7 #include <movit/effect_chain.h>
8 #include <movit/effect_util.h>
9 #include <movit/flat_input.h>
10 #include <movit/image_format.h>
11 #include <movit/init.h>
12 #include <movit/resource_pool.h>
17 #include <sys/resource.h>
20 #include <condition_variable>
31 #include "DeckLinkAPI.h"
33 #include "alsa_output.h"
34 #include "bmusb/bmusb.h"
35 #include "bmusb/fake_capture.h"
36 #include "chroma_subsampler.h"
38 #include "decklink_capture.h"
39 #include "decklink_output.h"
41 #include "disk_space_estimator.h"
43 #include "input_mapping.h"
44 #include "pbo_frame_allocator.h"
45 #include "ref_counted_gl_sync.h"
46 #include "resampling_queue.h"
48 #include "timecode_renderer.h"
49 #include "video_encoder.h"
54 using namespace movit;
56 using namespace std::chrono;
57 using namespace std::placeholders;
58 using namespace bmusb;
60 Mixer *global_mixer = nullptr;
61 bool uses_mlock = false;
65 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
68 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
69 input_state->buffered_frames[card_index][frame_num] =
70 input_state->buffered_frames[card_index][frame_num - 1];
72 input_state->buffered_frames[card_index][0] = { frame, field_num };
74 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
75 input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
82 void QueueLengthPolicy::update_policy(unsigned queue_length)
84 if (queue_length == 0) { // Starvation.
85 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
87 fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
88 card_index, safe_queue_length);
90 frames_with_at_least_one = 0;
91 been_at_safe_point_since_last_starvation = false;
94 if (queue_length >= 1) {
95 if (queue_length >= safe_queue_length) {
96 been_at_safe_point_since_last_starvation = true;
98 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
100 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
101 card_index, safe_queue_length);
102 frames_with_at_least_one = 0;
105 frames_with_at_least_one = 0;
109 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
111 num_cards(num_cards),
112 mixer_surface(create_surface(format)),
113 h264_encoder_surface(create_surface(format)),
114 decklink_output_surface(create_surface(format)),
115 audio_mixer(num_cards)
117 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
120 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
121 // will be halved when sampling them, and we need to compensate here.
122 movit_texel_subpixel_precision /= 2.0;
124 resource_pool.reset(new ResourcePool);
125 theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
126 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
127 output_channel[i].parent = this;
128 output_channel[i].channel = i;
131 ImageFormat inout_format;
132 inout_format.color_space = COLORSPACE_sRGB;
133 inout_format.gamma_curve = GAMMA_sRGB;
135 // Display chain; shows the live output produced by the main chain (its RGBA version).
136 display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
138 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, global_flags.width, global_flags.height); // FIXME: GL_UNSIGNED_BYTE is really wrong.
139 display_chain->add_input(display_input);
140 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
141 display_chain->set_dither_bits(0); // Don't bother.
142 display_chain->finalize();
144 video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
146 // Start listening for clients only once VideoEncoder has written its header, if any.
149 // First try initializing the then PCI devices, then USB, then
150 // fill up with fake cards until we have the desired number of cards.
151 unsigned num_pci_devices = 0;
152 unsigned card_index = 0;
155 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
156 if (decklink_iterator != nullptr) {
157 for ( ; card_index < num_cards; ++card_index) {
159 if (decklink_iterator->Next(&decklink) != S_OK) {
163 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
164 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
165 output->set_device(decklink);
166 configure_card(card_index, capture, /*is_fake_capture=*/false, output);
169 decklink_iterator->Release();
170 fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
172 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
176 unsigned num_usb_devices = BMUSBCapture::num_cards();
177 for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
178 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
179 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
180 configure_card(card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
182 fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
184 unsigned num_fake_cards = 0;
185 for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
186 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
187 configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
190 if (num_fake_cards > 0) {
191 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
194 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
195 BMUSBCapture::start_bm_thread();
197 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
198 cards[card_index].queue_length_policy.reset(card_index);
201 chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
203 timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
204 display_timecode_in_stream = global_flags.display_timecode_in_stream;
205 display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
207 if (global_flags.enable_alsa_output) {
208 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
210 if (global_flags.output_card != -1) {
211 desired_output_card_index = global_flags.output_card;
212 set_output_card_internal(global_flags.output_card);
218 BMUSBCapture::stop_bm_thread();
220 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
222 unique_lock<mutex> lock(card_mutex);
223 cards[card_index].should_quit = true; // Unblock thread.
