Merge pull request #8322 from jherico/deferred_batches

Deferred batches

interface/src/Application.cpp

@@ -1689,6 +1689,17 @@ void Application::paintGL() {
         renderArgs._context->syncCache();
     }
 
+    auto framebufferCache = DependencyManager::get<FramebufferCache>();
+    // Final framebuffer that will be handled to the display-plugin
+    auto finalFramebuffer = framebufferCache->getFramebuffer();
+
+    _gpuContext->beginFrame(finalFramebuffer, getHMDSensorPose());
+    // Reset the gpu::Context Stages
+    // Back to the default framebuffer;
+    gpu::doInBatch(_gpuContext, [&](gpu::Batch& batch) {
+        batch.resetStages();
+    });
+
     auto inputs = AvatarInputs::getInstance();
     if (inputs->mirrorVisible()) {
         PerformanceTimer perfTimer("Mirror");
@@ -1711,10 +1722,6 @@ void Application::paintGL() {
         QSize size = getDeviceSize();
         renderArgs._viewport = glm::ivec4(0, 0, size.width(), size.height());
         _applicationOverlay.renderOverlay(&renderArgs);
-        auto overlayTexture = _applicationOverlay.acquireOverlay();
-        if (overlayTexture) {
-            displayPlugin->submitOverlayTexture(overlayTexture);
-        }
     }
 
     glm::vec3 boomOffset;
@@ -1816,12 +1823,8 @@ void Application::paintGL() {
     getApplicationCompositor().setFrameInfo(_frameCount, _myCamera.getTransform());
 
     // Primary rendering pass
-    auto framebufferCache = DependencyManager::get<FramebufferCache>();
     const QSize size = framebufferCache->getFrameBufferSize();
 
-    // Final framebuffer that will be handled to the display-plugin
-    auto finalFramebuffer = framebufferCache->getFramebuffer();
-
     {
         PROFILE_RANGE(__FUNCTION__ "/mainRender");
         PerformanceTimer perfTimer("mainRender");
@@ -1880,6 +1883,13 @@ void Application::paintGL() {
         renderArgs._context->enableStereo(false);
     }
 
+    _gpuContext->endFrame();
+
+    gpu::TexturePointer overlayTexture = _applicationOverlay.acquireOverlay();
+    if (overlayTexture) {
+        displayPlugin->submitOverlayTexture(overlayTexture);
+    }
+
     // deliver final composited scene to the display plugin
     {
         PROFILE_RANGE(__FUNCTION__ "/pluginOutput");
@@ -1900,11 +1910,6 @@ void Application::paintGL() {
 
     {
         Stats::getInstance()->setRenderDetails(renderArgs._details);
-        // Reset the gpu::Context Stages
-        // Back to the default framebuffer;
-        gpu::doInBatch(renderArgs._context, [&](gpu::Batch& batch) {
-            batch.resetStages();
-        });
     }
 
     uint64_t lastPaintDuration = usecTimestampNow() - lastPaintBegin;

libraries/gpu/src/gpu/Batch.cpp

@@ -46,6 +46,33 @@ Batch::Batch() {
     _drawCallInfos.reserve(_drawCallInfosMax);
 }
 
+Batch::Batch(const Batch& batch_) {
+    Batch& batch = *const_cast<Batch*>(&batch_);
+    _commands.swap(batch._commands);
+    _commandOffsets.swap(batch._commandOffsets);
+    _params.swap(batch._params);
+    _data.swap(batch._data);
+    _invalidModel = batch._invalidModel;
+    _currentModel = batch._currentModel;
+    _objects.swap(batch._objects);
+    _currentNamedCall = batch._currentNamedCall;
+
+    _buffers._items.swap(batch._buffers._items);
+    _textures._items.swap(batch._textures._items);
+    _streamFormats._items.swap(batch._streamFormats._items);
+    _transforms._items.swap(batch._transforms._items);
+    _pipelines._items.swap(batch._pipelines._items);
+    _framebuffers._items.swap(batch._framebuffers._items);
+    _drawCallInfos.swap(batch._drawCallInfos);
+    _queries._items.swap(batch._queries._items);
+    _lambdas._items.swap(batch._lambdas._items);
+    _profileRanges._items.swap(batch._profileRanges._items);
+    _names._items.swap(batch._names._items);
+    _namedData.swap(batch._namedData);
+    _enableStereo = batch._enableStereo;
+    _enableSkybox = batch._enableSkybox;
+}
+
 Batch::~Batch() {
     _commandsMax = std::max(_commands.size(), _commandsMax);
     _commandOffsetsMax = std::max(_commandOffsets.size(), _commandOffsetsMax);

libraries/gpu/src/gpu/Context.cpp

@@ -9,7 +9,7 @@
 //  See the accompanying file LICENSE or http://www.apache.org/licenses/LICENSE-2.0.html
 //
 #include "Context.h"
-
+#include "Frame.h"
 using namespace gpu;
 
