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>
19 #include <condition_variable>
30 #include "DeckLinkAPI.h"
32 #include "alsa_output.h"
33 #include "basic_stats.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;
67 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
70 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
71 input_state->buffered_frames[card_index][frame_num] =
72 input_state->buffered_frames[card_index][frame_num - 1];
74 input_state->buffered_frames[card_index][0] = { frame, field_num };
76 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
77 input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
82 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
85 switch (userdata->pixel_format) {
86 case PixelFormat_10BitYCbCr:
87 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
89 case PixelFormat_8BitYCbCr:
90 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
92 case PixelFormat_8BitBGRA:
93 first = userdata->tex_rgba[field] == 0;
95 case PixelFormat_8BitYCbCrPlanar:
96 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
103 width != userdata->last_width[field] ||
104 height != userdata->last_height[field] ||
105 cbcr_width != userdata->last_cbcr_width[field] ||
106 cbcr_height != userdata->last_cbcr_height[field]) {
107 // We changed resolution since last use of this texture, so we need to create
108 // a new object. Note that this each card has its own PBOFrameAllocator,
109 // we don't need to worry about these flip-flopping between resolutions.
110 switch (userdata->pixel_format) {
111 case PixelFormat_10BitYCbCr:
112 glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
114 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
117 case PixelFormat_8BitYCbCr: {
118 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
120 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
122 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
124 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
128 case PixelFormat_8BitYCbCrPlanar: {
129 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
131 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
133 glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
135 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
137 glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
139 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
143 case PixelFormat_8BitBGRA:
144 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
146 if (global_flags.can_disable_srgb_decoder) { // See the comments in tweaked_inputs.h.
147 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
149 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
154 userdata->last_width[field] = width;
155 userdata->last_height[field] = height;
156 userdata->last_cbcr_width[field] = cbcr_width;
157 userdata->last_cbcr_height[field] = cbcr_height;
159 if (global_flags.ten_bit_input &&
160 (first || v210_width != userdata->last_v210_width[field])) {
161 // Same as above; we need to recreate the texture.
162 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
164 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
166 userdata->last_v210_width[field] = v210_width;
170 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
172 if (interlaced_stride) {
175 if (global_flags.flush_pbos) {
176 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
180 glBindTexture(GL_TEXTURE_2D, tex);
182 if (interlaced_stride) {
183 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
186 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
190 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
192 glBindTexture(GL_TEXTURE_2D, 0);
194 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
200 void JitterHistory::register_metrics(const vector<pair<string, string>> &labels)
202 global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames);
203 global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE);
206 void JitterHistory::unregister_metrics(const vector<pair<string, string>> &labels)
208 global_metrics.remove("input_underestimated_jitter_frames", labels);
209 global_metrics.remove("input_estimated_max_jitter_seconds", labels);
212 void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames)
214 if (expected_timestamp > steady_clock::time_point::min()) {
215 expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE);
216 double jitter_seconds = fabs(duration<double>(expected_timestamp - now).count());
217 history.push_back(orders.insert(jitter_seconds));
218 if (jitter_seconds > estimate_max_jitter()) {
219 ++metric_input_underestimated_jitter_frames;
222 metric_input_estimated_max_jitter_seconds = estimate_max_jitter();
224 if (history.size() > history_length) {
225 orders.erase(history.front());
228 assert(history.size() <= history_length);
230 expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE);
233 double JitterHistory::estimate_max_jitter() const
235 if (orders.empty()) {
238 size_t elem_idx = lrint((orders.size() - 1) * percentile);
239 if (percentile <= 0.5) {
240 return *next(orders.begin(), elem_idx) * multiplier;
242 return *prev(orders.end(), elem_idx + 1) * multiplier;
246 void QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
248 global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
251 void QueueLengthPolicy::unregister_metrics(const vector<pair<string, string>> &labels)
253 global_metrics.remove("input_queue_safe_length_frames", labels);
256 void QueueLengthPolicy::update_policy(steady_clock::time_point now,
257 steady_clock::time_point expected_next_frame,
258 int64_t input_frame_duration,
259 int64_t master_frame_duration,
260 double max_input_card_jitter_seconds,
261 double max_master_card_jitter_seconds)
263 double input_frame_duration_seconds = input_frame_duration / double(TIMEBASE);
264 double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
266 // Figure out when we can expect the next frame for this card, assuming
267 // worst-case jitter (ie., the frame is maximally late).
268 double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
270 // How many times are the master card expected to tick in that time?
271 // We assume the master clock has worst-case jitter but not any rate
272 // discrepancy, ie., it ticks as early as possible every time, but not
274 double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
276 // As a special case, if the master card ticks faster than the input card,
277 // we expect the queue to drain by itself even without dropping. But if
278 // the difference is small (e.g. 60 Hz master and 59.94 input), it would
279 // go slowly enough that the effect wouldn't really be appreciable.
280 // We account for this by looking at the situation five frames ahead,
281 // assuming everything else is the same.