224 cards[card_index].new_frames_changed.notify_all();
226 cards[card_index].capture->stop_dequeue_thread();
227 if (cards[card_index].output) {
228 cards[card_index].output->end_output();
229 cards[card_index].output.reset();
233 video_encoder.reset(nullptr);
236 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output)
238 printf("Configuring card %d...\n", card_index);
240 CaptureCard *card = &cards[card_index];
241 if (card->capture != nullptr) {
242 card->capture->stop_dequeue_thread();
244 card->capture.reset(capture);
245 card->is_fake_capture = is_fake_capture;
246 if (card->output.get() != output) {
247 card->output.reset(output);
249 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
250 if (card->frame_allocator == nullptr) {
251 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height)); // 8 MB.
253 card->capture->set_video_frame_allocator(card->frame_allocator.get());
254 if (card->surface == nullptr) {
255 card->surface = create_surface_with_same_format(mixer_surface);
257 while (!card->new_frames.empty()) card->new_frames.pop_front();
258 card->last_timecode = -1;
259 card->capture->configure_card();
261 // NOTE: start_bm_capture() happens in thread_func().
263 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
264 audio_mixer.reset_resampler(device);
265 audio_mixer.set_display_name(device, card->capture->get_description());
266 audio_mixer.trigger_state_changed_callback();
269 void Mixer::set_output_card_internal(int card_index)
271 // We don't really need to take card_mutex, since we're in the mixer
272 // thread and don't mess with any queues (which is the only thing that happens
273 // from other threads), but it's probably the safest in the long run.
274 unique_lock<mutex> lock(card_mutex);
275 if (output_card_index != -1) {
276 // Switch the old card from output to input.
277 CaptureCard *old_card = &cards[output_card_index];
278 old_card->output->end_output();
280 // Stop the fake card that we put into place.
281 // This needs to _not_ happen under the mutex, to avoid deadlock
282 // (delivering the last frame needs to take the mutex).
283 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
285 fake_capture->stop_dequeue_thread();
287 old_card->capture = move(old_card->parked_capture);
288 old_card->is_fake_capture = false;
289 old_card->capture->start_bm_capture();
291 if (card_index != -1) {
292 CaptureCard *card = &cards[card_index];
293 bmusb::CaptureInterface *capture = card->capture.get();
294 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
295 // several seconds to complete (blocking on DisableVideoInput);
296 // see if we can maybe do it asynchronously.
298 capture->stop_dequeue_thread();
300 card->parked_capture = move(card->capture);
301 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
302 configure_card(card_index, fake_capture, /*is_fake_capture=*/true, card->output.release());
303 card->queue_length_policy.reset(card_index);
304 card->capture->start_bm_capture();
305 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
306 card->output->start_output(desired_output_video_mode, pts_int);
308 output_card_index = card_index;
313 int unwrap_timecode(uint16_t current_wrapped, int last)
315 uint16_t last_wrapped = last & 0xffff;
316 if (current_wrapped > last_wrapped) {
317 return (last & ~0xffff) | current_wrapped;
319 return 0x10000 + ((last & ~0xffff) | current_wrapped);
325 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
326 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
327 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
329 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
330 CaptureCard *card = &cards[card_index];
332 if (is_mode_scanning[card_index]) {
333 if (video_format.has_signal) {
334 // Found a stable signal, so stop scanning.