 Context::CreateBackend Context::_createBackendCallback = nullptr;
@@ -20,6 +20,13 @@ Context::Context() {
     if (_createBackendCallback) {
         _backend.reset(_createBackendCallback());
     }
+
+    _frameHandler = [this](Frame& frame){
+        for (size_t i = 0; i < frame.batches.size(); ++i) {
+            _backend->_stereo = frame.stereoStates[i];
+            _backend->render(frame.batches[i]);
+        }
+    };
 }
 
 Context::Context(const Context& context) {
@@ -28,6 +35,43 @@ Context::Context(const Context& context) {
 Context::~Context() {
 }
 
+void Context::setFrameHandler(FrameHandler handler) {
+    _frameHandler = handler;
+}
+
+#define DEFERRED_RENDERING
+
+void Context::beginFrame(const FramebufferPointer& outputFramebuffer, const glm::mat4& renderPose) {
+    _currentFrame = Frame();
+    _currentFrame.framebuffer = outputFramebuffer;
+    _currentFrame.pose = renderPose;
+    _frameActive = true;
+}
+
+void Context::append(Batch& batch) {
+    if (!_frameActive) {
+        qWarning() << "Batch executed outside of frame boundaries";
+    }
+#ifdef DEFERRED_RENDERING
+    _currentFrame.batches.emplace_back(batch);
+    _currentFrame.stereoStates.emplace_back(_stereo);
+#else
+    _backend->_stereo = _stereo;
+    _backend->render(batch);
+#endif
+}
+
+void Context::endFrame() {
+#ifdef DEFERRED_RENDERING
+    if (_frameHandler) {
+        _frameHandler(_currentFrame);
+    }
+#endif
+    _currentFrame = Frame();
+    _frameActive = false;
+}
+
+
 bool Context::makeProgram(Shader& shader, const Shader::BindingSet& bindings) {
     if (shader.isProgram() && _makeProgramCallback) {
         return _makeProgramCallback(shader, bindings);
@@ -35,36 +79,38 @@ bool Context::makeProgram(Shader& shader, const Shader::BindingSet& bindings) {
     return false;
 }
 
-void Context::render(Batch& batch) {
-    PROFILE_RANGE(__FUNCTION__);
-    _backend->render(batch);
-}
-
 void Context::enableStereo(bool enable) {
-    _backend->enableStereo(enable);
+    _stereo._enable = enable;
 }
 
 bool Context::isStereo() {
-    return _backend->isStereo();
+    return _stereo._enable;
 }
 
 void Context::setStereoProjections(const mat4 eyeProjections[2]) {
-    _backend->setStereoProjections(eyeProjections);
+    for (int i = 0; i < 2; ++i) {
+        _stereo._eyeProjections[i] = eyeProjections[i];
+    }
 }
 
-void Context::setStereoViews(const mat4 eyeViews[2]) {
-    _backend->setStereoViews(eyeViews);
+void Context::setStereoViews(const mat4 views[2]) {
+    for (int i = 0; i < 2; ++i) {
+        _stereo._eyeViews[i] = views[i];
+    }
 }
 
 void Context::getStereoProjections(mat4* eyeProjections) const {
-    _backend->getStereoProjections(eyeProjections);
+    for (int i = 0; i < 2; ++i) {
+        eyeProjections[i] = _stereo._eyeProjections[i];
+    }
 }
 
 void Context::getStereoViews(mat4* eyeViews) const {
-    _backend->getStereoViews(eyeViews);
+    for (int i = 0; i < 2; ++i) {
+        eyeViews[i] = _stereo._eyeViews[i];
+    }
 }
 
