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 QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
203 global_metrics.add("input_queue_length_frames", labels, &metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
204 global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
205 global_metrics.add("input_queue_duped_frames", labels, &metric_input_duped_frames);
208 void QueueLengthPolicy::update_policy(unsigned queue_length)
210 metric_input_queue_length_frames = queue_length;
212 if (queue_length == 0) { // Starvation.
213 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
215 fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
216 card_index, safe_queue_length);
218 frames_with_at_least_one = 0;
219 been_at_safe_point_since_last_starvation = false;
220 ++metric_input_duped_frames;
221 metric_input_queue_safe_length_frames = safe_queue_length;
224 if (queue_length >= safe_queue_length) {
225 been_at_safe_point_since_last_starvation = true;
227 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
229 metric_input_queue_safe_length_frames = safe_queue_length;
230 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
231 card_index, safe_queue_length);
232 frames_with_at_least_one = 0;
236 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
238 num_cards(num_cards),
239 mixer_surface(create_surface(format)),
240 h264_encoder_surface(create_surface(format)),
241 decklink_output_surface(create_surface(format)),
242 ycbcr_interpretation(global_flags.ycbcr_interpretation),
243 audio_mixer(num_cards)
245 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
248 // This nearly always should be true.
249 global_flags.can_disable_srgb_decoder =
250 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
251 epoxy_has_gl_extension("GL_ARB_sampler_objects");
253 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
254 // will be halved when sampling them, and we need to compensate here.
255 movit_texel_subpixel_precision /= 2.0;
257 resource_pool.reset(new ResourcePool);
258 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
259 output_channel[i].parent = this;
260 output_channel[i].channel = i;
263 ImageFormat inout_format;
264 inout_format.color_space = COLORSPACE_sRGB;
265 inout_format.gamma_curve = GAMMA_sRGB;
267 // Matches the 4:2:0 format created by the main chain.
268 YCbCrFormat ycbcr_format;
269 ycbcr_format.chroma_subsampling_x = 2;
270 ycbcr_format.chroma_subsampling_y = 2;
271 if (global_flags.ycbcr_rec709_coefficients) {
272 ycbcr_format.luma_coefficients = YCBCR_REC_709;
274 ycbcr_format.luma_coefficients = YCBCR_REC_601;
276 ycbcr_format.full_range = false;
277 ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
278 ycbcr_format.cb_x_position = 0.0f;
279 ycbcr_format.cr_x_position = 0.0f;
280 ycbcr_format.cb_y_position = 0.5f;
281 ycbcr_format.cr_y_position = 0.5f;
283 // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
284 display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
286 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
287 display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
288 display_chain->add_input(display_input);
289 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
290 display_chain->set_dither_bits(0); // Don't bother.
291 display_chain->finalize();
293 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));
295 // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
296 theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
298 // Start listening for clients only once VideoEncoder has written its header, if any.
301 // First try initializing the then PCI devices, then USB, then
302 // fill up with fake cards until we have the desired number of cards.
303 unsigned num_pci_devices = 0;
304 unsigned card_index = 0;
307 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
308 if (decklink_iterator != nullptr) {
309 for ( ; card_index < num_cards; ++card_index) {
311 if (decklink_iterator->Next(&decklink) != S_OK) {
315 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
316 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
317 output->set_device(decklink);
318 configure_card(card_index, capture, CardType::LIVE_CARD, output);
321 decklink_iterator->Release();
322 fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
324 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
328 unsigned num_usb_devices = BMUSBCapture::num_cards();
329 for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
330 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
331 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
332 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
334 fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
336 unsigned num_fake_cards = 0;
337 for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
338 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
339 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
342 if (num_fake_cards > 0) {
343 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
346 // Initialize all video inputs the theme asked for. Note that these are
347 // all put _after_ the regular cards, which stop at <num_cards> - 1.
348 std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
349 for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
350 if (card_index >= MAX_VIDEO_CARDS) {
351 fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
354 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
355 video_inputs[video_card_index]->set_card_index(card_index);
357 num_video_inputs = video_inputs.size();
359 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
360 BMUSBCapture::start_bm_thread();
362 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
363 cards[card_index].queue_length_policy.reset(card_index);
366 chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
368 if (global_flags.ten_bit_input) {
369 if (!v210Converter::has_hardware_support()) {
370 fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
371 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
374 v210_converter.reset(new v210Converter());
376 // These are all the widths listed in the Blackmagic SDK documentation
377 // (section 2.7.3, “Display Modes”).