282 double frames_allowed;
283 if (master_frame_duration < input_frame_duration) {
284 frames_allowed = frames_needed + 5 * (input_frame_duration_seconds - master_frame_duration_seconds) / master_frame_duration_seconds;
286 frames_allowed = frames_needed;
289 safe_queue_length = max<int>(floor(frames_allowed), 0);
290 metric_input_queue_safe_length_frames = safe_queue_length;
293 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
295 num_cards(num_cards),
296 mixer_surface(create_surface(format)),
297 h264_encoder_surface(create_surface(format)),
298 decklink_output_surface(create_surface(format)),
299 audio_mixer(num_cards)
301 memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
302 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
305 // This nearly always should be true.
306 global_flags.can_disable_srgb_decoder =
307 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
308 epoxy_has_gl_extension("GL_ARB_sampler_objects");
310 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
311 // will be halved when sampling them, and we need to compensate here.
312 movit_texel_subpixel_precision /= 2.0;
314 resource_pool.reset(new ResourcePool);
315 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
316 output_channel[i].parent = this;
317 output_channel[i].channel = i;
320 ImageFormat inout_format;
321 inout_format.color_space = COLORSPACE_sRGB;
322 inout_format.gamma_curve = GAMMA_sRGB;
324 // Matches the 4:2:0 format created by the main chain.
325 YCbCrFormat ycbcr_format;
326 ycbcr_format.chroma_subsampling_x = 2;
327 ycbcr_format.chroma_subsampling_y = 2;
328 if (global_flags.ycbcr_rec709_coefficients) {
329 ycbcr_format.luma_coefficients = YCBCR_REC_709;
331 ycbcr_format.luma_coefficients = YCBCR_REC_601;
333 ycbcr_format.full_range = false;
334 ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
335 ycbcr_format.cb_x_position = 0.0f;
336 ycbcr_format.cr_x_position = 0.0f;
337 ycbcr_format.cb_y_position = 0.5f;
338 ycbcr_format.cr_y_position = 0.5f;
340 // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
341 display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
343 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
344 display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
345 display_chain->add_input(display_input);
346 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
347 display_chain->set_dither_bits(0); // Don't bother.
348 display_chain->finalize();
350 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));
352 // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
353 theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
355 // Start listening for clients only once VideoEncoder has written its header, if any.
358 // First try initializing the then PCI devices, then USB, then
359 // fill up with fake cards until we have the desired number of cards.
360 unsigned num_pci_devices = 0;
361 unsigned card_index = 0;
364 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
365 if (decklink_iterator != nullptr) {
366 for ( ; card_index < num_cards; ++card_index) {
368 if (decklink_iterator->Next(&decklink) != S_OK) {
372 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
373 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
374 if (!output->set_device(decklink)) {
378 configure_card(card_index, capture, CardType::LIVE_CARD, output);
381 decklink_iterator->Release();
382 fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
384 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
388 unsigned num_usb_devices = BMUSBCapture::num_cards();
389 for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
390 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
391 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
392 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
394 fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
396 unsigned num_fake_cards = 0;
397 for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
398 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
399 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
402 if (num_fake_cards > 0) {
403 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
406 // Initialize all video inputs the theme asked for. Note that these are
407 // all put _after_ the regular cards, which stop at <num_cards> - 1.
408 std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
409 for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
410 if (card_index >= MAX_VIDEO_CARDS) {
411 fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
414 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
415 video_inputs[video_card_index]->set_card_index(card_index);
417 num_video_inputs = video_inputs.size();
419 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
420 BMUSBCapture::start_bm_thread();
422 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
423 cards[card_index].queue_length_policy.reset(card_index);
426 chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
428 if (global_flags.ten_bit_input) {
429 if (!v210Converter::has_hardware_support()) {
430 fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
431 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
434 v210_converter.reset(new v210Converter());
436 // These are all the widths listed in the Blackmagic SDK documentation
437 // (section 2.7.3, “Display Modes”).
438 v210_converter->precompile_shader(720);
439 v210_converter->precompile_shader(1280);
440 v210_converter->precompile_shader(1920);
441 v210_converter->precompile_shader(2048);
442 v210_converter->precompile_shader(3840);
443 v210_converter->precompile_shader(4096);
445 if (global_flags.ten_bit_output) {
446 if (!v210Converter::has_hardware_support()) {
447 fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
448 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
453 timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
454 display_timecode_in_stream = global_flags.display_timecode_in_stream;
455 display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
457 if (global_flags.enable_alsa_output) {
458 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
460 if (global_flags.output_card != -1) {
461 desired_output_card_index = global_flags.output_card;
462 set_output_card_internal(global_flags.output_card);
465 output_jitter_history.register_metrics({{ "card", "output" }});
470 BMUSBCapture::stop_bm_thread();
472 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
474 unique_lock<mutex> lock(card_mutex);
475 cards[card_index].should_quit = true; // Unblock thread.
476 cards[card_index].new_frames_changed.notify_all();
478 cards[card_index].capture->stop_dequeue_thread();
479 if (cards[card_index].output) {
480 cards[card_index].output->end_output();
481 cards[card_index].output.reset();
485 video_encoder.reset(nullptr);
488 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
490 printf("Configuring card %d...\n", card_index);
492 CaptureCard *card = &cards[card_index];
493 if (card->capture != nullptr) {
494 card->capture->stop_dequeue_thread();
496 card->capture.reset(capture);
497 card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
498 card->type = card_type;
499 if (card->output.get() != output) {
500 card->output.reset(output);
503 PixelFormat pixel_format;
504 if (card_type == CardType::FFMPEG_INPUT) {
505 pixel_format = capture->get_current_pixel_format();
506 } else if (global_flags.ten_bit_input) {
507 pixel_format = PixelFormat_10BitYCbCr;
509 pixel_format = PixelFormat_8BitYCbCr;
512 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
513 if (card->frame_allocator == nullptr) {
514 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height)); // 8 MB.