335 is_mode_scanning[card_index] = false;
337 static constexpr double switch_time_s = 0.1; // Should be enough time for the signal to stabilize.
338 steady_clock::time_point now = steady_clock::now();
339 double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
340 if (sec_since_last_switch > switch_time_s) {
341 // It isn't this mode; try the next one.
342 mode_scanlist_index[card_index]++;
343 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
344 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
345 last_mode_scan_change[card_index] = now;
350 int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
351 assert(frame_length > 0);
353 size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
354 if (num_samples > OUTPUT_FREQUENCY / 10) {
355 printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
356 card_index, int(audio_frame.len), int(audio_offset),
357 timecode, int(video_frame.len), int(video_offset), video_format.id);
358 if (video_frame.owner) {
359 video_frame.owner->release_frame(video_frame);
361 if (audio_frame.owner) {
362 audio_frame.owner->release_frame(audio_frame);
367 int dropped_frames = 0;
368 if (card->last_timecode != -1) {
369 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
372 // Number of samples per frame if we need to insert silence.
373 // (Could be nonintegral, but resampling will save us then.)
374 const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
376 if (dropped_frames > MAX_FPS * 2) {
377 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
378 card_index, card->last_timecode, timecode);
379 audio_mixer.reset_resampler(device);
381 } else if (dropped_frames > 0) {
382 // Insert silence as needed.
383 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
384 card_index, dropped_frames, timecode);
388 success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
392 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
394 // Done with the audio, so release it.
395 if (audio_frame.owner) {
396 audio_frame.owner->release_frame(audio_frame);
399 card->last_timecode = timecode;
401 size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
402 if (video_frame.len - video_offset == 0 ||
403 video_frame.len - video_offset != expected_length) {
404 if (video_frame.len != 0) {
405 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
406 card_index, video_frame.len - video_offset, expected_length);
408 if (video_frame.owner) {
409 video_frame.owner->release_frame(video_frame);
412 // Still send on the information that we _had_ a frame, even though it's corrupted,
413 // so that pts can go up accordingly.
415 unique_lock<mutex> lock(card_mutex);
416 CaptureCard::NewFrame new_frame;
417 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
418 new_frame.length = frame_length;
419 new_frame.interlaced = false;
420 new_frame.dropped_frames = dropped_frames;
421 new_frame.received_timestamp = video_frame.received_timestamp;
422 card->new_frames.push_back(move(new_frame));
423 card->new_frames_changed.notify_all();
428 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
430 unsigned num_fields = video_format.interlaced ? 2 : 1;
431 steady_clock::time_point frame_upload_start;
432 bool interlaced_stride = false;
433 if (video_format.interlaced) {
434 // Send the two fields along as separate frames; the other side will need to add
435 // a deinterlacer to actually get this right.
436 assert(video_format.height % 2 == 0);
437 video_format.height /= 2;
438 assert(frame_length % 2 == 0);
441 if (video_format.second_field_start == 1) {
442 interlaced_stride = true;
444 frame_upload_start = steady_clock::now();
446 userdata->last_interlaced = video_format.interlaced;
447 userdata->last_has_signal = video_format.has_signal;
448 userdata->last_is_connected = video_format.is_connected;
449 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
450 userdata->last_frame_rate_den = video_format.frame_rate_den;
451 RefCountedFrame frame(video_frame);
453 // Upload the textures.
454 size_t cbcr_width = video_format.width / 2;
455 size_t cbcr_offset = video_offset / 2;
456 size_t y_offset = video_frame.size / 2 + video_offset / 2;
458 for (unsigned field = 0; field < num_fields; ++field) {
459 // Put the actual texture upload in a lambda that is executed in the main thread.
460 // It is entirely possible to do this in the same thread (and it might even be
461 // faster, depending on the GPU and driver), but it appears to be trickling
462 // driver bugs very easily.
464 // Note that this means we must hold on to the actual frame data in <userdata>
465 // until the upload command is run, but we hold on to <frame> much longer than that
466 // (in fact, all the way until we no longer use the texture in rendering).
