1 #include "jpeg_frame_view.h"
5 #include "jpeg_destroyer.h"
6 #include "jpeglib_error_wrapper.h"
8 #include "shared/context.h"
9 #include "shared/metrics.h"
10 #include "shared/post_to_main_thread.h"
11 #include "video_stream.h"
12 #include "ycbcr_converter.h"
14 #include <QMouseEvent>
18 #include <condition_variable>
21 #include <movit/init.h>
22 #include <movit/resource_pool.h>
23 #include <movit/util.h>
30 // Must come after the Qt stuff.
31 #include "vaapi_jpeg_decoder.h"
33 using namespace movit;
35 using namespace std::chrono;
39 // Just an arbitrary order for std::map.
40 struct FrameOnDiskLexicalOrder {
41 bool operator()(const FrameOnDisk &a, const FrameOnDisk &b) const
45 if (a.offset != b.offset)
46 return a.offset < b.offset;
47 if (a.filename_idx != b.filename_idx)
48 return a.filename_idx < b.filename_idx;
49 assert(a.size == b.size);
54 inline size_t frame_size(const Frame &frame)
56 size_t y_size = frame.width * frame.height;
57 size_t cbcr_size = y_size / frame.chroma_subsampling_x / frame.chroma_subsampling_y;
58 return y_size + cbcr_size * 2;
62 shared_ptr<Frame> frame;
66 // There can be multiple JPEGFrameView instances, so make all the metrics static.
67 once_flag jpeg_metrics_inited;
68 atomic<int64_t> metric_jpeg_cache_used_bytes{ 0 }; // Same value as cache_bytes_used.
69 atomic<int64_t> metric_jpeg_cache_limit_bytes{ size_t(CACHE_SIZE_MB) * 1024 * 1024 };
70 atomic<int64_t> metric_jpeg_cache_given_up_frames{ 0 };
71 atomic<int64_t> metric_jpeg_cache_hit_frames{ 0 };
72 atomic<int64_t> metric_jpeg_cache_miss_frames{ 0 };
73 atomic<int64_t> metric_jpeg_software_decode_frames{ 0 };
74 atomic<int64_t> metric_jpeg_software_fail_frames{ 0 };
75 atomic<int64_t> metric_jpeg_vaapi_decode_frames{ 0 };
76 atomic<int64_t> metric_jpeg_vaapi_fail_frames{ 0 };
77 atomic<int64_t> metric_jpeg_prepared_frames{ 0 };
78 atomic<int64_t> metric_jpeg_displayed_frames{ 0 };
79 Summary metric_jpeg_decode_time_seconds;
84 map<FrameOnDisk, LRUFrame, FrameOnDiskLexicalOrder> cache; // Under cache_mu.
85 size_t cache_bytes_used = 0; // Under cache_mu.
86 atomic<size_t> event_counter{ 0 };
87 extern QGLWidget *global_share_widget;
88 extern atomic<bool> should_quit;
90 shared_ptr<Frame> decode_jpeg(const string &jpeg)
92 steady_clock::time_point start = steady_clock::now();
93 shared_ptr<Frame> frame;
94 if (vaapi_jpeg_decoding_usable) {
95 frame = decode_jpeg_vaapi(jpeg);
96 if (frame != nullptr) {
97 ++metric_jpeg_vaapi_decode_frames;
98 steady_clock::time_point stop = steady_clock::now();
99 metric_jpeg_decode_time_seconds.count_event(duration<double>(stop - start).count());
102 fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
103 ++metric_jpeg_vaapi_fail_frames;
106 frame.reset(new Frame);
108 jpeg_decompress_struct dinfo;
109 JPEGWrapErrorManager error_mgr(&dinfo);
110 if (!error_mgr.run([&dinfo] { jpeg_create_decompress(&dinfo); })) {
111 return get_black_frame();
113 JPEGDestroyer destroy_dinfo(&dinfo);
115 jpeg_save_markers(&dinfo, JPEG_APP0 + 1, 0xFFFF);
117 if (!error_mgr.run([&dinfo, &jpeg] {
118 jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
119 jpeg_read_header(&dinfo, true);
121 return get_black_frame();
124 if (dinfo.num_components != 3) {
125 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
126 dinfo.num_components,
127 dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
128 dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
129 dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
130 return get_black_frame();
132 if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
133 dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor || // Y' must not be subsampled.
134 dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
135 dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor || // Cb and Cr must be identically subsampled.
136 (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
137 (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) { // No 2:3 subsampling or other weirdness.
