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 "v210_converter.h"
50 #include "video_encoder.h"
55 using namespace movit;
57 using namespace std::chrono;
58 using namespace std::placeholders;
59 using namespace bmusb;
61 Mixer *global_mixer = nullptr;
62 bool uses_mlock = false;
66 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
69 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
70 input_state->buffered_frames[card_index][frame_num] =
71 input_state->buffered_frames[card_index][frame_num - 1];
73 input_state->buffered_frames[card_index][0] = { frame, field_num };
75 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
76 input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
81 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned v210_width)
84 if (global_flags.ten_bit_input) {
85 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
87 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
91 width != userdata->last_width[field] ||
92 height != userdata->last_height[field]) {
93 // We changed resolution since last use of this texture, so we need to create
94 // a new object. Note that this each card has its own PBOFrameAllocator,
95 // we don't need to worry about these flip-flopping between resolutions.
96 if (global_flags.ten_bit_input) {
97 glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
99 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
102 size_t cbcr_width = width / 2;
104 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
106 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
108 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
110 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
113 userdata->last_width[field] = width;
114 userdata->last_height[field] = height;
116 if (global_flags.ten_bit_input &&
117 (first || v210_width != userdata->last_v210_width[field])) {
118 // Same as above; we need to recreate the texture.
119 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
121 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
123 userdata->last_v210_width[field] = v210_width;
127 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
129 if (interlaced_stride) {
132 if (global_flags.flush_pbos) {
133 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
137 glBindTexture(GL_TEXTURE_2D, tex);
139 if (interlaced_stride) {
140 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
143 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
147 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
149 glBindTexture(GL_TEXTURE_2D, 0);
151 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
157 void QueueLengthPolicy::update_policy(unsigned queue_length)
159 if (queue_length == 0) { // Starvation.
160 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
162 fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
163 card_index, safe_queue_length);
165 frames_with_at_least_one = 0;
166 been_at_safe_point_since_last_starvation = false;
169 if (queue_length >= 1) {
170 if (queue_length >= safe_queue_length) {
171 been_at_safe_point_since_last_starvation = true;
173 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
175 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
176 card_index, safe_queue_length);
177 frames_with_at_least_one = 0;
180 frames_with_at_least_one = 0;
184 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
186 num_cards(num_cards),
187 mixer_surface(create_surface(format)),
188 h264_encoder_surface(create_surface(format)),
189 decklink_output_surface(create_surface(format)),
190 audio_mixer(num_cards)
192 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
195 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
196 // will be halved when sampling them, and we need to compensate here.
197 movit_texel_subpixel_precision /= 2.0;
199 resource_pool.reset(new ResourcePool);
200 theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
201 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
202 output_channel[i].parent = this;
203 output_channel[i].channel = i;
206 ImageFormat inout_format;
207 inout_format.color_space = COLORSPACE_sRGB;
208 inout_format.gamma_curve = GAMMA_sRGB;
210 // Display chain; shows the live output produced by the main chain (its RGBA version).
211 display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
213 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, global_flags.width, global_flags.height); // FIXME: GL_UNSIGNED_BYTE is really wrong.
214 display_chain->add_input(display_input);
215 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
216 display_chain->set_dither_bits(0); // Don't bother.
217 display_chain->finalize();
219 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));
221 // Start listening for clients only once VideoEncoder has written its header, if any.
224 // First try initializing the then PCI devices, then USB, then
225 // fill up with fake cards until we have the desired number of cards.
