDisplayList) {
        willHaveFunctor = mDisplayList->hasFunctor();
    }
    bool childFunctorsNeedLayer =
            mProperties.prepareForFunctorPresence(willHaveFunctor, functorsNeedLayer);

    if (CC_UNLIKELY(mPositionListener.get())) {
        mPositionListener->onPositionUpdated(*this, info);
    }

    prepareLayer(info, animatorDirtyMask);
    if (info.mode == TreeInfo::MODE_FULL) {
        pushStagingDisplayListChanges(observer, info);
    }

    if (mDisplayList) {
        info.out.hasFunctors |= mDisplayList->hasFunctor();
        bool isDirty = mDisplayList->prepareListAndChildren(
                observer, info, childFunctorsNeedLayer,
                [](RenderNode* child, TreeObserver& observer, TreeInfo& info,
                   bool functorsNeedLayer) {
                    child->prepareTreeImpl(observer, info, functorsNeedLayer);
                });
        if (isDirty) {
            damageSelf(info);
        }
    }
    pushLayerUpdate(info);

    info.damageAccumulator->popTransform();
}

damageAccumulator是从CanvasContext中传过来的,是CanvasContext的成员,damage的累乘器。主要是用来标记,屏幕的那些区域被破坏了,需要重新绘制,所有的RenderNode累加起来,就是总的。

我们来看一眼pushTransform。

void DamageAccumulator::pushCommon() {
    if (!mHead->next) {
        DirtyStack* nextFrame = mAllocator.create_trivial<DirtyStack>();
        nextFrame->next = nullptr;
        nextFrame->prev = mHead;
        mHead->next = nextFrame;
    }
    mHead = mHead->next;
    mHead->pendingDirty.setEmpty();
}

void DamageAccumulator::pushTransform(const RenderNode* transform) {
    pushCommon();
    mHead->type = TransformRenderNode;
    mHead->renderNode = transform;
}

damage累加器中,每一个元素由DirtyStack描述,分两种类型:TransformMatrix4和TransformRenderNode。采用一个双向链表mHead进行管理。

pushStagingPropertiesChanges,property是对RenderNode的描述,也就是对View的描述,比如大小,位置等。有两个状态,正在使用的syncProperties和待处理的mStagingProperties。syncProperties时,将mStagingProperties赋值给syncProperties。这里,很多状态都是这样同步的。

pushStagingDisplayListChanges,和前面的Property一样的流程,只是这里是syncDisplayList。这样,前面录制好Ops,就通过mStagingDisplayList传给mDisplayList。

绘制的Ops都放在mDisplayList中,这边会去递归的调用每个RenderNode的prepareTreeImpl。

pushLayerUpdate,将要更新的RenderNode都加到TreeInfo的layerUpdateQueue中,还有其对应的damage大小。

累加器的popTransform,就是将该Node的DirtyStack生效。

Prepare完成,代码量还是非常多的,我们主要关心我们的数据流。DisplayList的数据,不是更新到了Context的mLayerUpdateQueue中?

绘制

CanvasContext Prepare完后,绘制一帧的数据就准备好了。绘制是在各自的pipeline中进行的。OpenGLPipeline的绘制流程如下:

bool OpenGLPipeline::draw(const Frame& frame, const SkRect& screenDirty, const SkRect& dirty,
                          const FrameBuilder::LightGeometry& lightGeometry,
                          LayerUpdateQueue* layerUpdateQueue, const Rect& contentDrawBounds,
                          bool opaque, bool wideColorGamut,
                          const BakedOpRenderer::LightInfo& lightInfo,
                          const std::vector<sp<RenderNode>>& renderNodes,
                          FrameInfoVisualizer* profiler) {
    mEglManager.damageFrame(frame, dirty);

    bool drew = false;

    auto& caches = Caches::getInstance();
    FrameBuilder frameBuilder(dirty, frame.width(), frame.height(), lightGeometry, caches);

    frameBuilder.deferLayers(*layerUpdateQueue);
    layerUpdateQueue->clear();

    frameBuilder.deferRenderNodeScene(renderNodes, contentDrawBounds);

