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"
42 #include "ffmpeg_capture.h"
44 #include "input_mapping.h"
46 #include "pbo_frame_allocator.h"
47 #include "ref_counted_gl_sync.h"
48 #include "resampling_queue.h"
50 #include "timecode_renderer.h"
51 #include "v210_converter.h"
52 #include "video_encoder.h"
57 using namespace movit;
59 using namespace std::chrono;
60 using namespace std::placeholders;
61 using namespace bmusb;
63 Mixer *global_mixer = nullptr;
64 bool uses_mlock = false;
68 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
71 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
72 input_state->buffered_frames[card_index][frame_num] =
73 input_state->buffered_frames[card_index][frame_num - 1];
75 input_state->buffered_frames[card_index][0] = { frame, field_num };
77 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
78 input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
83 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
86 switch (userdata->pixel_format) {
87 case PixelFormat_10BitYCbCr:
88 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
90 case PixelFormat_8BitYCbCr:
91 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
93 case PixelFormat_8BitBGRA:
94 first = userdata->tex_rgba[field] == 0;
96 case PixelFormat_8BitYCbCrPlanar:
97 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
104 width != userdata->last_width[field] ||
105 height != userdata->last_height[field] ||
106 cbcr_width != userdata->last_cbcr_width[field] ||
107 cbcr_height != userdata->last_cbcr_height[field]) {
108 // We changed resolution since last use of this texture, so we need to create
109 // a new object. Note that this each card has its own PBOFrameAllocator,
110 // we don't need to worry about these flip-flopping between resolutions.
111 switch (userdata->pixel_format) {
112 case PixelFormat_10BitYCbCr:
113 glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
115 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
118 case PixelFormat_8BitYCbCr: {
119 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
121 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
123 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
125 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
129 case PixelFormat_8BitYCbCrPlanar: {
130 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
132 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
134 glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
136 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
138 glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
140 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
144 case PixelFormat_8BitBGRA:
145 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
147 if (global_flags.can_disable_srgb_decoder) { // See the comments in tweaked_inputs.h.
148 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
150 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
155 userdata->last_width[field] = width;
156 userdata->last_height[field] = height;
157 userdata->last_cbcr_width[field] = cbcr_width;
158 userdata->last_cbcr_height[field] = cbcr_height;
160 if (global_flags.ten_bit_input &&
161 (first || v210_width != userdata->last_v210_width[field])) {
162 // Same as above; we need to recreate the texture.
163 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
165 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
167 userdata->last_v210_width[field] = v210_width;
171 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
173 if (interlaced_stride) {
176 if (global_flags.flush_pbos) {
177 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
181 glBindTexture(GL_TEXTURE_2D, tex);
183 if (interlaced_stride) {
184 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
187 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
191 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
193 glBindTexture(GL_TEXTURE_2D, 0);
195 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
201 void JitterHistory::register_metrics(const vector<pair<string, string>> &labels)
203 global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames);
204 global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE);
207 void JitterHistory::unregister_metrics(const vector<pair<string, string>> &labels)
209 global_metrics.remove("input_underestimated_jitter_frames", labels);
210 global_metrics.remove("input_estimated_max_jitter_seconds", labels);
213 void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames)
215 if (expected_timestamp > steady_clock::time_point::min()) {
216 expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE);
217 double jitter_seconds = fabs(duration<double>(expected_timestamp - now).count());
218 history.push_back(orders.insert(jitter_seconds));
219 if (jitter_seconds > estimate_max_jitter()) {
220 ++metric_input_underestimated_jitter_frames;
223 metric_input_estimated_max_jitter_seconds = estimate_max_jitter();
225 if (history.size() > history_length) {
226 orders.erase(history.front());
229 assert(history.size() <= history_length);
231 expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE);
234 double JitterHistory::estimate_max_jitter() const
236 if (orders.empty()) {
239 size_t elem_idx = lrint((orders.size() - 1) * percentile);
240 if (percentile <= 0.5) {
241 return *next(orders.begin(), elem_idx) * multiplier;
243 return *prev(orders.end(), elem_idx + 1) * multiplier;
247 void QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
249 global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
252 void QueueLengthPolicy::unregister_metrics(const vector<pair<string, string>> &labels)
254 global_metrics.remove("input_queue_safe_length_frames", labels);
257 void QueueLengthPolicy::update_policy(steady_clock::time_point now,
258 steady_clock::time_point expected_next_frame,
259 int64_t input_frame_duration,
260 int64_t master_frame_duration,
261 double max_input_card_jitter_seconds,
262 double max_master_card_jitter_seconds)
264 double input_frame_duration_seconds = input_frame_duration / double(TIMEBASE);
265 double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
267 // Figure out when we can expect the next frame for this card, assuming
268 // worst-case jitter (ie., the frame is maximally late).
269 double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
271 // How many times are the master card expected to tick in that time?
272 // We assume the master clock has worst-case jitter but not any rate
273 // discrepancy, ie., it ticks as early as possible every time, but not
275 double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
277 // As a special case, if the master card ticks faster than the input card,
278 // we expect the queue to drain by itself even without dropping. But if
279 // the difference is small (e.g. 60 Hz master and 59.94 input), it would
280 // go slowly enough that the effect wouldn't really be appreciable.
281 // We account for this by looking at the situation five frames ahead,
282 // assuming everything else is the same.
