X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=futatabi%2Fjpeg_frame_view.cpp;h=c1afafd765ca32f803c35c23b821c949bb807e91;hb=e0cb348ca42ae7057f8f5acee92a23e7eb26075f;hp=059a3410586285ff230d6ebcc469846f428a4dc3;hpb=52336c086b8bc355b55e2046e3a055b1b4c70ef7;p=nageru

diff --git a/futatabi/jpeg_frame_view.cpp b/futatabi/jpeg_frame_view.cpp
index 059a341..c1afafd 100644
--- a/futatabi/jpeg_frame_view.cpp
+++ b/futatabi/jpeg_frame_view.cpp
@@ -4,6 +4,7 @@
 #include "flags.h"
 #include "jpeg_destroyer.h"
 #include "jpeglib_error_wrapper.h"
+#include "shared/metrics.h"
 #include "shared/post_to_main_thread.h"
 #include "video_stream.h"
 #include "ycbcr_converter.h"
@@ -32,9 +33,8 @@ using namespace std;
 namespace {
 
 // Just an arbitrary order for std::map.
-struct FrameOnDiskLexicalOrder
-{
-	bool operator() (const FrameOnDisk &a, const FrameOnDisk &b) const
+struct FrameOnDiskLexicalOrder {
+	bool operator()(const FrameOnDisk &a, const FrameOnDisk &b) const
 	{
 		if (a.pts != b.pts)
 			return a.pts < b.pts;
@@ -59,28 +59,24 @@ 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;
-};
+// There can be multiple JPEGFrameView instances, so make all the metrics static.
+once_flag jpeg_metrics_inited;
+atomic<int64_t> metric_jpeg_cache_used_bytes{ 0 };  // Same value as cache_bytes_used.
+atomic<int64_t> metric_jpeg_cache_limit_bytes{ size_t(CACHE_SIZE_MB) * 1024 * 1024 };
+atomic<int64_t> metric_jpeg_cache_given_up_frames{ 0 };
+atomic<int64_t> metric_jpeg_cache_hit_frames{ 0 };
+atomic<int64_t> metric_jpeg_cache_miss_frames{ 0 };
+atomic<int64_t> metric_jpeg_software_decode_frames{ 0 };
+atomic<int64_t> metric_jpeg_software_fail_frames{ 0 };
+atomic<int64_t> metric_jpeg_vaapi_decode_frames{ 0 };
+atomic<int64_t> metric_jpeg_vaapi_fail_frames{ 0 };
 
 }  // namespace
 
-thread JPEGFrameView::jpeg_decoder_thread;
 mutex 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<PendingDecode> pending_decodes;  // Under cache_mu.
-atomic<size_t> event_counter{0};
+atomic<size_t> event_counter{ 0 };
 extern QGLWidget *global_share_widget;
 extern atomic<bool> should_quit;
 
@@ -90,33 +86,35 @@ shared_ptr<Frame> decode_jpeg(const string &jpeg)
 	if (vaapi_jpeg_decoding_usable) {
 		frame = decode_jpeg_vaapi(jpeg);
 		if (frame != nullptr) {
+			++metric_jpeg_vaapi_decode_frames;
 			return frame;
 		}
 		fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
+		++metric_jpeg_vaapi_fail_frames;
 	}
 
 	frame.reset(new Frame);
 
 	jpeg_decompress_struct dinfo;
 	JPEGWrapErrorManager error_mgr(&dinfo);
-	if (!error_mgr.run([&dinfo]{ jpeg_create_decompress(&dinfo); })) {
+	if (!error_mgr.run([&dinfo] { jpeg_create_decompress(&dinfo); })) {
 		return get_black_frame();
 	}
 	JPEGDestroyer destroy_dinfo(&dinfo);
 
-	if (!error_mgr.run([&dinfo, &jpeg]{
-		jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
-		jpeg_read_header(&dinfo, true);
-	})) {
+	if (!error_mgr.run([&dinfo, &jpeg] {
+		    jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
+		    jpeg_read_header(&dinfo, true);
+	    })) {
 		return get_black_frame();
 	}
 
 	if (dinfo.num_components != 3) {
 		fprintf(stderr, "Not a color 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);
+		        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);
 		return get_black_frame();
 	}
 	if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
@@ -126,16 +124,16 @@ shared_ptr<Frame> decode_jpeg(const string &jpeg)
 	    (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
 	    (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
 		fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
-			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.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);
+		abort();
 	}
 	dinfo.raw_data_out = true;
 