    BakedOpRenderer renderer(caches, mRenderThread.renderState(), opaque, wideColorGamut,
                             lightInfo);
    frameBuilder.replayBakedOps<BakedOpDispatcher>(renderer);
    ProfileRenderer profileRenderer(renderer);
    profiler->draw(profileRenderer);
    drew = renderer.didDraw();

    // post frame cleanup
    caches.clearGarbage();
    caches.pathCache.trim();
    caches.tessellationCache.trim();

#if DEBUG_MEMORY_USAGE
    caches.dumpMemoryUsage();
#else
    if (CC_UNLIKELY(Properties::debugLevel & kDebugMemory)) {
        caches.dumpMemoryUsage();
    }
#endif

    return drew;
}

Frame是描述一帧数据信息的,主要是宽,高,ufferAge,和Surface这几个属性。绘制开始时,由EglManager根据Surface的属性构建。

Frame EglManager::beginFrame(EGLSurface surface) {
    LOG_ALWAYS_FATAL_IF(surface == EGL_NO_SURFACE, "Tried to beginFrame on EGL_NO_SURFACE!");
    makeCurrent(surface);
    Frame frame;
    frame.mSurface = surface;
    eglQuerySurface(mEglDisplay, surface, EGL_WIDTH, &frame.mWidth);
    eglQuerySurface(mEglDisplay, surface, EGL_HEIGHT, &frame.mHeight);
    frame.mBufferAge = queryBufferAge(surface);
    eglBeginFrame(mEglDisplay, surface);
    return frame;
}

damageFrame主要是部分更新参数的设置,前面我们也damage的区域就是前面Prepare时累加器累加出来的。

FrameBuilder,用来创建一帧Frame,继承CanvasStateClient。

FrameBuilder::FrameBuilder(const SkRect& clip, uint32_t viewportWidth, uint32_t viewportHeight,
                           const LightGeometry& lightGeometry, Caches& caches)
        : mStdAllocator(mAllocator)
        , mLayerBuilders(mStdAllocator)
        , mLayerStack(mStdAllocator)
        , mCanvasState(*this)
        , mCaches(caches)
        , mLightRadius(lightGeometry.radius)
        , mDrawFbo0(true) {
    // Prepare to defer Fbo0
    auto fbo0 = mAllocator.create<LayerBuilder>(viewportWidth, viewportHeight, Rect(clip));
    mLayerBuilders.push_back(fbo0);
    mLayerStack.push_back(0);
    mCanvasState.initializeSaveStack(viewportWidth, viewportHeight, clip.fLeft, clip.fTop,
                                     clip.fRight, clip.fBottom, lightGeometry.center);
}

FrameBuilder创建一个LayerBuilder的List来记录Rendernode的绘制状态,然后以倒序的方式去replay录制的RenderNode。

deferLayers主要是做了一个倒序,所有的RenderNode进行倒序,RenderNode的Ops也进行倒序。

void FrameBuilder::deferLayers(const LayerUpdateQueue& layers) {
    // Render all layers to be updated, in order. Defer in reverse order, so that they'll be
    // updated in the order they're passed in (mLayerBuilders are issued to Renderer in reverse)
    for (int i = layers.entries().size() - 1; i >= 0; i--) {
        RenderNode* layerNode = layers.entries()[i].renderNode.get();
        // only schedule repaint if node still on layer - possible it may have been
        // removed during a dropped frame, but layers may still remain scheduled so
        // as not to lose info on what portion is damaged
        OffscreenBuffer* layer = layerNode->getLayer();
        if (CC_LIKELY(layer)) {
            ATRACE_FORMAT("Optimize HW Layer DisplayList %s %ux%u", layerNode->getName(),
                          layerNode->getWidth(), layerNode->getHeight());