283 double frames_allowed;
284 if (master_frame_duration < input_frame_duration) {
285 frames_allowed = frames_needed + 5 * (input_frame_duration_seconds - master_frame_duration_seconds) / master_frame_duration_seconds;
287 frames_allowed = frames_needed;
290 safe_queue_length = max<int>(floor(frames_allowed), 0);
291 metric_input_queue_safe_length_frames = safe_queue_length;
294 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
296 num_cards(num_cards),
297 mixer_surface(create_surface(format)),
298 h264_encoder_surface(create_surface(format)),
299 decklink_output_surface(create_surface(format)),
300 audio_mixer(num_cards)
302 memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
303 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
306 // This nearly always should be true.
307 global_flags.can_disable_srgb_decoder =
308 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
309 epoxy_has_gl_extension("GL_ARB_sampler_objects");
311 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
312 // will be halved when sampling them, and we need to compensate here.
313 movit_texel_subpixel_precision /= 2.0;
315 resource_pool.reset(new ResourcePool);
316 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
317 output_channel[i].parent = this;
318 output_channel[i].channel = i;
321 ImageFormat inout_format;
322 inout_format.color_space = COLORSPACE_sRGB;
323 inout_format.gamma_curve = GAMMA_sRGB;
325 // Matches the 4:2:0 format created by the main chain.
326 YCbCrFormat ycbcr_format;
327 ycbcr_format.chroma_subsampling_x = 2;
328 ycbcr_format.chroma_subsampling_y = 2;
329 if (global_flags.ycbcr_rec709_coefficients) {
330 ycbcr_format.luma_coefficients = YCBCR_REC_709;
332 ycbcr_format.luma_coefficients = YCBCR_REC_601;
334 ycbcr_format.full_range = false;
335 ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
336 ycbcr_format.cb_x_position = 0.0f;
337 ycbcr_format.cr_x_position = 0.0f;
338 ycbcr_format.cb_y_position = 0.5f;
339 ycbcr_format.cr_y_position = 0.5f;
341 // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
342 display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
344 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
345 display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
346 display_chain->add_input(display_input);
347 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
348 display_chain->set_dither_bits(0); // Don't bother.
349 display_chain->finalize();
351 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));
353 // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
354 theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
356 // Start listening for clients only once VideoEncoder has written its header, if any.
359 // First try initializing the then PCI devices, then USB, then
360 // fill up with fake cards until we have the desired number of cards.
361 unsigned num_pci_devices = 0;
362 unsigned card_index = 0;
365 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
366 if (decklink_iterator != nullptr) {
367 for ( ; card_index < num_cards; ++card_index) {
369 if (decklink_iterator->Next(&decklink) != S_OK) {
373 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
374 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
375 output->set_device(decklink);
376 configure_card(card_index, capture, CardType::LIVE_CARD, output);
379 decklink_iterator->Release();
380 fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
382 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
386 unsigned num_usb_devices = BMUSBCapture::num_cards();
387 for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
388 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
389 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
390 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
392 fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
394 unsigned num_fake_cards = 0;
395 for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
396 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
397 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
400 if (num_fake_cards > 0) {
401 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
404 // Initialize all video inputs the theme asked for. Note that these are
405 // all put _after_ the regular cards, which stop at <num_cards> - 1.
406 std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
407 for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
408 if (card_index >= MAX_VIDEO_CARDS) {
409 fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
412 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
413 video_inputs[video_card_index]->set_card_index(card_index);
415 num_video_inputs = video_inputs.size();
417 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
418 BMUSBCapture::start_bm_thread();
420 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
421 cards[card_index].queue_length_policy.reset(card_index);
424 chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
426 if (global_flags.ten_bit_input) {
427 if (!v210Converter::has_hardware_support()) {
428 fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
429 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
432 v210_converter.reset(new v210Converter());
434 // These are all the widths listed in the Blackmagic SDK documentation
435 // (section 2.7.3, “Display Modes”).
436 v210_converter->precompile_shader(720);
437 v210_converter->precompile_shader(1280);
438 v210_converter->precompile_shader(1920);
439 v210_converter->precompile_shader(2048);
440 v210_converter->precompile_shader(3840);
441 v210_converter->precompile_shader(4096);
443 if (global_flags.ten_bit_output) {
444 if (!v210Converter::has_hardware_support()) {
445 fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
446 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
451 timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
452 display_timecode_in_stream = global_flags.display_timecode_in_stream;
453 display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
455 if (global_flags.enable_alsa_output) {
456 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
458 if (global_flags.output_card != -1) {
459 desired_output_card_index = global_flags.output_card;
460 set_output_card_internal(global_flags.output_card);
463 metric_start_time_seconds = get_timestamp_for_metrics();
465 output_jitter_history.register_metrics({{ "card", "output" }});
466 global_metrics.add("frames_output_total", &metric_frames_output_total);
467 global_metrics.add("frames_output_dropped", &metric_frames_output_dropped);
468 global_metrics.add("start_time_seconds", &metric_start_time_seconds, Metrics::TYPE_GAUGE);
469 global_metrics.add("memory_used_bytes", &metrics_memory_used_bytes);
470 global_metrics.add("memory_locked_limit_bytes", &metrics_memory_locked_limit_bytes);
475 BMUSBCapture::stop_bm_thread();
477 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
479 unique_lock<mutex> lock(card_mutex);
480 cards[card_index].should_quit = true; // Unblock thread.