467 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
468 unsigned field_start_line;
470 field_start_line = video_format.second_field_start;
472 field_start_line = video_format.extra_lines_top;
475 if (userdata->tex_y[field] == 0 ||
476 userdata->tex_cbcr[field] == 0 ||
477 video_format.width != userdata->last_width[field] ||
478 video_format.height != userdata->last_height[field]) {
479 // We changed resolution since last use of this texture, so we need to create
480 // a new object. Note that this each card has its own PBOFrameAllocator,
481 // we don't need to worry about these flip-flopping between resolutions.
482 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
484 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
486 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
488 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
490 userdata->last_width[field] = video_format.width;
491 userdata->last_height[field] = video_format.height;
494 GLuint pbo = userdata->pbo;
496 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
499 size_t field_y_start = y_offset + video_format.width * field_start_line;
500 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
502 if (global_flags.flush_pbos) {
503 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
505 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
509 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
511 if (interlaced_stride) {
512 glPixelStorei(GL_UNPACK_ROW_LENGTH, cbcr_width * 2);
515 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
518 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
520 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
522 if (interlaced_stride) {
523 glPixelStorei(GL_UNPACK_ROW_LENGTH, video_format.width * 2);
526 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
529 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
531 glBindTexture(GL_TEXTURE_2D, 0);
533 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
535 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
540 // Don't upload the second field as fast as we can; wait until
541 // the field time has approximately passed. (Otherwise, we could
542 // get timing jitter against the other sources, and possibly also
543 // against the video display, although the latter is not as critical.)
544 // This requires our system clock to be reasonably close to the
545 // video clock, but that's not an unreasonable assumption.
546 steady_clock::time_point second_field_start = frame_upload_start +
547 nanoseconds(frame_length * 1000000000 / TIMEBASE);
548 this_thread::sleep_until(second_field_start);
552 unique_lock<mutex> lock(card_mutex);
553 CaptureCard::NewFrame new_frame;
554 new_frame.frame = frame;
555 new_frame.length = frame_length;
556 new_frame.field = field;
557 new_frame.interlaced = video_format.interlaced;
558 new_frame.upload_func = upload_func;
559 new_frame.dropped_frames = dropped_frames;
560 new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp.
561 card->new_frames.push_back(move(new_frame));
562 card->new_frames_changed.notify_all();
567 void Mixer::bm_hotplug_add(libusb_device *dev)
569 lock_guard<mutex> lock(hotplug_mutex);
570 hotplugged_cards.push_back(dev);
573 void Mixer::bm_hotplug_remove(unsigned card_index)
575 cards[card_index].new_frames_changed.notify_all();
578 void Mixer::thread_func()
580 pthread_setname_np(pthread_self(), "Mixer_OpenGL");
582 eglBindAPI(EGL_OPENGL_API);
583 QOpenGLContext *context = create_context(mixer_surface);
584 if (!make_current(context, mixer_surface)) {
589 // Start the actual capture. (We don't want to do it before we're actually ready
590 // to process output frames.)
591 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
592 if (int(card_index) != output_card_index) {
593 cards[card_index].capture->start_bm_capture();
597 steady_clock::time_point start, now;
598 start = steady_clock::now();
600 int stats_dropped_frames = 0;
602 while (!should_quit) {
603 if (desired_output_card_index != output_card_index) {
604 set_output_card_internal(desired_output_card_index);
606 if (output_card_index != -1 &&
607 desired_output_video_mode != output_video_mode) {
608 DeckLinkOutput *output = cards[output_card_index].output.get();
609 output->end_output();
610 desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
611 output->start_output(desired_output_video_mode, pts_int);
614 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
615 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
617 bool master_card_is_output;
618 unsigned master_card_index;
619 if (output_card_index != -1) {
620 master_card_is_output = true;
621 master_card_index = output_card_index;
623 master_card_is_output = false;
624 master_card_index = theme->map_signal(master_clock_channel);
625 assert(master_card_index < num_cards);
628 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
629 schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration, output_frame_info.is_preroll, output_frame_info.frame_timestamp);
630 stats_dropped_frames += output_frame_info.dropped_frames;
632 handle_hotplugged_cards();
634 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
635 if (card_index == master_card_index || !has_new_frame[card_index]) {
638 if (new_frames[card_index].frame->len == 0) {
639 ++new_frames[card_index].dropped_frames;
641 if (new_frames[card_index].dropped_frames > 0) {
642 printf("Card %u dropped %d frames before this\n",
643 card_index, int(new_frames[card_index].dropped_frames));
647 // If the first card is reporting a corrupted or otherwise dropped frame,
648 // just increase the pts (skipping over this frame) and don't try to compute anything new.