516 card->capture->set_video_frame_allocator(card->frame_allocator.get());
517 if (card->surface == nullptr) {
518 card->surface = create_surface_with_same_format(mixer_surface);
520 while (!card->new_frames.empty()) card->new_frames.pop_front();
521 card->last_timecode = -1;
522 card->capture->set_pixel_format(pixel_format);
523 card->capture->configure_card();
525 // NOTE: start_bm_capture() happens in thread_func().
527 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
528 audio_mixer.reset_resampler(device);
529 audio_mixer.set_display_name(device, card->capture->get_description());
530 audio_mixer.trigger_state_changed_callback();
532 // Unregister old metrics, if any.
533 if (!card->labels.empty()) {
534 const vector<pair<string, string>> &labels = card->labels;
535 card->jitter_history.unregister_metrics(labels);
536 card->queue_length_policy.unregister_metrics(labels);
537 global_metrics.remove("input_received_frames", labels);
538 global_metrics.remove("input_dropped_frames_jitter", labels);
539 global_metrics.remove("input_dropped_frames_error", labels);
540 global_metrics.remove("input_dropped_frames_resets", labels);
541 global_metrics.remove("input_queue_length_frames", labels);
542 global_metrics.remove("input_queue_duped_frames", labels);
544 global_metrics.remove("input_has_signal_bool", labels);
545 global_metrics.remove("input_is_connected_bool", labels);
546 global_metrics.remove("input_interlaced_bool", labels);
547 global_metrics.remove("input_width_pixels", labels);
548 global_metrics.remove("input_height_pixels", labels);
549 global_metrics.remove("input_frame_rate_nom", labels);
550 global_metrics.remove("input_frame_rate_den", labels);
551 global_metrics.remove("input_sample_rate_hz", labels);
555 vector<pair<string, string>> labels;
557 snprintf(card_name, sizeof(card_name), "%d", card_index);
558 labels.emplace_back("card", card_name);
561 case CardType::LIVE_CARD:
562 labels.emplace_back("cardtype", "live");
564 case CardType::FAKE_CAPTURE:
565 labels.emplace_back("cardtype", "fake");
567 case CardType::FFMPEG_INPUT:
568 labels.emplace_back("cardtype", "ffmpeg");
573 card->jitter_history.register_metrics(labels);
574 card->queue_length_policy.register_metrics(labels);
575 global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
576 global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
577 global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
578 global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
579 global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
580 global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
582 global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
583 global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
584 global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
585 global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
586 global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
587 global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
588 global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
589 global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
590 card->labels = labels;
593 void Mixer::set_output_card_internal(int card_index)
595 // We don't really need to take card_mutex, since we're in the mixer
596 // thread and don't mess with any queues (which is the only thing that happens
597 // from other threads), but it's probably the safest in the long run.
598 unique_lock<mutex> lock(card_mutex);
599 if (output_card_index != -1) {
600 // Switch the old card from output to input.
601 CaptureCard *old_card = &cards[output_card_index];
602 old_card->output->end_output();
604 // Stop the fake card that we put into place.
605 // This needs to _not_ happen under the mutex, to avoid deadlock
606 // (delivering the last frame needs to take the mutex).
607 CaptureInterface *fake_capture = old_card->capture.get();
609 fake_capture->stop_dequeue_thread();
611 old_card->capture = move(old_card->parked_capture); // TODO: reset the metrics
612 old_card->is_fake_capture = false;
613 old_card->capture->start_bm_capture();
615 if (card_index != -1) {
616 CaptureCard *card = &cards[card_index];
617 CaptureInterface *capture = card->capture.get();
618 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
619 // several seconds to complete (blocking on DisableVideoInput);
620 // see if we can maybe do it asynchronously.
622 capture->stop_dequeue_thread();
624 card->parked_capture = move(card->capture);
625 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
626 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
627 card->queue_length_policy.reset(card_index);
628 card->capture->start_bm_capture();
629 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
630 card->output->start_output(desired_output_video_mode, pts_int);
632 output_card_index = card_index;
633 output_jitter_history.clear();
638 int unwrap_timecode(uint16_t current_wrapped, int last)
640 uint16_t last_wrapped = last & 0xffff;
641 if (current_wrapped > last_wrapped) {
642 return (last & ~0xffff) | current_wrapped;
644 return 0x10000 + ((last & ~0xffff) | current_wrapped);
650 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
651 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
652 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
654 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
655 CaptureCard *card = &cards[card_index];
657 ++card->metric_input_received_frames;
658 card->metric_input_has_signal_bool = video_format.has_signal;
659 card->metric_input_is_connected_bool = video_format.is_connected;
660 card->metric_input_interlaced_bool = video_format.interlaced;
661 card->metric_input_width_pixels = video_format.width;
662 card->metric_input_height_pixels = video_format.height;
663 card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
664 card->metric_input_frame_rate_den = video_format.frame_rate_den;
665 card->metric_input_sample_rate_hz = audio_format.sample_rate;
667 if (is_mode_scanning[card_index]) {
668 if (video_format.has_signal) {
669 // Found a stable signal, so stop scanning.
