X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=jpeg_frame_view.cpp;h=be53688a03ee33bca3bb50688030978eb411ade2;hb=bdef311c334b674ba39a931805fb7d32ce8698da;hp=8e22f629541903fc2b2d20d1399aa18c33ec57fd;hpb=530f3b3aa10aafaf22da10e53d3d243df26bf174;p=nageru

diff --git a/jpeg_frame_view.cpp b/jpeg_frame_view.cpp
index 8e22f62..be53688 100644
--- a/jpeg_frame_view.cpp
+++ b/jpeg_frame_view.cpp
@@ -1,30 +1,55 @@
 #include "jpeg_frame_view.h"
 
-#include <jpeglib.h>
-#include <stdint.h>
+#include "defs.h"
+#include "jpeg_destroyer.h"
+#include "post_to_main_thread.h"
+#include "video_stream.h"
+#include "ycbcr_converter.h"
 
+#include <QMouseEvent>
+#include <QScreen>
 #include <atomic>
 #include <condition_variable>
 #include <deque>
+#include <jpeglib.h>
+#include <movit/init.h>
+#include <movit/resource_pool.h>
+#include <movit/util.h>
 #include <mutex>
+#include <stdint.h>
 #include <thread>
+#include <unistd.h>
 #include <utility>
 
-#include <QMouseEvent>
-
-#include <movit/resource_pool.h>
-#include <movit/init.h>
-#include <movit/util.h>
-
-#include "defs.h"
-#include "post_to_main_thread.h"
-#include "video_stream.h"
+// Must come after the Qt stuff.
+#include "vaapi_jpeg_decoder.h"
 
 using namespace movit;
 using namespace std;
 
-bool operator< (const JPEGID &a, const JPEGID &b) {
-	return make_pair(a.stream_idx, a.pts) < make_pair(b.stream_idx, b.pts);
+namespace {
+
+// Just an arbitrary order for std::map.
+struct FrameOnDiskLexicalOrder
+{
+	bool operator() (const FrameOnDisk &a, const FrameOnDisk &b) const
+	{
+		if (a.pts != b.pts)
+			return a.pts < b.pts;
+		if (a.offset != b.offset)
+			return a.offset < b.offset;
+		if (a.filename_idx != b.filename_idx)
+			return a.filename_idx < b.filename_idx;
+		assert(a.size == b.size);
+		return false;
+	}
+};
+
+inline size_t frame_size(const Frame &frame)
+{
+	size_t y_size = frame.width * frame.height;
+	size_t cbcr_size = y_size / frame.chroma_subsampling_x / frame.chroma_subsampling_y;
+	return y_size + cbcr_size * 2;
 }
 
 struct LRUFrame {
@@ -32,22 +57,49 @@ struct LRUFrame {
 	size_t last_used;
 };
 
+struct PendingDecode {
+	JPEGFrameView *destination;
+
+	// For actual decodes (only if frame below is nullptr).
+	FrameOnDisk primary, secondary;
+	float fade_alpha;  // Irrelevant if secondary.stream_idx == -1.
+
+	// Already-decoded frames are also sent through PendingDecode,
+	// so that they get drawn in the right order. If frame is nullptr,
+	// it's a real decode.
+	shared_ptr<Frame> frame;
+};
+
+}  // namespace
+
+thread JPEGFrameView::jpeg_decoder_thread;
 mutex cache_mu;
-map<JPEGID, LRUFrame> cache;  // Under cache_mu.
+map<FrameOnDisk, LRUFrame, FrameOnDiskLexicalOrder> cache;  // Under cache_mu.
+size_t cache_bytes_used = 0;  // Under cache_mu.
 condition_variable any_pending_decodes;
-deque<pair<JPEGID, JPEGFrameView *>> pending_decodes;  // Under cache_mu.
+deque<PendingDecode> pending_decodes;  // Under cache_mu.
 atomic<size_t> event_counter{0};
 extern QGLWidget *global_share_widget;
+extern atomic<bool> should_quit;
 