226 unsigned num_pci_devices = 0;
227 unsigned card_index = 0;
230 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
231 if (decklink_iterator != nullptr) {
232 for ( ; card_index < num_cards; ++card_index) {
234 if (decklink_iterator->Next(&decklink) != S_OK) {
238 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
239 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
240 output->set_device(decklink);
241 configure_card(card_index, capture, /*is_fake_capture=*/false, output);
244 decklink_iterator->Release();
245 fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
247 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
251 unsigned num_usb_devices = BMUSBCapture::num_cards();
252 for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
253 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
254 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
255 configure_card(card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
257 fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
259 unsigned num_fake_cards = 0;
260 for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
261 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
262 configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
265 if (num_fake_cards > 0) {
266 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
269 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
270 BMUSBCapture::start_bm_thread();
272 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
273 cards[card_index].queue_length_policy.reset(card_index);
276 chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
278 if (global_flags.ten_bit_input) {
279 if (!v210Converter::has_hardware_support()) {
280 fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
281 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
284 v210_converter.reset(new v210Converter());
286 // These are all the widths listed in the Blackmagic SDK documentation
287 // (section 2.7.3, “Display Modes”).
288 v210_converter->precompile_shader(720);
289 v210_converter->precompile_shader(1280);
290 v210_converter->precompile_shader(1920);
291 v210_converter->precompile_shader(2048);
292 v210_converter->precompile_shader(3840);
293 v210_converter->precompile_shader(4096);
296 timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
297 display_timecode_in_stream = global_flags.display_timecode_in_stream;
298 display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
300 if (global_flags.enable_alsa_output) {
301 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
303 if (global_flags.output_card != -1) {
304 desired_output_card_index = global_flags.output_card;
305 set_output_card_internal(global_flags.output_card);
311 BMUSBCapture::stop_bm_thread();
313 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
315 unique_lock<mutex> lock(card_mutex);
316 cards[card_index].should_quit = true; // Unblock thread.
317 cards[card_index].new_frames_changed.notify_all();
319 cards[card_index].capture->stop_dequeue_thread();
320 if (cards[card_index].output) {
321 cards[card_index].output->end_output();
322 cards[card_index].output.reset();
326 video_encoder.reset(nullptr);
329 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output)
331 printf("Configuring card %d...\n", card_index);
333 CaptureCard *card = &cards[card_index];
334 if (card->capture != nullptr) {
335 card->capture->stop_dequeue_thread();
337 card->capture.reset(capture);
338 card->is_fake_capture = is_fake_capture;
339 if (card->output.get() != output) {
340 card->output.reset(output);
342 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
343 if (card->frame_allocator == nullptr) {
344 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height)); // 8 MB.
346 card->capture->set_video_frame_allocator(card->frame_allocator.get());
347 if (card->surface == nullptr) {
348 card->surface = create_surface_with_same_format(mixer_surface);
350 while (!card->new_frames.empty()) card->new_frames.pop_front();
351 card->last_timecode = -1;
352 card->capture->set_pixel_format(global_flags.ten_bit_input ? PixelFormat_10BitYCbCr : PixelFormat_8BitYCbCr);
353 card->capture->configure_card();
355 // NOTE: start_bm_capture() happens in thread_func().
357 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
358 audio_mixer.reset_resampler(device);
359 audio_mixer.set_display_name(device, card->capture->get_description());
360 audio_mixer.trigger_state_changed_callback();
363 void Mixer::set_output_card_internal(int card_index)
365 // We don't really need to take card_mutex, since we're in the mixer
366 // thread and don't mess with any queues (which is the only thing that happens
367 // from other threads), but it's probably the safest in the long run.
368 unique_lock<mutex> lock(card_mutex);
369 if (output_card_index != -1) {
370 // Switch the old card from output to input.
371 CaptureCard *old_card = &cards[output_card_index];
372 old_card->output->end_output();
374 // Stop the fake card that we put into place.
375 // This needs to _not_ happen under the mutex, to avoid deadlock
376 // (delivering the last frame needs to take the mutex).
377 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
379 fake_capture->stop_dequeue_thread();
381 old_card->capture = move(old_card->parked_capture);
382 old_card->is_fake_capture = false;
383 old_card->capture->start_bm_capture();
385 if (card_index != -1) {
386 CaptureCard *card = &cards[card_index];
387 bmusb::CaptureInterface *capture = card->capture.get();
388 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
389 // several seconds to complete (blocking on DisableVideoInput);
390 // see if we can maybe do it asynchronously.