670 is_mode_scanning[card_index] = false;
672 static constexpr double switch_time_s = 0.1; // Should be enough time for the signal to stabilize.
673 steady_clock::time_point now = steady_clock::now();
674 double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
675 if (sec_since_last_switch > switch_time_s) {
676 // It isn't this mode; try the next one.
677 mode_scanlist_index[card_index]++;
678 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
679 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
680 last_mode_scan_change[card_index] = now;
685 int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
686 assert(frame_length > 0);
688 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;
689 if (num_samples > OUTPUT_FREQUENCY / 10) {
690 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",
691 card_index, int(audio_frame.len), int(audio_offset),
692 timecode, int(video_frame.len), int(video_offset), video_format.id);
693 if (video_frame.owner) {
694 video_frame.owner->release_frame(video_frame);
696 if (audio_frame.owner) {
697 audio_frame.owner->release_frame(audio_frame);
702 int dropped_frames = 0;
703 if (card->last_timecode != -1) {
704 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
707 // Number of samples per frame if we need to insert silence.
708 // (Could be nonintegral, but resampling will save us then.)
709 const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
711 if (dropped_frames > MAX_FPS * 2) {
712 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
713 card_index, card->last_timecode, timecode);
714 audio_mixer.reset_resampler(device);
716 ++card->metric_input_resets;
717 } else if (dropped_frames > 0) {
718 // Insert silence as needed.
719 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
720 card_index, dropped_frames, timecode);
721 card->metric_input_dropped_frames_error += dropped_frames;
725 success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
729 if (num_samples > 0) {
730 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
733 // Done with the audio, so release it.
734 if (audio_frame.owner) {
735 audio_frame.owner->release_frame(audio_frame);
738 card->last_timecode = timecode;
740 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
742 size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
743 size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
744 if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
745 // The calculation above is wrong for planar Y'CbCr, so just override it.
746 assert(card->type == CardType::FFMPEG_INPUT);
747 assert(video_offset == 0);
748 expected_length = video_frame.len;
750 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
751 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
752 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
753 cbcr_offset = video_format.width * video_format.height;
756 // All the other Y'CbCr formats are 4:2:2.
757 cbcr_width = video_format.width / 2;
758 cbcr_height = video_format.height;
759 cbcr_offset = video_offset / 2;
760 y_offset = video_frame.size / 2 + video_offset / 2;
762 if (video_frame.len - video_offset == 0 ||
763 video_frame.len - video_offset != expected_length) {
764 if (video_frame.len != 0) {
765 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
766 card_index, video_frame.len - video_offset, expected_length);
768 if (video_frame.owner) {
769 video_frame.owner->release_frame(video_frame);
772 // Still send on the information that we _had_ a frame, even though it's corrupted,
773 // so that pts can go up accordingly.
775 unique_lock<mutex> lock(card_mutex);
776 CaptureCard::NewFrame new_frame;
777 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
778 new_frame.length = frame_length;
779 new_frame.interlaced = false;
780 new_frame.dropped_frames = dropped_frames;
781 new_frame.received_timestamp = video_frame.received_timestamp;
782 card->new_frames.push_back(move(new_frame));
783 card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
785 card->new_frames_changed.notify_all();
789 unsigned num_fields = video_format.interlaced ? 2 : 1;
790 steady_clock::time_point frame_upload_start;
791 bool interlaced_stride = false;
792 if (video_format.interlaced) {
793 // Send the two fields along as separate frames; the other side will need to add
794 // a deinterlacer to actually get this right.
795 assert(video_format.height % 2 == 0);
796 video_format.height /= 2;
798 assert(frame_length % 2 == 0);
801 if (video_format.second_field_start == 1) {
802 interlaced_stride = true;
804 frame_upload_start = steady_clock::now();
806 userdata->last_interlaced = video_format.interlaced;
807 userdata->last_has_signal = video_format.has_signal;
808 userdata->last_is_connected = video_format.is_connected;
809 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
810 userdata->last_frame_rate_den = video_format.frame_rate_den;
811 RefCountedFrame frame(video_frame);
813 // Upload the textures.
814 for (unsigned field = 0; field < num_fields; ++field) {
815 // Put the actual texture upload in a lambda that is executed in the main thread.
816 // It is entirely possible to do this in the same thread (and it might even be
817 // faster, depending on the GPU and driver), but it appears to be trickling
818 // driver bugs very easily.
820 // Note that this means we must hold on to the actual frame data in <userdata>
821 // until the upload command is run, but we hold on to <frame> much longer than that
822 // (in fact, all the way until we no longer use the texture in rendering).
823 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
824 unsigned field_start_line;
826 field_start_line = video_format.second_field_start;
828 field_start_line = video_format.extra_lines_top;
831 // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
832 size_t v210_width = video_format.stride / sizeof(uint32_t);
833 ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
835 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
838 switch (userdata->pixel_format) {
839 case PixelFormat_10BitYCbCr: {
840 size_t field_start = video_offset + video_format.stride * field_start_line;
841 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);
842 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
845 case PixelFormat_8BitYCbCr: {
846 size_t field_y_start = y_offset + video_format.width * field_start_line;
847 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
849 // Make up our own strides, since we are interleaving.
