From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Sat, 16 Jun 2018 08:36:12 +0000 (+0200)
Subject: Convert Y'CbCr using Movit instead of in software with Qt. Saves a fair amount of... 
X-Git-Tag: 1.8.0~76^2~272
X-Git-Url: https://git.sesse.net/?a=commitdiff_plain;h=1b8e3294b89ec9131f507bbb53ee9b8ad9abafee;p=nageru

Convert Y'CbCr using Movit instead of in software with Qt. Saves a fair amount of CPU.
---

diff --git a/Makefile b/Makefile
index 42b5ea0..beebb67 100644
--- a/Makefile
+++ b/Makefile
@@ -1,9 +1,9 @@
 CXX=g++
-PKG_MODULES := Qt5Core Qt5Gui Qt5Widgets Qt5PrintSupport libjpeg
+PKG_MODULES := Qt5Core Qt5Gui Qt5Widgets Qt5OpenGLExtensions Qt5OpenGL Qt5PrintSupport libjpeg movit
 CXXFLAGS ?= -O2 -g -Wall  # Will be overridden by environment.
-CXXFLAGS += -std=gnu++11 -fPIC $(shell pkg-config --cflags $(PKG_MODULES)) -pthread
+CXXFLAGS += -std=gnu++11 -fPIC $(shell pkg-config --cflags $(PKG_MODULES)) -DMOVIT_SHADER_DIR=\"$(shell pkg-config --variable=shaderdir movit)\" -pthread
 
-LDLIBS=$(shell pkg-config --libs $(PKG_MODULES)) -pthread -lavformat -lavcodec -lavutil -lswscale
+LDLIBS=$(shell pkg-config --libs $(PKG_MODULES)) -pthread -lavformat -lavcodec -lavutil -lswscale -lGL
 
 # Qt objects
 OBJS_WITH_MOC = mainwindow.o jpeg_frame_view.o clip_list.o
diff --git a/jpeg_frame_view.cpp b/jpeg_frame_view.cpp
index e7ee5f1..65fccbb 100644
--- a/jpeg_frame_view.cpp
+++ b/jpeg_frame_view.cpp
@@ -1,5 +1,6 @@
 #include "jpeg_frame_view.h"
 
+#include <jpeglib.h>
 #include <stdint.h>
 
 #include <atomic>
@@ -8,12 +9,15 @@
 #include <mutex>
 #include <thread>
 #include <utility>
-#include <QGraphicsPixmapItem>
-#include <QPixmap>
+
+#include <movit/resource_pool.h>
+#include <movit/init.h>
+#include <movit/util.h>
 
 #include "defs.h"
 #include "post_to_main_thread.h"
 
+using namespace movit;
 using namespace std;
 
 string filename_for_frame(unsigned stream_idx, int64_t pts);
@@ -26,16 +30,85 @@ bool operator< (const JPEGID &a, const JPEGID &b) {
 	return make_pair(a.stream_idx, a.pts) < make_pair(b.stream_idx, b.pts);
 }
 
-struct LRUPixmap {
-	shared_ptr<QPixmap> pixmap;
+struct LRUFrame {
+	shared_ptr<Frame> frame;
 	size_t last_used;
 };
 
