]> git.sesse.net Git - nageru/blob - futatabi/jpeg_frame_view.cpp
Add a Futatabi metric for JPEG decode times.
[nageru] / futatabi / jpeg_frame_view.cpp
1 #include "jpeg_frame_view.h"
2
3 #include "defs.h"
4 #include "flags.h"
5 #include "jpeg_destroyer.h"
6 #include "jpeglib_error_wrapper.h"
7 #include "shared/metrics.h"
8 #include "shared/post_to_main_thread.h"
9 #include "video_stream.h"
10 #include "ycbcr_converter.h"
11
12 #include <QMouseEvent>
13 #include <QScreen>
14 #include <atomic>
15 #include <chrono>
16 #include <condition_variable>
17 #include <deque>
18 #include <jpeglib.h>
19 #include <movit/init.h>
20 #include <movit/resource_pool.h>
21 #include <movit/util.h>
22 #include <mutex>
23 #include <stdint.h>
24 #include <thread>
25 #include <unistd.h>
26 #include <utility>
27
28 // Must come after the Qt stuff.
29 #include "vaapi_jpeg_decoder.h"
30
31 using namespace movit;
32 using namespace std;
33 using namespace std::chrono;
34
35 namespace {
36
37 // Just an arbitrary order for std::map.
38 struct FrameOnDiskLexicalOrder {
39         bool operator()(const FrameOnDisk &a, const FrameOnDisk &b) const
40         {
41                 if (a.pts != b.pts)
42                         return a.pts < b.pts;
43                 if (a.offset != b.offset)
44                         return a.offset < b.offset;
45                 if (a.filename_idx != b.filename_idx)
46                         return a.filename_idx < b.filename_idx;
47                 assert(a.size == b.size);
48                 return false;
49         }
50 };
51
52 inline size_t frame_size(const Frame &frame)
53 {
54         size_t y_size = frame.width * frame.height;
55         size_t cbcr_size = y_size / frame.chroma_subsampling_x / frame.chroma_subsampling_y;
56         return y_size + cbcr_size * 2;
57 }
58
59 struct LRUFrame {
60         shared_ptr<Frame> frame;
61         size_t last_used;
62 };
63
64 // There can be multiple JPEGFrameView instances, so make all the metrics static.
65 once_flag jpeg_metrics_inited;
66 atomic<int64_t> metric_jpeg_cache_used_bytes{ 0 };  // Same value as cache_bytes_used.
67 atomic<int64_t> metric_jpeg_cache_limit_bytes{ size_t(CACHE_SIZE_MB) * 1024 * 1024 };
68 atomic<int64_t> metric_jpeg_cache_given_up_frames{ 0 };
69 atomic<int64_t> metric_jpeg_cache_hit_frames{ 0 };
70 atomic<int64_t> metric_jpeg_cache_miss_frames{ 0 };
71 atomic<int64_t> metric_jpeg_software_decode_frames{ 0 };
72 atomic<int64_t> metric_jpeg_software_fail_frames{ 0 };
73 atomic<int64_t> metric_jpeg_vaapi_decode_frames{ 0 };
74 atomic<int64_t> metric_jpeg_vaapi_fail_frames{ 0 };
75 Summary metric_jpeg_decode_time_seconds;
76
77 }  // namespace
78
79 mutex cache_mu;
80 map<FrameOnDisk, LRUFrame, FrameOnDiskLexicalOrder> cache;  // Under cache_mu.
81 size_t cache_bytes_used = 0;  // Under cache_mu.