481 cards[card_index].new_frames_changed.notify_all();
483 cards[card_index].capture->stop_dequeue_thread();
484 if (cards[card_index].output) {
485 cards[card_index].output->end_output();
486 cards[card_index].output.reset();
490 video_encoder.reset(nullptr);
493 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
495 printf("Configuring card %d...\n", card_index);
497 CaptureCard *card = &cards[card_index];
498 if (card->capture != nullptr) {
499 card->capture->stop_dequeue_thread();
501 card->capture.reset(capture);
502 card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
503 card->type = card_type;
504 if (card->output.get() != output) {
505 card->output.reset(output);
508 PixelFormat pixel_format;
509 if (card_type == CardType::FFMPEG_INPUT) {
510 pixel_format = capture->get_current_pixel_format();
511 } else if (global_flags.ten_bit_input) {
512 pixel_format = PixelFormat_10BitYCbCr;
514 pixel_format = PixelFormat_8BitYCbCr;
517 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
518 if (card->frame_allocator == nullptr) {
519 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height)); // 8 MB.
521 card->capture->set_video_frame_allocator(card->frame_allocator.get());
522 if (card->surface == nullptr) {
523 card->surface = create_surface_with_same_format(mixer_surface);
525 while (!card->new_frames.empty()) card->new_frames.pop_front();
526 card->last_timecode = -1;
527 card->capture->set_pixel_format(pixel_format);
528 card->capture->configure_card();
530 // NOTE: start_bm_capture() happens in thread_func().
532 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
533 audio_mixer.reset_resampler(device);
534 audio_mixer.set_display_name(device, card->capture->get_description());
535 audio_mixer.trigger_state_changed_callback();
537 // Unregister old metrics, if any.
538 if (!card->labels.empty()) {
539 const vector<pair<string, string>> &labels = card->labels;
540 card->jitter_history.unregister_metrics(labels);
541 card->queue_length_policy.unregister_metrics(labels);
542 global_metrics.remove("input_received_frames", labels);
543 global_metrics.remove("input_dropped_frames_jitter", labels);
544 global_metrics.remove("input_dropped_frames_error", labels);
545 global_metrics.remove("input_dropped_frames_resets", labels);
546 global_metrics.remove("input_queue_length_frames", labels);
547 global_metrics.remove("input_queue_duped_frames", labels);
549 global_metrics.remove("input_has_signal_bool", labels);
550 global_metrics.remove("input_is_connected_bool", labels);
551 global_metrics.remove("input_interlaced_bool", labels);
552 global_metrics.remove("input_width_pixels", labels);
553 global_metrics.remove("input_height_pixels", labels);
554 global_metrics.remove("input_frame_rate_nom", labels);
555 global_metrics.remove("input_frame_rate_den", labels);
556 global_metrics.remove("input_sample_rate_hz", labels);
560 vector<pair<string, string>> labels;
562 snprintf(card_name, sizeof(card_name), "%d", card_index);
563 labels.emplace_back("card", card_name);
566 case CardType::LIVE_CARD:
567 labels.emplace_back("cardtype", "live");
569 case CardType::FAKE_CAPTURE:
570 labels.emplace_back("cardtype", "fake");
572 case CardType::FFMPEG_INPUT:
573 labels.emplace_back("cardtype", "ffmpeg");
578 card->jitter_history.register_metrics(labels);
579 card->queue_length_policy.register_metrics(labels);
580 global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
581 global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
582 global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
583 global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
584 global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
585 global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
587 global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
588 global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
589 global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
590 global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
591 global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
592 global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
593 global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
594 global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
595 card->labels = labels;
598 void Mixer::set_output_card_internal(int card_index)
600 // We don't really need to take card_mutex, since we're in the mixer
601 // thread and don't mess with any queues (which is the only thing that happens
602 // from other threads), but it's probably the safest in the long run.
603 unique_lock<mutex> lock(card_mutex);
604 if (output_card_index != -1) {
605 // Switch the old card from output to input.
606 CaptureCard *old_card = &cards[output_card_index];
607 old_card->output->end_output();
609 // Stop the fake card that we put into place.
610 // This needs to _not_ happen under the mutex, to avoid deadlock
611 // (delivering the last frame needs to take the mutex).
612 CaptureInterface *fake_capture = old_card->capture.get();
614 fake_capture->stop_dequeue_thread();
616 old_card->capture = move(old_card->parked_capture); // TODO: reset the metrics
617 old_card->is_fake_capture = false;
618 old_card->capture->start_bm_capture();
620 if (card_index != -1) {
621 CaptureCard *card = &cards[card_index];
622 CaptureInterface *capture = card->capture.get();
623 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
624 // several seconds to complete (blocking on DisableVideoInput);
625 // see if we can maybe do it asynchronously.
627 capture->stop_dequeue_thread();
629 card->parked_capture = move(card->capture);
630 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
631 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
632 card->queue_length_policy.reset(card_index);
633 card->capture->start_bm_capture();
634 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
635 card->output->start_output(desired_output_video_mode, pts_int);
637 output_card_index = card_index;
638 output_jitter_history.clear();
643 int unwrap_timecode(uint16_t current_wrapped, int last)
645 uint16_t last_wrapped = last & 0xffff;
646 if (current_wrapped > last_wrapped) {
647 return (last & ~0xffff) | current_wrapped;
649 return 0x10000 + ((last & ~0xffff) | current_wrapped);
655 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
656 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
657 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
659 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
660 CaptureCard *card = &cards[card_index];
662 ++card->metric_input_received_frames;
663 card->metric_input_has_signal_bool = video_format.has_signal;
664 card->metric_input_is_connected_bool = video_format.is_connected;
665 card->metric_input_interlaced_bool = video_format.interlaced;
666 card->metric_input_width_pixels = video_format.width;
667 card->metric_input_height_pixels = video_format.height;
668 card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
669 card->metric_input_frame_rate_den = video_format.frame_rate_den;
670 card->metric_input_sample_rate_hz = audio_format.sample_rate;
672 if (is_mode_scanning[card_index]) {
673 if (video_format.has_signal) {
674 // Found a stable signal, so stop scanning.