378 v210_converter->precompile_shader(720);
379 v210_converter->precompile_shader(1280);
380 v210_converter->precompile_shader(1920);
381 v210_converter->precompile_shader(2048);
382 v210_converter->precompile_shader(3840);
383 v210_converter->precompile_shader(4096);
385 if (global_flags.ten_bit_output) {
386 if (!v210Converter::has_hardware_support()) {
387 fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
388 fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
393 timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
394 display_timecode_in_stream = global_flags.display_timecode_in_stream;
395 display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
397 if (global_flags.enable_alsa_output) {
398 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
400 if (global_flags.output_card != -1) {
401 desired_output_card_index = global_flags.output_card;
402 set_output_card_internal(global_flags.output_card);
405 global_metrics.add("frames_output_total", &metric_frames_output_total);
406 global_metrics.add("frames_output_dropped", &metric_frames_output_dropped);
407 global_metrics.add("uptime_seconds", &metric_uptime_seconds);
408 global_metrics.add("memory_used_bytes", &metrics_memory_used_bytes);
409 global_metrics.add("metrics_memory_locked_limit_bytes", &metrics_memory_locked_limit_bytes);
414 BMUSBCapture::stop_bm_thread();
416 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
418 unique_lock<mutex> lock(card_mutex);
419 cards[card_index].should_quit = true; // Unblock thread.
420 cards[card_index].new_frames_changed.notify_all();
422 cards[card_index].capture->stop_dequeue_thread();
423 if (cards[card_index].output) {
424 cards[card_index].output->end_output();
425 cards[card_index].output.reset();
429 video_encoder.reset(nullptr);
432 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
434 printf("Configuring card %d...\n", card_index);
436 CaptureCard *card = &cards[card_index];
437 if (card->capture != nullptr) {
438 card->capture->stop_dequeue_thread();
440 card->capture.reset(capture);
441 card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
442 card->type = card_type;
443 if (card->output.get() != output) {
444 card->output.reset(output);
447 PixelFormat pixel_format;
448 if (card_type == CardType::FFMPEG_INPUT) {
449 pixel_format = capture->get_current_pixel_format();
450 } else if (global_flags.ten_bit_input) {
451 pixel_format = PixelFormat_10BitYCbCr;
453 pixel_format = PixelFormat_8BitYCbCr;
456 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
457 if (card->frame_allocator == nullptr) {
458 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height)); // 8 MB.
460 card->capture->set_video_frame_allocator(card->frame_allocator.get());
461 if (card->surface == nullptr) {
462 card->surface = create_surface_with_same_format(mixer_surface);
464 while (!card->new_frames.empty()) card->new_frames.pop_front();
465 card->last_timecode = -1;
466 card->capture->set_pixel_format(pixel_format);
467 card->capture->configure_card();
469 // NOTE: start_bm_capture() happens in thread_func().
471 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
472 audio_mixer.reset_resampler(device);
473 audio_mixer.set_display_name(device, card->capture->get_description());
474 audio_mixer.trigger_state_changed_callback();
477 vector<pair<string, string>> labels;
479 snprintf(card_name, sizeof(card_name), "%d", card_index);
480 labels.emplace_back("card", card_name);
483 case CardType::LIVE_CARD:
484 labels.emplace_back("cardtype", "live");
486 case CardType::FAKE_CAPTURE:
487 labels.emplace_back("cardtype", "fake");
489 case CardType::FFMPEG_INPUT:
490 labels.emplace_back("cardtype", "ffmpeg");
495 card->queue_length_policy.register_metrics(labels);
496 global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
497 global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
498 global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
500 global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
501 global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
502 global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
503 global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
504 global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
505 global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
506 global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
507 global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
510 void Mixer::set_output_card_internal(int card_index)
512 // We don't really need to take card_mutex, since we're in the mixer
513 // thread and don't mess with any queues (which is the only thing that happens
514 // from other threads), but it's probably the safest in the long run.
515 unique_lock<mutex> lock(card_mutex);
516 if (output_card_index != -1) {
517 // Switch the old card from output to input.
518 CaptureCard *old_card = &cards[output_card_index];
519 old_card->output->end_output();
521 // Stop the fake card that we put into place.
522 // This needs to _not_ happen under the mutex, to avoid deadlock
523 // (delivering the last frame needs to take the mutex).
524 CaptureInterface *fake_capture = old_card->capture.get();
526 fake_capture->stop_dequeue_thread();
528 old_card->capture = move(old_card->parked_capture);
529 old_card->is_fake_capture = false;
530 old_card->capture->start_bm_capture();
532 if (card_index != -1) {
533 CaptureCard *card = &cards[card_index];
534 CaptureInterface *capture = card->capture.get();
535 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
536 // several seconds to complete (blocking on DisableVideoInput);
537 // see if we can maybe do it asynchronously.