82 atomic<size_t> event_counter{ 0 };
83 extern QGLWidget *global_share_widget;
84 extern atomic<bool> should_quit;
85
86 shared_ptr<Frame> decode_jpeg(const string &jpeg)
87 {
88         steady_clock::time_point start = steady_clock::now();
89         shared_ptr<Frame> frame;
90         if (vaapi_jpeg_decoding_usable) {
91                 frame = decode_jpeg_vaapi(jpeg);
92                 if (frame != nullptr) {
93                         ++metric_jpeg_vaapi_decode_frames;
94                         steady_clock::time_point stop = steady_clock::now();
95                         metric_jpeg_decode_time_seconds.count_event(duration<double>(stop - start).count());
96                         return frame;
97                 }
98                 fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
99                 ++metric_jpeg_vaapi_fail_frames;
100         }
101
102         frame.reset(new Frame);
103
104         jpeg_decompress_struct dinfo;
105         JPEGWrapErrorManager error_mgr(&dinfo);
106         if (!error_mgr.run([&dinfo] { jpeg_create_decompress(&dinfo); })) {
107                 return get_black_frame();
108         }
109         JPEGDestroyer destroy_dinfo(&dinfo);
110
111         if (!error_mgr.run([&dinfo, &jpeg] {
112                     jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
113                     jpeg_read_header(&dinfo, true);
114             })) {
115                 return get_black_frame();
116         }
117
118         jpeg_save_markers(&dinfo, JPEG_APP0 + 1, 0xFFFF);
119
120         if (dinfo.num_components != 3) {
121                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
122                         dinfo.num_components,
123                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
124                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
125                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
126                 return get_black_frame();
127         }
128         if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
129             dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor ||  // Y' must not be subsampled.
130             dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
131             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor ||  // Cb and Cr must be identically subsampled.
132             (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
133             (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) {  // No 2:3 subsampling or other weirdness.
134                 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
135                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
136                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
137                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
138                 abort();
139         }
140         dinfo.raw_data_out = true;
141
142         if (!error_mgr.run([&dinfo] {
143                     jpeg_start_decompress(&dinfo);
144             })) {
145                 return get_black_frame();
146         }
147
148         frame->width = dinfo.output_width;
149         frame->height = dinfo.output_height;
150         frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
151         frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
152
153         unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
154         unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
155         unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
156         unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
157
158         unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
159         unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
160         unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
161         unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
162
163         // TODO: Decode into a PBO.
164         frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
165         frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
166         frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
167         frame->pitch_y = luma_width_blocks * DCTSIZE;
168         frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
169
170         if (dinfo.marker_list != nullptr &&
171             dinfo.marker_list->marker == JPEG_APP0 + 1 &&
172             dinfo.marker_list->data_length >= 4 &&
173             memcmp(dinfo.marker_list->data, "Exif", 4) == 0) {
174                 frame->exif_data.assign(reinterpret_cast<char *>(dinfo.marker_list->data),
175                         dinfo.marker_list->data_length);
176         }
177
178         if (!error_mgr.run([&dinfo, &frame, v_mcu_size, mcu_height_blocks] {
179                     JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
180                     JSAMPARRAY data[3] = { yptr, cbptr, crptr };
181                     for (unsigned y = 0; y < mcu_height_blocks; ++y) {
182                             // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
183                             for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
184                                     yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
185                                     cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
186                                     crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
187                             }
188
189                             jpeg_read_raw_data(&dinfo, data, v_mcu_size);
190                     }
191
192                     (void)jpeg_finish_decompress(&dinfo);
193             })) {
194                 return get_black_frame();
195         }
196
197         ++metric_jpeg_software_decode_frames;
198         steady_clock::time_point stop = steady_clock::now();
199         metric_jpeg_decode_time_seconds.count_event(duration<double>(stop - start).count());
200         return frame;
201 }
202
203 void prune_cache()
204 {
205         // Assumes cache_mu is held.
206         int64_t bytes_still_to_remove = cache_bytes_used - (size_t(CACHE_SIZE_MB) * 1024 * 1024) * 9 / 10;
207         if (bytes_still_to_remove <= 0)
208                 return;
209
210         vector<pair<size_t, size_t>> lru_timestamps_and_size;
211         for (const auto &key_and_value : cache) {
212                 lru_timestamps_and_size.emplace_back(
213                         key_and_value.second.last_used,
214                         frame_size(*key_and_value.second.frame));
215         }
216         sort(lru_timestamps_and_size.begin(), lru_timestamps_and_size.end());
217
218         // Remove the oldest ones until we are below 90% of the cache used.