675 is_mode_scanning[card_index] = false;
677 static constexpr double switch_time_s = 0.1; // Should be enough time for the signal to stabilize.
678 steady_clock::time_point now = steady_clock::now();
679 double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
680 if (sec_since_last_switch > switch_time_s) {
681 // It isn't this mode; try the next one.
682 mode_scanlist_index[card_index]++;
683 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
684 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
685 last_mode_scan_change[card_index] = now;
690 int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
691 assert(frame_length > 0);
693 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;
694 if (num_samples > OUTPUT_FREQUENCY / 10) {
695 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",
696 card_index, int(audio_frame.len), int(audio_offset),
697 timecode, int(video_frame.len), int(video_offset), video_format.id);
698 if (video_frame.owner) {
699 video_frame.owner->release_frame(video_frame);
701 if (audio_frame.owner) {
702 audio_frame.owner->release_frame(audio_frame);
707 int dropped_frames = 0;
708 if (card->last_timecode != -1) {
709 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
712 // Number of samples per frame if we need to insert silence.
713 // (Could be nonintegral, but resampling will save us then.)
714 const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
716 if (dropped_frames > MAX_FPS * 2) {
717 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
718 card_index, card->last_timecode, timecode);
719 audio_mixer.reset_resampler(device);
721 ++card->metric_input_resets;
722 } else if (dropped_frames > 0) {
723 // Insert silence as needed.
724 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
725 card_index, dropped_frames, timecode);
726 card->metric_input_dropped_frames_error += dropped_frames;
730 success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
734 if (num_samples > 0) {
735 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
738 // Done with the audio, so release it.
739 if (audio_frame.owner) {
740 audio_frame.owner->release_frame(audio_frame);
743 card->last_timecode = timecode;
745 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
747 size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
748 size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
749 if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
750 // The calculation above is wrong for planar Y'CbCr, so just override it.
751 assert(card->type == CardType::FFMPEG_INPUT);
752 assert(video_offset == 0);
753 expected_length = video_frame.len;
755 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
756 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
757 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
758 cbcr_offset = video_format.width * video_format.height;
761 // All the other Y'CbCr formats are 4:2:2.
762 cbcr_width = video_format.width / 2;
763 cbcr_height = video_format.height;
764 cbcr_offset = video_offset / 2;
765 y_offset = video_frame.size / 2 + video_offset / 2;
767 if (video_frame.len - video_offset == 0 ||
768 video_frame.len - video_offset != expected_length) {
769 if (video_frame.len != 0) {
770 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
771 card_index, video_frame.len - video_offset, expected_length);
773 if (video_frame.owner) {
774 video_frame.owner->release_frame(video_frame);
777 // Still send on the information that we _had_ a frame, even though it's corrupted,
778 // so that pts can go up accordingly.
780 unique_lock<mutex> lock(card_mutex);
781 CaptureCard::NewFrame new_frame;
782 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
783 new_frame.length = frame_length;
784 new_frame.interlaced = false;
785 new_frame.dropped_frames = dropped_frames;
786 new_frame.received_timestamp = video_frame.received_timestamp;
787 card->new_frames.push_back(move(new_frame));
788 card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
790 card->new_frames_changed.notify_all();
794 unsigned num_fields = video_format.interlaced ? 2 : 1;
795 steady_clock::time_point frame_upload_start;
796 bool interlaced_stride = false;
797 if (video_format.interlaced) {
798 // Send the two fields along as separate frames; the other side will need to add
799 // a deinterlacer to actually get this right.
800 assert(video_format.height % 2 == 0);
801 video_format.height /= 2;
803 assert(frame_length % 2 == 0);
806 if (video_format.second_field_start == 1) {
807 interlaced_stride = true;
809 frame_upload_start = steady_clock::now();
811 userdata->last_interlaced = video_format.interlaced;
812 userdata->last_has_signal = video_format.has_signal;
813 userdata->last_is_connected = video_format.is_connected;
814 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
815 userdata->last_frame_rate_den = video_format.frame_rate_den;
816 RefCountedFrame frame(video_frame);
818 // Upload the textures.
819 for (unsigned field = 0; field < num_fields; ++field) {
820 // Put the actual texture upload in a lambda that is executed in the main thread.
821 // It is entirely possible to do this in the same thread (and it might even be
822 // faster, depending on the GPU and driver), but it appears to be trickling
823 // driver bugs very easily.
825 // Note that this means we must hold on to the actual frame data in <userdata>
826 // until the upload command is run, but we hold on to <frame> much longer than that
827 // (in fact, all the way until we no longer use the texture in rendering).
828 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
829 unsigned field_start_line;
831 field_start_line = video_format.second_field_start;
833 field_start_line = video_format.extra_lines_top;
836 // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
837 size_t v210_width = video_format.stride / sizeof(uint32_t);
838 ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
840 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
843 switch (userdata->pixel_format) {
844 case PixelFormat_10BitYCbCr: {
845 size_t field_start = video_offset + video_format.stride * field_start_line;
846 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);
847 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
850 case PixelFormat_8BitYCbCr: {
851 size_t field_y_start = y_offset + video_format.width * field_start_line;
852 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
854 // Make up our own strides, since we are interleaving.
855 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);
856 upload_texture(userdata->tex_cbcr[field], cbcr_width, cbcr_height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
859 case PixelFormat_8BitYCbCrPlanar: {
860 assert(field_start_line == 0); // We don't really support interlaced here.
