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 master_frame_duration,
260 double max_input_card_jitter_seconds,
261 double max_master_card_jitter_seconds)
263 double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
265 // Figure out when we can expect the next frame for this card, assuming
266 // worst-case jitter (ie., the frame is maximally late).
267 double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
269 // How many times are the master card expected to tick in that time?
270 // We assume the master clock has worst-case jitter but not any rate
271 // discrepancy, ie., it ticks as early as possible every time, but not
273 double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
275 // As a special case, if the master card ticks faster than the input card,
276 // we expect the queue to drain by itself even without dropping. But if
277 // the difference is small (e.g. 60 Hz master and 59.94 input), it would
278 // go slowly enough that the effect wouldn't really be appreciable.
279 // We account for this by looking at the situation five frames ahead,
280 // assuming everything else is the same.
281 double frames_allowed;
282 if (max_master_card_jitter_seconds < max_input_card_jitter_seconds) {
283 frames_allowed = frames_needed + 5 * (max_input_card_jitter_seconds - max_master_card_jitter_seconds) / master_frame_duration_seconds;
285 frames_allowed = frames_needed;
288 safe_queue_length = max<int>(floor(frames_allowed), 0);
289 metric_input_queue_safe_length_frames = safe_queue_length;
292 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
294 num_cards(num_cards),
295 mixer_surface(create_surface(format)),
296 h264_encoder_surface(create_surface(format)),
297 decklink_output_surface(create_surface(format)),
298 audio_mixer(num_cards)
300 memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
301 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
304 // This nearly always should be true.
305 global_flags.can_disable_srgb_decoder =
306 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
307 epoxy_has_gl_extension("GL_ARB_sampler_objects");
309 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
310 // will be halved when sampling them, and we need to compensate here.
311 movit_texel_subpixel_precision /= 2.0;
313 resource_pool.reset(new ResourcePool);
314 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
315 output_channel[i].parent = this;
316 output_channel[i].channel = i;
319 ImageFormat inout_format;
320 inout_format.color_space = COLORSPACE_sRGB;
321 inout_format.gamma_curve = GAMMA_sRGB;
323 // Matches the 4:2:0 format created by the main chain.
324 YCbCrFormat ycbcr_format;
325 ycbcr_format.chroma_subsampling_x = 2;
326 ycbcr_format.chroma_subsampling_y = 2;
327 if (global_flags.ycbcr_rec709_coefficients) {
328 ycbcr_format.luma_coefficients = YCBCR_REC_709;
330 ycbcr_format.luma_coefficients = YCBCR_REC_601;
332 ycbcr_format.full_range = false;
333 ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
334 ycbcr_format.cb_x_position = 0.0f;
335 ycbcr_format.cr_x_position = 0.0f;
336 ycbcr_format.cb_y_position = 0.5f;
337 ycbcr_format.cr_y_position = 0.5f;
339 // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
340 display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
342 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
343 display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
344 display_chain->add_input(display_input);
345 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
346 display_chain->set_dither_bits(0); // Don't bother.
347 display_chain->finalize();
349 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));
351 // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
352 theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
354 // Start listening for clients only once VideoEncoder has written its header, if any.
357 // First try initializing the then PCI devices, then USB, then
358 // fill up with fake cards until we have the desired number of cards.
359 unsigned num_pci_devices = 0;
360 unsigned card_index = 0;
363 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
364 if (decklink_iterator != nullptr) {
365 for ( ; card_index < num_cards; ++card_index) {
367 if (decklink_iterator->Next(&decklink) != S_OK) {
371 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
372 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
373 output->set_device(decklink);
374 configure_card(card_index, capture, CardType::LIVE_CARD, output);
377 decklink_iterator->Release();
378 fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
380 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
384 unsigned num_usb_devices = BMUSBCapture::num_cards();
385 for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
386 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
387 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
388 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
390 fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
392 unsigned num_fake_cards = 0;
393 for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
394 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
395 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
398 if (num_fake_cards > 0) {
399 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
402 // Initialize all video inputs the theme asked for. Note that these are
403 // all put _after_ the regular cards, which stop at <num_cards> - 1.
404 std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
405 for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
406 if (card_index >= MAX_VIDEO_CARDS) {
407 fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
410 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
411 video_inputs[video_card_index]->set_card_index(card_index);
413 num_video_inputs = video_inputs.size();
415 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
416 BMUSBCapture::start_bm_thread();
418 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
419 cards[card_index].queue_length_policy.reset(card_index);
422 chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
424 if (global_flags.ten_bit_input) {
425 if (!v210Converter::has_hardware_support()) {
426 fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
427 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
430 v210_converter.reset(new v210Converter());
432 // These are all the widths listed in the Blackmagic SDK documentation
433 // (section 2.7.3, “Display Modes”).
