4 // The actual video mixer, running in its own separate background thread.
15 #include <condition_variable>
26 #include <movit/effect.h>
27 #include <movit/image_format.h>
29 #include "audio_mixer.h"
30 #include "bmusb/bmusb.h"
32 #include "shared/httpd.h"
33 #include "input_state.h"
35 #include "pbo_frame_allocator.h"
36 #include "ref_counted_frame.h"
37 #include "shared/ref_counted_gl_sync.h"
39 #include "shared/timebase.h"
40 #include "video_encoder.h"
41 #include "ycbcr_interpretation.h"
44 class ChromaSubsampler;
49 class TimecodeRenderer;
59 // A class to estimate the future jitter. Used in QueueLengthPolicy (see below).
61 // There are many ways to estimate jitter; I've tested a few ones (and also
62 // some algorithms that don't explicitly model jitter) with different
63 // parameters on some real-life data in experiments/queue_drop_policy.cpp.
64 // This is one based on simple order statistics where I've added some margin in
65 // the number of starvation events; I believe that about one every hour would
66 // probably be acceptable, but this one typically goes lower than that, at the
67 // cost of 2–3 ms extra latency. (If the queue is hard-limited to one frame, it's
68 // possible to get ~10 ms further down, but this would mean framedrops every
69 // second or so.) The general strategy is: Take the 99.9-percentile jitter over
70 // last 5000 frames, multiply by two, and that's our worst-case jitter
71 // estimate. The fact that we're not using the max value means that we could
72 // actually even throw away very late frames immediately, which means we only
73 // get one user-visible event instead of seeing something both when the frame
74 // arrives late (duplicate frame) and then again when we drop.
77 static constexpr size_t history_length = 5000;
78 static constexpr double percentile = 0.999;
79 static constexpr double multiplier = 2.0;
82 void register_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
83 void unregister_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
89 void frame_arrived(std::chrono::steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames);
90 std::chrono::steady_clock::time_point get_expected_next_frame() const { return expected_timestamp; }
91 double estimate_max_jitter() const;
94 // A simple O(k) based algorithm for getting the k-th largest or
95 // smallest element from our window; we simply keep the multiset
96 // ordered (insertions and deletions are O(n) as always) and then
97 // iterate from one of the sides. If we had larger values of k,
98 // we could go for a more complicated setup with two sets or heaps
99 // (one increasing and one decreasing) that we keep balanced around
100 // the point, or it is possible to reimplement std::set with
101 // counts in each node. However, since k=5, we don't need this.
102 std::multiset<double> orders;
103 std::deque<std::multiset<double>::iterator> history;
105 std::chrono::steady_clock::time_point expected_timestamp = std::chrono::steady_clock::time_point::min();
107 // Metrics. There are no direct summaries for jitter, since we already have latency summaries.
108 std::atomic<int64_t> metric_input_underestimated_jitter_frames{0};
109 std::atomic<double> metric_input_estimated_max_jitter_seconds{0.0 / 0.0};
112 // For any card that's not the master (where we pick out the frames as they
113 // come, as fast as we can process), there's going to be a queue. The question
114 // is when we should drop frames from that queue (apart from the obvious
115 // dropping if the 16-frame queue should become full), especially given that
116 // the frame rate could be lower or higher than the master (either subtly or
117 // dramatically). We have two (conflicting) demands:
119 // 1. We want to avoid starving the queue.
120 // 2. We don't want to add more delay than is needed.
122 // Our general strategy is to drop as many frames as we can (helping for #2)
123 // that we think is safe for #1 given jitter. To this end, we measure the
124 // deviation from the expected arrival time for all cards, and use that for
125 // continuous jitter estimation.
127 // We then drop everything from the queue that we're sure we won't need to
128 // serve the output in the time before the next frame arrives. Typically,
129 // this means the queue will contain 0 or 1 frames, although more is also
130 // possible if the jitter is very high.
131 class QueueLengthPolicy {
133 QueueLengthPolicy() {}
134 void reset(unsigned card_index) {
135 this->card_index = card_index;
138 void register_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
139 void unregister_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
141 // Call after picking out a frame, so 0 means starvation.