861 size_t field_y_start = y_offset;
862 size_t field_cb_start = cbcr_offset;
863 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
865 // Make up our own strides, since we are interleaving.
866 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);
867 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
868 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
871 case PixelFormat_8BitBGRA: {
872 size_t field_start = video_offset + video_format.stride * field_start_line;
873 upload_texture(userdata->tex_rgba[field], video_format.width, video_format.height, video_format.stride, interlaced_stride, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, field_start);
874 // These could be asked to deliver mipmaps at any time.
875 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
877 glGenerateMipmap(GL_TEXTURE_2D);
879 glBindTexture(GL_TEXTURE_2D, 0);
887 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
892 // Don't upload the second field as fast as we can; wait until
893 // the field time has approximately passed. (Otherwise, we could
894 // get timing jitter against the other sources, and possibly also
895 // against the video display, although the latter is not as critical.)
896 // This requires our system clock to be reasonably close to the
897 // video clock, but that's not an unreasonable assumption.
898 steady_clock::time_point second_field_start = frame_upload_start +
899 nanoseconds(frame_length * 1000000000 / TIMEBASE);
900 this_thread::sleep_until(second_field_start);
904 unique_lock<mutex> lock(card_mutex);
905 CaptureCard::NewFrame new_frame;
906 new_frame.frame = frame;
907 new_frame.length = frame_length;
908 new_frame.field = field;
909 new_frame.interlaced = video_format.interlaced;
910 new_frame.upload_func = upload_func;
911 new_frame.dropped_frames = dropped_frames;
912 new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp.
913 card->new_frames.push_back(move(new_frame));
914 card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
916 card->new_frames_changed.notify_all();
920 void Mixer::bm_hotplug_add(libusb_device *dev)
922 lock_guard<mutex> lock(hotplug_mutex);
923 hotplugged_cards.push_back(dev);
926 void Mixer::bm_hotplug_remove(unsigned card_index)
928 cards[card_index].new_frames_changed.notify_all();
931 void Mixer::thread_func()
933 pthread_setname_np(pthread_self(), "Mixer_OpenGL");
935 eglBindAPI(EGL_OPENGL_API);
936 QOpenGLContext *context = create_context(mixer_surface);
937 if (!make_current(context, mixer_surface)) {
942 // Start the actual capture. (We don't want to do it before we're actually ready
943 // to process output frames.)
944 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
945 if (int(card_index) != output_card_index) {
946 cards[card_index].capture->start_bm_capture();
950 steady_clock::time_point start, now;
951 start = steady_clock::now();
953 int stats_dropped_frames = 0;
955 while (!should_quit) {
956 if (desired_output_card_index != output_card_index) {
957 set_output_card_internal(desired_output_card_index);
959 if (output_card_index != -1 &&
960 desired_output_video_mode != output_video_mode) {
961 DeckLinkOutput *output = cards[output_card_index].output.get();
962 output->end_output();
963 desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
964 output->start_output(desired_output_video_mode, pts_int);
967 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
968 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
970 bool master_card_is_output;
971 unsigned master_card_index;
972 if (output_card_index != -1) {
973 master_card_is_output = true;
974 master_card_index = output_card_index;
976 master_card_is_output = false;
977 master_card_index = theme->map_signal(master_clock_channel);
978 assert(master_card_index < num_cards);
981 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
982 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);
983 stats_dropped_frames += output_frame_info.dropped_frames;
985 handle_hotplugged_cards();
987 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
988 if (card_index == master_card_index || !has_new_frame[card_index]) {
991 if (new_frames[card_index].frame->len == 0) {
992 ++new_frames[card_index].dropped_frames;
994 if (new_frames[card_index].dropped_frames > 0) {
995 printf("Card %u dropped %d frames before this\n",
996 card_index, int(new_frames[card_index].dropped_frames));
1000 // If the first card is reporting a corrupted or otherwise dropped frame,
1001 // just increase the pts (skipping over this frame) and don't try to compute anything new.
1002 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
1003 ++stats_dropped_frames;
1004 pts_int += new_frames[master_card_index].length;
1008 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1009 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1012 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1013 assert(new_frame->frame != nullptr);
1014 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1017 // The new texture might need uploading before use.
1018 if (new_frame->upload_func) {
1019 new_frame->upload_func();
1020 new_frame->upload_func = nullptr;
1024 int64_t frame_duration = output_frame_info.frame_duration;
1025 render_one_frame(frame_duration);
1027 pts_int += frame_duration;
1029 now = steady_clock::now();
1030 double elapsed = duration<double>(now - start).count();
1032 metric_frames_output_total = frame_num;
1033 metric_frames_output_dropped = stats_dropped_frames;
1035 if (frame_num % 100 == 0) {
1036 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
1037 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
1038 1e3 * elapsed / frame_num);
1039 // chain->print_phase_timing();
1041 // Check our memory usage, to see if we are close to our mlockall()
1042 // limit (if at all set).
1044 if (getrusage(RUSAGE_SELF, &used) == -1) {
1045 perror("getrusage(RUSAGE_SELF)");
1051 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
1052 perror("getrlimit(RLIMIT_MEMLOCK)");
1056 if (limit.rlim_cur == 0) {
1057 printf(", using %ld MB memory (locked)",
1058 long(used.ru_maxrss / 1024));
1060 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
1061 long(used.ru_maxrss / 1024),
1062 long(limit.rlim_cur / 1048576),
1063 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
1065 metrics_memory_locked_limit_bytes = limit.rlim_cur;
1067 printf(", using %ld MB memory (not locked)",
1068 long(used.ru_maxrss / 1024));
1069 metrics_memory_locked_limit_bytes = 0.0 / 0.0;
1074 metrics_memory_used_bytes = used.ru_maxrss * 1024;
1078 if (should_cut.exchange(false)) { // Test and clear.