219         size_t lru_cutoff_point = 0;
220         for (const pair<size_t, size_t> &it : lru_timestamps_and_size) {
221                 lru_cutoff_point = it.first;
222                 bytes_still_to_remove -= it.second;
223                 if (bytes_still_to_remove <= 0)
224                         break;
225         }
226
227         for (auto it = cache.begin(); it != cache.end();) {
228                 if (it->second.last_used <= lru_cutoff_point) {
229                         cache_bytes_used -= frame_size(*it->second.frame);
230                         metric_jpeg_cache_used_bytes = cache_bytes_used;
231                         it = cache.erase(it);
232                 } else {
233                         ++it;
234                 }
235         }
236 }
237
238 shared_ptr<Frame> decode_jpeg_with_cache(FrameOnDisk frame_spec, CacheMissBehavior cache_miss_behavior, FrameReader *frame_reader, bool *did_decode)
239 {
240         *did_decode = false;
241         {
242                 lock_guard<mutex> lock(cache_mu);
243                 auto it = cache.find(frame_spec);
244                 if (it != cache.end()) {
245                         ++metric_jpeg_cache_hit_frames;
246                         it->second.last_used = event_counter++;
247                         return it->second.frame;
248                 }
249         }
250
251         if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
252                 ++metric_jpeg_cache_given_up_frames;
253                 return nullptr;
254         }
255
256         ++metric_jpeg_cache_miss_frames;
257
258         *did_decode = true;
259         shared_ptr<Frame> frame = decode_jpeg(frame_reader->read_frame(frame_spec, /*read_video=*/true, /*read_audio=*/false).video);
260
261         lock_guard<mutex> lock(cache_mu);
262         cache_bytes_used += frame_size(*frame);
263         metric_jpeg_cache_used_bytes = cache_bytes_used;
264         cache[frame_spec] = LRUFrame{ frame, event_counter++ };
265
266         if (cache_bytes_used > size_t(CACHE_SIZE_MB) * 1024 * 1024) {
267                 prune_cache();
268         }
269         return frame;
270 }
271
272 void JPEGFrameView::jpeg_decoder_thread_func()
273 {
274         size_t num_decoded = 0, num_dropped = 0;
275
276         pthread_setname_np(pthread_self(), "JPEGDecoder");
277         while (!should_quit.load()) {
278                 PendingDecode decode;
279                 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
280                 {
281                         unique_lock<mutex> lock(cache_mu);  // TODO: Perhaps under another lock?
282                         any_pending_decodes.wait(lock, [this] {
283                                 return !pending_decodes.empty() || should_quit.load();
284                         });
285                         if (should_quit.load())
286                                 break;
287                         decode = pending_decodes.front();
288                         pending_decodes.pop_front();
289
290                         if (pending_decodes.size() > 3) {
291                                 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
292                         }
293                 }
294
295                 if (decode.frame != nullptr) {
296                         // Already decoded, so just show it.
297                         setDecodedFrame(decode.frame, nullptr, 1.0f);
298                         continue;
299                 }
300
301                 shared_ptr<Frame> primary_frame, secondary_frame;
302                 bool drop = false;
303                 for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
304                         const FrameOnDisk &frame_spec = (subframe_idx == 0 ? decode.primary : decode.secondary);
305                         if (frame_spec.pts == -1) {
306                                 // No secondary frame.