 mutex cache_mu;
-map<JPEGID, LRUPixmap> cache;  // Under cache_mu.
+map<JPEGID, LRUFrame> cache;  // Under cache_mu.
 condition_variable any_pending_decodes;
 deque<pair<JPEGID, JPEGFrameView *>> pending_decodes;  // Under cache_mu.
 atomic<size_t> event_counter{0};
+extern QGLWidget *global_share_widget;
+
+// TODO: Decode using VA-API if available.
+shared_ptr<Frame> decode_jpeg(const string &filename)
+{
+	shared_ptr<Frame> frame(new Frame);
+
+	jpeg_decompress_struct dinfo;
+	jpeg_error_mgr jerr;
+	dinfo.err = jpeg_std_error(&jerr);
+	jpeg_create_decompress(&dinfo);
+
+	FILE *fp = fopen(filename.c_str(), "rb");
+	if (fp == nullptr) {
+		perror(filename.c_str());
+		exit(1);
+	}
+	jpeg_stdio_src(&dinfo, fp);
+
+	jpeg_read_header(&dinfo, true);
+
+	if (dinfo.num_components != 3 ||
+            dinfo.comp_info[0].h_samp_factor != 2 ||
+            dinfo.comp_info[0].v_samp_factor != 2 ||
+            dinfo.comp_info[1].h_samp_factor != 1 ||
+            dinfo.comp_info[1].v_samp_factor != 2 ||
+            dinfo.comp_info[2].h_samp_factor != 1 ||
+            dinfo.comp_info[2].v_samp_factor != 2) {
+		fprintf(stderr, "Not 4:2:2 JPEG! (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
+			dinfo.num_components,
+			dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
+			dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
+			dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
+		exit(1);
+	}
+	dinfo.raw_data_out = true;
+
+	jpeg_start_decompress(&dinfo);
+
+	frame->width = dinfo.output_width;
+	frame->height = dinfo.output_height;
+
+	unsigned chroma_width_blocks = (dinfo.output_width + 15) / 16;
+	unsigned width_blocks = chroma_width_blocks * 2;
+	unsigned height_blocks = (dinfo.output_height + 15) / 16;
+
+	// TODO: Decode into a PBO.
+	frame->y.reset(new uint8_t[width_blocks * (height_blocks * 2) * DCTSIZE2]);
+	frame->cb.reset(new uint8_t[chroma_width_blocks * (height_blocks * 2) * DCTSIZE2]);
+	frame->cr.reset(new uint8_t[chroma_width_blocks * (height_blocks * 2) * DCTSIZE2]);
+
+	JSAMPROW yptr[16], cbptr[16], crptr[16];
+	JSAMPARRAY data[3] = { yptr, cbptr, crptr };
+	for (unsigned y = 0; y < height_blocks; ++y) {
+		for (unsigned yy = 0; yy < DCTSIZE * 2; ++yy) {
+			yptr[yy] = frame->y.get() + (y * DCTSIZE * 2 + yy) * width_blocks * DCTSIZE;
+			cbptr[yy] = frame->cb.get() + (y * DCTSIZE * 2 + yy) * chroma_width_blocks * DCTSIZE;
+			crptr[yy] = frame->cr.get() + (y * DCTSIZE * 2 + yy) * chroma_width_blocks * DCTSIZE;
+		}
+
+		jpeg_read_raw_data(&dinfo, data, /*num_lines=*/16);
+	}
+
+	(void) jpeg_finish_decompress(&dinfo);
+	jpeg_destroy_decompress(&dinfo);
+	fclose(fp);
+
+	return frame;
+}
 
 void prune_cache()
 {
@@ -65,7 +138,7 @@ void jpeg_decoder_thread()
 	for ( ;; ) {
 		JPEGID id;
 		JPEGFrameView *dest;
-		shared_ptr<QPixmap> pixmap;
+		shared_ptr<Frame> frame;
 		{
 			unique_lock<mutex> lock(cache_mu);
 			any_pending_decodes.wait(lock, [] {
@@ -77,12 +150,12 @@ void jpeg_decoder_thread()
 
 			auto it = cache.find(id);
 			if (it != cache.end()) {
-				pixmap = it->second.pixmap;
+				frame = it->second.frame;
 				it->second.last_used = event_counter++;
 			}
 		}
 
-		if (pixmap == nullptr) {
+		if (frame == nullptr) {
 			// Not found in the cache, so we need to do a decode or drop the request.
 			// Prune the queue if there are too many pending for this destination.
 			// TODO: Could we get starvation here?
@@ -97,11 +170,10 @@ void jpeg_decoder_thread()
 				continue;
 			}
 