1079 video_encoder->do_cut(frame_num);
1083 // Reset every 100 frames, so that local variations in frame times
1084 // (especially for the first few frames, when the shaders are
1085 // compiled etc.) don't make it hard to measure for the entire
1086 // remaining duration of the program.
1087 if (frame == 10000) {
1095 resource_pool->clean_context();
1098 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1100 if (output_card_index != -1) {
1101 // The output card (ie., cards[output_card_index].output) is the master clock,
1102 // so no input card (ie., cards[card_index].capture) is.
1105 return (card_index == master_card_index);
1108 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1110 // Count the number of frames in the queue, including any frames
1111 // we dropped. It's hard to know exactly how we should deal with
1112 // dropped (corrupted) input frames; they don't help our goal of
1113 // avoiding starvation, but they still add to the problem of latency.
1114 // Since dropped frames is going to mean a bump in the signal anyway,
1115 // we err on the side of having more stable latency instead.
1116 unsigned queue_length = 0;
1117 for (const CaptureCard::NewFrame &frame : card->new_frames) {
1118 queue_length += frame.dropped_frames + 1;
1121 // If needed, drop frames until the queue is below the safe limit.
1122 // We prefer to drop from the head, because all else being equal,
1123 // we'd like more recent frames (less latency).
1124 unsigned dropped_frames = 0;
1125 while (queue_length > safe_queue_length) {
1126 assert(!card->new_frames.empty());
1127 assert(queue_length > card->new_frames.front().dropped_frames);
1128 queue_length -= card->new_frames.front().dropped_frames;
1130 if (queue_length <= safe_queue_length) {
1131 // No need to drop anything.
1135 card->new_frames.pop_front();
1136 card->new_frames_changed.notify_all();
1141 card->metric_input_dropped_frames_jitter += dropped_frames;
1142 card->metric_input_queue_length_frames = queue_length;
1145 if (dropped_frames > 0) {
1146 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1147 card_index, dropped_frames);
1153 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])
1155 OutputFrameInfo output_frame_info;
1157 unique_lock<mutex> lock(card_mutex, defer_lock);
1158 if (master_card_is_output) {
1159 // Clocked to the output, so wait for it to be ready for the next frame.
1160 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);
1163 // Wait for the master card to have a new frame.
1164 // TODO: Add a timeout.
1165 output_frame_info.is_preroll = false;
1167 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(); });
1170 if (master_card_is_output) {
1171 handle_hotplugged_cards();
1172 } else if (cards[master_card_index].new_frames.empty()) {
1173 // We were woken up, but not due to a new frame. Deal with it
1174 // and then restart.
1175 assert(cards[master_card_index].capture->get_disconnected());
1176 handle_hotplugged_cards();
1181 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1182 CaptureCard *card = &cards[card_index];
1183 if (card->new_frames.empty()) { // Starvation.
1184 ++card->metric_input_duped_frames;
1186 new_frames[card_index] = move(card->new_frames.front());
1187 has_new_frame[card_index] = true;
1188 card->new_frames.pop_front();
1189 card->new_frames_changed.notify_all();
1193 if (!master_card_is_output) {
1194 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1195 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1196 output_frame_info.frame_duration = new_frames[master_card_index].length;
1199 if (!output_frame_info.is_preroll) {
1200 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1203 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1204 CaptureCard *card = &cards[card_index];
1205 if (has_new_frame[card_index] &&
1206 !input_card_is_master_clock(card_index, master_card_index) &&
1207 !output_frame_info.is_preroll) {
1208 card->queue_length_policy.update_policy(
1209 output_frame_info.frame_timestamp,
1210 card->jitter_history.get_expected_next_frame(),
1211 new_frames[master_card_index].length,
1212 output_frame_info.frame_duration,
1213 card->jitter_history.estimate_max_jitter(),
1214 output_jitter_history.estimate_max_jitter());
1215 trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1216 card->queue_length_policy.get_safe_queue_length()));
1220 // This might get off by a fractional sample when changing master card
1221 // between ones with different frame rates, but that's fine.
1222 int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1223 output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1224 fractional_samples = num_samples_times_timebase % TIMEBASE;
1225 assert(output_frame_info.num_samples >= 0);
1227 return output_frame_info;
1230 void Mixer::handle_hotplugged_cards()
1232 // Check for cards that have been disconnected since last frame.
1233 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1234 CaptureCard *card = &cards[card_index];
1235 if (card->capture->get_disconnected()) {
1236 fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1237 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1238 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1239 card->queue_length_policy.reset(card_index);
1240 card->capture->start_bm_capture();
1244 // Check for cards that have been connected since last frame.
1245 vector<libusb_device *> hotplugged_cards_copy;
1247 lock_guard<mutex> lock(hotplug_mutex);
1248 swap(hotplugged_cards, hotplugged_cards_copy);
1250 for (libusb_device *new_dev : hotplugged_cards_copy) {
1251 // Look for a fake capture card where we can stick this in.
1252 int free_card_index = -1;
1253 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1254 if (cards[card_index].is_fake_capture) {
1255 free_card_index = card_index;
1260 if (free_card_index == -1) {
1261 fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1262 libusb_unref_device(new_dev);
1264 // BMUSBCapture takes ownership.