850 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);
851 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);
854 case PixelFormat_8BitYCbCrPlanar: {
855 assert(field_start_line == 0); // We don't really support interlaced here.
856 size_t field_y_start = y_offset;
857 size_t field_cb_start = cbcr_offset;
858 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
860 // Make up our own strides, since we are interleaving.
861 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);
862 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
863 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
866 case PixelFormat_8BitBGRA: {
867 size_t field_start = video_offset + video_format.stride * field_start_line;
868 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);
869 // These could be asked to deliver mipmaps at any time.
870 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
872 glGenerateMipmap(GL_TEXTURE_2D);
874 glBindTexture(GL_TEXTURE_2D, 0);
882 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
887 // Don't upload the second field as fast as we can; wait until
888 // the field time has approximately passed. (Otherwise, we could
889 // get timing jitter against the other sources, and possibly also
890 // against the video display, although the latter is not as critical.)
891 // This requires our system clock to be reasonably close to the
892 // video clock, but that's not an unreasonable assumption.
893 steady_clock::time_point second_field_start = frame_upload_start +
894 nanoseconds(frame_length * 1000000000 / TIMEBASE);
895 this_thread::sleep_until(second_field_start);
899 unique_lock<mutex> lock(card_mutex);
900 CaptureCard::NewFrame new_frame;
901 new_frame.frame = frame;
902 new_frame.length = frame_length;
903 new_frame.field = field;
904 new_frame.interlaced = video_format.interlaced;
905 new_frame.upload_func = upload_func;
906 new_frame.dropped_frames = dropped_frames;
907 new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp.
908 card->new_frames.push_back(move(new_frame));
909 card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
911 card->new_frames_changed.notify_all();
915 void Mixer::bm_hotplug_add(libusb_device *dev)
917 lock_guard<mutex> lock(hotplug_mutex);
918 hotplugged_cards.push_back(dev);
921 void Mixer::bm_hotplug_remove(unsigned card_index)
923 cards[card_index].new_frames_changed.notify_all();
926 void Mixer::thread_func()
928 pthread_setname_np(pthread_self(), "Mixer_OpenGL");
930 eglBindAPI(EGL_OPENGL_API);
931 QOpenGLContext *context = create_context(mixer_surface);
932 if (!make_current(context, mixer_surface)) {
937 // Start the actual capture. (We don't want to do it before we're actually ready
938 // to process output frames.)
939 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
940 if (int(card_index) != output_card_index) {
941 cards[card_index].capture->start_bm_capture();
945 BasicStats basic_stats(/*verbose=*/true);
946 int stats_dropped_frames = 0;
948 while (!should_quit) {
949 if (desired_output_card_index != output_card_index) {
950 set_output_card_internal(desired_output_card_index);
952 if (output_card_index != -1 &&
953 desired_output_video_mode != output_video_mode) {
954 DeckLinkOutput *output = cards[output_card_index].output.get();
955 output->end_output();
956 desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
957 output->start_output(desired_output_video_mode, pts_int);
960 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
961 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
963 bool master_card_is_output;
964 unsigned master_card_index;
965 if (output_card_index != -1) {
966 master_card_is_output = true;
967 master_card_index = output_card_index;
969 master_card_is_output = false;
970 master_card_index = theme->map_signal(master_clock_channel);
971 assert(master_card_index < num_cards);
974 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
975 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);
976 stats_dropped_frames += output_frame_info.dropped_frames;
978 handle_hotplugged_cards();
980 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
981 if (card_index == master_card_index || !has_new_frame[card_index]) {
984 if (new_frames[card_index].frame->len == 0) {
985 ++new_frames[card_index].dropped_frames;
987 if (new_frames[card_index].dropped_frames > 0) {
988 printf("Card %u dropped %d frames before this\n",
989 card_index, int(new_frames[card_index].dropped_frames));
993 // If the first card is reporting a corrupted or otherwise dropped frame,
994 // just increase the pts (skipping over this frame) and don't try to compute anything new.
995 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
996 ++stats_dropped_frames;
997 pts_int += new_frames[master_card_index].length;
1001 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1002 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1005 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1006 assert(new_frame->frame != nullptr);
1007 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1010 // The new texture might need uploading before use.
1011 if (new_frame->upload_func) {
1012 new_frame->upload_func();
1013 new_frame->upload_func = nullptr;
1017 int64_t frame_duration = output_frame_info.frame_duration;
1018 render_one_frame(frame_duration);
1020 pts_int += frame_duration;
1022 basic_stats.update(frame_num, stats_dropped_frames);
1023 // if (frame_num % 100 == 0) chain->print_phase_timing();
1025 if (should_cut.exchange(false)) { // Test and clear.
1026 video_encoder->do_cut(frame_num);
1030 // Reset every 100 frames, so that local variations in frame times
1031 // (especially for the first few frames, when the shaders are
1032 // compiled etc.) don't make it hard to measure for the entire
1033 // remaining duration of the program.
1034 if (frame == 10000) {
1042 resource_pool->clean_context();
1045 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1047 if (output_card_index != -1) {
1048 // The output card (ie., cards[output_card_index].output) is the master clock,
1049 // so no input card (ie., cards[card_index].capture) is.