434 v210_converter->precompile_shader(720);
435 v210_converter->precompile_shader(1280);
436 v210_converter->precompile_shader(1920);
437 v210_converter->precompile_shader(2048);
438 v210_converter->precompile_shader(3840);
439 v210_converter->precompile_shader(4096);
441 if (global_flags.ten_bit_output) {
442 if (!v210Converter::has_hardware_support()) {
443 fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
444 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
449 timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
450 display_timecode_in_stream = global_flags.display_timecode_in_stream;
451 display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
453 if (global_flags.enable_alsa_output) {
454 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
456 if (global_flags.output_card != -1) {
457 desired_output_card_index = global_flags.output_card;
458 set_output_card_internal(global_flags.output_card);
461 metric_start_time_seconds = get_timestamp_for_metrics();
463 output_jitter_history.register_metrics({{ "card", "output" }});
464 global_metrics.add("frames_output_total", &metric_frames_output_total);
465 global_metrics.add("frames_output_dropped", &metric_frames_output_dropped);
466 global_metrics.add("start_time_seconds", &metric_start_time_seconds, Metrics::TYPE_GAUGE);
467 global_metrics.add("memory_used_bytes", &metrics_memory_used_bytes);
468 global_metrics.add("memory_locked_limit_bytes", &metrics_memory_locked_limit_bytes);
473 BMUSBCapture::stop_bm_thread();
475 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
477 unique_lock<mutex> lock(card_mutex);
478 cards[card_index].should_quit = true; // Unblock thread.
479 cards[card_index].new_frames_changed.notify_all();
481 cards[card_index].capture->stop_dequeue_thread();
482 if (cards[card_index].output) {
483 cards[card_index].output->end_output();
484 cards[card_index].output.reset();
488 video_encoder.reset(nullptr);
491 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
493 printf("Configuring card %d...\n", card_index);
495 CaptureCard *card = &cards[card_index];
496 if (card->capture != nullptr) {
497 card->capture->stop_dequeue_thread();
499 card->capture.reset(capture);
500 card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
501 card->type = card_type;
502 if (card->output.get() != output) {
503 card->output.reset(output);
506 PixelFormat pixel_format;
507 if (card_type == CardType::FFMPEG_INPUT) {
508 pixel_format = capture->get_current_pixel_format();
509 } else if (global_flags.ten_bit_input) {
510 pixel_format = PixelFormat_10BitYCbCr;
512 pixel_format = PixelFormat_8BitYCbCr;
515 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
516 if (card->frame_allocator == nullptr) {
517 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height)); // 8 MB.
519 card->capture->set_video_frame_allocator(card->frame_allocator.get());
520 if (card->surface == nullptr) {
521 card->surface = create_surface_with_same_format(mixer_surface);
523 while (!card->new_frames.empty()) card->new_frames.pop_front();
524 card->last_timecode = -1;
525 card->capture->set_pixel_format(pixel_format);
526 card->capture->configure_card();
528 // NOTE: start_bm_capture() happens in thread_func().
530 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
531 audio_mixer.reset_resampler(device);
532 audio_mixer.set_display_name(device, card->capture->get_description());
533 audio_mixer.trigger_state_changed_callback();
535 // Unregister old metrics, if any.
536 if (!card->labels.empty()) {
537 const vector<pair<string, string>> &labels = card->labels;
538 card->jitter_history.unregister_metrics(labels);
539 card->queue_length_policy.unregister_metrics(labels);
540 global_metrics.remove("input_received_frames", labels);
541 global_metrics.remove("input_dropped_frames_jitter", labels);
542 global_metrics.remove("input_dropped_frames_error", labels);
543 global_metrics.remove("input_dropped_frames_resets", labels);
544 global_metrics.remove("input_queue_length_frames", labels);
545 global_metrics.remove("input_queue_duped_frames", labels);
547 global_metrics.remove("input_has_signal_bool", labels);
548 global_metrics.remove("input_is_connected_bool", labels);
549 global_metrics.remove("input_interlaced_bool", labels);
550 global_metrics.remove("input_width_pixels", labels);
551 global_metrics.remove("input_height_pixels", labels);
552 global_metrics.remove("input_frame_rate_nom", labels);
553 global_metrics.remove("input_frame_rate_den", labels);
554 global_metrics.remove("input_sample_rate_hz", labels);
558 vector<pair<string, string>> labels;
560 snprintf(card_name, sizeof(card_name), "%d", card_index);
561 labels.emplace_back("card", card_name);
564 case CardType::LIVE_CARD:
565 labels.emplace_back("cardtype", "live");
567 case CardType::FAKE_CAPTURE:
568 labels.emplace_back("cardtype", "fake");
570 case CardType::FFMPEG_INPUT:
571 labels.emplace_back("cardtype", "ffmpeg");
576 card->jitter_history.register_metrics(labels);
577 card->queue_length_policy.register_metrics(labels);
578 global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
579 global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
580 global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
581 global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
582 global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
583 global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
585 global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
586 global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
587 global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
588 global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
589 global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
590 global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
591 global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
592 global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
593 card->labels = labels;
596 void Mixer::set_output_card_internal(int card_index)
598 // We don't really need to take card_mutex, since we're in the mixer
599 // thread and don't mess with any queues (which is the only thing that happens
600 // from other threads), but it's probably the safest in the long run.