-// TODO: Decode using VA-API if available.
 shared_ptr<Frame> decode_jpeg(const string &filename)
 {
-	shared_ptr<Frame> frame(new Frame);
+	shared_ptr<Frame> frame;
+	if (vaapi_jpeg_decoding_usable) {
+		frame = decode_jpeg_vaapi(filename);
+		if (frame != nullptr) {
+			return frame;
+		}
+		fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
+	}
+
+	frame.reset(new Frame);
 
 	jpeg_decompress_struct dinfo;
 	jpeg_error_mgr jerr;
 	dinfo.err = jpeg_std_error(&jerr);
 	jpeg_create_decompress(&dinfo);
+	JPEGDestroyer destroy_dinfo(&dinfo);
 
 	FILE *fp = fopen(filename.c_str(), "rb");
 	if (fp == nullptr) {
@@ -117,8 +169,7 @@ shared_ptr<Frame> decode_jpeg(const string &filename)
 		jpeg_read_raw_data(&dinfo, data, v_mcu_size);
 	}
 
-	(void) jpeg_finish_decompress(&dinfo);
-	jpeg_destroy_decompress(&dinfo);
+	(void)jpeg_finish_decompress(&dinfo);
 	fclose(fp);
 
 	return frame;
@@ -127,16 +178,28 @@ shared_ptr<Frame> decode_jpeg(const string &filename)
 void prune_cache()
 {
 	// Assumes cache_mu is held.
-	vector<size_t> lru_timestamps;
+	int64_t bytes_still_to_remove = cache_bytes_used - (size_t(CACHE_SIZE_MB) * 1024 * 1024) * 9 / 10;
+	if (bytes_still_to_remove <= 0) return;
+
+	vector<pair<size_t, size_t>> lru_timestamps_and_size;
 	for (const auto &key_and_value : cache) {
-		lru_timestamps.push_back(key_and_value.second.last_used);
+		lru_timestamps_and_size.emplace_back(
+			key_and_value.second.last_used,
+			frame_size(*key_and_value.second.frame));
+	}
+	sort(lru_timestamps_and_size.begin(), lru_timestamps_and_size.end());
+
+	// Remove the oldest ones until we are below 90% of the cache used.
+	size_t lru_cutoff_point = 0;
+	for (const pair<size_t, size_t> &it : lru_timestamps_and_size) {
+		lru_cutoff_point = it.first;
+		bytes_still_to_remove -= it.second;
+		if (bytes_still_to_remove <= 0) break;
 	}
 
-	size_t cutoff_point = CACHE_SIZE / 10;  // Prune away the 10% oldest ones.
-	nth_element(lru_timestamps.begin(), lru_timestamps.begin() + cutoff_point, lru_timestamps.end());
-	size_t must_be_used_after = lru_timestamps[cutoff_point];
 	for (auto it = cache.begin(); it != cache.end(); ) {
-		if (it->second.last_used < must_be_used_after) {
+		if (it->second.last_used <= lru_cutoff_point) {
+			cache_bytes_used -= frame_size(*it->second.frame);
 			it = cache.erase(it);
 		} else {
 			++it;
@@ -144,12 +207,12 @@ void prune_cache()
 	}
 }
 
-shared_ptr<Frame> decode_jpeg_with_cache(JPEGID id, CacheMissBehavior cache_miss_behavior, bool *did_decode)
+shared_ptr<Frame> decode_jpeg_with_cache(FrameOnDisk frame_spec, CacheMissBehavior cache_miss_behavior, bool *did_decode)
 {
 	*did_decode = false;
 	{
 		unique_lock<mutex> lock(cache_mu);
-		auto it = cache.find(id);
+		auto it = cache.find(frame_spec);
 		if (it != cache.end()) {
 			it->second.last_used = event_counter++;
 			return it->second.frame;
@@ -161,38 +224,39 @@ shared_ptr<Frame> decode_jpeg_with_cache(JPEGID id, CacheMissBehavior cache_miss
 	}
 
 	*did_decode = true;
-	shared_ptr<Frame> frame = decode_jpeg(filename_for_frame(id.stream_idx, id.pts));
+	shared_ptr<Frame> frame = decode_jpeg(read_frame(frame_spec));
 
 	unique_lock<mutex> lock(cache_mu);
-	cache[id] = LRUFrame{ frame, event_counter++ };
+	cache_bytes_used += frame_size(*frame);
+	cache[frame_spec] = LRUFrame{ frame, event_counter++ };
 