307                                 continue;
308                         }
309
310                         bool found_in_cache;
311                         shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, cache_miss_behavior, &frame_reader, &found_in_cache);
312
313                         if (frame == nullptr) {
314                                 assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
315                                 drop = true;
316                                 break;
317                         }
318
319                         if (!found_in_cache) {
320                                 ++num_decoded;
321                                 if (num_decoded % 1000 == 0) {
322                                         fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
323                                                 num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
324                                 }
325                         }
326                         if (subframe_idx == 0) {
327                                 primary_frame = std::move(frame);
328                         } else {
329                                 secondary_frame = std::move(frame);
330                         }
331                 }
332                 if (drop) {
333                         ++num_dropped;
334                         continue;
335                 }
336
337                 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
338                 setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
339         }
340 }
341
342 JPEGFrameView::~JPEGFrameView()
343 {
344         any_pending_decodes.notify_all();
345         jpeg_decoder_thread.join();
346 }
347
348 JPEGFrameView::JPEGFrameView(QWidget *parent)
349         : QGLWidget(parent, global_share_widget)
350 {
351         call_once(jpeg_metrics_inited, [] {
352                 global_metrics.add("jpeg_cache_used_bytes", &metric_jpeg_cache_used_bytes, Metrics::TYPE_GAUGE);
353                 global_metrics.add("jpeg_cache_limit_bytes", &metric_jpeg_cache_limit_bytes, Metrics::TYPE_GAUGE);
354                 global_metrics.add("jpeg_cache_frames", { { "action", "given_up" } }, &metric_jpeg_cache_given_up_frames);
355                 global_metrics.add("jpeg_cache_frames", { { "action", "hit" } }, &metric_jpeg_cache_hit_frames);
356                 global_metrics.add("jpeg_cache_frames", { { "action", "miss" } }, &metric_jpeg_cache_miss_frames);
357                 global_metrics.add("jpeg_decode_frames", { { "decoder", "software" }, { "result", "decode" } }, &metric_jpeg_software_decode_frames);
358                 global_metrics.add("jpeg_decode_frames", { { "decoder", "software" }, { "result", "fail" } }, &metric_jpeg_software_fail_frames);
359                 global_metrics.add("jpeg_decode_frames", { { "decoder", "vaapi" }, { "result", "decode" } }, &metric_jpeg_vaapi_decode_frames);
360                 global_metrics.add("jpeg_decode_frames", { { "decoder", "vaapi" }, { "result", "fail" } }, &metric_jpeg_vaapi_fail_frames);
361                 vector<double> quantiles{ 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99 };
362                 metric_jpeg_decode_time_seconds.init(quantiles, 60.0);
363                 global_metrics.add("jpeg_decode_time_seconds", &metric_jpeg_decode_time_seconds);
364         });
365 }
366
367 void JPEGFrameView::setFrame(unsigned stream_idx, FrameOnDisk frame, FrameOnDisk secondary_frame, float fade_alpha)
368 {
369         current_stream_idx = stream_idx;  // TODO: Does this interact with fades?
370
371         lock_guard<mutex> lock(cache_mu);
372         PendingDecode decode;
373         decode.primary = frame;
374         decode.secondary = secondary_frame;
375         decode.fade_alpha = fade_alpha;
376         pending_decodes.push_back(decode);
377         any_pending_decodes.notify_all();
378 }
379
380 void JPEGFrameView::setFrame(shared_ptr<Frame> frame)
381 {
382         lock_guard<mutex> lock(cache_mu);
383         PendingDecode decode;
384         decode.frame = std::move(frame);
385         pending_decodes.push_back(decode);
386         any_pending_decodes.notify_all();
387 }
388
389 void JPEGFrameView::initializeGL()
390 {
391         glDisable(GL_BLEND);
392         glDisable(GL_DEPTH_TEST);
393         check_error();
394
395         resource_pool = new ResourcePool;
396         jpeg_decoder_thread = std::thread(&JPEGFrameView::jpeg_decoder_thread_func, this);
397
398         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
399
400         ImageFormat inout_format;
401         inout_format.color_space = COLORSPACE_sRGB;
402         inout_format.gamma_curve = GAMMA_sRGB;
403
404         overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
405         overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
406
407         overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
408         overlay_chain->finalize();
409 }
410
411 void JPEGFrameView::resizeGL(int width, int height)
412 {
413         check_error();
414         glViewport(0, 0, width, height);
415         check_error();
416
417         // Save these, as width() and height() will lie with DPI scaling.