601 unique_lock<mutex> lock(card_mutex);
602 if (output_card_index != -1) {
603 // Switch the old card from output to input.
604 CaptureCard *old_card = &cards[output_card_index];
605 old_card->output->end_output();
607 // Stop the fake card that we put into place.
608 // This needs to _not_ happen under the mutex, to avoid deadlock
609 // (delivering the last frame needs to take the mutex).
610 CaptureInterface *fake_capture = old_card->capture.get();
612 fake_capture->stop_dequeue_thread();
614 old_card->capture = move(old_card->parked_capture); // TODO: reset the metrics
615 old_card->is_fake_capture = false;
616 old_card->capture->start_bm_capture();
618 if (card_index != -1) {
619 CaptureCard *card = &cards[card_index];
620 CaptureInterface *capture = card->capture.get();
621 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
622 // several seconds to complete (blocking on DisableVideoInput);
623 // see if we can maybe do it asynchronously.
625 capture->stop_dequeue_thread();
627 card->parked_capture = move(card->capture);
628 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
629 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
630 card->queue_length_policy.reset(card_index);
631 card->capture->start_bm_capture();
632 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
633 card->output->start_output(desired_output_video_mode, pts_int);
635 output_card_index = card_index;
636 output_jitter_history.clear();
641 int unwrap_timecode(uint16_t current_wrapped, int last)
643 uint16_t last_wrapped = last & 0xffff;
644 if (current_wrapped > last_wrapped) {
645 return (last & ~0xffff) | current_wrapped;
647 return 0x10000 + ((last & ~0xffff) | current_wrapped);
653 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
654 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
655 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
657 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
658 CaptureCard *card = &cards[card_index];
660 ++card->metric_input_received_frames;
661 card->metric_input_has_signal_bool = video_format.has_signal;
662 card->metric_input_is_connected_bool = video_format.is_connected;
663 card->metric_input_interlaced_bool = video_format.interlaced;
664 card->metric_input_width_pixels = video_format.width;
665 card->metric_input_height_pixels = video_format.height;
666 card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
667 card->metric_input_frame_rate_den = video_format.frame_rate_den;
668 card->metric_input_sample_rate_hz = audio_format.sample_rate;
670 if (is_mode_scanning[card_index]) {
671 if (video_format.has_signal) {
672 // Found a stable signal, so stop scanning.
673 is_mode_scanning[card_index] = false;
675 static constexpr double switch_time_s = 0.1; // Should be enough time for the signal to stabilize.
676 steady_clock::time_point now = steady_clock::now();
677 double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
678 if (sec_since_last_switch > switch_time_s) {
679 // It isn't this mode; try the next one.
680 mode_scanlist_index[card_index]++;
681 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
682 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
683 last_mode_scan_change[card_index] = now;
688 int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
689 assert(frame_length > 0);
691 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;
692 if (num_samples > OUTPUT_FREQUENCY / 10) {
693 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",
694 card_index, int(audio_frame.len), int(audio_offset),
695 timecode, int(video_frame.len), int(video_offset), video_format.id);
696 if (video_frame.owner) {
697 video_frame.owner->release_frame(video_frame);
699 if (audio_frame.owner) {
700 audio_frame.owner->release_frame(audio_frame);
705 int dropped_frames = 0;
706 if (card->last_timecode != -1) {
707 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
710 // Number of samples per frame if we need to insert silence.
711 // (Could be nonintegral, but resampling will save us then.)