1052 return (card_index == master_card_index);
1055 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1057 // Count the number of frames in the queue, including any frames
1058 // we dropped. It's hard to know exactly how we should deal with
1059 // dropped (corrupted) input frames; they don't help our goal of
1060 // avoiding starvation, but they still add to the problem of latency.
1061 // Since dropped frames is going to mean a bump in the signal anyway,
1062 // we err on the side of having more stable latency instead.
1063 unsigned queue_length = 0;
1064 for (const CaptureCard::NewFrame &frame : card->new_frames) {
1065 queue_length += frame.dropped_frames + 1;
1068 // If needed, drop frames until the queue is below the safe limit.
1069 // We prefer to drop from the head, because all else being equal,
1070 // we'd like more recent frames (less latency).
1071 unsigned dropped_frames = 0;
1072 while (queue_length > safe_queue_length) {
1073 assert(!card->new_frames.empty());
1074 assert(queue_length > card->new_frames.front().dropped_frames);
1075 queue_length -= card->new_frames.front().dropped_frames;
1077 if (queue_length <= safe_queue_length) {
1078 // No need to drop anything.
1082 card->new_frames.pop_front();
1083 card->new_frames_changed.notify_all();
1088 card->metric_input_dropped_frames_jitter += dropped_frames;
1089 card->metric_input_queue_length_frames = queue_length;
1092 if (dropped_frames > 0) {
1093 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1094 card_index, dropped_frames);
1100 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])
1102 OutputFrameInfo output_frame_info;
1104 unique_lock<mutex> lock(card_mutex, defer_lock);
1105 if (master_card_is_output) {
1106 // Clocked to the output, so wait for it to be ready for the next frame.
1107 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);
1110 // Wait for the master card to have a new frame.
1111 // TODO: Add a timeout.
1112 output_frame_info.is_preroll = false;
1114 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(); });
1117 if (master_card_is_output) {
1118 handle_hotplugged_cards();
1119 } else if (cards[master_card_index].new_frames.empty()) {
1120 // We were woken up, but not due to a new frame. Deal with it
1121 // and then restart.
1122 assert(cards[master_card_index].capture->get_disconnected());
1123 handle_hotplugged_cards();
1128 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1129 CaptureCard *card = &cards[card_index];
1130 if (card->new_frames.empty()) { // Starvation.
1131 ++card->metric_input_duped_frames;
1133 new_frames[card_index] = move(card->new_frames.front());
1134 has_new_frame[card_index] = true;
1135 card->new_frames.pop_front();
1136 card->new_frames_changed.notify_all();
1140 if (!master_card_is_output) {
1141 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1142 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1143 output_frame_info.frame_duration = new_frames[master_card_index].length;
1146 if (!output_frame_info.is_preroll) {
1147 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1150 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1151 CaptureCard *card = &cards[card_index];
1152 if (has_new_frame[card_index] &&
1153 !input_card_is_master_clock(card_index, master_card_index) &&
1154 !output_frame_info.is_preroll) {
1155 card->queue_length_policy.update_policy(
1156 output_frame_info.frame_timestamp,
1157 card->jitter_history.get_expected_next_frame(),
1158 new_frames[master_card_index].length,
1159 output_frame_info.frame_duration,
1160 card->jitter_history.estimate_max_jitter(),
1161 output_jitter_history.estimate_max_jitter());
1162 trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1163 card->queue_length_policy.get_safe_queue_length()));
1167 // This might get off by a fractional sample when changing master card
1168 // between ones with different frame rates, but that's fine.
1169 int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1170 output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1171 fractional_samples = num_samples_times_timebase % TIMEBASE;
1172 assert(output_frame_info.num_samples >= 0);
1174 return output_frame_info;
1177 void Mixer::handle_hotplugged_cards()
1179 // Check for cards that have been disconnected since last frame.
1180 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1181 CaptureCard *card = &cards[card_index];
1182 if (card->capture->get_disconnected()) {
1183 fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1184 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1185 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1186 card->queue_length_policy.reset(card_index);
1187 card->capture->start_bm_capture();
1191 // Check for cards that have been connected since last frame.
1192 vector<libusb_device *> hotplugged_cards_copy;
1194 lock_guard<mutex> lock(hotplug_mutex);
1195 swap(hotplugged_cards, hotplugged_cards_copy);
1197 for (libusb_device *new_dev : hotplugged_cards_copy) {
1198 // Look for a fake capture card where we can stick this in.
1199 int free_card_index = -1;
1200 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1201 if (cards[card_index].is_fake_capture) {
1202 free_card_index = card_index;
1207 if (free_card_index == -1) {
1208 fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1209 libusb_unref_device(new_dev);
1211 // BMUSBCapture takes ownership.
1212 fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1213 CaptureCard *card = &cards[free_card_index];
1214 BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1215 configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1216 card->queue_length_policy.reset(free_card_index);
1217 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1218 capture->start_bm_capture();
1224 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)
1226 // Resample the audio as needed, including from previously dropped frames.
1227 assert(num_cards > 0);
1228 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1229 const bool dropped_frame = (frame_num != dropped_frames);
1231 // Signal to the audio thread to process this frame.