-
 void Context::syncCache() {
     PROFILE_RANGE(__FUNCTION__);
     _backend->syncCache();
@@ -103,12 +149,12 @@ Backend::TransformCamera Backend::TransformCamera::getEyeCamera(int eye, const S
     if (!_stereo._skybox) {
         offsetTransform.postTranslate(-Vec3(_stereo._eyeViews[eye][3]));
     } else {
-       // FIXME: If "skybox" the ipd is set to 0 for now, let s try to propose a better solution for this in the future
+        // FIXME: If "skybox" the ipd is set to 0 for now, let s try to propose a better solution for this in the future
     }
     result._projection = _stereo._eyeProjections[eye];
     result.recomputeDerived(offsetTransform);
 
-    result._stereoInfo = Vec4(1.0f, (float) eye, 0.0f, 0.0f);
+    result._stereoInfo = Vec4(1.0f, (float)eye, 0.0f, 0.0f);
 
     return result;
 }
@@ -125,7 +171,7 @@ std::atomic<uint32_t> Context::_textureGPUTransferCount{ 0 };
 void Context::incrementBufferGPUCount() {
     _bufferGPUCount++;
 }
-void Context::decrementBufferGPUCount() { 
+void Context::decrementBufferGPUCount() {
     _bufferGPUCount--;
 }
 void Context::updateBufferGPUMemoryUsage(Size prevObjectSize, Size newObjectSize) {

libraries/gpu/src/gpu/Context.h

@@ -16,12 +16,13 @@
 
 #include <GLMHelpers.h>
 
+#include "Forward.h"
 #include "Batch.h"
-
 #include "Resource.h"
 #include "Texture.h"
 #include "Pipeline.h"
 #include "Framebuffer.h"
+#include "Frame.h"
 
 class QImage;
 
@@ -46,51 +47,11 @@ public:
     ContextStats(const ContextStats& stats) = default;
 };
 
-struct StereoState {
-    bool _enable{ false };
-    bool _skybox{ false };
-    // 0 for left eye, 1 for right eye
-    uint8_t _pass{ 0 };
-    mat4 _eyeViews[2];
-    mat4 _eyeProjections[2];
-};
-
 class Backend {
 public:
     virtual~ Backend() {};
 
     virtual void render(Batch& batch) = 0;
-    virtual void enableStereo(bool enable) {
-        _stereo._enable = enable;
-    }
-
-    virtual bool isStereo() {
-        return _stereo._enable;
-    }
-
-    void setStereoProjections(const mat4 eyeProjections[2]) {
-        for (int i = 0; i < 2; ++i) {
-            _stereo._eyeProjections[i] = eyeProjections[i];
-        }
-    }
-
-    void setStereoViews(const mat4 views[2]) {
-        for (int i = 0; i < 2; ++i) {
-            _stereo._eyeViews[i] = views[i];
-        }
-    }
-
-    void getStereoProjections(mat4* eyeProjections) const {
-        for (int i = 0; i < 2; ++i) {
-            eyeProjections[i] = _stereo._eyeProjections[i];
-        }
-    }
-
-    void getStereoViews(mat4* eyeViews) const {
-        for (int i = 0; i < 2; ++i) {
-            eyeViews[i] = _stereo._eyeViews[i];
-        }
-    }
 
     virtual void syncCache() = 0;
     virtual void downloadFramebuffer(const FramebufferPointer& srcFramebuffer, const Vec4i& region, QImage& destImage) = 0;
@@ -137,8 +98,25 @@ public:
     static void decrementTextureGPUTransferCount();
 
 protected:
-    StereoState  _stereo;
+    virtual bool isStereo() {
+        return _stereo._enable;
+    }
+
+    void getStereoProjections(mat4* eyeProjections) const {
+        for (int i = 0; i < 2; ++i) {
+            eyeProjections[i] = _stereo._eyeProjections[i];
+        }
+    }
+
+    void getStereoViews(mat4* eyeViews) const {
+        for (int i = 0; i < 2; ++i) {
+            eyeViews[i] = _stereo._eyeViews[i];
+        }
+    }
+
+    friend class Context;
     ContextStats _stats;
+    StereoState _stereo;
 };
 
 class Context {
@@ -161,7 +139,10 @@ public:
     Context();
     ~Context();
 