138 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
139 dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
140 dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
141 dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
144 dinfo.raw_data_out = true;
146 if (!error_mgr.run([&dinfo] {
147 jpeg_start_decompress(&dinfo);
149 return get_black_frame();
152 frame->width = dinfo.output_width;
153 frame->height = dinfo.output_height;
154 frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
155 frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
157 unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
158 unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
159 unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
160 unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
162 unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
163 unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
165 PBO pbo = global_pbo_pool->alloc_pbo();
166 const size_t chroma_width = dinfo.image_width / frame->chroma_subsampling_x;
167 const size_t chroma_height = dinfo.image_height / frame->chroma_subsampling_y;
168 size_t cb_offset = dinfo.image_width * dinfo.image_height;
169 size_t cr_offset = cb_offset + chroma_width * chroma_height;
170 uint8_t *y_pix = pbo.ptr;
171 uint8_t *cb_pix = pbo.ptr + cb_offset;
172 uint8_t *cr_pix = pbo.ptr + cr_offset;
173 unsigned pitch_y = luma_width_blocks * DCTSIZE;
174 unsigned pitch_chroma = chroma_width_blocks * DCTSIZE;
176 if (dinfo.marker_list != nullptr &&
177 dinfo.marker_list->marker == JPEG_APP0 + 1 &&
178 dinfo.marker_list->data_length >= 4 &&
179 memcmp(dinfo.marker_list->data, "Exif", 4) == 0) {
180 frame->exif_data.assign(reinterpret_cast<char *>(dinfo.marker_list->data),
181 dinfo.marker_list->data_length);
184 if (!error_mgr.run([&dinfo, &y_pix, &cb_pix, &cr_pix, pitch_y, pitch_chroma, v_mcu_size, mcu_height_blocks] {
185 unique_ptr<JSAMPROW[]> yptr(new JSAMPROW[v_mcu_size]);
186 unique_ptr<JSAMPROW[]> cbptr(new JSAMPROW[v_mcu_size]);
187 unique_ptr<JSAMPROW[]> crptr(new JSAMPROW[v_mcu_size]);
188 JSAMPARRAY data[3] = { yptr.get(), cbptr.get(), crptr.get() };
189 for (unsigned y = 0; y < mcu_height_blocks; ++y) {
190 // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
191 for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
192 yptr[yy] = y_pix + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * pitch_y;
193 cbptr[yy] = cb_pix + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * pitch_chroma;
194 crptr[yy] = cr_pix + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * pitch_chroma;
197 jpeg_read_raw_data(&dinfo, data, v_mcu_size);
200 (void)jpeg_finish_decompress(&dinfo);
202 return get_black_frame();
205 // FIXME: what about resolutions that are not divisible by the block factor?
206 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo.pbo);
207 frame->y = create_texture_2d(frame->width, frame->height, GL_R8, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
208 frame->cb = create_texture_2d(chroma_width, chroma_height, GL_R8, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cb_offset));
209 frame->cr = create_texture_2d(chroma_width, chroma_height, GL_R8, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cr_offset));
210 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
212 glFlushMappedNamedBufferRange(pbo.pbo, 0, dinfo.image_width * dinfo.image_height + chroma_width * chroma_height * 2);
213 glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT);
214 pbo.upload_done = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
216 frame->uploaded_ui_thread = pbo.upload_done;
217 frame->uploaded_interpolation = pbo.upload_done;
218 global_pbo_pool->release_pbo(move(pbo));
220 ++metric_jpeg_software_decode_frames;
221 steady_clock::time_point stop = steady_clock::now();
222 metric_jpeg_decode_time_seconds.count_event(duration<double>(stop - start).count());
228 // Assumes cache_mu is held.
229 int64_t bytes_still_to_remove = cache_bytes_used - (size_t(CACHE_SIZE_MB) * 1024 * 1024) * 9 / 10;
230 if (bytes_still_to_remove <= 0)
233 vector<pair<size_t, size_t>> lru_timestamps_and_size;
234 for (const auto &key_and_value : cache) {
235 lru_timestamps_and_size.emplace_back(
236 key_and_value.second.last_used,
237 frame_size(*key_and_value.second.frame));
239 sort(lru_timestamps_and_size.begin(), lru_timestamps_and_size.end());
241 // Remove the oldest ones until we are below 90% of the cache used.