392 capture->stop_dequeue_thread();
394 card->parked_capture = move(card->capture);
395 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
396 configure_card(card_index, fake_capture, /*is_fake_capture=*/true, card->output.release());
397 card->queue_length_policy.reset(card_index);
398 card->capture->start_bm_capture();
399 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
400 card->output->start_output(desired_output_video_mode, pts_int);
402 output_card_index = card_index;
407 int unwrap_timecode(uint16_t current_wrapped, int last)
409 uint16_t last_wrapped = last & 0xffff;
410 if (current_wrapped > last_wrapped) {
411 return (last & ~0xffff) | current_wrapped;
413 return 0x10000 + ((last & ~0xffff) | current_wrapped);
419 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
420 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
421 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
423 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
424 CaptureCard *card = &cards[card_index];
426 if (is_mode_scanning[card_index]) {
427 if (video_format.has_signal) {
428 // Found a stable signal, so stop scanning.
429 is_mode_scanning[card_index] = false;
431 static constexpr double switch_time_s = 0.1; // Should be enough time for the signal to stabilize.
432 steady_clock::time_point now = steady_clock::now();
433 double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
434 if (sec_since_last_switch > switch_time_s) {
435 // It isn't this mode; try the next one.
436 mode_scanlist_index[card_index]++;
437 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
438 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
439 last_mode_scan_change[card_index] = now;
444 int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
445 assert(frame_length > 0);
447 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;
448 if (num_samples > OUTPUT_FREQUENCY / 10) {
449 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",
450 card_index, int(audio_frame.len), int(audio_offset),
451 timecode, int(video_frame.len), int(video_offset), video_format.id);
452 if (video_frame.owner) {
453 video_frame.owner->release_frame(video_frame);
455 if (audio_frame.owner) {
456 audio_frame.owner->release_frame(audio_frame);
461 int dropped_frames = 0;
462 if (card->last_timecode != -1) {
463 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
466 // Number of samples per frame if we need to insert silence.
467 // (Could be nonintegral, but resampling will save us then.)
468 const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
470 if (dropped_frames > MAX_FPS * 2) {
471 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
472 card_index, card->last_timecode, timecode);
473 audio_mixer.reset_resampler(device);
475 } else if (dropped_frames > 0) {
476 // Insert silence as needed.
477 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
478 card_index, dropped_frames, timecode);
482 success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
486 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
488 // Done with the audio, so release it.
489 if (audio_frame.owner) {
490 audio_frame.owner->release_frame(audio_frame);
493 card->last_timecode = timecode;
495 size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
496 if (video_frame.len - video_offset == 0 ||
497 video_frame.len - video_offset != expected_length) {
498 if (video_frame.len != 0) {
499 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
500 card_index, video_frame.len - video_offset, expected_length);
502 if (video_frame.owner) {
503 video_frame.owner->release_frame(video_frame);
506 // Still send on the information that we _had_ a frame, even though it's corrupted,
507 // so that pts can go up accordingly.
509 unique_lock<mutex> lock(card_mutex);
510 CaptureCard::NewFrame new_frame;
511 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
512 new_frame.length = frame_length;
513 new_frame.interlaced = false;
514 new_frame.dropped_frames = dropped_frames;
515 new_frame.received_timestamp = video_frame.received_timestamp;
516 card->new_frames.push_back(move(new_frame));
517 card->new_frames_changed.notify_all();
522 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
524 unsigned num_fields = video_format.interlaced ? 2 : 1;
525 steady_clock::time_point frame_upload_start;
526 bool interlaced_stride = false;
527 if (video_format.interlaced) {
528 // Send the two fields along as separate frames; the other side will need to add
529 // a deinterlacer to actually get this right.