-    void render(Batch& batch);
+    void setFrameHandler(FrameHandler handler);
+    void beginFrame(const FramebufferPointer& outputFramebuffer, const glm::mat4& renderPose = glm::mat4());
+    void append(Batch& batch);
+    void endFrame();
 
     void enableStereo(bool enable = true);
     bool isStereo();
@@ -191,6 +172,10 @@ protected:
     Context(const Context& context);
 
     std::unique_ptr<Backend> _backend;
+    bool _frameActive { false };
+    Frame _currentFrame;
+    FrameHandler _frameHandler;
+    StereoState  _stereo;
 
     // This function can only be called by "static Shader::makeProgram()"
     // makeProgramShader(...) make a program shader ready to be used in a Batch.
@@ -234,7 +219,7 @@ template<typename F>
 void doInBatch(std::shared_ptr<gpu::Context> context, F f) {
     gpu::Batch batch;
     f(batch);
-    context->render(batch);
+    context->append(batch);
 }
 
 };

libraries/gpu/src/gpu/Forward.h

@@ -12,6 +12,7 @@
 #include <stdint.h>
 #include <memory>
 #include <vector>
+#include <functional>
 
 #include <glm/glm.hpp>
 
@@ -21,6 +22,9 @@ namespace gpu {
     class Context;
     using ContextPointer = std::shared_ptr<Context>;
     class GPUObject;
+    class Frame;
+    using FramePointer = std::shared_ptr<Frame>;
+    using FrameHandler = std::function<void(Frame& frame)>;
 
     using Stamp = int;
     using uint32 = uint32_t;
@@ -82,6 +86,15 @@ namespace gpu {
     class TextureView;
     using TextureViews = std::vector<TextureView>;
 
+    struct StereoState {
+        bool _enable{ false };
+        bool _skybox{ false };
+        // 0 for left eye, 1 for right eye
+        uint8 _pass{ 0 };
+        Mat4 _eyeViews[2];
+        Mat4 _eyeProjections[2];
+    };
+
     namespace gl {
         class GLBuffer;
     }

libraries/gpu/src/gpu/Frame.h

@@ -0,0 +1,29 @@
+//
+//  Created by Bradley Austin Davis on 2016/07/26
+//  Copyright 2013-2016 High Fidelity, Inc.
+//
+//  Distributed under the Apache License, Version 2.0.
+//  See the accompanying file LICENSE or http://www.apache.org/licenses/LICENSE-2.0.html
+//
+#ifndef hifi_gpu_Frame_h
+#define hifi_gpu_Frame_h
+
+#include "Forward.h"
+
+namespace gpu {
+
+class Frame {
+public:
+    /// The sensor pose used for rendering the frame, only applicable for HMDs
+    glm::mat4 pose;
+    /// The collection of batches which make up the frame
+    std::vector<Batch> batches;
+    std::vector<StereoState> stereoStates;
+    /// The destination framebuffer in which the results will be placed
+    FramebufferPointer framebuffer;
+};
+
+};
+
+
+#endif

libraries/render-utils/src/FramebufferCache.cpp

@@ -11,34 +11,25 @@
 
 #include "FramebufferCache.h"
 
-#include <mutex>
-
 #include <glm/glm.hpp>
+#include <gpu/Format.h>
+#include <gpu/Framebuffer.h>
 
-#include <QMap>
-#include <QQueue>
-#include <gpu/Batch.h>
 #include "RenderUtilsLogging.h"
 
-static QQueue<gpu::FramebufferPointer> _cachedFramebuffers;
-
-FramebufferCache::FramebufferCache() {
-}
-
-FramebufferCache::~FramebufferCache() {
-    _cachedFramebuffers.clear();
-}
-
 void FramebufferCache::setFrameBufferSize(QSize frameBufferSize) {
     //If the size changed, we need to delete our FBOs
     if (_frameBufferSize != frameBufferSize) {
         _frameBufferSize = frameBufferSize;
         _selfieFramebuffer.reset();
-        _cachedFramebuffers.clear();
         _occlusionFramebuffer.reset();
         _occlusionTexture.reset();
         _occlusionBlurredFramebuffer.reset();
         _occlusionBlurredTexture.reset();
+        {
+            std::unique_lock<std::mutex> lock(_mutex);
+            _cachedFramebuffers.clear();
+        }
     }
 }
 
@@ -55,8 +46,6 @@ void FramebufferCache::createPrimaryFramebuffer() {
 
     auto smoothSampler = gpu::Sampler(gpu::Sampler::FILTER_MIN_MAG_MIP_LINEAR);
 