242 size_t lru_cutoff_point = 0;
243 for (const pair<size_t, size_t> &it : lru_timestamps_and_size) {
244 lru_cutoff_point = it.first;
245 bytes_still_to_remove -= it.second;
246 if (bytes_still_to_remove <= 0)
250 for (auto it = cache.begin(); it != cache.end();) {
251 if (it->second.last_used <= lru_cutoff_point) {
252 cache_bytes_used -= frame_size(*it->second.frame);
253 metric_jpeg_cache_used_bytes = cache_bytes_used;
254 it = cache.erase(it);
261 shared_ptr<Frame> decode_jpeg_with_cache(FrameOnDisk frame_spec, CacheMissBehavior cache_miss_behavior, FrameReader *frame_reader, bool *did_decode)
265 lock_guard<mutex> lock(cache_mu);
266 auto it = cache.find(frame_spec);
267 if (it != cache.end()) {
268 ++metric_jpeg_cache_hit_frames;
269 it->second.last_used = event_counter++;
270 return it->second.frame;
274 if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
275 ++metric_jpeg_cache_given_up_frames;
279 ++metric_jpeg_cache_miss_frames;
282 shared_ptr<Frame> frame = decode_jpeg(frame_reader->read_frame(frame_spec, /*read_video=*/true, /*read_audio=*/false).video);
284 lock_guard<mutex> lock(cache_mu);
285 cache_bytes_used += frame_size(*frame);
286 metric_jpeg_cache_used_bytes = cache_bytes_used;
287 cache[frame_spec] = LRUFrame{ frame, event_counter++ };
289 if (cache_bytes_used > size_t(CACHE_SIZE_MB) * 1024 * 1024) {
295 void JPEGFrameView::jpeg_decoder_thread_func()
297 size_t num_decoded = 0, num_dropped = 0;
299 pthread_setname_np(pthread_self(), "JPEGDecoder");
300 QSurface *surface = create_surface();
301 QOpenGLContext *context = create_context(surface);
302 bool ok = make_current(context, surface);
304 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
307 while (!should_quit.load()) {
308 PendingDecode decode;
309 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
311 unique_lock<mutex> lock(cache_mu); // TODO: Perhaps under another lock?
312 any_pending_decodes.wait(lock, [this] {
313 return !pending_decodes.empty() || should_quit.load();
315 if (should_quit.load())
317 decode = pending_decodes.front();
318 pending_decodes.pop_front();
320 if (pending_decodes.size() > 3) {
321 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
325 if (decode.frame != nullptr) {
326 // Already decoded, so just show it.
327 setDecodedFrame(decode.frame, nullptr, 1.0f);
331 shared_ptr<Frame> primary_frame, secondary_frame;
333 for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
334 const FrameOnDisk &frame_spec = (subframe_idx == 0 ? decode.primary : decode.secondary);
335 if (frame_spec.pts == -1) {
336 // No secondary frame.
341 shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, cache_miss_behavior, &frame_reader, &found_in_cache);
343 if (frame == nullptr) {
344 assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
349 if (!found_in_cache) {
351 if (num_decoded % 1000 == 0) {
352 fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
353 num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
356 if (subframe_idx == 0) {
357 primary_frame = std::move(frame);
359 secondary_frame = std::move(frame);
367 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
368 setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
372 JPEGFrameView::~JPEGFrameView()
374 any_pending_decodes.notify_all();
375 jpeg_decoder_thread.join();
378 JPEGFrameView::JPEGFrameView(QWidget *parent)
379 : QGLWidget(parent, global_share_widget)
381 call_once(jpeg_metrics_inited, [] {
382 global_metrics.add("jpeg_cache_used_bytes", &metric_jpeg_cache_used_bytes, Metrics::TYPE_GAUGE);
383 global_metrics.add("jpeg_cache_limit_bytes", &metric_jpeg_cache_limit_bytes, Metrics::TYPE_GAUGE);
384 global_metrics.add("jpeg_cache_frames", { { "action", "given_up" } }, &metric_jpeg_cache_given_up_frames);
385 global_metrics.add("jpeg_cache_frames", { { "action", "hit" } }, &metric_jpeg_cache_hit_frames);
386 global_metrics.add("jpeg_cache_frames", { { "action", "miss" } }, &metric_jpeg_cache_miss_frames);
387 global_metrics.add("jpeg_decode_frames", { { "decoder", "software" }, { "result", "decode" } }, &metric_jpeg_software_decode_frames);
388 global_metrics.add("jpeg_decode_frames", { { "decoder", "software" }, { "result", "fail" } }, &metric_jpeg_software_fail_frames);
389 global_metrics.add("jpeg_decode_frames", { { "decoder", "vaapi" }, { "result", "decode" } }, &metric_jpeg_vaapi_decode_frames);
390 global_metrics.add("jpeg_decode_frames", { { "decoder", "vaapi" }, { "result", "fail" } }, &metric_jpeg_vaapi_fail_frames);
391 global_metrics.add("jpeg_frames", { { "action", "prepared" } }, &metric_jpeg_prepared_frames);
392 global_metrics.add("jpeg_frames", { { "action", "displayed" } }, &metric_jpeg_displayed_frames);
393 vector<double> quantiles{ 0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99 };
394 metric_jpeg_decode_time_seconds.init(quantiles, 60.0);
395 global_metrics.add("jpeg_decode_time_seconds", &metric_jpeg_decode_time_seconds);
399 void JPEGFrameView::setFrame(unsigned stream_idx, FrameOnDisk frame, FrameOnDisk secondary_frame, float fade_alpha)
401 current_stream_idx = stream_idx; // TODO: Does this interact with fades?