649 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
650 ++stats_dropped_frames;
651 pts_int += new_frames[master_card_index].length;
655 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
656 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
659 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
660 assert(new_frame->frame != nullptr);
661 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
664 // The new texture might need uploading before use.
665 if (new_frame->upload_func) {
666 new_frame->upload_func();
667 new_frame->upload_func = nullptr;
671 int64_t frame_duration = output_frame_info.frame_duration;
672 render_one_frame(frame_duration);
674 pts_int += frame_duration;
676 now = steady_clock::now();
677 double elapsed = duration<double>(now - start).count();
678 if (frame_num % 100 == 0) {
679 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
680 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
681 1e3 * elapsed / frame_num);
682 // chain->print_phase_timing();
684 // Check our memory usage, to see if we are close to our mlockall()
685 // limit (if at all set).
687 if (getrusage(RUSAGE_SELF, &used) == -1) {
688 perror("getrusage(RUSAGE_SELF)");
694 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
695 perror("getrlimit(RLIMIT_MEMLOCK)");
699 if (limit.rlim_cur == 0) {
700 printf(", using %ld MB memory (locked)",
701 long(used.ru_maxrss / 1024));
703 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
704 long(used.ru_maxrss / 1024),
705 long(limit.rlim_cur / 1048576),
706 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
709 printf(", using %ld MB memory (not locked)",
710 long(used.ru_maxrss / 1024));
717 if (should_cut.exchange(false)) { // Test and clear.
718 video_encoder->do_cut(frame_num);
722 // Reset every 100 frames, so that local variations in frame times
723 // (especially for the first few frames, when the shaders are
724 // compiled etc.) don't make it hard to measure for the entire
725 // remaining duration of the program.
726 if (frame == 10000) {
734 resource_pool->clean_context();
737 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
739 if (output_card_index != -1) {
740 // The output card (ie., cards[output_card_index].output) is the master clock,
741 // so no input card (ie., cards[card_index].capture) is.
744 return (card_index == master_card_index);
747 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
749 // Count the number of frames in the queue, including any frames
750 // we dropped. It's hard to know exactly how we should deal with
751 // dropped (corrupted) input frames; they don't help our goal of
752 // avoiding starvation, but they still add to the problem of latency.
753 // Since dropped frames is going to mean a bump in the signal anyway,
754 // we err on the side of having more stable latency instead.
755 unsigned queue_length = 0;
756 for (const CaptureCard::NewFrame &frame : card->new_frames) {
757 queue_length += frame.dropped_frames + 1;
759 card->queue_length_policy.update_policy(queue_length);
761 // If needed, drop frames until the queue is below the safe limit.
762 // We prefer to drop from the head, because all else being equal,
763 // we'd like more recent frames (less latency).
764 unsigned dropped_frames = 0;
765 while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
766 assert(!card->new_frames.empty());
767 assert(queue_length > card->new_frames.front().dropped_frames);
768 queue_length -= card->new_frames.front().dropped_frames;
770 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
771 // No need to drop anything.