            Rect layerDamage = layers.entries()[i].damage;
            // TODO: ensure layer damage can't be larger than layer
            layerDamage.doIntersect(0, 0, layer->viewportWidth, layer->viewportHeight);
            layerNode->computeOrdering();

            // map current light center into RenderNode's coordinate space
            Vector3 lightCenter = mCanvasState.currentSnapshot()->getRelativeLightCenter();
            layer->inverseTransformInWindow.mapPoint3d(lightCenter);

            saveForLayer(layerNode->getWidth(), layerNode->getHeight(), 0, 0, layerDamage,
                         lightCenter, nullptr, layerNode);

            if (layerNode->getDisplayList()) {
                deferNodeOps(*layerNode);
            }
            restoreForLayer();
        }
    }
}

倒序的目的,其实就是解决谁先画,谁后画的问题。Node都是Tree结构,如果子tree先绘制,父tree后绘制,这样后绘制的就会将前面绘制的遮盖住,看不见了。注意我们的数据流,倒序后的Layer放在mLayerBuilders中。

BakedOpRenderer是渲染器Renderer。它是主要的渲染管理者,用以管理渲染的任务集合,比如一帧数据,和包含的FBO。管理着他们的生命周期,绑定FrameBuffer。这是FBO创建,销毁等的唯一的地方。而所有的渲染操纵都是通过Dispatcher进行传递。

    BakedOpRenderer(Caches& caches, RenderState& renderState, bool opaque, bool wideColorGamut,
                    const LightInfo& lightInfo)
            : mGlopReceiver(DefaultGlopReceiver)
            , mRenderState(renderState)
            , mCaches(caches)
            , mOpaque(opaque)
            , mWideColorGamut(wideColorGamut)
            , mLightInfo(lightInfo) {}

mGlopReceiver是一个函数指针,默认为DefaultGlopReceiver。

    static void DefaultGlopReceiver(BakedOpRenderer& renderer, const Rect* dirtyBounds,
                                    const ClipBase* clip, const Glop& glop) {
        renderer.renderGlopImpl(dirtyBounds, clip, glop);
    }

replayBakedOps是一个模板函数,这样就可以自由决定录制Ops被replay的地方。它包含一个lambdas数组,通过这个数组,replay时,,录制的BakeOpState就能够通过state->op->opId找到对应的接收者进行replay。

replayBakedOps函数实现如下:

    template <typename StaticDispatcher, typename Renderer>
    void replayBakedOps(Renderer& renderer) {
        std::vector<OffscreenBuffer*> temporaryLayers;
        finishDefer();

#define X(Type)                                                                   \
    [](void* renderer, const BakedOpState& state) {                               \
        StaticDispatcher::on##Type(*(static_cast<Renderer*>(renderer)),           \
                                   static_cast<const Type&>(*(state.op)), state); \
    },
        static BakedOpReceiver unmergedReceivers[] = BUILD_RENDERABLE_OP_LUT(X);
#undef X

#define X(Type)                                                                           \
    [](void* renderer, const MergedBakedOpList& opList) {                                 \
        StaticDispatcher::onMerged##Type##s(*(static_cast<Renderer*>(renderer)), opList); \
    },
        static MergedOpReceiver mergedReceivers[] = BUILD_MERGEABLE_OP_LUT(X);
#undef X

        // Relay through layers in reverse order, since layers
        // later in the list will be drawn by earlier ones
        for (int i = mLayerBuilders.size() - 1; i >= 1; i--) {
            GL_CHECKPOINT(MODERATE);
            LayerBuilder& layer = *(mLayerBuilders[i]);
            if (layer.renderNode) {
                // cached HW layer - can't skip layer if empty
                renderer.startRepaintLayer(layer.offscreenBuffer, layer.repaintRect);
                GL_CHECKPOINT(MODERATE);
                layer.replayBakedOpsImpl((void*)&renderer, unmergedReceivers, mergedReceivers);
                GL_CHECKPOINT(MODERATE);
                renderer.endLayer();
            } else if (!layer.empty()) {
                // save layer - skip entire layer if empty (in which case, LayerOp has null layer).
                layer.offscreenBuffer = renderer.startTemporaryLayer(layer.width, layer.height);
                temporaryLayers.push_back(layer.offscreenBuffer);
                GL_CHECKPOINT(MODERATE);
                layer.replayBakedOpsImpl((void*)&renderer, unmergedReceivers, mergedReceivers);
                GL_CHECKPOINT(MODERATE);
                renderer.endLayer();
            }
        }