-	if (cache.size() > CACHE_SIZE) {
+	if (cache_bytes_used > size_t(CACHE_SIZE_MB) * 1024 * 1024) {
 		prune_cache();
 	}
 	return frame;
 }
 
-void jpeg_decoder_thread()
+void jpeg_decoder_thread_func()
 {
 	size_t num_decoded = 0, num_dropped = 0;
 
 	pthread_setname_np(pthread_self(), "JPEGDecoder");
-	for ( ;; ) {
-		JPEGID id;
-		JPEGFrameView *dest;
+	while (!should_quit.load()) {
+		PendingDecode decode;
 		CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
 		{
 			unique_lock<mutex> lock(cache_mu);  // TODO: Perhaps under another lock?
 			any_pending_decodes.wait(lock, [] {
-				return !pending_decodes.empty();
+				return !pending_decodes.empty() || should_quit.load();
 			});
-			id = pending_decodes.front().first;
-			dest = pending_decodes.front().second;
+			if (should_quit.load())
+				break;
+			decode = pending_decodes.front();
 			pending_decodes.pop_front();
 
 			size_t num_pending = 0;
-			for (const pair<JPEGID, JPEGFrameView *> &decode : pending_decodes) {
-				if (decode.second == dest) {
+			for (const PendingDecode &other_decode : pending_decodes) {
+				if (other_decode.destination == decode.destination) {
 					++num_pending;
 				}
 			}
@@ -201,35 +265,85 @@ void jpeg_decoder_thread()
 			}
 		}
 
-		bool found_in_cache;
-		shared_ptr<Frame> frame = decode_jpeg_with_cache(id, cache_miss_behavior, &found_in_cache);
-
-		if (frame == nullptr) {
-			assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
-			++num_dropped;
+		if (decode.frame != nullptr) {
+			// Already decoded, so just show it.
+			decode.destination->setDecodedFrame(decode.frame, nullptr, 1.0f);
 			continue;
 		}
 
-		if (!found_in_cache) {
-			++num_decoded;
-			if (num_decoded % 1000 == 0) {
-				fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
-					num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
+		shared_ptr<Frame> primary_frame, secondary_frame;
+		bool drop = false;
+		for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
+			const FrameOnDisk &frame_spec = (subframe_idx == 0 ? decode.primary : decode.secondary);
+			if (frame_spec.pts == -1) {
+				// No secondary frame.
+				continue;
+			}
+
+			bool found_in_cache;
+			shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, cache_miss_behavior, &found_in_cache);
+
+			if (frame == nullptr) {
+				assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
+				drop = true;
+				break;
+			}
+
+			if (!found_in_cache) {
+				++num_decoded;
+				if (num_decoded % 1000 == 0) {
+					fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
+						num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
+				}
+			}
+			if (subframe_idx == 0) {
+				primary_frame = move(frame);
+			} else {
+				secondary_frame = move(frame);
 			}
 		}
+		if (drop) {
+			++num_dropped;
+			continue;
+		}
 
-		dest->setDecodedFrame(frame);
+		// TODO: Could we get jitter between non-interpolated and interpolated frames here?
+		decode.destination->setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
 	}
 }
 
+void JPEGFrameView::shutdown()
+{
+	any_pending_decodes.notify_all();
+	jpeg_decoder_thread.join();
+}
+
 JPEGFrameView::JPEGFrameView(QWidget *parent)
-	: QGLWidget(parent, global_share_widget) {
+	: QGLWidget(parent, global_share_widget)
+{
 }
 