712 const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
714 if (dropped_frames > MAX_FPS * 2) {
715 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
716 card_index, card->last_timecode, timecode);
717 audio_mixer.reset_resampler(device);
719 ++card->metric_input_resets;
720 } else if (dropped_frames > 0) {
721 // Insert silence as needed.
722 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
723 card_index, dropped_frames, timecode);
724 card->metric_input_dropped_frames_error += dropped_frames;
728 success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
732 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
734 // Done with the audio, so release it.
735 if (audio_frame.owner) {
736 audio_frame.owner->release_frame(audio_frame);
739 card->last_timecode = timecode;
741 // Calculate jitter for this card here. We do it on arrival so that we
742 // make sure every frame counts, even the dropped ones -- and it will also
743 // make sure the jitter number is as recent as possible, should it change.
744 card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
746 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
748 size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
749 size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
750 if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
751 // The calculation above is wrong for planar Y'CbCr, so just override it.
752 assert(card->type == CardType::FFMPEG_INPUT);
753 assert(video_offset == 0);
754 expected_length = video_frame.len;
756 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
757 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
758 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
759 cbcr_offset = video_format.width * video_format.height;
762 // All the other Y'CbCr formats are 4:2:2.
763 cbcr_width = video_format.width / 2;
764 cbcr_height = video_format.height;
765 cbcr_offset = video_offset / 2;
766 y_offset = video_frame.size / 2 + video_offset / 2;
768 if (video_frame.len - video_offset == 0 ||
769 video_frame.len - video_offset != expected_length) {
770 if (video_frame.len != 0) {
771 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
772 card_index, video_frame.len - video_offset, expected_length);
774 if (video_frame.owner) {
775 video_frame.owner->release_frame(video_frame);
778 // Still send on the information that we _had_ a frame, even though it's corrupted,
779 // so that pts can go up accordingly.
781 unique_lock<mutex> lock(card_mutex);
782 CaptureCard::NewFrame new_frame;
783 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
784 new_frame.length = frame_length;
785 new_frame.interlaced = false;
786 new_frame.dropped_frames = dropped_frames;
787 new_frame.received_timestamp = video_frame.received_timestamp;
788 card->new_frames.push_back(move(new_frame));
789 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->new_frames_changed.notify_all();
919 void Mixer::bm_hotplug_add(libusb_device *dev)
921 lock_guard<mutex> lock(hotplug_mutex);
922 hotplugged_cards.push_back(dev);
925 void Mixer::bm_hotplug_remove(unsigned card_index)
927 cards[card_index].new_frames_changed.notify_all();
930 void Mixer::thread_func()
932 pthread_setname_np(pthread_self(), "Mixer_OpenGL");
934 eglBindAPI(EGL_OPENGL_API);
935 QOpenGLContext *context = create_context(mixer_surface);
936 if (!make_current(context, mixer_surface)) {
941 // Start the actual capture. (We don't want to do it before we're actually ready
942 // to process output frames.)
943 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
944 if (int(card_index) != output_card_index) {
945 cards[card_index].capture->start_bm_capture();
949 steady_clock::time_point start, now;
950 start = steady_clock::now();
952 int stats_dropped_frames = 0;
954 while (!should_quit) {
955 if (desired_output_card_index != output_card_index) {
956 set_output_card_internal(desired_output_card_index);
958 if (output_card_index != -1 &&
959 desired_output_video_mode != output_video_mode) {
960 DeckLinkOutput *output = cards[output_card_index].output.get();
961 output->end_output();
962 desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
963 output->start_output(desired_output_video_mode, pts_int);
966 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
967 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
969 bool master_card_is_output;
970 unsigned master_card_index;
971 if (output_card_index != -1) {
972 master_card_is_output = true;
973 master_card_index = output_card_index;
975 master_card_is_output = false;
976 master_card_index = theme->map_signal(master_clock_channel);
977 assert(master_card_index < num_cards);
980 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
981 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);
982 stats_dropped_frames += output_frame_info.dropped_frames;
984 handle_hotplugged_cards();
986 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
987 if (card_index == master_card_index || !has_new_frame[card_index]) {
990 if (new_frames[card_index].frame->len == 0) {
991 ++new_frames[card_index].dropped_frames;
993 if (new_frames[card_index].dropped_frames > 0) {
994 printf("Card %u dropped %d frames before this\n",
995 card_index, int(new_frames[card_index].dropped_frames));
999 // If the first card is reporting a corrupted or otherwise dropped frame,
1000 // just increase the pts (skipping over this frame) and don't try to compute anything new.