530 assert(video_format.height % 2 == 0);
531 video_format.height /= 2;
532 assert(frame_length % 2 == 0);
535 if (video_format.second_field_start == 1) {
536 interlaced_stride = true;
538 frame_upload_start = steady_clock::now();
540 userdata->last_interlaced = video_format.interlaced;
541 userdata->last_has_signal = video_format.has_signal;
542 userdata->last_is_connected = video_format.is_connected;
543 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
544 userdata->last_frame_rate_den = video_format.frame_rate_den;
545 RefCountedFrame frame(video_frame);
547 // Upload the textures.
548 const size_t cbcr_width = video_format.width / 2;
549 const size_t cbcr_offset = video_offset / 2;
550 const size_t y_offset = video_frame.size / 2 + video_offset / 2;
552 for (unsigned field = 0; field < num_fields; ++field) {
553 // Put the actual texture upload in a lambda that is executed in the main thread.
554 // It is entirely possible to do this in the same thread (and it might even be
555 // faster, depending on the GPU and driver), but it appears to be trickling
556 // driver bugs very easily.
558 // Note that this means we must hold on to the actual frame data in <userdata>
559 // until the upload command is run, but we hold on to <frame> much longer than that
560 // (in fact, all the way until we no longer use the texture in rendering).
561 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
562 unsigned field_start_line;
564 field_start_line = video_format.second_field_start;
566 field_start_line = video_format.extra_lines_top;
569 // For 8-bit input, v210_width will be nonsensical but not used.
570 size_t v210_width = video_format.stride / sizeof(uint32_t);
571 ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
573 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
576 if (global_flags.ten_bit_input) {
577 size_t field_start = video_offset + video_format.stride * field_start_line;
578 upload_texture(userdata->tex_v210[field], v210_width, video_format.height, video_format.stride, interlaced_stride, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, field_start);
579 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
581 size_t field_y_start = y_offset + video_format.width * field_start_line;
582 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
584 // Make up our own strides, since we are interleaving.
585 upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start);
586 upload_texture(userdata->tex_cbcr[field], cbcr_width, video_format.height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
589 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
594 // Don't upload the second field as fast as we can; wait until
595 // the field time has approximately passed. (Otherwise, we could
596 // get timing jitter against the other sources, and possibly also
597 // against the video display, although the latter is not as critical.)
598 // This requires our system clock to be reasonably close to the
599 // video clock, but that's not an unreasonable assumption.
600 steady_clock::time_point second_field_start = frame_upload_start +
601 nanoseconds(frame_length * 1000000000 / TIMEBASE);
602 this_thread::sleep_until(second_field_start);
606 unique_lock<mutex> lock(card_mutex);
607 CaptureCard::NewFrame new_frame;
608 new_frame.frame = frame;
609 new_frame.length = frame_length;
610 new_frame.field = field;
611 new_frame.interlaced = video_format.interlaced;
612 new_frame.upload_func = upload_func;
613 new_frame.dropped_frames = dropped_frames;
614 new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp.
615 card->new_frames.push_back(move(new_frame));
616 card->new_frames_changed.notify_all();
621 void Mixer::bm_hotplug_add(libusb_device *dev)
623 lock_guard<mutex> lock(hotplug_mutex);
624 hotplugged_cards.push_back(dev);
627 void Mixer::bm_hotplug_remove(unsigned card_index)
629 cards[card_index].new_frames_changed.notify_all();
632 void Mixer::thread_func()
634 pthread_setname_np(pthread_self(), "Mixer_OpenGL");
636 eglBindAPI(EGL_OPENGL_API);
637 QOpenGLContext *context = create_context(mixer_surface);
638 if (!make_current(context, mixer_surface)) {
643 // Start the actual capture. (We don't want to do it before we're actually ready
644 // to process output frames.)