539 capture->stop_dequeue_thread();
541 card->parked_capture = move(card->capture);
542 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
543 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
544 card->queue_length_policy.reset(card_index);
545 card->capture->start_bm_capture();
546 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
547 card->output->start_output(desired_output_video_mode, pts_int);
549 output_card_index = card_index;
554 int unwrap_timecode(uint16_t current_wrapped, int last)
556 uint16_t last_wrapped = last & 0xffff;
557 if (current_wrapped > last_wrapped) {
558 return (last & ~0xffff) | current_wrapped;
560 return 0x10000 + ((last & ~0xffff) | current_wrapped);
566 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
567 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
568 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
570 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
571 CaptureCard *card = &cards[card_index];
573 card->metric_input_has_signal_bool = video_format.has_signal;
574 card->metric_input_is_connected_bool = video_format.is_connected;
575 card->metric_input_interlaced_bool = video_format.interlaced;
576 card->metric_input_width_pixels = video_format.width;
577 card->metric_input_height_pixels = video_format.height;
578 card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
579 card->metric_input_frame_rate_den = video_format.frame_rate_den;
580 card->metric_input_sample_rate_hz = audio_format.sample_rate;
582 if (is_mode_scanning[card_index]) {
583 if (video_format.has_signal) {
584 // Found a stable signal, so stop scanning.
585 is_mode_scanning[card_index] = false;
587 static constexpr double switch_time_s = 0.1; // Should be enough time for the signal to stabilize.
588 steady_clock::time_point now = steady_clock::now();
589 double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
590 if (sec_since_last_switch > switch_time_s) {
591 // It isn't this mode; try the next one.
592 mode_scanlist_index[card_index]++;
593 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
594 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
595 last_mode_scan_change[card_index] = now;
600 int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
601 assert(frame_length > 0);
603 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;
604 if (num_samples > OUTPUT_FREQUENCY / 10) {
605 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",
606 card_index, int(audio_frame.len), int(audio_offset),
607 timecode, int(video_frame.len), int(video_offset), video_format.id);
608 if (video_frame.owner) {
609 video_frame.owner->release_frame(video_frame);
611 if (audio_frame.owner) {
612 audio_frame.owner->release_frame(audio_frame);
617 int dropped_frames = 0;
618 if (card->last_timecode != -1) {
619 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
622 // Number of samples per frame if we need to insert silence.
623 // (Could be nonintegral, but resampling will save us then.)
624 const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
626 if (dropped_frames > MAX_FPS * 2) {
627 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
628 card_index, card->last_timecode, timecode);
629 audio_mixer.reset_resampler(device);
631 ++card->metric_input_resets;
632 } else if (dropped_frames > 0) {
633 // Insert silence as needed.
634 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
635 card_index, dropped_frames, timecode);
636 card->metric_input_dropped_frames_error += dropped_frames;
640 success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
644 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
646 // Done with the audio, so release it.
647 if (audio_frame.owner) {
648 audio_frame.owner->release_frame(audio_frame);
651 card->last_timecode = timecode;
653 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
655 size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
656 size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
657 if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
658 // The calculation above is wrong for planar Y'CbCr, so just override it.
659 assert(card->type == CardType::FFMPEG_INPUT);
660 assert(video_offset == 0);
661 expected_length = video_frame.len;
663 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
664 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
665 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
666 cbcr_offset = video_format.width * video_format.height;
669 // All the other Y'CbCr formats are 4:2:2.
670 cbcr_width = video_format.width / 2;
671 cbcr_height = video_format.height;
672 cbcr_offset = video_offset / 2;
673 y_offset = video_frame.size / 2 + video_offset / 2;
675 if (video_frame.len - video_offset == 0 ||
676 video_frame.len - video_offset != expected_length) {
677 if (video_frame.len != 0) {
678 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
679 card_index, video_frame.len - video_offset, expected_length);
681 if (video_frame.owner) {
682 video_frame.owner->release_frame(video_frame);
685 // Still send on the information that we _had_ a frame, even though it's corrupted,
686 // so that pts can go up accordingly.
688 unique_lock<mutex> lock(card_mutex);
689 CaptureCard::NewFrame new_frame;
690 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
691 new_frame.length = frame_length;
692 new_frame.interlaced = false;
693 new_frame.dropped_frames = dropped_frames;
694 new_frame.received_timestamp = video_frame.received_timestamp;
695 card->new_frames.push_back(move(new_frame));
696 card->new_frames_changed.notify_all();
701 unsigned num_fields = video_format.interlaced ? 2 : 1;
702 steady_clock::time_point frame_upload_start;
703 bool interlaced_stride = false;
704 if (video_format.interlaced) {
705 // Send the two fields along as separate frames; the other side will need to add
706 // a deinterlacer to actually get this right.