1001 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
1002 ++stats_dropped_frames;
1003 pts_int += new_frames[master_card_index].length;
1007 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1008 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1011 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1012 assert(new_frame->frame != nullptr);
1013 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1016 // The new texture might need uploading before use.
1017 if (new_frame->upload_func) {
1018 new_frame->upload_func();
1019 new_frame->upload_func = nullptr;
1023 int64_t frame_duration = output_frame_info.frame_duration;
1024 render_one_frame(frame_duration);
1026 pts_int += frame_duration;
1028 now = steady_clock::now();
1029 double elapsed = duration<double>(now - start).count();
1031 metric_frames_output_total = frame_num;
1032 metric_frames_output_dropped = stats_dropped_frames;
1034 if (frame_num % 100 == 0) {
1035 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
1036 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
1037 1e3 * elapsed / frame_num);
1038 // chain->print_phase_timing();
1040 // Check our memory usage, to see if we are close to our mlockall()
1041 // limit (if at all set).
1043 if (getrusage(RUSAGE_SELF, &used) == -1) {
1044 perror("getrusage(RUSAGE_SELF)");
1050 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
1051 perror("getrlimit(RLIMIT_MEMLOCK)");
1055 if (limit.rlim_cur == 0) {
1056 printf(", using %ld MB memory (locked)",
1057 long(used.ru_maxrss / 1024));
1059 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
1060 long(used.ru_maxrss / 1024),
1061 long(limit.rlim_cur / 1048576),
1062 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
1064 metrics_memory_locked_limit_bytes = limit.rlim_cur;
1066 printf(", using %ld MB memory (not locked)",
1067 long(used.ru_maxrss / 1024));
1068 metrics_memory_locked_limit_bytes = 0.0 / 0.0;
1073 metrics_memory_used_bytes = used.ru_maxrss * 1024;
1077 if (should_cut.exchange(false)) { // Test and clear.
1078 video_encoder->do_cut(frame_num);
1082 // Reset every 100 frames, so that local variations in frame times
1083 // (especially for the first few frames, when the shaders are
1084 // compiled etc.) don't make it hard to measure for the entire
1085 // remaining duration of the program.
1086 if (frame == 10000) {
1094 resource_pool->clean_context();
1097 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1099 if (output_card_index != -1) {
1100 // The output card (ie., cards[output_card_index].output) is the master clock,
1101 // so no input card (ie., cards[card_index].capture) is.
1104 return (card_index == master_card_index);
1107 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1109 // Count the number of frames in the queue, including any frames
1110 // we dropped. It's hard to know exactly how we should deal with
1111 // dropped (corrupted) input frames; they don't help our goal of
1112 // avoiding starvation, but they still add to the problem of latency.
1113 // Since dropped frames is going to mean a bump in the signal anyway,
1114 // we err on the side of having more stable latency instead.
1115 unsigned queue_length = 0;
1116 for (const CaptureCard::NewFrame &frame : card->new_frames) {
1117 queue_length += frame.dropped_frames + 1;
1120 // If needed, drop frames until the queue is below the safe limit.
1121 // We prefer to drop from the head, because all else being equal,
1122 // we'd like more recent frames (less latency).
1123 unsigned dropped_frames = 0;
1124 while (queue_length > safe_queue_length) {
1125 assert(!card->new_frames.empty());
1126 assert(queue_length > card->new_frames.front().dropped_frames);
1127 queue_length -= card->new_frames.front().dropped_frames;
1129 if (queue_length <= safe_queue_length) {
1130 // No need to drop anything.
1134 card->new_frames.pop_front();
1135 card->new_frames_changed.notify_all();
1140 card->metric_input_dropped_frames_jitter += dropped_frames;
1141 card->metric_input_queue_length_frames = queue_length;
1144 if (dropped_frames > 0) {
1145 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1146 card_index, dropped_frames);
1152 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])
1154 OutputFrameInfo output_frame_info;
1156 unique_lock<mutex> lock(card_mutex, defer_lock);
1157 if (master_card_is_output) {
1158 // Clocked to the output, so wait for it to be ready for the next frame.
1159 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);
1162 // Wait for the master card to have a new frame.
1163 // TODO: Add a timeout.
1164 output_frame_info.is_preroll = false;
1166 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(); });
1169 if (master_card_is_output) {
1170 handle_hotplugged_cards();
1171 } else if (cards[master_card_index].new_frames.empty()) {
1172 // We were woken up, but not due to a new frame. Deal with it
1173 // and then restart.
1174 assert(cards[master_card_index].capture->get_disconnected());
1175 handle_hotplugged_cards();
1180 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1181 CaptureCard *card = &cards[card_index];
1182 if (card->new_frames.empty()) { // Starvation.