142 void update_policy(std::chrono::steady_clock::time_point now,
143 std::chrono::steady_clock::time_point expected_next_frame,
144 int64_t input_frame_duration,
145 int64_t master_frame_duration,
146 double max_input_card_jitter_seconds,
147 double max_master_card_jitter_seconds);
148 unsigned get_safe_queue_length() const { return safe_queue_length; }
151 unsigned card_index; // For debugging and metrics only.
152 unsigned safe_queue_length = 0; // Can never go below zero.
155 std::atomic<int64_t> metric_input_queue_safe_length_frames{1};
160 // The surface format is used for offscreen destinations for OpenGL contexts we need.
161 Mixer(const QSurfaceFormat &format, unsigned num_cards);
166 void transition_clicked(int transition_num);
167 void channel_clicked(int preview_num);
172 OUTPUT_INPUT0, // 1, 2, 3, up to 15 follow numerically.
176 struct DisplayFrame {
177 // The chain for rendering this frame. To render a display frame,
178 // first wait for <ready_fence>, then call <setup_chain>
179 // to wire up all the inputs, and then finally call
180 // chain->render_to_screen() or similar.
181 movit::EffectChain *chain;
182 std::function<void()> setup_chain;
184 // Asserted when all the inputs are ready; you cannot render the chain
186 RefCountedGLsync ready_fence;
188 // Holds on to all the input frames needed for this display frame,
189 // so they are not released while still rendering.
190 std::vector<RefCountedFrame> input_frames;
192 // Textures that should be released back to the resource pool
193 // when this frame disappears, if any.
194 // TODO: Refcount these as well?
195 std::vector<GLuint> temp_textures;
197 // Implicitly frees the previous one if there's a new frame available.
198 bool get_display_frame(Output output, DisplayFrame *frame) {
199 return output_channel[output].get_display_frame(frame);
202 // NOTE: Callbacks will be called with a mutex held, so you should probably
203 // not do real work in them.
204 typedef std::function<void()> new_frame_ready_callback_t;
205 void add_frame_ready_callback(Output output, void *key, new_frame_ready_callback_t callback)
207 output_channel[output].add_frame_ready_callback(key, callback);
210 void remove_frame_ready_callback(Output output, void *key)
212 output_channel[output].remove_frame_ready_callback(key);
215 // TODO: Should this really be per-channel? Shouldn't it just be called for e.g. the live output?
216 typedef std::function<void(const std::vector<std::string> &)> transition_names_updated_callback_t;
217 void set_transition_names_updated_callback(Output output, transition_names_updated_callback_t callback)
219 output_channel[output].set_transition_names_updated_callback(callback);
222 typedef std::function<void(const std::string &)> name_updated_callback_t;
223 void set_name_updated_callback(Output output, name_updated_callback_t callback)
225 output_channel[output].set_name_updated_callback(callback);
228 typedef std::function<void(const std::string &)> color_updated_callback_t;
229 void set_color_updated_callback(Output output, color_updated_callback_t callback)
231 output_channel[output].set_color_updated_callback(callback);
234 std::vector<std::string> get_transition_names()
236 return theme->get_transition_names(pts());
239 unsigned get_num_channels() const
241 return theme->get_num_channels();
244 std::string get_channel_name(unsigned channel) const
246 return theme->get_channel_name(channel);
249 std::string get_channel_color(unsigned channel) const
251 return theme->get_channel_color(channel);
254 int map_channel_to_signal(unsigned channel) const
256 return theme->map_channel_to_signal(channel);
259 int map_signal_to_card(int signal)
261 return theme->map_signal_to_card(signal);
264 unsigned get_master_clock() const
266 return master_clock_channel;
269 void set_master_clock(unsigned channel)
271 master_clock_channel = channel;
274 void set_signal_mapping(int signal, int card)
276 return theme->set_signal_mapping(signal, card);
279 YCbCrInterpretation get_input_ycbcr_interpretation(unsigned card_index) const;
280 void set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation);
282 bool get_supports_set_wb(unsigned channel) const
284 return theme->get_supports_set_wb(channel);
287 void set_wb(unsigned channel, double r, double g, double b) const
289 theme->set_wb(channel, r, g, b);
292 std::string format_status_line(const std::string &disk_space_left_text, double file_length_seconds)
294 return theme->format_status_line(disk_space_left_text, file_length_seconds);
297 // Note: You can also get this through the global variable global_audio_mixer.