707 assert(video_format.height % 2 == 0);
708 video_format.height /= 2;
710 assert(frame_length % 2 == 0);
713 if (video_format.second_field_start == 1) {
714 interlaced_stride = true;
716 frame_upload_start = steady_clock::now();
718 userdata->last_interlaced = video_format.interlaced;
719 userdata->last_has_signal = video_format.has_signal;
720 userdata->last_is_connected = video_format.is_connected;
721 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
722 userdata->last_frame_rate_den = video_format.frame_rate_den;
723 RefCountedFrame frame(video_frame);
725 // Upload the textures.
726 for (unsigned field = 0; field < num_fields; ++field) {
727 // Put the actual texture upload in a lambda that is executed in the main thread.
728 // It is entirely possible to do this in the same thread (and it might even be
729 // faster, depending on the GPU and driver), but it appears to be trickling
730 // driver bugs very easily.
732 // Note that this means we must hold on to the actual frame data in <userdata>
733 // until the upload command is run, but we hold on to <frame> much longer than that
734 // (in fact, all the way until we no longer use the texture in rendering).
735 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
736 unsigned field_start_line;
738 field_start_line = video_format.second_field_start;
740 field_start_line = video_format.extra_lines_top;
743 // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
744 size_t v210_width = video_format.stride / sizeof(uint32_t);
745 ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
747 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
750 switch (userdata->pixel_format) {
751 case PixelFormat_10BitYCbCr: {
752 size_t field_start = video_offset + video_format.stride * field_start_line;
753 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);
754 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
757 case PixelFormat_8BitYCbCr: {
758 size_t field_y_start = y_offset + video_format.width * field_start_line;
759 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
761 // Make up our own strides, since we are interleaving.
762 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);
763 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);
766 case PixelFormat_8BitYCbCrPlanar: {
767 assert(field_start_line == 0); // We don't really support interlaced here.
768 size_t field_y_start = y_offset;
769 size_t field_cb_start = cbcr_offset;
770 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
772 // Make up our own strides, since we are interleaving.
773 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);
774 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
775 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
778 case PixelFormat_8BitBGRA: {
779 size_t field_start = video_offset + video_format.stride * field_start_line;
780 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);
781 // These could be asked to deliver mipmaps at any time.
782 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
784 glGenerateMipmap(GL_TEXTURE_2D);
786 glBindTexture(GL_TEXTURE_2D, 0);
794 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
799 // Don't upload the second field as fast as we can; wait until
800 // the field time has approximately passed. (Otherwise, we could
801 // get timing jitter against the other sources, and possibly also
802 // against the video display, although the latter is not as critical.)
803 // This requires our system clock to be reasonably close to the
804 // video clock, but that's not an unreasonable assumption.
805 steady_clock::time_point second_field_start = frame_upload_start +
806 nanoseconds(frame_length * 1000000000 / TIMEBASE);
807 this_thread::sleep_until(second_field_start);
811 unique_lock<mutex> lock(card_mutex);
812 CaptureCard::NewFrame new_frame;
813 new_frame.frame = frame;
814 new_frame.length = frame_length;
815 new_frame.field = field;
816 new_frame.interlaced = video_format.interlaced;
817 new_frame.upload_func = upload_func;
818 new_frame.dropped_frames = dropped_frames;
819 new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp.
820 card->new_frames.push_back(move(new_frame));
821 card->new_frames_changed.notify_all();
826 void Mixer::bm_hotplug_add(libusb_device *dev)
828 lock_guard<mutex> lock(hotplug_mutex);
829 hotplugged_cards.push_back(dev);
832 void Mixer::bm_hotplug_remove(unsigned card_index)
834 cards[card_index].new_frames_changed.notify_all();
837 void Mixer::thread_func()
839 pthread_setname_np(pthread_self(), "Mixer_OpenGL");
841 eglBindAPI(EGL_OPENGL_API);
842 QOpenGLContext *context = create_context(mixer_surface);
843 if (!make_current(context, mixer_surface)) {
848 // Start the actual capture. (We don't want to do it before we're actually ready
849 // to process output frames.)