-	if (!error_mgr.run([&dinfo]{
-		jpeg_start_decompress(&dinfo);
-	})) {
+	if (!error_mgr.run([&dinfo] {
+		    jpeg_start_decompress(&dinfo);
+	    })) {
 		return get_black_frame();
 	}
 
@@ -162,24 +160,25 @@ shared_ptr<Frame> decode_jpeg(const string &jpeg)
 	frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
 
 	if (!error_mgr.run([&dinfo, &frame, v_mcu_size, mcu_height_blocks] {
-		JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
-		JSAMPARRAY data[3] = { yptr, cbptr, crptr };
-		for (unsigned y = 0; y < mcu_height_blocks; ++y) {
-			// NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
-			for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
-				yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
-				cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
-				crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
-			}
-
-			jpeg_read_raw_data(&dinfo, data, v_mcu_size);
-		}
-
-		(void)jpeg_finish_decompress(&dinfo);
-	})) {
+		    JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
+		    JSAMPARRAY data[3] = { yptr, cbptr, crptr };
+		    for (unsigned y = 0; y < mcu_height_blocks; ++y) {
+			    // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
+			    for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
+				    yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
+				    cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
+				    crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
+			    }
+
+			    jpeg_read_raw_data(&dinfo, data, v_mcu_size);
+		    }
+
+		    (void)jpeg_finish_decompress(&dinfo);
+	    })) {
 		return get_black_frame();
 	}
 
+	++metric_jpeg_software_decode_frames;
 	return frame;
 }
 
@@ -187,7 +186,8 @@ void prune_cache()
 {
 	// Assumes cache_mu is held.
 	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;
+	if (bytes_still_to_remove <= 0)
+		return;
 
 	vector<pair<size_t, size_t>> lru_timestamps_and_size;
 	for (const auto &key_and_value : cache) {
@@ -202,12 +202,14 @@ void prune_cache()
 	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;
+		if (bytes_still_to_remove <= 0)
+			break;
 	}
 
-	for (auto it = cache.begin(); it != cache.end(); ) {
+	for (auto it = cache.begin(); it != cache.end();) {
 		if (it->second.last_used <= lru_cutoff_point) {
 			cache_bytes_used -= frame_size(*it->second.frame);
+			metric_jpeg_cache_used_bytes = cache_bytes_used;
 			it = cache.erase(it);
 		} else {
 			++it;
@@ -219,23 +221,28 @@ shared_ptr<Frame> decode_jpeg_with_cache(FrameOnDisk frame_spec, CacheMissBehavi
 {
 	*did_decode = false;
 	{
-		unique_lock<mutex> lock(cache_mu);
+		lock_guard<mutex> lock(cache_mu);
 		auto it = cache.find(frame_spec);
 		if (it != cache.end()) {
+			++metric_jpeg_cache_hit_frames;
 			it->second.last_used = event_counter++;
 			return it->second.frame;
 		}
 	}
 
 	if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
+		++metric_jpeg_cache_given_up_frames;
 		return nullptr;
 	}
 
+	++metric_jpeg_cache_miss_frames;
+
 	*did_decode = true;
-	shared_ptr<Frame> frame = decode_jpeg(frame_reader->read_frame(frame_spec));
+	shared_ptr<Frame> frame = decode_jpeg(frame_reader->read_frame(frame_spec, /*read_audio=*/false).video);
 
-	unique_lock<mutex> lock(cache_mu);
+	lock_guard<mutex> lock(cache_mu);
 	cache_bytes_used += frame_size(*frame);
+	metric_jpeg_cache_used_bytes = cache_bytes_used;
 	cache[frame_spec] = LRUFrame{ frame, event_counter++ };
 
 	if (cache_bytes_used > size_t(CACHE_SIZE_MB) * 1024 * 1024) {
@@ -254,7 +261,7 @@ void JPEGFrameView::jpeg_decoder_thread_func()
 		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, [] {
+			any_pending_decodes.wait(lock, [this] {
 				return !pending_decodes.empty() || should_quit.load();
 			});
 			if (should_quit.load())
@@ -262,20 +269,14 @@ void JPEGFrameView::jpeg_decoder_thread_func()
 			decode = pending_decodes.front();
 			pending_decodes.pop_front();
 