1183 ++card->metric_input_duped_frames;
1185 new_frames[card_index] = move(card->new_frames.front());
1186 has_new_frame[card_index] = true;
1187 card->new_frames.pop_front();
1188 card->new_frames_changed.notify_all();
1192 if (!master_card_is_output) {
1193 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1194 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1195 output_frame_info.frame_duration = new_frames[master_card_index].length;
1198 if (!output_frame_info.is_preroll) {
1199 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1202 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1203 CaptureCard *card = &cards[card_index];
1204 if (has_new_frame[card_index] &&
1205 !input_card_is_master_clock(card_index, master_card_index) &&
1206 !output_frame_info.is_preroll) {
1207 card->queue_length_policy.update_policy(
1208 output_frame_info.frame_timestamp,
1209 card->jitter_history.get_expected_next_frame(),
1210 output_frame_info.frame_duration,
1211 card->jitter_history.estimate_max_jitter(),
1212 output_jitter_history.estimate_max_jitter());
1213 trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1214 card->queue_length_policy.get_safe_queue_length()));
1218 // This might get off by a fractional sample when changing master card
1219 // between ones with different frame rates, but that's fine.
1220 int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1221 output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1222 fractional_samples = num_samples_times_timebase % TIMEBASE;
1223 assert(output_frame_info.num_samples >= 0);
1225 return output_frame_info;
1228 void Mixer::handle_hotplugged_cards()
1230 // Check for cards that have been disconnected since last frame.
1231 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1232 CaptureCard *card = &cards[card_index];
1233 if (card->capture->get_disconnected()) {
1234 fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1235 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1236 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1237 card->queue_length_policy.reset(card_index);
1238 card->capture->start_bm_capture();
1242 // Check for cards that have been connected since last frame.
1243 vector<libusb_device *> hotplugged_cards_copy;
1245 lock_guard<mutex> lock(hotplug_mutex);
1246 swap(hotplugged_cards, hotplugged_cards_copy);
1248 for (libusb_device *new_dev : hotplugged_cards_copy) {
1249 // Look for a fake capture card where we can stick this in.
1250 int free_card_index = -1;
1251 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1252 if (cards[card_index].is_fake_capture) {
1253 free_card_index = card_index;
1258 if (free_card_index == -1) {
1259 fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1260 libusb_unref_device(new_dev);
1262 // BMUSBCapture takes ownership.
1263 fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1264 CaptureCard *card = &cards[free_card_index];
1265 BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1266 configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1267 card->queue_length_policy.reset(free_card_index);
1268 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1269 capture->start_bm_capture();
1275 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)
1277 // Resample the audio as needed, including from previously dropped frames.
1278 assert(num_cards > 0);
1279 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1280 const bool dropped_frame = (frame_num != dropped_frames);
1282 // Signal to the audio thread to process this frame.
1283 // Note that if the frame is a dropped frame, we signal that
1284 // we don't want to use this frame as base for adjusting
1285 // the resampler rate. The reason for this is that the timing
1286 // of these frames is often way too late; they typically don't
1287 // “arrive” before we synthesize them. Thus, we could end up
1288 // in a situation where we have inserted e.g. five audio frames
1289 // into the queue before we then start pulling five of them
1290 // back out. This makes ResamplingQueue overestimate the delay,
1291 // causing undue resampler changes. (We _do_ use the last,
1292 // non-dropped frame; perhaps we should just discard that as well,
1293 // since dropped frames are expected to be rare, and it might be
1294 // better to just wait until we have a slightly more normal situation).
1295 unique_lock<mutex> lock(audio_mutex);
1296 bool adjust_rate = !dropped_frame && !is_preroll;
1297 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1298 audio_task_queue_changed.notify_one();
1300 if (dropped_frame) {
1301 // For dropped frames, increase the pts. Note that if the format changed
1302 // in the meantime, we have no way of detecting that; we just have to
1303 // assume the frame length is always the same.
1304 pts_int += length_per_frame;
1309 void Mixer::render_one_frame(int64_t duration)
1311 // Determine the time code for this frame before we start rendering.
1312 string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1313 if (display_timecode_on_stdout) {
1314 printf("Timecode: '%s'\n", timecode_text.c_str());
1317 // Update Y'CbCr settings for all cards.
1319 unique_lock<mutex> lock(card_mutex);
1320 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1321 YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1322 input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1323 input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1324 input_state.full_range[card_index] = interpretation->full_range;
1328 // Get the main chain from the theme, and set its state immediately.