850 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
851 if (int(card_index) != output_card_index) {
852 cards[card_index].capture->start_bm_capture();
856 steady_clock::time_point start, now;
857 start = steady_clock::now();
859 int stats_dropped_frames = 0;
861 while (!should_quit) {
862 if (desired_output_card_index != output_card_index) {
863 set_output_card_internal(desired_output_card_index);
865 if (output_card_index != -1 &&
866 desired_output_video_mode != output_video_mode) {
867 DeckLinkOutput *output = cards[output_card_index].output.get();
868 output->end_output();
869 desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
870 output->start_output(desired_output_video_mode, pts_int);
873 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
874 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
876 bool master_card_is_output;
877 unsigned master_card_index;
878 if (output_card_index != -1) {
879 master_card_is_output = true;
880 master_card_index = output_card_index;
882 master_card_is_output = false;
883 master_card_index = theme->map_signal(master_clock_channel);
884 assert(master_card_index < num_cards);
887 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
888 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);
889 stats_dropped_frames += output_frame_info.dropped_frames;
891 handle_hotplugged_cards();
893 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
894 if (card_index == master_card_index || !has_new_frame[card_index]) {
897 if (new_frames[card_index].frame->len == 0) {
898 ++new_frames[card_index].dropped_frames;
900 if (new_frames[card_index].dropped_frames > 0) {
901 printf("Card %u dropped %d frames before this\n",
902 card_index, int(new_frames[card_index].dropped_frames));
906 // If the first card is reporting a corrupted or otherwise dropped frame,
907 // just increase the pts (skipping over this frame) and don't try to compute anything new.
908 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
909 ++stats_dropped_frames;
910 pts_int += new_frames[master_card_index].length;
914 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
915 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
918 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
919 assert(new_frame->frame != nullptr);
920 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
923 // The new texture might need uploading before use.
924 if (new_frame->upload_func) {
925 new_frame->upload_func();
926 new_frame->upload_func = nullptr;
930 int64_t frame_duration = output_frame_info.frame_duration;
931 render_one_frame(frame_duration);
933 pts_int += frame_duration;
935 now = steady_clock::now();
936 double elapsed = duration<double>(now - start).count();
938 metric_frames_output_total = frame_num;
939 metric_frames_output_dropped = stats_dropped_frames;
940 metric_uptime_seconds = elapsed;
942 if (frame_num % 100 == 0) {
943 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
944 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
945 1e3 * elapsed / frame_num);
946 // chain->print_phase_timing();
948 // Check our memory usage, to see if we are close to our mlockall()
949 // limit (if at all set).
951 if (getrusage(RUSAGE_SELF, &used) == -1) {
952 perror("getrusage(RUSAGE_SELF)");
958 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
959 perror("getrlimit(RLIMIT_MEMLOCK)");
963 if (limit.rlim_cur == 0) {
964 printf(", using %ld MB memory (locked)",
965 long(used.ru_maxrss / 1024));
967 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
968 long(used.ru_maxrss / 1024),
969 long(limit.rlim_cur / 1048576),
970 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
972 metrics_memory_locked_limit_bytes = limit.rlim_cur;
974 printf(", using %ld MB memory (not locked)",
975 long(used.ru_maxrss / 1024));
976 metrics_memory_locked_limit_bytes = 0.0 / 0.0;
981 metrics_memory_used_bytes = used.ru_maxrss;
985 if (should_cut.exchange(false)) { // Test and clear.
986 video_encoder->do_cut(frame_num);
990 // Reset every 100 frames, so that local variations in frame times
991 // (especially for the first few frames, when the shaders are
992 // compiled etc.) don't make it hard to measure for the entire
993 // remaining duration of the program.
994 if (frame == 10000) {
1002 resource_pool->clean_context();
1005 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1007 if (output_card_index != -1) {
1008 // The output card (ie., cards[output_card_index].output) is the master clock,
1009 // so no input card (ie., cards[card_index].capture) is.
1012 return (card_index == master_card_index);
1015 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
1017 // Count the number of frames in the queue, including any frames
1018 // we dropped. It's hard to know exactly how we should deal with
1019 // dropped (corrupted) input frames; they don't help our goal of
1020 // avoiding starvation, but they still add to the problem of latency.
1021 // Since dropped frames is going to mean a bump in the signal anyway,
1022 // we err on the side of having more stable latency instead.
1023 unsigned queue_length = 0;
1024 for (const CaptureCard::NewFrame &frame : card->new_frames) {
1025 queue_length += frame.dropped_frames + 1;
1027 card->queue_length_policy.update_policy(queue_length);
1029 // If needed, drop frames until the queue is below the safe limit.
1030 // We prefer to drop from the head, because all else being equal,
1031 // we'd like more recent frames (less latency).
1032 unsigned dropped_frames = 0;
1033 while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
1034 assert(!card->new_frames.empty());
1035 assert(queue_length > card->new_frames.front().dropped_frames);
1036 queue_length -= card->new_frames.front().dropped_frames;
1038 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
1039 // No need to drop anything.
1043 card->new_frames.pop_front();
1044 card->new_frames_changed.notify_all();
1049 card->metric_input_dropped_frames_jitter += dropped_frames;
1052 if (dropped_frames > 0) {
1053 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1054 card_index, dropped_frames);
1060 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])
1062 OutputFrameInfo output_frame_info;
1064 unique_lock<mutex> lock(card_mutex, defer_lock);
1065 if (master_card_is_output) {
1066 // Clocked to the output, so wait for it to be ready for the next frame.