-
-
     resizeAmbientOcclusionBuffers();
 }
 
@@ -87,7 +76,8 @@ void FramebufferCache::resizeAmbientOcclusionBuffers() {
 
 
 gpu::FramebufferPointer FramebufferCache::getFramebuffer() {
-    if (_cachedFramebuffers.isEmpty()) {
+    std::unique_lock<std::mutex> lock(_mutex);
+    if (_cachedFramebuffers.empty()) {
         _cachedFramebuffers.push_back(gpu::FramebufferPointer(gpu::Framebuffer::create(gpu::Element::COLOR_SRGBA_32, _frameBufferSize.width(), _frameBufferSize.height())));
     }
     gpu::FramebufferPointer result = _cachedFramebuffers.front();
@@ -96,6 +86,7 @@ gpu::FramebufferPointer FramebufferCache::getFramebuffer() {
 }
 
 void FramebufferCache::releaseFramebuffer(const gpu::FramebufferPointer& framebuffer) {
+    std::unique_lock<std::mutex> lock(_mutex);
     if (QSize(framebuffer->getSize().x, framebuffer->getSize().y) == _frameBufferSize) {
         _cachedFramebuffers.push_back(framebuffer);
     }

libraries/render-utils/src/FramebufferCache.h

@@ -11,13 +11,10 @@
 
 #include <QSize>
 
-#include <gpu/Framebuffer.h>
+#include <mutex>
+#include <gpu/Forward.h>
 #include <DependencyManager.h>
 
-namespace gpu {
-class Batch;
-}
-
 /// Stores cached textures, including render-to-texture targets.
 class FramebufferCache : public Dependency {
     SINGLETON_DEPENDENCY
@@ -47,9 +44,6 @@ public:
     void releaseFramebuffer(const gpu::FramebufferPointer& framebuffer);
 
 private:
-    FramebufferCache();
-    virtual ~FramebufferCache();
-
     void createPrimaryFramebuffer();
 
     gpu::FramebufferPointer _shadowFramebuffer;
@@ -65,6 +59,9 @@ private:
     QSize _frameBufferSize{ 100, 100 };
     int _AOResolutionLevel = 1; // AO perform at half res
 
+    std::mutex _mutex;
+    std::list<gpu::FramebufferPointer> _cachedFramebuffers;
+
     // Resize/reallocate the buffers used for AO
     // the size of the AO buffers is scaled by the AOResolutionScale;
     void resizeAmbientOcclusionBuffers();

libraries/render-utils/src/RenderDeferredTask.cpp

@@ -347,10 +347,10 @@ void DrawOverlay3D::run(const SceneContextPointer& sceneContext, const RenderCon
         // Needs to be distinct from the other batch because using the clear call 
         // while stereo is enabled triggers a warning
         if (_opaquePass) {
-            gpu::Batch batch;
-            batch.enableStereo(false);
-            batch.clearFramebuffer(gpu::Framebuffer::BUFFER_DEPTH, glm::vec4(), 1.f, 0, true);
-            args->_context->render(batch);
+            gpu::doInBatch(args->_context, [&](gpu::Batch& batch){
+                batch.enableStereo(false);
+                batch.clearFramebuffer(gpu::Framebuffer::BUFFER_DEPTH, glm::vec4(), 1.f, 0, true);
+            });
         }
 
         // Render the items

tests/render-perf/src/main.cpp

@@ -403,7 +403,12 @@ private:
             renderArgs._blitFramebuffer = finalFramebuffer;
         }
 
+        _gpuContext->beginFrame(renderArgs._blitFramebuffer);
+        gpu::doInBatch(renderArgs._context, [&](gpu::Batch& batch) {
+            batch.resetStages();
+        });
         render(&renderArgs);
+        _gpuContext->endFrame();
         GLuint glTex;
         {
             auto gpuTex = renderArgs._blitFramebuffer->getRenderBuffer(0);
@@ -428,9 +433,6 @@ private:
         _offscreenContext->makeCurrent();
         framebufferCache->releaseFramebuffer(renderArgs._blitFramebuffer);
         renderArgs._blitFramebuffer.reset();
-        gpu::doInBatch(renderArgs._context, [&](gpu::Batch& batch) {
-            batch.resetStages();
-        });
         _fpsCounter.increment();
         static size_t _frameCount { 0 };
         ++_frameCount;