1265 fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1266 CaptureCard *card = &cards[free_card_index];
1267 BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1268 configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1269 card->queue_length_policy.reset(free_card_index);
1270 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1271 capture->start_bm_capture();
1277 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)
1279 // Resample the audio as needed, including from previously dropped frames.
1280 assert(num_cards > 0);
1281 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1282 const bool dropped_frame = (frame_num != dropped_frames);
1284 // Signal to the audio thread to process this frame.
1285 // Note that if the frame is a dropped frame, we signal that
1286 // we don't want to use this frame as base for adjusting
1287 // the resampler rate. The reason for this is that the timing
1288 // of these frames is often way too late; they typically don't
1289 // “arrive” before we synthesize them. Thus, we could end up
1290 // in a situation where we have inserted e.g. five audio frames
1291 // into the queue before we then start pulling five of them
1292 // back out. This makes ResamplingQueue overestimate the delay,
1293 // causing undue resampler changes. (We _do_ use the last,
1294 // non-dropped frame; perhaps we should just discard that as well,
1295 // since dropped frames are expected to be rare, and it might be
1296 // better to just wait until we have a slightly more normal situation).
1297 unique_lock<mutex> lock(audio_mutex);
1298 bool adjust_rate = !dropped_frame && !is_preroll;
1299 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1300 audio_task_queue_changed.notify_one();
1302 if (dropped_frame) {
1303 // For dropped frames, increase the pts. Note that if the format changed
1304 // in the meantime, we have no way of detecting that; we just have to
1305 // assume the frame length is always the same.
1306 pts_int += length_per_frame;
1311 void Mixer::render_one_frame(int64_t duration)
1313 // Determine the time code for this frame before we start rendering.
1314 string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1315 if (display_timecode_on_stdout) {
1316 printf("Timecode: '%s'\n", timecode_text.c_str());
1319 // Update Y'CbCr settings for all cards.
1321 unique_lock<mutex> lock(card_mutex);
1322 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1323 YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1324 input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1325 input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1326 input_state.full_range[card_index] = interpretation->full_range;
1330 // Get the main chain from the theme, and set its state immediately.
1331 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1332 EffectChain *chain = theme_main_chain.chain;
1333 theme_main_chain.setup_chain();
1334 //theme_main_chain.chain->enable_phase_timing(true);
1336 // The theme can't (or at least shouldn't!) call connect_signal() on
1337 // each FFmpeg input, so we'll do it here.
1338 for (const pair<LiveInputWrapper *, FFmpegCapture *> &conn : theme->get_signal_connections()) {
1339 conn.first->connect_signal_raw(conn.second->get_card_index(), input_state);
1342 // If HDMI/SDI output is active and the user has requested auto mode,
1343 // its mode overrides the existing Y'CbCr setting for the chain.
1344 YCbCrLumaCoefficients ycbcr_output_coefficients;
1345 if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1346 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1348 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1351 // TODO: Reduce the duplication against theme.cpp.
1352 YCbCrFormat output_ycbcr_format;
1353 output_ycbcr_format.chroma_subsampling_x = 1;
1354 output_ycbcr_format.chroma_subsampling_y = 1;
1355 output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1356 output_ycbcr_format.full_range = false;
1357 output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1358 chain->change_ycbcr_output_format(output_ycbcr_format);
1360 // Render main chain. If we're using zerocopy Quick Sync encoding
1361 // (the default case), we take an extra copy of the created outputs,
1362 // so that we can display it back to the screen later (it's less memory
1363 // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1364 // Ideally, we'd like to avoid taking copies and just use the main textures
1365 // for display as well, but they're just views into VA-API memory and must be
1366 // unmapped during encoding, so we can't use them for display, unfortunately.
1367 GLuint y_tex, cbcr_full_tex, cbcr_tex;
1368 GLuint y_copy_tex, cbcr_copy_tex = 0;
1369 GLuint y_display_tex, cbcr_display_tex;
1370 GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1371 GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1372 const bool is_zerocopy = video_encoder->is_zerocopy();
1374 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1375 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1376 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1378 y_display_tex = y_copy_tex;
1379 cbcr_display_tex = cbcr_copy_tex;
1381 // y_tex and cbcr_tex will be given by VideoEncoder.
1383 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1384 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1385 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1387 y_display_tex = y_tex;
1388 cbcr_display_tex = cbcr_tex;
1391 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1392 bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1397 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1399 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1402 chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1404 if (display_timecode_in_stream) {
1405 // Render the timecode on top.
1406 timecode_renderer->render_timecode(fbo, timecode_text);
1409 resource_pool->release_fbo(fbo);
1412 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1414 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1416 if (output_card_index != -1) {
1417 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1419 resource_pool->release_2d_texture(cbcr_full_tex);
1421 // Set the right state for the Y' and CbCr textures we use for display.
1422 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1423 glBindTexture(GL_TEXTURE_2D, y_display_tex);
1424 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1425 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1426 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1428 glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1429 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1430 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1431 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1433 RefCountedGLsync fence = video_encoder->end_frame();
1435 // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1436 // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1437 DisplayFrame live_frame;
1438 live_frame.chain = display_chain.get();
1439 live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1440 display_input->set_texture_num(0, y_display_tex);
1441 display_input->set_texture_num(1, cbcr_display_tex);
1443 live_frame.ready_fence = fence;
1444 live_frame.input_frames = {};
1445 live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1446 output_channel[OUTPUT_LIVE].output_frame(live_frame);
1448 // Set up preview and any additional channels.