1067 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);
1070 // Wait for the master card to have a new frame.
1071 // TODO: Add a timeout.
1072 output_frame_info.is_preroll = false;
1074 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(); });
1077 if (master_card_is_output) {
1078 handle_hotplugged_cards();
1079 } else if (cards[master_card_index].new_frames.empty()) {
1080 // We were woken up, but not due to a new frame. Deal with it
1081 // and then restart.
1082 assert(cards[master_card_index].capture->get_disconnected());
1083 handle_hotplugged_cards();
1087 if (!master_card_is_output) {
1088 output_frame_info.frame_timestamp =
1089 cards[master_card_index].new_frames.front().received_timestamp;
1092 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1093 CaptureCard *card = &cards[card_index];
1094 if (input_card_is_master_clock(card_index, master_card_index)) {
1095 // We don't use the queue length policy for the master card,
1096 // but we will if it stops being the master. Thus, clear out
1097 // the policy in case we switch in the future.
1098 card->queue_length_policy.reset(card_index);
1099 assert(!card->new_frames.empty());
1101 trim_queue(card, card_index);
1103 if (!card->new_frames.empty()) {
1104 new_frames[card_index] = move(card->new_frames.front());
1105 has_new_frame[card_index] = true;
1106 card->new_frames.pop_front();
1107 card->new_frames_changed.notify_all();
1111 if (!master_card_is_output) {
1112 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1113 output_frame_info.frame_duration = new_frames[master_card_index].length;
1116 // This might get off by a fractional sample when changing master card
1117 // between ones with different frame rates, but that's fine.
1118 int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1119 output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1120 fractional_samples = num_samples_times_timebase % TIMEBASE;
1121 assert(output_frame_info.num_samples >= 0);
1123 return output_frame_info;
1126 void Mixer::handle_hotplugged_cards()
1128 // Check for cards that have been disconnected since last frame.
1129 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1130 CaptureCard *card = &cards[card_index];
1131 if (card->capture->get_disconnected()) {
1132 fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1133 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1134 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1135 card->queue_length_policy.reset(card_index);
1136 card->capture->start_bm_capture();
1140 // Check for cards that have been connected since last frame.
1141 vector<libusb_device *> hotplugged_cards_copy;
1143 lock_guard<mutex> lock(hotplug_mutex);
1144 swap(hotplugged_cards, hotplugged_cards_copy);
1146 for (libusb_device *new_dev : hotplugged_cards_copy) {
1147 // Look for a fake capture card where we can stick this in.
1148 int free_card_index = -1;
1149 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1150 if (cards[card_index].is_fake_capture) {
1151 free_card_index = card_index;
1156 if (free_card_index == -1) {
1157 fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1158 libusb_unref_device(new_dev);
1160 // BMUSBCapture takes ownership.
1161 fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1162 CaptureCard *card = &cards[free_card_index];
1163 BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1164 configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1165 card->queue_length_policy.reset(free_card_index);
1166 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1167 capture->start_bm_capture();
1173 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)
1175 // Resample the audio as needed, including from previously dropped frames.
1176 assert(num_cards > 0);
1177 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1178 const bool dropped_frame = (frame_num != dropped_frames);
1180 // Signal to the audio thread to process this frame.
1181 // Note that if the frame is a dropped frame, we signal that
1182 // we don't want to use this frame as base for adjusting
1183 // the resampler rate. The reason for this is that the timing
1184 // of these frames is often way too late; they typically don't
1185 // “arrive” before we synthesize them. Thus, we could end up
1186 // in a situation where we have inserted e.g. five audio frames
1187 // into the queue before we then start pulling five of them
1188 // back out. This makes ResamplingQueue overestimate the delay,
1189 // causing undue resampler changes. (We _do_ use the last,
1190 // non-dropped frame; perhaps we should just discard that as well,
1191 // since dropped frames are expected to be rare, and it might be
1192 // better to just wait until we have a slightly more normal situation).
1193 unique_lock<mutex> lock(audio_mutex);
1194 bool adjust_rate = !dropped_frame && !is_preroll;
1195 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1196 audio_task_queue_changed.notify_one();
1198 if (dropped_frame) {
1199 // For dropped frames, increase the pts. Note that if the format changed
1200 // in the meantime, we have no way of detecting that; we just have to
1201 // assume the frame length is always the same.
1202 pts_int += length_per_frame;
1207 void Mixer::render_one_frame(int64_t duration)
1209 // Determine the time code for this frame before we start rendering.
1210 string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1211 if (display_timecode_on_stdout) {
1212 printf("Timecode: '%s'\n", timecode_text.c_str());
1215 // Update Y'CbCr settings for all cards.