403 lock_guard<mutex> lock(cache_mu);
404 PendingDecode decode;
405 decode.primary = frame;
406 decode.secondary = secondary_frame;
407 decode.fade_alpha = fade_alpha;
408 pending_decodes.push_back(decode);
409 any_pending_decodes.notify_all();
412 void JPEGFrameView::setFrame(shared_ptr<Frame> frame)
414 lock_guard<mutex> lock(cache_mu);
415 PendingDecode decode;
416 decode.frame = std::move(frame);
417 decode.fade_alpha = 0.0f;
418 pending_decodes.push_back(decode);
419 any_pending_decodes.notify_all();
422 void JPEGFrameView::initializeGL()
427 glDisable(GL_DEPTH_TEST);
430 resource_pool = new ResourcePool;
431 jpeg_decoder_thread = std::thread(&JPEGFrameView::jpeg_decoder_thread_func, this);
433 ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
435 ImageFormat inout_format;
436 inout_format.color_space = COLORSPACE_sRGB;
437 inout_format.gamma_curve = GAMMA_sRGB;
439 overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
440 overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
442 overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
443 overlay_chain->finalize();
446 void JPEGFrameView::resizeGL(int width, int height)
449 glViewport(0, 0, width, height);
452 // Save these, as width() and height() will lie with DPI scaling.
457 void JPEGFrameView::paintGL()
459 glViewport(0, 0, gl_width, gl_height);
460 if (current_frame == nullptr) {
461 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
462 glClear(GL_COLOR_BUFFER_BIT);
466 if (!displayed_this_frame) {
467 ++metric_jpeg_displayed_frames;
468 displayed_this_frame = true;
470 if (current_frame->uploaded_ui_thread != nullptr) {
471 glWaitSync(current_frame->uploaded_ui_thread.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
472 current_frame->uploaded_ui_thread.reset();
474 if (current_secondary_frame != nullptr && current_secondary_frame->uploaded_ui_thread != nullptr) {
475 glWaitSync(current_secondary_frame->uploaded_ui_thread.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
476 current_secondary_frame->uploaded_ui_thread.reset();
480 current_chain->render_to_screen();
482 if (overlay_image != nullptr) {
483 if (overlay_input_needs_refresh) {
484 overlay_input->set_width(overlay_width);
485 overlay_input->set_height(overlay_height);
486 overlay_input->set_pixel_data(overlay_image->bits());
487 overlay_input_needs_refresh = false;
489 glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
490 overlay_chain->render_to_screen();
498 void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
500 post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
501 current_frame = frame;
502 current_secondary_frame = secondary_frame;
504 if (secondary_frame != nullptr) {
505 current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
507 current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
509 ++metric_jpeg_prepared_frames;
510 displayed_this_frame = false;
515 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
517 if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
522 void JPEGFrameView::set_overlay(const string &text)
525 overlay_image.reset();
529 // Figure out how large the texture needs to be.
531 QImage img(overlay_width, overlay_height, QImage::Format_Grayscale8);
532 QPainter painter(&img);
533 QFont font = painter.font();
534 font.setPointSize(12);
535 QFontMetrics metrics(font);
536 overlay_base_width = lrint(metrics.boundingRect(QString::fromStdString(text)).width() + 8.0);
537 overlay_base_height = lrint(metrics.height());
540 float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
541 overlay_width = lrint(overlay_base_width * dpr);
542 overlay_height = lrint(overlay_base_height * dpr);
544 // Work around OpenGL alignment issues.
545 while (overlay_width % 4 != 0) ++overlay_width;
547 // Now do the actual drawing.
548 overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
549 overlay_image->setDevicePixelRatio(dpr);
550 overlay_image->fill(0);
551 QPainter painter(overlay_image.get());
553 painter.setPen(Qt::white);
554 QFont font = painter.font();
555 font.setPointSize(12);
556 painter.setFont(font);
558 painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
560 // Don't refresh immediately; we might not have an OpenGL context here.
561 overlay_input_needs_refresh = true;
564 shared_ptr<Frame> get_black_frame()
566 static shared_ptr<Frame> black_frame;
567 static once_flag flag;
569 // Not really black, but whatever. :-)
570 uint8_t black[] = { 0, 0, 0, 255 };
571 RefCountedTexture black_tex = create_texture_2d(1, 1, GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, black);
573 black_frame->y = black_tex;
574 black_frame->cb = black_tex;
575 black_frame->cr = move(black_tex);
576 black_frame->width = 1;
577 black_frame->height = 1;
578 black_frame->chroma_subsampling_x = 1;
579 black_frame->chroma_subsampling_y = 1;
581 ++metric_jpeg_software_fail_frames;