298 AudioMixer *get_audio_mixer() { return audio_mixer.get(); }
299 const AudioMixer *get_audio_mixer() const { return audio_mixer.get(); }
306 unsigned get_num_cards() const { return num_cards; }
308 std::string get_card_description(unsigned card_index) const {
309 assert(card_index < num_cards);
310 return cards[card_index].capture->get_description();
313 // The difference between this and the previous function is that if a card
314 // is used as the current output, get_card_description() will return the
315 // fake card that's replacing it for input, whereas this function will return
316 // the card's actual name.
317 std::string get_output_card_description(unsigned card_index) const {
318 assert(card_can_be_used_as_output(card_index));
319 assert(card_index < num_cards);
320 if (cards[card_index].parked_capture) {
321 return cards[card_index].parked_capture->get_description();
323 return cards[card_index].capture->get_description();
327 bool card_can_be_used_as_output(unsigned card_index) const {
328 assert(card_index < num_cards);
329 return cards[card_index].output != nullptr;
332 bool card_is_ffmpeg(unsigned card_index) const {
333 assert(card_index < num_cards + num_video_inputs);
334 return cards[card_index].type == CardType::FFMPEG_INPUT;
337 std::map<uint32_t, bmusb::VideoMode> get_available_video_modes(unsigned card_index) const {
338 assert(card_index < num_cards);
339 return cards[card_index].capture->get_available_video_modes();
342 uint32_t get_current_video_mode(unsigned card_index) const {
343 assert(card_index < num_cards);
344 return cards[card_index].capture->get_current_video_mode();
347 void set_video_mode(unsigned card_index, uint32_t mode) {
348 assert(card_index < num_cards);
349 cards[card_index].capture->set_video_mode(mode);
352 void start_mode_scanning(unsigned card_index);
354 std::map<uint32_t, std::string> get_available_video_inputs(unsigned card_index) const {
355 assert(card_index < num_cards);
356 return cards[card_index].capture->get_available_video_inputs();
359 uint32_t get_current_video_input(unsigned card_index) const {
360 assert(card_index < num_cards);
361 return cards[card_index].capture->get_current_video_input();
364 void set_video_input(unsigned card_index, uint32_t input) {
365 assert(card_index < num_cards);
366 cards[card_index].capture->set_video_input(input);
369 std::map<uint32_t, std::string> get_available_audio_inputs(unsigned card_index) const {
370 assert(card_index < num_cards);
371 return cards[card_index].capture->get_available_audio_inputs();
374 uint32_t get_current_audio_input(unsigned card_index) const {
375 assert(card_index < num_cards);
376 return cards[card_index].capture->get_current_audio_input();
379 void set_audio_input(unsigned card_index, uint32_t input) {
380 assert(card_index < num_cards);
381 cards[card_index].capture->set_audio_input(input);
384 std::string get_ffmpeg_filename(unsigned card_index) const;
386 void set_ffmpeg_filename(unsigned card_index, const std::string &filename);
388 void change_x264_bitrate(unsigned rate_kbit) {
389 video_encoder->change_x264_bitrate(rate_kbit);
392 int get_output_card_index() const { // -1 = no output, just stream.
393 return desired_output_card_index;
396 void set_output_card(int card_index) { // -1 = no output, just stream.
397 desired_output_card_index = card_index;
400 std::map<uint32_t, bmusb::VideoMode> get_available_output_video_modes() const;
402 uint32_t get_output_video_mode() const {
403 return desired_output_video_mode;
406 void set_output_video_mode(uint32_t mode) {
407 desired_output_video_mode = mode;
410 void set_display_timecode_in_stream(bool enable) {
411 display_timecode_in_stream = enable;
414 void set_display_timecode_on_stdout(bool enable) {
415 display_timecode_on_stdout = enable;
418 int64_t get_num_connected_clients() const {
419 return httpd.get_num_connected_clients();
422 Theme::MenuEntry *get_theme_menu() { return theme->get_theme_menu(); }
424 void theme_menu_entry_clicked(int lua_ref) { return theme->theme_menu_entry_clicked(lua_ref); }
426 void set_theme_menu_callback(std::function<void()> callback)
428 theme->set_theme_menu_callback(callback);
431 void wait_for_next_frame();
436 enum class CardType {
442 void configure_card(unsigned card_index, bmusb::CaptureInterface *capture, CardType card_type, DeckLinkOutput *output);
443 void set_output_card_internal(int card_index); // Should only be called from the mixer thread.