775 card->new_frames.pop_front();
776 card->new_frames_changed.notify_all();
781 if (dropped_frames > 0) {
782 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
783 card_index, dropped_frames);
788 Mixer::OutputFrameInfo Mixer::get_one_frame_from_each_card(unsigned master_card_index, bool master_card_is_output, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS])
790 OutputFrameInfo output_frame_info;
792 unique_lock<mutex> lock(card_mutex, defer_lock);
793 if (master_card_is_output) {
794 // Clocked to the output, so wait for it to be ready for the next frame.
795 cards[master_card_index].output->wait_for_frame(pts_int, &output_frame_info.dropped_frames, &output_frame_info.frame_duration, &output_frame_info.is_preroll, &output_frame_info.frame_timestamp);
798 // Wait for the master card to have a new frame.
799 // TODO: Add a timeout.
800 output_frame_info.is_preroll = false;
802 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
805 if (master_card_is_output) {
806 handle_hotplugged_cards();
807 } else if (cards[master_card_index].new_frames.empty()) {
808 // We were woken up, but not due to a new frame. Deal with it
810 assert(cards[master_card_index].capture->get_disconnected());
811 handle_hotplugged_cards();
815 if (!master_card_is_output) {
816 output_frame_info.frame_timestamp =
817 cards[master_card_index].new_frames.front().received_timestamp;
820 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
821 CaptureCard *card = &cards[card_index];
822 if (input_card_is_master_clock(card_index, master_card_index)) {
823 // We don't use the queue length policy for the master card,
824 // but we will if it stops being the master. Thus, clear out
825 // the policy in case we switch in the future.
826 card->queue_length_policy.reset(card_index);
827 assert(!card->new_frames.empty());
829 trim_queue(card, card_index);
831 if (!card->new_frames.empty()) {
832 new_frames[card_index] = move(card->new_frames.front());
833 has_new_frame[card_index] = true;
834 card->new_frames.pop_front();
835 card->new_frames_changed.notify_all();
839 if (!master_card_is_output) {
840 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
841 output_frame_info.frame_duration = new_frames[master_card_index].length;
844 // This might get off by a fractional sample when changing master card
845 // between ones with different frame rates, but that's fine.
846 int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
847 output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
848 fractional_samples = num_samples_times_timebase % TIMEBASE;
849 assert(output_frame_info.num_samples >= 0);
851 return output_frame_info;
854 void Mixer::handle_hotplugged_cards()
856 // Check for cards that have been disconnected since last frame.
857 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
858 CaptureCard *card = &cards[card_index];
859 if (card->capture->get_disconnected()) {
860 fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
861 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
862 configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
863 card->queue_length_policy.reset(card_index);
864 card->capture->start_bm_capture();
868 // Check for cards that have been connected since last frame.
869 vector<libusb_device *> hotplugged_cards_copy;
871 lock_guard<mutex> lock(hotplug_mutex);
872 swap(hotplugged_cards, hotplugged_cards_copy);
874 for (libusb_device *new_dev : hotplugged_cards_copy) {
875 // Look for a fake capture card where we can stick this in.
876 int free_card_index = -1;
877 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
878 if (cards[card_index].is_fake_capture) {
879 free_card_index = card_index;
884 if (free_card_index == -1) {
885 fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
886 libusb_unref_device(new_dev);
888 // BMUSBCapture takes ownership.
889 fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
890 CaptureCard *card = &cards[free_card_index];
891 BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
892 configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
893 card->queue_length_policy.reset(free_card_index);
894 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
895 capture->start_bm_capture();
901 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame, bool is_preroll, steady_clock::time_point frame_timestamp)
903 // Resample the audio as needed, including from previously dropped frames.
904 assert(num_cards > 0);
905 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
906 const bool dropped_frame = (frame_num != dropped_frames);
908 // Signal to the audio thread to process this frame.