645 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
646 if (int(card_index) != output_card_index) {
647 cards[card_index].capture->start_bm_capture();
651 steady_clock::time_point start, now;
652 start = steady_clock::now();
654 int stats_dropped_frames = 0;
656 while (!should_quit) {
657 if (desired_output_card_index != output_card_index) {
658 set_output_card_internal(desired_output_card_index);
660 if (output_card_index != -1 &&
661 desired_output_video_mode != output_video_mode) {
662 DeckLinkOutput *output = cards[output_card_index].output.get();
663 output->end_output();
664 desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
665 output->start_output(desired_output_video_mode, pts_int);
668 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
669 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
671 bool master_card_is_output;
672 unsigned master_card_index;
673 if (output_card_index != -1) {
674 master_card_is_output = true;
675 master_card_index = output_card_index;
677 master_card_is_output = false;
678 master_card_index = theme->map_signal(master_clock_channel);
679 assert(master_card_index < num_cards);
682 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
683 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);
684 stats_dropped_frames += output_frame_info.dropped_frames;
686 handle_hotplugged_cards();
688 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
689 if (card_index == master_card_index || !has_new_frame[card_index]) {
692 if (new_frames[card_index].frame->len == 0) {
693 ++new_frames[card_index].dropped_frames;
695 if (new_frames[card_index].dropped_frames > 0) {
696 printf("Card %u dropped %d frames before this\n",
697 card_index, int(new_frames[card_index].dropped_frames));
701 // If the first card is reporting a corrupted or otherwise dropped frame,
702 // just increase the pts (skipping over this frame) and don't try to compute anything new.
703 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
704 ++stats_dropped_frames;
705 pts_int += new_frames[master_card_index].length;
709 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
710 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
713 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
714 assert(new_frame->frame != nullptr);
715 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
718 // The new texture might need uploading before use.
719 if (new_frame->upload_func) {
720 new_frame->upload_func();
721 new_frame->upload_func = nullptr;
725 int64_t frame_duration = output_frame_info.frame_duration;
726 render_one_frame(frame_duration);
728 pts_int += frame_duration;
730 now = steady_clock::now();
731 double elapsed = duration<double>(now - start).count();
732 if (frame_num % 100 == 0) {
733 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
734 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
735 1e3 * elapsed / frame_num);
736 // chain->print_phase_timing();
738 // Check our memory usage, to see if we are close to our mlockall()
739 // limit (if at all set).
741 if (getrusage(RUSAGE_SELF, &used) == -1) {
742 perror("getrusage(RUSAGE_SELF)");
748 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
749 perror("getrlimit(RLIMIT_MEMLOCK)");
753 if (limit.rlim_cur == 0) {
754 printf(", using %ld MB memory (locked)",
755 long(used.ru_maxrss / 1024));
757 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
758 long(used.ru_maxrss / 1024),
759 long(limit.rlim_cur / 1048576),
760 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
763 printf(", using %ld MB memory (not locked)",
764 long(used.ru_maxrss / 1024));
771 if (should_cut.exchange(false)) { // Test and clear.
772 video_encoder->do_cut(frame_num);
776 // Reset every 100 frames, so that local variations in frame times
777 // (especially for the first few frames, when the shaders are
778 // compiled etc.) don't make it hard to measure for the entire
779 // remaining duration of the program.
780 if (frame == 10000) {
788 resource_pool->clean_context();
791 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
793 if (output_card_index != -1) {
794 // The output card (ie., cards[output_card_index].output) is the master clock,
795 // so no input card (ie., cards[card_index].capture) is.
798 return (card_index == master_card_index);
801 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
803 // Count the number of frames in the queue, including any frames
804 // we dropped. It's hard to know exactly how we should deal with
805 // dropped (corrupted) input frames; they don't help our goal of
806 // avoiding starvation, but they still add to the problem of latency.
807 // Since dropped frames is going to mean a bump in the signal anyway,
808 // we err on the side of having more stable latency instead.
809 unsigned queue_length = 0;
810 for (const CaptureCard::NewFrame &frame : card->new_frames) {
811 queue_length += frame.dropped_frames + 1;
813 card->queue_length_policy.update_policy(queue_length);
815 // If needed, drop frames until the queue is below the safe limit.
816 // We prefer to drop from the head, because all else being equal,
817 // we'd like more recent frames (less latency).