-			size_t num_pending = 0;
-			for (const PendingDecode &other_decode : pending_decodes) {
-				if (other_decode.destination == decode.destination) {
-					++num_pending;
-				}
-			}
-			if (num_pending > 3) {
+			if (pending_decodes.size() > 3) {
 				cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
 			}
 		}
 
 		if (decode.frame != nullptr) {
 			// Already decoded, so just show it.
-			decode.destination->setDecodedFrame(decode.frame, nullptr, 1.0f);
+			setDecodedFrame(decode.frame, nullptr, 1.0f);
 			continue;
 		}
 
@@ -289,7 +290,7 @@ void JPEGFrameView::jpeg_decoder_thread_func()
 			}
 
 			bool found_in_cache;
-			shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, cache_miss_behavior, &decode.destination->frame_reader, &found_in_cache);
+			shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, cache_miss_behavior, &frame_reader, &found_in_cache);
 
 			if (frame == nullptr) {
 				assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
@@ -301,7 +302,7 @@ void JPEGFrameView::jpeg_decoder_thread_func()
 				++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));
+					        num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
 				}
 			}
 			if (subframe_idx == 0) {
@@ -316,11 +317,11 @@ void JPEGFrameView::jpeg_decoder_thread_func()
 		}
 
 		// TODO: Could we get jitter between non-interpolated and interpolated frames here?
-		decode.destination->setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
+		setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
 	}
 }
 
-void JPEGFrameView::shutdown()
+JPEGFrameView::~JPEGFrameView()
 {
 	any_pending_decodes.notify_all();
 	jpeg_decoder_thread.join();
@@ -329,45 +330,49 @@ void JPEGFrameView::shutdown()
 JPEGFrameView::JPEGFrameView(QWidget *parent)
 	: QGLWidget(parent, global_share_widget)
 {
+	call_once(jpeg_metrics_inited, [] {
+		global_metrics.add("jpeg_cache_used_bytes", &metric_jpeg_cache_used_bytes, Metrics::TYPE_GAUGE);
+		global_metrics.add("jpeg_cache_limit_bytes", &metric_jpeg_cache_limit_bytes, Metrics::TYPE_GAUGE);
+		global_metrics.add("jpeg_cache_frames", { { "action", "given_up" } }, &metric_jpeg_cache_given_up_frames);
+		global_metrics.add("jpeg_cache_frames", { { "action", "hit" } }, &metric_jpeg_cache_hit_frames);
+		global_metrics.add("jpeg_cache_frames", { { "action", "miss" } }, &metric_jpeg_cache_miss_frames);
+		global_metrics.add("jpeg_decode_frames", { { "decoder", "software" }, { "result", "decode" } }, &metric_jpeg_software_decode_frames);
+		global_metrics.add("jpeg_decode_frames", { { "decoder", "software" }, { "result", "fail" } }, &metric_jpeg_software_fail_frames);
+		global_metrics.add("jpeg_decode_frames", { { "decoder", "vaapi" }, { "result", "decode" } }, &metric_jpeg_vaapi_decode_frames);
+		global_metrics.add("jpeg_decode_frames", { { "decoder", "vaapi" }, { "result", "fail" } }, &metric_jpeg_vaapi_fail_frames);
+	});
 }
 
 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);
+	lock_guard<mutex> lock(cache_mu);
 	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);
+	lock_guard<mutex> lock(cache_mu);
 	PendingDecode decode;
 	decode.frame = std::move(frame);
-	decode.destination = this;
 	pending_decodes.push_back(decode);
 	any_pending_decodes.notify_all();
 }
 
-ResourcePool *resource_pool = nullptr;
-
 void JPEGFrameView::initializeGL()
 {
 	glDisable(GL_BLEND);
 	glDisable(GL_DEPTH_TEST);
 	check_error();
 
-	static once_flag once;
-	call_once(once, [] {
-		resource_pool = new ResourcePool;
-		jpeg_decoder_thread = std::thread(jpeg_decoder_thread_func);
-	});
+	resource_pool = new ResourcePool;
+	jpeg_decoder_thread = std::thread(&JPEGFrameView::jpeg_decoder_thread_func, this);
 
 	ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
 
@@ -485,5 +490,6 @@ shared_ptr<Frame> get_black_frame()
 		black_frame->pitch_y = global_flags.width;
 		black_frame->pitch_chroma = global_flags.width / 2;
 	});
+	++metric_jpeg_software_fail_frames;
 	return black_frame;
 }