1217 unique_lock<mutex> lock(card_mutex);
1218 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1219 YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1220 input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1221 input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1222 input_state.full_range[card_index] = interpretation->full_range;
1226 // Get the main chain from the theme, and set its state immediately.
1227 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1228 EffectChain *chain = theme_main_chain.chain;
1229 theme_main_chain.setup_chain();
1230 //theme_main_chain.chain->enable_phase_timing(true);
1232 // The theme can't (or at least shouldn't!) call connect_signal() on
1233 // each FFmpeg input, so we'll do it here.
1234 for (const pair<LiveInputWrapper *, FFmpegCapture *> &conn : theme->get_signal_connections()) {
1235 conn.first->connect_signal_raw(conn.second->get_card_index());
1238 // If HDMI/SDI output is active and the user has requested auto mode,
1239 // its mode overrides the existing Y'CbCr setting for the chain.
1240 YCbCrLumaCoefficients ycbcr_output_coefficients;
1241 if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1242 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1244 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1247 // TODO: Reduce the duplication against theme.cpp.
1248 YCbCrFormat output_ycbcr_format;
1249 output_ycbcr_format.chroma_subsampling_x = 1;
1250 output_ycbcr_format.chroma_subsampling_y = 1;
1251 output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1252 output_ycbcr_format.full_range = false;
1253 output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1254 chain->change_ycbcr_output_format(output_ycbcr_format);
1256 // Render main chain. If we're using zerocopy Quick Sync encoding
1257 // (the default case), we take an extra copy of the created outputs,
1258 // so that we can display it back to the screen later (it's less memory
1259 // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1260 // Ideally, we'd like to avoid taking copies and just use the main textures
1261 // for display as well, but they're just views into VA-API memory and must be
1262 // unmapped during encoding, so we can't use them for display, unfortunately.
1263 GLuint y_tex, cbcr_full_tex, cbcr_tex;
1264 GLuint y_copy_tex, cbcr_copy_tex = 0;
1265 GLuint y_display_tex, cbcr_display_tex;
1266 GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1267 GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1268 const bool is_zerocopy = video_encoder->is_zerocopy();
1270 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1271 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1272 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1274 y_display_tex = y_copy_tex;
1275 cbcr_display_tex = cbcr_copy_tex;
1277 // y_tex and cbcr_tex will be given by VideoEncoder.
1279 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1280 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1281 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1283 y_display_tex = y_tex;
1284 cbcr_display_tex = cbcr_tex;
1287 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1288 bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1293 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1295 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1298 chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1300 if (display_timecode_in_stream) {
1301 // Render the timecode on top.
1302 timecode_renderer->render_timecode(fbo, timecode_text);
1305 resource_pool->release_fbo(fbo);
1308 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1310 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1312 if (output_card_index != -1) {
1313 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1315 resource_pool->release_2d_texture(cbcr_full_tex);
1317 // Set the right state for the Y' and CbCr textures we use for display.
1318 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1319 glBindTexture(GL_TEXTURE_2D, y_display_tex);
1320 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1321 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1322 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1324 glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1325 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1326 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1327 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1329 RefCountedGLsync fence = video_encoder->end_frame();
1331 // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1332 // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1333 DisplayFrame live_frame;
1334 live_frame.chain = display_chain.get();
1335 live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1336 display_input->set_texture_num(0, y_display_tex);
1337 display_input->set_texture_num(1, cbcr_display_tex);
1339 live_frame.ready_fence = fence;
1340 live_frame.input_frames = {};
1341 live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1342 output_channel[OUTPUT_LIVE].output_frame(live_frame);
1344 // Set up preview and any additional channels.
1345 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1346 DisplayFrame display_frame;
1347 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions
1348 display_frame.chain = chain.chain;
1349 display_frame.setup_chain = chain.setup_chain;
1350 display_frame.ready_fence = fence;
1351 display_frame.input_frames = chain.input_frames;
1352 display_frame.temp_textures = {};
1353 output_channel[i].output_frame(display_frame);
1357 void Mixer::audio_thread_func()
1359 pthread_setname_np(pthread_self(), "Mixer_Audio");
1361 while (!should_quit) {
1365 unique_lock<mutex> lock(audio_mutex);
1366 audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1370 task = audio_task_queue.front();
1371 audio_task_queue.pop();
1374 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1375 task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1376 vector<float> samples_out = audio_mixer.get_output(
1377 task.frame_timestamp,
1379 rate_adjustment_policy);
1381 // Send the samples to the sound card, then add them to the output.
1383 alsa->write(samples_out);
1385 if (output_card_index != -1) {
1386 const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue.