818 unsigned dropped_frames = 0;
819 while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
820 assert(!card->new_frames.empty());
821 assert(queue_length > card->new_frames.front().dropped_frames);
822 queue_length -= card->new_frames.front().dropped_frames;
824 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
825 // No need to drop anything.
829 card->new_frames.pop_front();
830 card->new_frames_changed.notify_all();
835 if (dropped_frames > 0) {
836 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
837 card_index, dropped_frames);
842 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])
844 OutputFrameInfo output_frame_info;
846 unique_lock<mutex> lock(card_mutex, defer_lock);
847 if (master_card_is_output) {
848 // Clocked to the output, so wait for it to be ready for the next frame.
849 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);
852 // Wait for the master card to have a new frame.
853 // TODO: Add a timeout.
854 output_frame_info.is_preroll = false;
856 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(); });
859 if (master_card_is_output) {
860 handle_hotplugged_cards();
861 } else if (cards[master_card_index].new_frames.empty()) {
862 // We were woken up, but not due to a new frame. Deal with it
864 assert(cards[master_card_index].capture->get_disconnected());
865 handle_hotplugged_cards();
869 if (!master_card_is_output) {
870 output_frame_info.frame_timestamp =
871 cards[master_card_index].new_frames.front().received_timestamp;
874 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
875 CaptureCard *card = &cards[card_index];
876 if (input_card_is_master_clock(card_index, master_card_index)) {
877 // We don't use the queue length policy for the master card,
878 // but we will if it stops being the master. Thus, clear out
879 // the policy in case we switch in the future.
880 card->queue_length_policy.reset(card_index);
881 assert(!card->new_frames.empty());
883 trim_queue(card, card_index);
885 if (!card->new_frames.empty()) {
886 new_frames[card_index] = move(card->new_frames.front());
887 has_new_frame[card_index] = true;
888 card->new_frames.pop_front();
889 card->new_frames_changed.notify_all();
893 if (!master_card_is_output) {
894 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
895 output_frame_info.frame_duration = new_frames[master_card_index].length;
898 // This might get off by a fractional sample when changing master card
899 // between ones with different frame rates, but that's fine.
900 int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
901 output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
902 fractional_samples = num_samples_times_timebase % TIMEBASE;
903 assert(output_frame_info.num_samples >= 0);
905 return output_frame_info;
908 void Mixer::handle_hotplugged_cards()
910 // Check for cards that have been disconnected since last frame.
911 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
912 CaptureCard *card = &cards[card_index];
913 if (card->capture->get_disconnected()) {
914 fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
915 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
916 configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
917 card->queue_length_policy.reset(card_index);
918 card->capture->start_bm_capture();
922 // Check for cards that have been connected since last frame.
923 vector<libusb_device *> hotplugged_cards_copy;
925 lock_guard<mutex> lock(hotplug_mutex);
926 swap(hotplugged_cards, hotplugged_cards_copy);
928 for (libusb_device *new_dev : hotplugged_cards_copy) {
929 // Look for a fake capture card where we can stick this in.
930 int free_card_index = -1;
931 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
932 if (cards[card_index].is_fake_capture) {
933 free_card_index = card_index;
938 if (free_card_index == -1) {
939 fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
940 libusb_unref_device(new_dev);
942 // BMUSBCapture takes ownership.
943 fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
944 CaptureCard *card = &cards[free_card_index];
945 BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
946 configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
947 card->queue_length_policy.reset(free_card_index);
948 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
949 capture->start_bm_capture();
955 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)
957 // Resample the audio as needed, including from previously dropped frames.
958 assert(num_cards > 0);
959 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
960 const bool dropped_frame = (frame_num != dropped_frames);
962 // Signal to the audio thread to process this frame.