418         gl_width = width;
419         gl_height = height;
420 }
421
422 void JPEGFrameView::paintGL()
423 {
424         glViewport(0, 0, gl_width, gl_height);
425         if (current_frame == nullptr) {
426                 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
427                 glClear(GL_COLOR_BUFFER_BIT);
428                 return;
429         }
430
431         check_error();
432         current_chain->render_to_screen();
433
434         if (overlay_image != nullptr) {
435                 if (overlay_input_needs_refresh) {
436                         overlay_input->set_width(overlay_width);
437                         overlay_input->set_height(overlay_height);
438                         overlay_input->set_pixel_data(overlay_image->bits());
439                         overlay_input_needs_refresh = false;
440                 }
441                 glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
442                 overlay_chain->render_to_screen();
443         }
444 }
445
446 namespace {
447
448 }  // namespace
449
450 void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
451 {
452         post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
453                 current_frame = frame;
454                 current_secondary_frame = secondary_frame;
455
456                 if (secondary_frame != nullptr) {
457                         current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
458                 } else {
459                         current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
460                 }
461                 update();
462         });
463 }
464
465 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
466 {
467         if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
468                 emit clicked();
469         }
470 }
471
472 void JPEGFrameView::set_overlay(const string &text)
473 {
474         if (text.empty()) {
475                 overlay_image.reset();
476                 return;
477         }
478
479         // Figure out how large the texture needs to be.
480         {
481                 QImage img(overlay_width, overlay_height, QImage::Format_Grayscale8);
482                 QPainter painter(&img);
483                 QFont font = painter.font();
484                 font.setPointSize(12);
485                 QFontMetrics metrics(font);
486                 overlay_base_width = lrint(metrics.boundingRect(QString::fromStdString(text)).width() + 8.0);
487                 overlay_base_height = lrint(metrics.height());
488         }
489
490         float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
491         overlay_width = lrint(overlay_base_width * dpr);
492         overlay_height = lrint(overlay_base_height * dpr);
493
494         // Work around OpenGL alignment issues.
495         while (overlay_width % 4 != 0) ++overlay_width;
496
497         // Now do the actual drawing.
498         overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
499         overlay_image->setDevicePixelRatio(dpr);
500         overlay_image->fill(0);
501         QPainter painter(overlay_image.get());
502
503         painter.setPen(Qt::white);
504         QFont font = painter.font();
505         font.setPointSize(12);
506         painter.setFont(font);
507
508         painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
509
510         // Don't refresh immediately; we might not have an OpenGL context here.
511         overlay_input_needs_refresh = true;
512 }
513
514 shared_ptr<Frame> get_black_frame()
515 {
516         static shared_ptr<Frame> black_frame;
517         static once_flag flag;
518         call_once(flag, [] {
519                 black_frame.reset(new Frame);
520                 black_frame->y.reset(new uint8_t[global_flags.width * global_flags.height]);
521                 black_frame->cb.reset(new uint8_t[(global_flags.width / 2) * (global_flags.height / 2)]);
522                 black_frame->cr.reset(new uint8_t[(global_flags.width / 2) * (global_flags.height / 2)]);
523                 black_frame->width = global_flags.width;
524                 black_frame->height = global_flags.height;
525                 black_frame->chroma_subsampling_x = 2;
526                 black_frame->chroma_subsampling_y = 2;
527                 black_frame->pitch_y = global_flags.width;
528                 black_frame->pitch_chroma = global_flags.width / 2;
529         });
530         ++metric_jpeg_software_fail_frames;
531         return black_frame;
532 }