-			pixmap.reset(
-				new QPixmap(QString::fromStdString(filename_for_frame(id.stream_idx, id.pts))));
+			frame = decode_jpeg(filename_for_frame(id.stream_idx, id.pts));
 
 			unique_lock<mutex> lock(cache_mu);
-			cache[id] = LRUPixmap{ pixmap, event_counter++ };
+			cache[id] = LRUFrame{ frame, event_counter++ };
 
 			if (cache.size() > CACHE_SIZE) {
 				prune_cache();
@@ -113,39 +185,96 @@ void jpeg_decoder_thread()
 			}
 		}
 
-		dest->setPixmap(pixmap);
+		dest->setDecodedFrame(frame);
 	}
 }
 
 JPEGFrameView::JPEGFrameView(QWidget *parent)
-	: QGraphicsView(parent) {
-	scene.addItem(&item);
-	setScene(&scene);
-	setHorizontalScrollBarPolicy(Qt::ScrollBarAlwaysOff);
-	setVerticalScrollBarPolicy(Qt::ScrollBarAlwaysOff);
+	: QGLWidget(parent, global_share_widget) {
+}
+
+void JPEGFrameView::update_frame()
+{
+	unique_lock<mutex> lock(cache_mu);
+	pending_decodes.emplace_back(JPEGID{ stream_idx, pts }, this);
+	any_pending_decodes.notify_all();
+}
+
+ResourcePool *resource_pool = nullptr;
+
+void JPEGFrameView::initializeGL()
+{
+	glDisable(GL_BLEND);
+	glDisable(GL_DEPTH_TEST);
+	glDepthMask(GL_FALSE);
+	check_error();
 
 	static once_flag once;
 	call_once(once, [] {
+		CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
+		resource_pool = new ResourcePool;
+
 		std::thread(&jpeg_decoder_thread).detach();
 	});
+
+	chain.reset(new EffectChain(1280, 720, resource_pool));
+	ImageFormat image_format;
+	image_format.color_space = COLORSPACE_sRGB;
+	image_format.gamma_curve = GAMMA_sRGB;
+	YCbCrFormat ycbcr_format;
+	ycbcr_format.luma_coefficients = YCBCR_REC_709;
+	ycbcr_format.full_range = false;
+	ycbcr_format.num_levels = 256;
+	ycbcr_format.chroma_subsampling_x = 2;
+	ycbcr_format.chroma_subsampling_y = 1;
+	ycbcr_format.cb_x_position = 0.0f;  // H.264 -- _not_ JPEG, even though our input is MJPEG-encoded
+	ycbcr_format.cb_y_position = 0.5f;  // Irrelevant.
+	ycbcr_format.cr_x_position = 0.0f;
+	ycbcr_format.cr_y_position = 0.5f;
+	ycbcr_input = (movit::YCbCrInput *)chain->add_input(new YCbCrInput(image_format, ycbcr_format, 1280, 720));
+
+	ImageFormat inout_format;
+        inout_format.color_space = COLORSPACE_sRGB;
+        inout_format.gamma_curve = GAMMA_sRGB;
+
+	check_error();
+	chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
+	check_error();
+	chain->set_dither_bits(8);
+	check_error();
+	chain->finalize();
+	check_error();
 }
 
-void JPEGFrameView::update_frame()
+void JPEGFrameView::resizeGL(int width, int height)
 {
-	unique_lock<mutex> lock(cache_mu);
-	pending_decodes.emplace_back(JPEGID{ stream_idx, pts }, this);
-	any_pending_decodes.notify_all();
+	check_error();
+	glViewport(0, 0, width, height);
+	check_error();
 }
 
-void JPEGFrameView::resizeEvent(QResizeEvent *event)
+void JPEGFrameView::paintGL()
 {
-	fitInView(&item, Qt::KeepAspectRatio);
+	//glClearColor(0.0f, 1.0f, 0.0f, 1.0f);
+	//glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
+
+	check_error();
+	chain->render_to_screen();
 }
 