1232 // Note that if the frame is a dropped frame, we signal that
1233 // we don't want to use this frame as base for adjusting
1234 // the resampler rate. The reason for this is that the timing
1235 // of these frames is often way too late; they typically don't
1236 // “arrive” before we synthesize them. Thus, we could end up
1237 // in a situation where we have inserted e.g. five audio frames
1238 // into the queue before we then start pulling five of them
1239 // back out. This makes ResamplingQueue overestimate the delay,
1240 // causing undue resampler changes. (We _do_ use the last,
1241 // non-dropped frame; perhaps we should just discard that as well,
1242 // since dropped frames are expected to be rare, and it might be
1243 // better to just wait until we have a slightly more normal situation).
1244 unique_lock<mutex> lock(audio_mutex);
1245 bool adjust_rate = !dropped_frame && !is_preroll;
1246 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1247 audio_task_queue_changed.notify_one();
1249 if (dropped_frame) {
1250 // For dropped frames, increase the pts. Note that if the format changed
1251 // in the meantime, we have no way of detecting that; we just have to
1252 // assume the frame length is always the same.
1253 pts_int += length_per_frame;
1258 void Mixer::render_one_frame(int64_t duration)
1260 // Determine the time code for this frame before we start rendering.
1261 string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1262 if (display_timecode_on_stdout) {
1263 printf("Timecode: '%s'\n", timecode_text.c_str());
1266 // Update Y'CbCr settings for all cards.
1268 unique_lock<mutex> lock(card_mutex);
1269 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1270 YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1271 input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1272 input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1273 input_state.full_range[card_index] = interpretation->full_range;
1277 // Get the main chain from the theme, and set its state immediately.
1278 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1279 EffectChain *chain = theme_main_chain.chain;
1280 theme_main_chain.setup_chain();
1281 //theme_main_chain.chain->enable_phase_timing(true);
1283 // The theme can't (or at least shouldn't!) call connect_signal() on
1284 // each FFmpeg input, so we'll do it here.
1285 for (const pair<LiveInputWrapper *, FFmpegCapture *> &conn : theme->get_signal_connections()) {
1286 conn.first->connect_signal_raw(conn.second->get_card_index(), input_state);
1289 // If HDMI/SDI output is active and the user has requested auto mode,
1290 // its mode overrides the existing Y'CbCr setting for the chain.
1291 YCbCrLumaCoefficients ycbcr_output_coefficients;
1292 if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1293 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1295 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1298 // TODO: Reduce the duplication against theme.cpp.
1299 YCbCrFormat output_ycbcr_format;
1300 output_ycbcr_format.chroma_subsampling_x = 1;
1301 output_ycbcr_format.chroma_subsampling_y = 1;
1302 output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1303 output_ycbcr_format.full_range = false;
1304 output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1305 chain->change_ycbcr_output_format(output_ycbcr_format);
1307 // Render main chain. If we're using zerocopy Quick Sync encoding
1308 // (the default case), we take an extra copy of the created outputs,
1309 // so that we can display it back to the screen later (it's less memory
1310 // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1311 // Ideally, we'd like to avoid taking copies and just use the main textures
1312 // for display as well, but they're just views into VA-API memory and must be
1313 // unmapped during encoding, so we can't use them for display, unfortunately.
1314 GLuint y_tex, cbcr_full_tex, cbcr_tex;
1315 GLuint y_copy_tex, cbcr_copy_tex = 0;
1316 GLuint y_display_tex, cbcr_display_tex;
1317 GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1318 GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1319 const bool is_zerocopy = video_encoder->is_zerocopy();
1321 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1322 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1323 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1325 y_display_tex = y_copy_tex;
1326 cbcr_display_tex = cbcr_copy_tex;
1328 // y_tex and cbcr_tex will be given by VideoEncoder.
1330 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1331 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1332 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1334 y_display_tex = y_tex;
1335 cbcr_display_tex = cbcr_tex;
1338 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1339 bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1344 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1346 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1349 chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1351 if (display_timecode_in_stream) {
1352 // Render the timecode on top.
1353 timecode_renderer->render_timecode(fbo, timecode_text);
1356 resource_pool->release_fbo(fbo);
1359 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1361 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1363 if (output_card_index != -1) {
1364 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1366 resource_pool->release_2d_texture(cbcr_full_tex);
1368 // Set the right state for the Y' and CbCr textures we use for display.
1369 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1370 glBindTexture(GL_TEXTURE_2D, y_display_tex);
1371 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1372 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1373 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1375 glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1376 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1377 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1378 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1380 RefCountedGLsync fence = video_encoder->end_frame();
1382 // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1383 // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1384 DisplayFrame live_frame;
1385 live_frame.chain = display_chain.get();
1386 live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1387 display_input->set_texture_num(0, y_display_tex);
1388 display_input->set_texture_num(1, cbcr_display_tex);
1390 live_frame.ready_fence = fence;
1391 live_frame.input_frames = {};
1392 live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1393 output_channel[OUTPUT_LIVE].output_frame(live_frame);
1395 // Set up preview and any additional channels.