909 // Note that if the frame is a dropped frame, we signal that
910 // we don't want to use this frame as base for adjusting
911 // the resampler rate. The reason for this is that the timing
912 // of these frames is often way too late; they typically don't
913 // “arrive” before we synthesize them. Thus, we could end up
914 // in a situation where we have inserted e.g. five audio frames
915 // into the queue before we then start pulling five of them
916 // back out. This makes ResamplingQueue overestimate the delay,
917 // causing undue resampler changes. (We _do_ use the last,
918 // non-dropped frame; perhaps we should just discard that as well,
919 // since dropped frames are expected to be rare, and it might be
920 // better to just wait until we have a slightly more normal situation).
921 unique_lock<mutex> lock(audio_mutex);
922 bool adjust_rate = !dropped_frame && !is_preroll;
923 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
924 audio_task_queue_changed.notify_one();
927 // For dropped frames, increase the pts. Note that if the format changed
928 // in the meantime, we have no way of detecting that; we just have to
929 // assume the frame length is always the same.
930 pts_int += length_per_frame;
935 void Mixer::render_one_frame(int64_t duration)
937 // Determine the time code for this frame before we start rendering.
938 string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
939 if (display_timecode_on_stdout) {
940 printf("Timecode: '%s'\n", timecode_text.c_str());
943 // Get the main chain from the theme, and set its state immediately.
944 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
945 EffectChain *chain = theme_main_chain.chain;
946 theme_main_chain.setup_chain();
947 //theme_main_chain.chain->enable_phase_timing(true);
949 GLuint y_tex, cbcr_tex;
950 bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
953 // Render main chain.
954 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
955 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, global_flags.width, global_flags.height); // Saves texture bandwidth, although dithering gets messed up.
956 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
958 chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
960 if (display_timecode_in_stream) {
961 // Render the timecode on top.
962 timecode_renderer->render_timecode(fbo, timecode_text);
965 resource_pool->release_fbo(fbo);
967 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
968 if (output_card_index != -1) {
969 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, theme_main_chain.input_frames, pts_int, duration);
971 resource_pool->release_2d_texture(cbcr_full_tex);
973 // Set the right state for rgba_tex.
974 glBindFramebuffer(GL_FRAMEBUFFER, 0);
975 glBindTexture(GL_TEXTURE_2D, rgba_tex);
976 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
977 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
978 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
980 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
981 RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
983 // The live frame just shows the RGBA texture we just rendered.
984 // It owns rgba_tex now.
985 DisplayFrame live_frame;
986 live_frame.chain = display_chain.get();
987 live_frame.setup_chain = [this, rgba_tex]{
988 display_input->set_texture_num(rgba_tex);
990 live_frame.ready_fence = fence;
991 live_frame.input_frames = {};
992 live_frame.temp_textures = { rgba_tex };
993 output_channel[OUTPUT_LIVE].output_frame(live_frame);
995 // Set up preview and any additional channels.
996 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
997 DisplayFrame display_frame;
998 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions
999 display_frame.chain = chain.chain;
1000 display_frame.setup_chain = chain.setup_chain;
1001 display_frame.ready_fence = fence;
1002 display_frame.input_frames = chain.input_frames;
1003 display_frame.temp_textures = {};
1004 output_channel[i].output_frame(display_frame);
1008 void Mixer::audio_thread_func()
1010 pthread_setname_np(pthread_self(), "Mixer_Audio");
1012 while (!should_quit) {
1016 unique_lock<mutex> lock(audio_mutex);
1017 audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1021 task = audio_task_queue.front();
1022 audio_task_queue.pop();
1025 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1026 task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1027 vector<float> samples_out = audio_mixer.get_output(
1028 task.frame_timestamp,
1030 rate_adjustment_policy);
1032 // Send the samples to the sound card, then add them to the output.