        GL_CHECKPOINT(MODERATE);
        if (CC_LIKELY(mDrawFbo0)) {
            const LayerBuilder& fbo0 = *(mLayerBuilders[0]);
            renderer.startFrame(fbo0.width, fbo0.height, fbo0.repaintRect);
            GL_CHECKPOINT(MODERATE);
            fbo0.replayBakedOpsImpl((void*)&renderer, unmergedReceivers, mergedReceivers);
            GL_CHECKPOINT(MODERATE);
            renderer.endFrame(fbo0.repaintRect);
        }

        for (auto& temporaryLayer : temporaryLayers) {
            renderer.recycleTemporaryLayer(temporaryLayer);
        }
    }

这个表和前面我们在录制的流程中说的LUT就对应起来了,unmergedReceivers和mergedReceivers分别和对应的LUT表对应。比如我们的ColorOp,就调的BakedOpDispatcher::onColorOp。另外要注意的是,我们的drawColor是从fbo0这里调的。

void BakedOpDispatcher::onColorOp(BakedOpRenderer& renderer, const ColorOp& op,
                                  const BakedOpState& state) {
    SkPaint paint;
    paint.setColor(op.color);
    paint.setBlendMode(op.mode);

    Glop glop;
    GlopBuilder(renderer.renderState(), renderer.caches(), &glop)
            .setRoundRectClipState(state.roundRectClipState)
            .setMeshUnitQuad()
            .setFillPaint(paint, state.alpha)
            .setTransform(Matrix4::identity(), TransformFlags::None)
            .setModelViewMapUnitToRect(state.computedState.clipState->rect)
            .build();
    renderer.renderGlop(state, glop);
}

我们需要绘制的color值,直接设置到画笔paint,blend模式也设置到paint。

这部分的逻辑在LayerBuilder的replayBakedOpsImpl函数中。

void LayerBuilder::replayBakedOpsImpl(void* arg, BakedOpReceiver* unmergedReceivers,
                                      MergedOpReceiver* mergedReceivers) const {
    if (renderNode) {
        ATRACE_FORMAT_BEGIN("Issue HW Layer DisplayList %s %ux%u", renderNode->getName(), width,
                            height);
    } else {
        ATRACE_BEGIN("flush drawing commands");
    }

    for (const BatchBase* batch : mBatches) {
        size_t size = batch->getOps().size();
        if (size > 1 && batch->isMerging()) {
            int opId = batch->getOps()[0]->op->opId;
            const MergingOpBatch* mergingBatch = static_cast<const MergingOpBatch*>(batch);
            MergedBakedOpList data = {batch->getOps().data(), size,
                                      mergingBatch->getClipSideFlags(),
                                      mergingBatch->getClipRect()};
            mergedReceivers[opId](arg, data);
        } else {
            for (const BakedOpState* op : batch->getOps()) {
                unmergedReceivers[op->op->opId](arg, *op);
            }
        }
    }
    ATRACE_END();
}

我们的drawcolor是从unmergedReceivers调的!