-void JPEGFrameView::setPixmap(std::shared_ptr<QPixmap> pixmap)
+void JPEGFrameView::setDecodedFrame(std::shared_ptr<Frame> frame)
 {
-	post_to_main_thread([this, pixmap] {
-		item.setPixmap(*pixmap);
-		fitInView(&item, Qt::KeepAspectRatio);
+	post_to_main_thread([this, frame] {
+		current_frame = frame;
+		int width_blocks = (frame->width + 15) / 16;
+		ycbcr_input->set_width(frame->width);
+		ycbcr_input->set_height(frame->height);
+		ycbcr_input->set_pixel_data(0, frame->y.get());
+		ycbcr_input->set_pixel_data(1, frame->cb.get());
+		ycbcr_input->set_pixel_data(2, frame->cr.get());
+		ycbcr_input->set_pitch(0, width_blocks * 16);
+		ycbcr_input->set_pitch(1, width_blocks * 8);
+		ycbcr_input->set_pitch(2, width_blocks * 8);
+		update();
 	});
 }
diff --git a/jpeg_frame_view.h b/jpeg_frame_view.h
index e5665cd..4341819 100644
--- a/jpeg_frame_view.h
+++ b/jpeg_frame_view.h
@@ -1,15 +1,22 @@
 #ifndef _JPEG_FRAME_VIEW_H
 #define _JPEG_FRAME_VIEW_H 1
 
-#include <QGraphicsView>
-#include <QGraphicsPixmapItem>
-#include <QGraphicsScene>
+#include <epoxy/gl.h>
+#include <QGLWidget>
 
 #include <stdint.h>
 
+#include <movit/effect_chain.h>
+#include <movit/ycbcr_input.h>
+
 #include <memory>
 
-class JPEGFrameView : public QGraphicsView {
+struct Frame {
+	std::unique_ptr<uint8_t[]> y, cb, cr;
+	unsigned width, height;
+};
+
+class JPEGFrameView : public QGLWidget {
 	Q_OBJECT
 
 public:
@@ -22,20 +29,22 @@ public:
 		update_frame();
 	}
 
-	void setPixmap(std::shared_ptr<QPixmap> pixmap);
+	void setDecodedFrame(std::shared_ptr<Frame> frame);
 
 protected:
-	void resizeEvent(QResizeEvent *event) override;
+	void initializeGL() override;
+	void resizeGL(int width, int height) override;
+	void paintGL() override;
 
 private:
 	void update_frame();
 
-	QGraphicsPixmapItem item;
-	QGraphicsScene scene;
-
 	unsigned stream_idx;
 	int64_t pts;
-	bool dirty = false;
+
+	std::unique_ptr<movit::EffectChain> chain;
+	std::shared_ptr<Frame> current_frame;  // So that we hold on to the pixels.
+	movit::YCbCrInput *ycbcr_input;
 };
 
 #endif  // !defined(_JPEG_FRAME_VIEW_H)
diff --git a/main.cpp b/main.cpp
index d916f44..0f1e6fa 100644
--- a/main.cpp
+++ b/main.cpp
@@ -39,6 +39,7 @@ string filename_for_frame(unsigned stream_idx, int64_t pts)
 
 mutex frame_mu;
 vector<int64_t> frames[MAX_STREAMS];
+QGLWidget *global_share_widget;
 
 int record_thread_func();
 
@@ -47,7 +48,29 @@ int main(int argc, char **argv)
 	av_register_all();
 	avformat_network_init();
 
+	QCoreApplication::setAttribute(Qt::AA_ShareOpenGLContexts, true);
+
+	QSurfaceFormat fmt;
+	fmt.setDepthBufferSize(0);
+	fmt.setStencilBufferSize(0);
+	fmt.setProfile(QSurfaceFormat::CoreProfile);
+	fmt.setMajorVersion(3);
+	fmt.setMinorVersion(1);
+
+	// Turn off vsync, since Qt generally gives us at most frame rate
+	// (display frequency) / (number of QGLWidgets active).
+	fmt.setSwapInterval(0);
+
+	QSurfaceFormat::setDefaultFormat(fmt);
+
+	QGLFormat::setDefaultFormat(QGLFormat::fromSurfaceFormat(fmt));
+
 	QApplication app(argc, argv);
+	global_share_widget = new QGLWidget();
+	if (!global_share_widget->isValid()) {
+		fprintf(stderr, "Failed to initialize OpenGL. Futatabi needs at least OpenGL 3.1 to function properly.\n");
+		exit(1);
+	}
 	MainWindow mainWindow;
 	mainWindow.show();
 