1034 alsa->write(samples_out);
1036 if (output_card_index != -1) {
1037 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1038 cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1040 video_encoder->add_audio(task.pts_int, move(samples_out));
1044 void Mixer::release_display_frame(DisplayFrame *frame)
1046 for (GLuint texnum : frame->temp_textures) {
1047 resource_pool->release_2d_texture(texnum);
1049 frame->temp_textures.clear();
1050 frame->ready_fence.reset();
1051 frame->input_frames.clear();
1056 mixer_thread = thread(&Mixer::thread_func, this);
1057 audio_thread = thread(&Mixer::audio_thread_func, this);
1063 audio_task_queue_changed.notify_one();
1064 mixer_thread.join();
1065 audio_thread.join();
1068 void Mixer::transition_clicked(int transition_num)
1070 theme->transition_clicked(transition_num, pts());
1073 void Mixer::channel_clicked(int preview_num)
1075 theme->channel_clicked(preview_num);
1078 void Mixer::start_mode_scanning(unsigned card_index)
1080 assert(card_index < num_cards);
1081 if (is_mode_scanning[card_index]) {
1084 is_mode_scanning[card_index] = true;
1085 mode_scanlist[card_index].clear();
1086 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1087 mode_scanlist[card_index].push_back(mode.first);
1089 assert(!mode_scanlist[card_index].empty());
1090 mode_scanlist_index[card_index] = 0;
1091 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1092 last_mode_scan_change[card_index] = steady_clock::now();
1095 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1097 assert(desired_output_card_index != -1);
1098 unique_lock<mutex> lock(card_mutex);
1099 return cards[desired_output_card_index].output->get_available_video_modes();
1102 Mixer::OutputChannel::~OutputChannel()
1104 if (has_current_frame) {
1105 parent->release_display_frame(¤t_frame);
1107 if (has_ready_frame) {
1108 parent->release_display_frame(&ready_frame);
1112 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1114 // Store this frame for display. Remove the ready frame if any
1115 // (it was seemingly never used).
1117 unique_lock<mutex> lock(frame_mutex);
1118 if (has_ready_frame) {
1119 parent->release_display_frame(&ready_frame);
1121 ready_frame = frame;
1122 has_ready_frame = true;
1125 if (new_frame_ready_callback) {
1126 new_frame_ready_callback();
1129 // Reduce the number of callbacks by filtering duplicates. The reason
1130 // why we bother doing this is that Qt seemingly can get into a state
1131 // where its builds up an essentially unbounded queue of signals,
1132 // consuming more and more memory, and there's no good way of collapsing
1133 // user-defined signals or limiting the length of the queue.
1134 if (transition_names_updated_callback) {
1135 vector<string> transition_names = global_mixer->get_transition_names();
1136 bool changed = false;
1137 if (transition_names.size() != last_transition_names.size()) {
1140 for (unsigned i = 0; i < transition_names.size(); ++i) {
1141 if (transition_names[i] != last_transition_names[i]) {
1148 transition_names_updated_callback(transition_names);
1149 last_transition_names = transition_names;
1152 if (name_updated_callback) {
1153 string name = global_mixer->get_channel_name(channel);
1154 if (name != last_name) {
1155 name_updated_callback(name);
1159 if (color_updated_callback) {
1160 string color = global_mixer->get_channel_color(channel);
1161 if (color != last_color) {
1162 color_updated_callback(color);
1168 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1170 unique_lock<mutex> lock(frame_mutex);
1171 if (!has_current_frame && !has_ready_frame) {
1175 if (has_current_frame && has_ready_frame) {
1176 // We have a new ready frame. Toss the current one.
1177 parent->release_display_frame(¤t_frame);
1178 has_current_frame = false;
1180 if (has_ready_frame) {
1181 assert(!has_current_frame);
1182 current_frame = ready_frame;
1183 ready_frame.ready_fence.reset(); // Drop the refcount.
1184 ready_frame.input_frames.clear(); // Drop the refcounts.
1185 has_current_frame = true;
1186 has_ready_frame = false;
1189 *frame = current_frame;
1193 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1195 new_frame_ready_callback = callback;
1198 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1200 transition_names_updated_callback = callback;
1203 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1205 name_updated_callback = callback;
1208 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1210 color_updated_callback = callback;
1213 mutex RefCountedGLsync::fence_lock;