963 // Note that if the frame is a dropped frame, we signal that
964 // we don't want to use this frame as base for adjusting
965 // the resampler rate. The reason for this is that the timing
966 // of these frames is often way too late; they typically don't
967 // “arrive” before we synthesize them. Thus, we could end up
968 // in a situation where we have inserted e.g. five audio frames
969 // into the queue before we then start pulling five of them
970 // back out. This makes ResamplingQueue overestimate the delay,
971 // causing undue resampler changes. (We _do_ use the last,
972 // non-dropped frame; perhaps we should just discard that as well,
973 // since dropped frames are expected to be rare, and it might be
974 // better to just wait until we have a slightly more normal situation).
975 unique_lock<mutex> lock(audio_mutex);
976 bool adjust_rate = !dropped_frame && !is_preroll;
977 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
978 audio_task_queue_changed.notify_one();
981 // For dropped frames, increase the pts. Note that if the format changed
982 // in the meantime, we have no way of detecting that; we just have to
983 // assume the frame length is always the same.
984 pts_int += length_per_frame;
989 void Mixer::render_one_frame(int64_t duration)
991 // Determine the time code for this frame before we start rendering.
992 string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
993 if (display_timecode_on_stdout) {
994 printf("Timecode: '%s'\n", timecode_text.c_str());
997 // Get the main chain from the theme, and set its state immediately.
998 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
999 EffectChain *chain = theme_main_chain.chain;
1000 theme_main_chain.setup_chain();
1001 //theme_main_chain.chain->enable_phase_timing(true);
1003 GLuint y_tex, cbcr_tex;
1004 bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
1007 // Render main chain.
1008 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
1009 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, global_flags.width, global_flags.height); // Saves texture bandwidth, although dithering gets messed up.
1010 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
1012 chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1014 if (display_timecode_in_stream) {
1015 // Render the timecode on top.
1016 timecode_renderer->render_timecode(fbo, timecode_text);
1019 resource_pool->release_fbo(fbo);
1021 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1022 if (output_card_index != -1) {
1023 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, theme_main_chain.input_frames, pts_int, duration);
1025 resource_pool->release_2d_texture(cbcr_full_tex);
1027 // Set the right state for rgba_tex.
1028 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1029 glBindTexture(GL_TEXTURE_2D, rgba_tex);
1030 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1031 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1032 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1034 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1035 RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
1037 // The live frame just shows the RGBA texture we just rendered.
1038 // It owns rgba_tex now.
1039 DisplayFrame live_frame;
1040 live_frame.chain = display_chain.get();
1041 live_frame.setup_chain = [this, rgba_tex]{
1042 display_input->set_texture_num(rgba_tex);
1044 live_frame.ready_fence = fence;
1045 live_frame.input_frames = {};
1046 live_frame.temp_textures = { rgba_tex };
1047 output_channel[OUTPUT_LIVE].output_frame(live_frame);
1049 // Set up preview and any additional channels.
1050 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1051 DisplayFrame display_frame;
1052 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions
1053 display_frame.chain = chain.chain;
1054 display_frame.setup_chain = chain.setup_chain;
1055 display_frame.ready_fence = fence;
1056 display_frame.input_frames = chain.input_frames;
1057 display_frame.temp_textures = {};
1058 output_channel[i].output_frame(display_frame);
1062 void Mixer::audio_thread_func()
1064 pthread_setname_np(pthread_self(), "Mixer_Audio");
1066 while (!should_quit) {
1070 unique_lock<mutex> lock(audio_mutex);
1071 audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1075 task = audio_task_queue.front();
1076 audio_task_queue.pop();
1079 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1080 task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1081 vector<float> samples_out = audio_mixer.get_output(
1082 task.frame_timestamp,
1084 rate_adjustment_policy);
1086 // Send the samples to the sound card, then add them to the output.