1329 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1330 EffectChain *chain = theme_main_chain.chain;
1331 theme_main_chain.setup_chain();
1332 //theme_main_chain.chain->enable_phase_timing(true);
1334 // The theme can't (or at least shouldn't!) call connect_signal() on
1335 // each FFmpeg input, so we'll do it here.
1336 for (const pair<LiveInputWrapper *, FFmpegCapture *> &conn : theme->get_signal_connections()) {
1337 conn.first->connect_signal_raw(conn.second->get_card_index());
1340 // If HDMI/SDI output is active and the user has requested auto mode,
1341 // its mode overrides the existing Y'CbCr setting for the chain.
1342 YCbCrLumaCoefficients ycbcr_output_coefficients;
1343 if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1344 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1346 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1349 // TODO: Reduce the duplication against theme.cpp.
1350 YCbCrFormat output_ycbcr_format;
1351 output_ycbcr_format.chroma_subsampling_x = 1;
1352 output_ycbcr_format.chroma_subsampling_y = 1;
1353 output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1354 output_ycbcr_format.full_range = false;
1355 output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1356 chain->change_ycbcr_output_format(output_ycbcr_format);
1358 // Render main chain. If we're using zerocopy Quick Sync encoding
1359 // (the default case), we take an extra copy of the created outputs,
1360 // so that we can display it back to the screen later (it's less memory
1361 // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1362 // Ideally, we'd like to avoid taking copies and just use the main textures
1363 // for display as well, but they're just views into VA-API memory and must be
1364 // unmapped during encoding, so we can't use them for display, unfortunately.
1365 GLuint y_tex, cbcr_full_tex, cbcr_tex;
1366 GLuint y_copy_tex, cbcr_copy_tex = 0;
1367 GLuint y_display_tex, cbcr_display_tex;
1368 GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1369 GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1370 const bool is_zerocopy = video_encoder->is_zerocopy();
1372 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1373 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1374 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1376 y_display_tex = y_copy_tex;
1377 cbcr_display_tex = cbcr_copy_tex;
1379 // y_tex and cbcr_tex will be given by VideoEncoder.
1381 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1382 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1383 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1385 y_display_tex = y_tex;
1386 cbcr_display_tex = cbcr_tex;
1389 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1390 bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1395 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1397 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1400 chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1402 if (display_timecode_in_stream) {
1403 // Render the timecode on top.
1404 timecode_renderer->render_timecode(fbo, timecode_text);
1407 resource_pool->release_fbo(fbo);
1410 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1412 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1414 if (output_card_index != -1) {
1415 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1417 resource_pool->release_2d_texture(cbcr_full_tex);
1419 // Set the right state for the Y' and CbCr textures we use for display.
1420 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1421 glBindTexture(GL_TEXTURE_2D, y_display_tex);
1422 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1423 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1424 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1426 glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1427 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1428 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1429 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1431 RefCountedGLsync fence = video_encoder->end_frame();
1433 // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1434 // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1435 DisplayFrame live_frame;
1436 live_frame.chain = display_chain.get();
1437 live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1438 display_input->set_texture_num(0, y_display_tex);
1439 display_input->set_texture_num(1, cbcr_display_tex);
1441 live_frame.ready_fence = fence;
1442 live_frame.input_frames = {};
1443 live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1444 output_channel[OUTPUT_LIVE].output_frame(live_frame);
1446 // Set up preview and any additional channels.
1447 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1448 DisplayFrame display_frame;
1449 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions
1450 display_frame.chain = chain.chain;
1451 display_frame.setup_chain = chain.setup_chain;
1452 display_frame.ready_fence = fence;
1453 display_frame.input_frames = chain.input_frames;
1454 display_frame.temp_textures = {};
1455 output_channel[i].output_frame(display_frame);
1459 void Mixer::audio_thread_func()
1461 pthread_setname_np(pthread_self(), "Mixer_Audio");
1463 while (!should_quit) {
1467 unique_lock<mutex> lock(audio_mutex);
1468 audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1472 task = audio_task_queue.front();
1473 audio_task_queue.pop();
1476 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1477 task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1478 vector<float> samples_out = audio_mixer.get_output(
1479 task.frame_timestamp,
1481 rate_adjustment_policy);
1483 // Send the samples to the sound card, then add them to the output.