-void JPEGFrameView::update_frame()
+void JPEGFrameView::setFrame(unsigned stream_idx, FrameOnDisk frame, FrameOnDisk secondary_frame, float fade_alpha)
 {
+	current_stream_idx = stream_idx;  // TODO: Does this interact with fades?
+
 	unique_lock<mutex> lock(cache_mu);
-	pending_decodes.emplace_back(JPEGID{ stream_idx, pts }, this);
+	PendingDecode decode;
+	decode.primary = frame;
+	decode.secondary = secondary_frame;
+	decode.fade_alpha = fade_alpha;
+	decode.destination = this;
+	pending_decodes.push_back(decode);
+	any_pending_decodes.notify_all();
+}
+
+void JPEGFrameView::setFrame(shared_ptr<Frame> frame)
+{
+	unique_lock<mutex> lock(cache_mu);
+	PendingDecode decode;
+	decode.frame = std::move(frame);
+	decode.destination = this;
+	pending_decodes.push_back(decode);
 	any_pending_decodes.notify_all();
 }
 
@@ -244,35 +358,20 @@ void JPEGFrameView::initializeGL()
 	static once_flag once;
 	call_once(once, [] {
 		resource_pool = new ResourcePool;
-		std::thread(&jpeg_decoder_thread).detach();
+		jpeg_decoder_thread = std::thread(jpeg_decoder_thread_func);
 	});
 
-	chain.reset(new EffectChain(1280, 720, resource_pool));
-	ImageFormat image_format;
-	image_format.color_space = COLORSPACE_sRGB;
-	image_format.gamma_curve = GAMMA_sRGB;
-	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));
+	ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
 
 	ImageFormat inout_format;
-        inout_format.color_space = COLORSPACE_sRGB;
-        inout_format.gamma_curve = GAMMA_sRGB;
+	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();
+	overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
+	overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
+
+	overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
+	overlay_chain->finalize();
 }
 
 void JPEGFrameView::resizeGL(int width, int height)
@@ -280,10 +379,15 @@ void JPEGFrameView::resizeGL(int width, int height)
 	check_error();
 	glViewport(0, 0, width, height);
 	check_error();
+
+	// Save these, as width() and height() will lie with DPI scaling.
+	gl_width = width;
+	gl_height = height;
 }
 
 void JPEGFrameView::paintGL()
 {
+	glViewport(0, 0, gl_width, gl_height);
 	if (current_frame == nullptr) {
 		glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
 		glClear(GL_COLOR_BUFFER_BIT);
@@ -291,24 +395,34 @@ void JPEGFrameView::paintGL()
 	}
 
 	check_error();
-	chain->render_to_screen();
+	current_chain->render_to_screen();
+
+	if (overlay_image != nullptr) {
+		if (overlay_input_needs_refresh) {
+			overlay_input->set_width(overlay_width);
+			overlay_input->set_height(overlay_height);
+			overlay_input->set_pixel_data(overlay_image->bits());
+		}
+		glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
+		overlay_chain->render_to_screen();
+	}
 }
 
-void JPEGFrameView::setDecodedFrame(std::shared_ptr<Frame> frame)
+namespace {
+
+}  // namespace
+
+void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
 {
-	post_to_main_thread([this, frame] {
+	post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
 		current_frame = frame;
-		ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x;
-		ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y;
-		ycbcr_input->change_ycbcr_format(ycbcr_format);
-		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, frame->pitch_y);
-		ycbcr_input->set_pitch(1, frame->pitch_chroma);
-		ycbcr_input->set_pitch(2, frame->pitch_chroma);
+		current_secondary_frame = secondary_frame;
+
+		if (secondary_frame != nullptr) {
+			current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
+		} else {
+			current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
+		}
 		update();
 	});
 }
@@ -319,3 +433,30 @@ void JPEGFrameView::mousePressEvent(QMouseEvent *event)
 		emit clicked();
 	}
 }
+
+void JPEGFrameView::set_overlay(const string &text)
+{
+	if (text.empty()) {
+		overlay_image.reset();
+		return;
+	}
+
+	float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
+	overlay_width = lrint(overlay_base_width * dpr);
+	overlay_height = lrint(overlay_base_height * dpr);
+
+	overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
+	overlay_image->setDevicePixelRatio(dpr);
+	overlay_image->fill(0);
+	QPainter painter(overlay_image.get());
+
+	painter.setPen(Qt::white);
+	QFont font = painter.font();
+	font.setPointSize(12);
+	painter.setFont(font);
+
+	painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
+
+	// Don't refresh immediately; we might not have an OpenGL context here.
+	overlay_input_needs_refresh = true;
+}