diff --git a/mainwindow.cpp b/mainwindow.cpp
index 84231b7..9a2c5f1 100644
--- a/mainwindow.cpp
+++ b/mainwindow.cpp
@@ -133,3 +133,17 @@ void MainWindow::live_player_clip_done()
 		playlist_clips->set_currently_playing(-1);
 	}
 }
+
+void MainWindow::resizeEvent(QResizeEvent *event)
+{
+	QMainWindow::resizeEvent(event);
+
+	// Ask for a relayout, but only after the event loop is done doing relayout
+	// on everything else.
+	QMetaObject::invokeMethod(this, "relayout", Qt::QueuedConnection);
+}
+
+void MainWindow::relayout()
+{
+	ui->live_display->setMinimumHeight(ui->live_display->width() * 9 / 16);
+}
diff --git a/mainwindow.h b/mainwindow.h
index c037c77..4d09cb3 100644
--- a/mainwindow.h
+++ b/mainwindow.h
@@ -28,6 +28,11 @@ private:
 	void preview_clicked();
 	void play_clicked();
 	void live_player_clip_done();
+
+	void resizeEvent(QResizeEvent *event) override;
+
+private slots:
+	void relayout();
 };
 
 extern MainWindow *global_mainwindow;
diff --git a/ui_mainwindow.ui b/ui_mainwindow.ui
index 7e51adc..c3597d9 100644
--- a/ui_mainwindow.ui
+++ b/ui_mainwindow.ui
@@ -23,20 +23,20 @@
       <widget class="QWidget" name="horizontalLayoutWidget">
        <layout class="QHBoxLayout" name="horizontalLayout">
         <item>
-         <layout class="QVBoxLayout" name="verticalLayout_3">
+         <layout class="QVBoxLayout" name="verticalLayout_3" stretch="1,0">
           <item>
            <layout class="QGridLayout" name="gridLayout_2">
             <item row="1" column="0">
-             <widget class="JPEGFrameView" name="input3_display"/>
+             <widget class="JPEGFrameView" name="input3_display" native="true"/>
             </item>
             <item row="0" column="1">
-             <widget class="JPEGFrameView" name="input2_display"/>
+             <widget class="JPEGFrameView" name="input2_display" native="true"/>
             </item>
             <item row="1" column="1">
-             <widget class="JPEGFrameView" name="input4_display"/>
+             <widget class="JPEGFrameView" name="input4_display" native="true"/>
             </item>
             <item row="0" column="0">
-             <widget class="JPEGFrameView" name="input1_display"/>
+             <widget class="JPEGFrameView" name="input1_display" native="true"/>
             </item>
            </layout>
           </item>
@@ -53,9 +53,9 @@
          </layout>
         </item>
         <item>
-         <layout class="QVBoxLayout" name="verticalLayout">
+         <layout class="QVBoxLayout" name="verticalLayout" stretch="1,0">
           <item>
-           <widget class="JPEGFrameView" name="preview_display"/>
+           <widget class="JPEGFrameView" name="preview_display" native="true"/>
           </item>
           <item>
            <widget class="QLabel" name="label_2">
@@ -70,9 +70,9 @@
          </layout>
         </item>
         <item>
-         <layout class="QVBoxLayout" name="verticalLayout_2">
+         <layout class="QVBoxLayout" name="verticalLayout_2" stretch="1,0">
           <item>
-           <widget class="JPEGFrameView" name="live_display"/>
+           <widget class="JPEGFrameView" name="live_display" native="true"/>
           </item>
           <item>
            <widget class="QLabel" name="label_3">
@@ -152,7 +152,7 @@
  <customwidgets>
   <customwidget>
    <class>JPEGFrameView</class>
-   <extends>QGraphicsView</extends>
+   <extends>QWidget</extends>
    <header>jpeg_frame_view.h</header>
   </customwidget>
  </customwidgets>