1088 alsa->write(samples_out);
1090 if (output_card_index != -1) {
1091 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1092 cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1094 video_encoder->add_audio(task.pts_int, move(samples_out));
1098 void Mixer::release_display_frame(DisplayFrame *frame)
1100 for (GLuint texnum : frame->temp_textures) {
1101 resource_pool->release_2d_texture(texnum);
1103 frame->temp_textures.clear();
1104 frame->ready_fence.reset();
1105 frame->input_frames.clear();
1110 mixer_thread = thread(&Mixer::thread_func, this);
1111 audio_thread = thread(&Mixer::audio_thread_func, this);
1117 audio_task_queue_changed.notify_one();
1118 mixer_thread.join();
1119 audio_thread.join();
1122 void Mixer::transition_clicked(int transition_num)
1124 theme->transition_clicked(transition_num, pts());
1127 void Mixer::channel_clicked(int preview_num)
1129 theme->channel_clicked(preview_num);
1132 void Mixer::start_mode_scanning(unsigned card_index)
1134 assert(card_index < num_cards);
1135 if (is_mode_scanning[card_index]) {
1138 is_mode_scanning[card_index] = true;
1139 mode_scanlist[card_index].clear();
1140 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1141 mode_scanlist[card_index].push_back(mode.first);
1143 assert(!mode_scanlist[card_index].empty());
1144 mode_scanlist_index[card_index] = 0;
1145 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1146 last_mode_scan_change[card_index] = steady_clock::now();
1149 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1151 assert(desired_output_card_index != -1);
1152 unique_lock<mutex> lock(card_mutex);
1153 return cards[desired_output_card_index].output->get_available_video_modes();
1156 Mixer::OutputChannel::~OutputChannel()
1158 if (has_current_frame) {
1159 parent->release_display_frame(¤t_frame);
1161 if (has_ready_frame) {
1162 parent->release_display_frame(&ready_frame);
1166 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1168 // Store this frame for display. Remove the ready frame if any
1169 // (it was seemingly never used).
1171 unique_lock<mutex> lock(frame_mutex);
1172 if (has_ready_frame) {
1173 parent->release_display_frame(&ready_frame);
1175 ready_frame = frame;
1176 has_ready_frame = true;
1179 if (new_frame_ready_callback) {
1180 new_frame_ready_callback();
1183 // Reduce the number of callbacks by filtering duplicates. The reason
1184 // why we bother doing this is that Qt seemingly can get into a state
1185 // where its builds up an essentially unbounded queue of signals,
1186 // consuming more and more memory, and there's no good way of collapsing
1187 // user-defined signals or limiting the length of the queue.
1188 if (transition_names_updated_callback) {
1189 vector<string> transition_names = global_mixer->get_transition_names();
1190 bool changed = false;
1191 if (transition_names.size() != last_transition_names.size()) {
1194 for (unsigned i = 0; i < transition_names.size(); ++i) {
1195 if (transition_names[i] != last_transition_names[i]) {
1202 transition_names_updated_callback(transition_names);
1203 last_transition_names = transition_names;
1206 if (name_updated_callback) {
1207 string name = global_mixer->get_channel_name(channel);
1208 if (name != last_name) {
1209 name_updated_callback(name);
1213 if (color_updated_callback) {
1214 string color = global_mixer->get_channel_color(channel);
1215 if (color != last_color) {
1216 color_updated_callback(color);
1222 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1224 unique_lock<mutex> lock(frame_mutex);
1225 if (!has_current_frame && !has_ready_frame) {
1229 if (has_current_frame && has_ready_frame) {
1230 // We have a new ready frame. Toss the current one.
1231 parent->release_display_frame(¤t_frame);
1232 has_current_frame = false;
1234 if (has_ready_frame) {
1235 assert(!has_current_frame);
1236 current_frame = ready_frame;
1237 ready_frame.ready_fence.reset(); // Drop the refcount.
1238 ready_frame.input_frames.clear(); // Drop the refcounts.
1239 has_current_frame = true;
1240 has_ready_frame = false;
1243 *frame = current_frame;
1247 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1249 new_frame_ready_callback = callback;
1252 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1254 transition_names_updated_callback = callback;
1257 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1259 name_updated_callback = callback;
1262 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1264 color_updated_callback = callback;
1267 mutex RefCountedGLsync::fence_lock;