代码写的确实复杂,得慢慢的看,看明白后,有以后就可以跳过这一块的逻辑了,直接去看Ops绘制的地方~

渲染Ops的时,又被封装了一次,都被封装成Glop。Glop由GlopBuilder统一构建。构建完后,由renderGlop进行渲染。

    void renderGlop(const BakedOpState& state, const Glop& glop) {
        renderGlop(&state.computedState.clippedBounds, state.computedState.getClipIfNeeded(), glop);
    }

    void renderGlop(const Rect* dirtyBounds, const ClipBase* clip, const Glop& glop) {
        mGlopReceiver(*this, dirtyBounds, clip, glop);
    }

mGlopReceiver是一个函数指针,指向的是DefaultGlopReceiver。封装一下,最后的实现为BakedOpRenderer的renderGlopImpl。

renderGlopImpl函数如下:

void BakedOpRenderer::renderGlopImpl(const Rect* dirtyBounds, const ClipBase* clip,
                                     const Glop& glop) {
    prepareRender(dirtyBounds, clip);
    // Disable blending if this is the first draw to the main framebuffer, in case app has defined
    // transparency where it doesn't make sense - as first draw in opaque window. Note that we only
    // apply this improvement when the blend mode is SRC_OVER - other modes (e.g. CLEAR) can be
    // valid draws that affect other content (e.g. draw CLEAR, then draw DST_OVER)
    bool overrideDisableBlending = !mHasDrawn && mOpaque && !mRenderTarget.frameBufferId &&
                                   glop.blend.src == GL_ONE &&
                                   glop.blend.dst == GL_ONE_MINUS_SRC_ALPHA;
    mRenderState.render(glop, mRenderTarget.orthoMatrix, overrideDisableBlending);
    if (!mRenderTarget.frameBufferId) mHasDrawn = true;
}

在renderGlopImpl中,准备了一个Render,最终是通过mRenderState的render进行渲染。在RenderState的render中,直接调用OpenGLES的接口,需绘制我们的Ops了。具体怎么绘制的,就是OpenGL的问题了,这里就不看了,交给OpenGL去吧。

waitOnFences等待所有的task已经绘制完成,这里的fence和BufferQueue那边的Fence不是同一个概念。绘制完后,通过swapBuffers函数,交换buffer,将绘制完的数据送去显示。

另外,hwui中还做了很多Jank的跟踪,便于debug性能

小结

测试代码才几行,底层却折腾了这么多,我们来总结一下:

  • 硬件绘制,或硬件加速,就是通过hwui,将2D的绘图操纵转换为3D的绘图
  • 每一个绘制采用一个RecordedOp进行描述,复杂的绘图将被拆分成简单的基本绘图,并利用RecordingCanvas进行录制。
  • 每个View都对应RenderNode,而每个界面有一个DisplayList,用以保存录制的Ops。
  • 每个进程只有一个RenderThread,所有的绘图都在RenderThread中完成,因此,其他线程的操纵都通过Task或WorkItem的形式post到RenderThread中完成。DrawFrameTask是RenderThread中比较特殊的一个task,是用以绘制整个界面的,跟随Vync而触发。
  • OpenGL是单线程的,所以每个RenderThread都有各自的上下文,CanvasContext,通过Preparetree,将DisplayList中Ops都同步到CanvasContext的layerUpdateQueue中,准备好绘制帧的数据。
  • 绘制是由具体的Pipeline完成的,目前有3中类型的Pipeline,OpenGLPipeline是默认的Pipeline。
  • OpenGLPipeline绘制时,通过FrameBuilder和LayerBuilder,将DisplayList的数据进一步封装。在replayBakedOps时,将Opo的操纵转换为具体的绘制操纵,通过BakedOpDispatcher分发给BakedOpRenderer进行渲染。而真正的渲染是在mRenderState完成,直接调用OpenGL的接口。

这中间,只要抓住数据流,Ops和DisplayList,这条主线,理解起来就轻松些。总的来说,可以分为以下几个部分,我们用一张总体的图来描述:


HWUI相关类图
  • Recording部分,这部分主要是2D到3D的转换,录制绘图操纵Ops
  • Draw 控制部分,这部分主要和上层应用和显示系统同步,控制绘制的进行,包括动画的处理
  • Draw的执行部分,这部分主要和具体的加速系统交互,采用具体的加速API进行界面的绘制

以上就是结合测试代码,讲解的hwui的具体内容。

推荐阅读更多精彩内容