1449 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1450 DisplayFrame display_frame;
1451 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions
1452 display_frame.chain = chain.chain;
1453 display_frame.setup_chain = chain.setup_chain;
1454 display_frame.ready_fence = fence;
1455 display_frame.input_frames = chain.input_frames;
1456 display_frame.temp_textures = {};
1457 output_channel[i].output_frame(display_frame);
1461 void Mixer::audio_thread_func()
1463 pthread_setname_np(pthread_self(), "Mixer_Audio");
1465 while (!should_quit) {
1469 unique_lock<mutex> lock(audio_mutex);
1470 audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1474 task = audio_task_queue.front();
1475 audio_task_queue.pop();
1478 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1479 task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1480 vector<float> samples_out = audio_mixer.get_output(
1481 task.frame_timestamp,
1483 rate_adjustment_policy);
1485 // Send the samples to the sound card, then add them to the output.
1487 alsa->write(samples_out);
1489 if (output_card_index != -1) {
1490 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1491 cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1493 video_encoder->add_audio(task.pts_int, move(samples_out));
1497 void Mixer::release_display_frame(DisplayFrame *frame)
1499 for (GLuint texnum : frame->temp_textures) {
1500 resource_pool->release_2d_texture(texnum);
1502 frame->temp_textures.clear();
1503 frame->ready_fence.reset();
1504 frame->input_frames.clear();
1509 mixer_thread = thread(&Mixer::thread_func, this);
1510 audio_thread = thread(&Mixer::audio_thread_func, this);
1516 audio_task_queue_changed.notify_one();
1517 mixer_thread.join();
1518 audio_thread.join();
1521 void Mixer::transition_clicked(int transition_num)
1523 theme->transition_clicked(transition_num, pts());
1526 void Mixer::channel_clicked(int preview_num)
1528 theme->channel_clicked(preview_num);
1531 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1533 unique_lock<mutex> lock(card_mutex);
1534 return ycbcr_interpretation[card_index];
1537 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1539 unique_lock<mutex> lock(card_mutex);
1540 ycbcr_interpretation[card_index] = interpretation;
1543 void Mixer::start_mode_scanning(unsigned card_index)
1545 assert(card_index < num_cards);
1546 if (is_mode_scanning[card_index]) {
1549 is_mode_scanning[card_index] = true;
1550 mode_scanlist[card_index].clear();
1551 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1552 mode_scanlist[card_index].push_back(mode.first);
1554 assert(!mode_scanlist[card_index].empty());
1555 mode_scanlist_index[card_index] = 0;
1556 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1557 last_mode_scan_change[card_index] = steady_clock::now();
1560 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1562 assert(desired_output_card_index != -1);
1563 unique_lock<mutex> lock(card_mutex);
1564 return cards[desired_output_card_index].output->get_available_video_modes();
1567 Mixer::OutputChannel::~OutputChannel()
1569 if (has_current_frame) {
1570 parent->release_display_frame(¤t_frame);
1572 if (has_ready_frame) {
1573 parent->release_display_frame(&ready_frame);
1577 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1579 // Store this frame for display. Remove the ready frame if any
1580 // (it was seemingly never used).
1582 unique_lock<mutex> lock(frame_mutex);
1583 if (has_ready_frame) {
1584 parent->release_display_frame(&ready_frame);
1586 ready_frame = frame;
1587 has_ready_frame = true;
1589 // Call the callbacks under the mutex (they should be short),
1590 // so that we don't race against a callback removal.
1591 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1592 key_and_callback.second();
1596 // Reduce the number of callbacks by filtering duplicates. The reason
1597 // why we bother doing this is that Qt seemingly can get into a state
1598 // where its builds up an essentially unbounded queue of signals,
1599 // consuming more and more memory, and there's no good way of collapsing
1600 // user-defined signals or limiting the length of the queue.
1601 if (transition_names_updated_callback) {
1602 vector<string> transition_names = global_mixer->get_transition_names();
1603 bool changed = false;
1604 if (transition_names.size() != last_transition_names.size()) {
1607 for (unsigned i = 0; i < transition_names.size(); ++i) {
1608 if (transition_names[i] != last_transition_names[i]) {
1615 transition_names_updated_callback(transition_names);
1616 last_transition_names = transition_names;
1619 if (name_updated_callback) {
1620 string name = global_mixer->get_channel_name(channel);
1621 if (name != last_name) {
1622 name_updated_callback(name);
1626 if (color_updated_callback) {
1627 string color = global_mixer->get_channel_color(channel);
1628 if (color != last_color) {
1629 color_updated_callback(color);
1635 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1637 unique_lock<mutex> lock(frame_mutex);
1638 if (!has_current_frame && !has_ready_frame) {
1642 if (has_current_frame && has_ready_frame) {
1643 // We have a new ready frame. Toss the current one.
1644 parent->release_display_frame(¤t_frame);
1645 has_current_frame = false;
1647 if (has_ready_frame) {
1648 assert(!has_current_frame);
1649 current_frame = ready_frame;
1650 ready_frame.ready_fence.reset(); // Drop the refcount.
1651 ready_frame.input_frames.clear(); // Drop the refcounts.
1652 has_current_frame = true;
1653 has_ready_frame = false;
1656 *frame = current_frame;
1660 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1662 unique_lock<mutex> lock(frame_mutex);
1663 new_frame_ready_callbacks[key] = callback;
1666 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1668 unique_lock<mutex> lock(frame_mutex);
1669 new_frame_ready_callbacks.erase(key);
1672 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1674 transition_names_updated_callback = callback;
1677 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1679 name_updated_callback = callback;
1682 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1684 color_updated_callback = callback;
1687 mutex RefCountedGLsync::fence_lock;