1485 alsa->write(samples_out);
1487 if (output_card_index != -1) {
1488 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1489 cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1491 video_encoder->add_audio(task.pts_int, move(samples_out));
1495 void Mixer::release_display_frame(DisplayFrame *frame)
1497 for (GLuint texnum : frame->temp_textures) {
1498 resource_pool->release_2d_texture(texnum);
1500 frame->temp_textures.clear();
1501 frame->ready_fence.reset();
1502 frame->input_frames.clear();
1507 mixer_thread = thread(&Mixer::thread_func, this);
1508 audio_thread = thread(&Mixer::audio_thread_func, this);
1514 audio_task_queue_changed.notify_one();
1515 mixer_thread.join();
1516 audio_thread.join();
1519 void Mixer::transition_clicked(int transition_num)
1521 theme->transition_clicked(transition_num, pts());
1524 void Mixer::channel_clicked(int preview_num)
1526 theme->channel_clicked(preview_num);
1529 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1531 unique_lock<mutex> lock(card_mutex);
1532 return ycbcr_interpretation[card_index];
1535 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1537 unique_lock<mutex> lock(card_mutex);
1538 ycbcr_interpretation[card_index] = interpretation;
1541 void Mixer::start_mode_scanning(unsigned card_index)
1543 assert(card_index < num_cards);
1544 if (is_mode_scanning[card_index]) {
1547 is_mode_scanning[card_index] = true;
1548 mode_scanlist[card_index].clear();
1549 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1550 mode_scanlist[card_index].push_back(mode.first);
1552 assert(!mode_scanlist[card_index].empty());
1553 mode_scanlist_index[card_index] = 0;
1554 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1555 last_mode_scan_change[card_index] = steady_clock::now();
1558 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1560 assert(desired_output_card_index != -1);
1561 unique_lock<mutex> lock(card_mutex);
1562 return cards[desired_output_card_index].output->get_available_video_modes();
1565 Mixer::OutputChannel::~OutputChannel()
1567 if (has_current_frame) {
1568 parent->release_display_frame(¤t_frame);
1570 if (has_ready_frame) {
1571 parent->release_display_frame(&ready_frame);
1575 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1577 // Store this frame for display. Remove the ready frame if any
1578 // (it was seemingly never used).
1580 unique_lock<mutex> lock(frame_mutex);
1581 if (has_ready_frame) {
1582 parent->release_display_frame(&ready_frame);
1584 ready_frame = frame;
1585 has_ready_frame = true;
1587 // Call the callbacks under the mutex (they should be short),
1588 // so that we don't race against a callback removal.
1589 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1590 key_and_callback.second();
1594 // Reduce the number of callbacks by filtering duplicates. The reason
1595 // why we bother doing this is that Qt seemingly can get into a state
1596 // where its builds up an essentially unbounded queue of signals,
1597 // consuming more and more memory, and there's no good way of collapsing
1598 // user-defined signals or limiting the length of the queue.
1599 if (transition_names_updated_callback) {
1600 vector<string> transition_names = global_mixer->get_transition_names();
1601 bool changed = false;
1602 if (transition_names.size() != last_transition_names.size()) {
1605 for (unsigned i = 0; i < transition_names.size(); ++i) {
1606 if (transition_names[i] != last_transition_names[i]) {
1613 transition_names_updated_callback(transition_names);
1614 last_transition_names = transition_names;
1617 if (name_updated_callback) {
1618 string name = global_mixer->get_channel_name(channel);
1619 if (name != last_name) {
1620 name_updated_callback(name);
1624 if (color_updated_callback) {
1625 string color = global_mixer->get_channel_color(channel);
1626 if (color != last_color) {
1627 color_updated_callback(color);
1633 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1635 unique_lock<mutex> lock(frame_mutex);
1636 if (!has_current_frame && !has_ready_frame) {
1640 if (has_current_frame && has_ready_frame) {
1641 // We have a new ready frame. Toss the current one.
1642 parent->release_display_frame(¤t_frame);
1643 has_current_frame = false;
1645 if (has_ready_frame) {
1646 assert(!has_current_frame);
1647 current_frame = ready_frame;
1648 ready_frame.ready_fence.reset(); // Drop the refcount.
1649 ready_frame.input_frames.clear(); // Drop the refcounts.
1650 has_current_frame = true;
1651 has_ready_frame = false;
1654 *frame = current_frame;
1658 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1660 unique_lock<mutex> lock(frame_mutex);
1661 new_frame_ready_callbacks[key] = callback;
1664 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1666 unique_lock<mutex> lock(frame_mutex);
1667 new_frame_ready_callbacks.erase(key);
1670 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1672 transition_names_updated_callback = callback;
1675 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1677 name_updated_callback = callback;
1680 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1682 color_updated_callback = callback;
1685 mutex RefCountedGLsync::fence_lock;