1396 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1397 DisplayFrame display_frame;
1398 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions
1399 display_frame.chain = chain.chain;
1400 display_frame.setup_chain = chain.setup_chain;
1401 display_frame.ready_fence = fence;
1402 display_frame.input_frames = chain.input_frames;
1403 display_frame.temp_textures = {};
1404 output_channel[i].output_frame(display_frame);
1408 void Mixer::audio_thread_func()
1410 pthread_setname_np(pthread_self(), "Mixer_Audio");
1412 while (!should_quit) {
1416 unique_lock<mutex> lock(audio_mutex);
1417 audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1421 task = audio_task_queue.front();
1422 audio_task_queue.pop();
1425 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1426 task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1427 vector<float> samples_out = audio_mixer.get_output(
1428 task.frame_timestamp,
1430 rate_adjustment_policy);
1432 // Send the samples to the sound card, then add them to the output.
1434 alsa->write(samples_out);
1436 if (output_card_index != -1) {
1437 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1438 cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1440 video_encoder->add_audio(task.pts_int, move(samples_out));
1444 void Mixer::release_display_frame(DisplayFrame *frame)
1446 for (GLuint texnum : frame->temp_textures) {
1447 resource_pool->release_2d_texture(texnum);
1449 frame->temp_textures.clear();
1450 frame->ready_fence.reset();
1451 frame->input_frames.clear();
1456 mixer_thread = thread(&Mixer::thread_func, this);
1457 audio_thread = thread(&Mixer::audio_thread_func, this);
1463 audio_task_queue_changed.notify_one();
1464 mixer_thread.join();
1465 audio_thread.join();
1468 void Mixer::transition_clicked(int transition_num)
1470 theme->transition_clicked(transition_num, pts());
1473 void Mixer::channel_clicked(int preview_num)
1475 theme->channel_clicked(preview_num);
1478 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1480 unique_lock<mutex> lock(card_mutex);
1481 return ycbcr_interpretation[card_index];
1484 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1486 unique_lock<mutex> lock(card_mutex);
1487 ycbcr_interpretation[card_index] = interpretation;
1490 void Mixer::start_mode_scanning(unsigned card_index)
1492 assert(card_index < num_cards);
1493 if (is_mode_scanning[card_index]) {
1496 is_mode_scanning[card_index] = true;
1497 mode_scanlist[card_index].clear();
1498 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1499 mode_scanlist[card_index].push_back(mode.first);
1501 assert(!mode_scanlist[card_index].empty());
1502 mode_scanlist_index[card_index] = 0;
1503 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1504 last_mode_scan_change[card_index] = steady_clock::now();
1507 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1509 assert(desired_output_card_index != -1);
1510 unique_lock<mutex> lock(card_mutex);
1511 return cards[desired_output_card_index].output->get_available_video_modes();
1514 Mixer::OutputChannel::~OutputChannel()
1516 if (has_current_frame) {
1517 parent->release_display_frame(¤t_frame);
1519 if (has_ready_frame) {
1520 parent->release_display_frame(&ready_frame);
1524 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1526 // Store this frame for display. Remove the ready frame if any
1527 // (it was seemingly never used).
1529 unique_lock<mutex> lock(frame_mutex);
1530 if (has_ready_frame) {
1531 parent->release_display_frame(&ready_frame);
1533 ready_frame = frame;
1534 has_ready_frame = true;
1536 // Call the callbacks under the mutex (they should be short),
1537 // so that we don't race against a callback removal.
1538 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1539 key_and_callback.second();
1543 // Reduce the number of callbacks by filtering duplicates. The reason
1544 // why we bother doing this is that Qt seemingly can get into a state
1545 // where its builds up an essentially unbounded queue of signals,
1546 // consuming more and more memory, and there's no good way of collapsing
1547 // user-defined signals or limiting the length of the queue.
1548 if (transition_names_updated_callback) {
1549 vector<string> transition_names = global_mixer->get_transition_names();
1550 bool changed = false;
1551 if (transition_names.size() != last_transition_names.size()) {
1554 for (unsigned i = 0; i < transition_names.size(); ++i) {
1555 if (transition_names[i] != last_transition_names[i]) {
1562 transition_names_updated_callback(transition_names);
1563 last_transition_names = transition_names;
1566 if (name_updated_callback) {
1567 string name = global_mixer->get_channel_name(channel);
1568 if (name != last_name) {
1569 name_updated_callback(name);
1573 if (color_updated_callback) {
1574 string color = global_mixer->get_channel_color(channel);
1575 if (color != last_color) {
1576 color_updated_callback(color);
1582 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1584 unique_lock<mutex> lock(frame_mutex);
1585 if (!has_current_frame && !has_ready_frame) {
1589 if (has_current_frame && has_ready_frame) {
1590 // We have a new ready frame. Toss the current one.
1591 parent->release_display_frame(¤t_frame);
1592 has_current_frame = false;
1594 if (has_ready_frame) {
1595 assert(!has_current_frame);
1596 current_frame = ready_frame;
1597 ready_frame.ready_fence.reset(); // Drop the refcount.
1598 ready_frame.input_frames.clear(); // Drop the refcounts.
1599 has_current_frame = true;
1600 has_ready_frame = false;
1603 *frame = current_frame;
1607 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1609 unique_lock<mutex> lock(frame_mutex);
1610 new_frame_ready_callbacks[key] = callback;
1613 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1615 unique_lock<mutex> lock(frame_mutex);
1616 new_frame_ready_callbacks.erase(key);
1619 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1621 transition_names_updated_callback = callback;
1624 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1626 name_updated_callback = callback;
1629 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1631 color_updated_callback = callback;
1634 mutex RefCountedGLsync::fence_lock;