1387 cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1389 video_encoder->add_audio(task.pts_int, move(samples_out));
1393 void Mixer::release_display_frame(DisplayFrame *frame)
1395 for (GLuint texnum : frame->temp_textures) {
1396 resource_pool->release_2d_texture(texnum);
1398 frame->temp_textures.clear();
1399 frame->ready_fence.reset();
1400 frame->input_frames.clear();
1405 mixer_thread = thread(&Mixer::thread_func, this);
1406 audio_thread = thread(&Mixer::audio_thread_func, this);
1412 audio_task_queue_changed.notify_one();
1413 mixer_thread.join();
1414 audio_thread.join();
1417 void Mixer::transition_clicked(int transition_num)
1419 theme->transition_clicked(transition_num, pts());
1422 void Mixer::channel_clicked(int preview_num)
1424 theme->channel_clicked(preview_num);
1427 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1429 unique_lock<mutex> lock(card_mutex);
1430 return ycbcr_interpretation[card_index];
1433 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1435 unique_lock<mutex> lock(card_mutex);
1436 ycbcr_interpretation[card_index] = interpretation;
1439 void Mixer::start_mode_scanning(unsigned card_index)
1441 assert(card_index < num_cards);
1442 if (is_mode_scanning[card_index]) {
1445 is_mode_scanning[card_index] = true;
1446 mode_scanlist[card_index].clear();
1447 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1448 mode_scanlist[card_index].push_back(mode.first);
1450 assert(!mode_scanlist[card_index].empty());
1451 mode_scanlist_index[card_index] = 0;
1452 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1453 last_mode_scan_change[card_index] = steady_clock::now();
1456 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1458 assert(desired_output_card_index != -1);
1459 unique_lock<mutex> lock(card_mutex);
1460 return cards[desired_output_card_index].output->get_available_video_modes();
1463 Mixer::OutputChannel::~OutputChannel()
1465 if (has_current_frame) {
1466 parent->release_display_frame(¤t_frame);
1468 if (has_ready_frame) {
1469 parent->release_display_frame(&ready_frame);
1473 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1475 // Store this frame for display. Remove the ready frame if any
1476 // (it was seemingly never used).
1478 unique_lock<mutex> lock(frame_mutex);
1479 if (has_ready_frame) {
1480 parent->release_display_frame(&ready_frame);
1482 ready_frame = frame;
1483 has_ready_frame = true;
1485 // Call the callbacks under the mutex (they should be short),
1486 // so that we don't race against a callback removal.
1487 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1488 key_and_callback.second();
1492 // Reduce the number of callbacks by filtering duplicates. The reason
1493 // why we bother doing this is that Qt seemingly can get into a state
1494 // where its builds up an essentially unbounded queue of signals,
1495 // consuming more and more memory, and there's no good way of collapsing
1496 // user-defined signals or limiting the length of the queue.
1497 if (transition_names_updated_callback) {
1498 vector<string> transition_names = global_mixer->get_transition_names();
1499 bool changed = false;
1500 if (transition_names.size() != last_transition_names.size()) {
1503 for (unsigned i = 0; i < transition_names.size(); ++i) {
1504 if (transition_names[i] != last_transition_names[i]) {
1511 transition_names_updated_callback(transition_names);
1512 last_transition_names = transition_names;
1515 if (name_updated_callback) {
1516 string name = global_mixer->get_channel_name(channel);
1517 if (name != last_name) {
1518 name_updated_callback(name);
1522 if (color_updated_callback) {
1523 string color = global_mixer->get_channel_color(channel);
1524 if (color != last_color) {
1525 color_updated_callback(color);
1531 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1533 unique_lock<mutex> lock(frame_mutex);
1534 if (!has_current_frame && !has_ready_frame) {
1538 if (has_current_frame && has_ready_frame) {
1539 // We have a new ready frame. Toss the current one.
1540 parent->release_display_frame(¤t_frame);
1541 has_current_frame = false;
1543 if (has_ready_frame) {
1544 assert(!has_current_frame);
1545 current_frame = ready_frame;
1546 ready_frame.ready_fence.reset(); // Drop the refcount.
1547 ready_frame.input_frames.clear(); // Drop the refcounts.
1548 has_current_frame = true;
1549 has_ready_frame = false;
1552 *frame = current_frame;
1556 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1558 unique_lock<mutex> lock(frame_mutex);
1559 new_frame_ready_callbacks[key] = callback;
1562 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1564 unique_lock<mutex> lock(frame_mutex);
1565 new_frame_ready_callbacks.erase(key);
1568 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1570 transition_names_updated_callback = callback;
1573 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1575 name_updated_callback = callback;
1578 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1580 color_updated_callback = callback;
1583 mutex RefCountedGLsync::fence_lock;