444 void bm_frame(unsigned card_index, uint16_t timecode,
445 bmusb::FrameAllocator::Frame video_frame, size_t video_offset, bmusb::VideoFormat video_format,
446 bmusb::FrameAllocator::Frame audio_frame, size_t audio_offset, bmusb::AudioFormat audio_format);
447 void bm_hotplug_add(libusb_device *dev);
448 void bm_hotplug_remove(unsigned card_index);
449 void place_rectangle(movit::Effect *resample_effect, movit::Effect *padding_effect, float x0, float y0, float x1, float y1);
451 void handle_hotplugged_cards();
452 void schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame, bool is_preroll, std::chrono::steady_clock::time_point frame_timestamp);
453 std::string get_timecode_text() const;
454 void render_one_frame(int64_t duration);
455 void audio_thread_func();
456 void release_display_frame(DisplayFrame *frame);
457 double pts() { return double(pts_int) / TIMEBASE; }
458 void trim_queue(CaptureCard *card, size_t safe_queue_length);
459 std::pair<std::string, std::string> get_channels_json();
460 std::pair<std::string, std::string> get_channel_color_http(unsigned channel_idx);
463 unsigned num_cards, num_video_inputs, num_html_inputs = 0;
465 QSurface *mixer_surface, *h264_encoder_surface, *decklink_output_surface, *image_update_surface;
466 std::unique_ptr<movit::ResourcePool> resource_pool;
467 std::unique_ptr<Theme> theme;
468 std::atomic<unsigned> audio_source_channel{0};
469 std::atomic<int> master_clock_channel{0}; // Gets overridden by <output_card_index> if set.
470 int output_card_index = -1; // -1 for none.
471 uint32_t output_video_mode = -1;
473 // The mechanics of changing the output card and modes are so intricately connected
474 // with the work the mixer thread is doing. Thus, we don't change it directly,
475 // we just set this variable instead, which signals to the mixer thread that
476 // it should do the change before the next frame. This simplifies locking
477 // considerations immensely.
478 std::atomic<int> desired_output_card_index{-1};
479 std::atomic<uint32_t> desired_output_video_mode{0};
481 std::unique_ptr<movit::EffectChain> display_chain;
482 std::unique_ptr<ChromaSubsampler> chroma_subsampler;
483 std::unique_ptr<v210Converter> v210_converter;
484 std::unique_ptr<VideoEncoder> video_encoder;
485 std::unique_ptr<MJPEGEncoder> mjpeg_encoder;
487 std::unique_ptr<TimecodeRenderer> timecode_renderer;
488 std::atomic<bool> display_timecode_in_stream{false};
489 std::atomic<bool> display_timecode_on_stdout{false};
491 // Effects part of <display_chain>. Owned by <display_chain>.
492 movit::YCbCrInput *display_input;
494 int64_t pts_int = 0; // In TIMEBASE units.
496 mutable std::mutex frame_num_mutex;
497 std::condition_variable frame_num_updated;
498 unsigned frame_num = 0; // Under <frame_num_mutex>.
500 // Accumulated errors in number of 1/TIMEBASE audio samples. If OUTPUT_FREQUENCY divided by
501 // frame rate is integer, will always stay zero.
502 unsigned fractional_samples = 0;
504 mutable std::mutex card_mutex;
505 bool has_bmusb_thread = false;
507 std::unique_ptr<bmusb::CaptureInterface> capture;
508 bool is_fake_capture;
510 std::unique_ptr<DeckLinkOutput> output;
512 // CEF only delivers frames when it actually has a change.
513 // If we trim the queue for latency reasons, we could thus
514 // end up in a situation trimming a frame that was meant to
515 // be displayed for a long time, which is really suboptimal.
516 // Thus, if we drop the last frame we have, may_have_dropped_last_frame
517 // is set to true, and the next starvation event will trigger
518 // us requestin a CEF repaint.
519 bool is_cef_capture, may_have_dropped_last_frame = false;
521 // If this card is used for output (ie., output_card_index points to it),
522 // it cannot simultaneously be uesd for capture, so <capture> gets replaced
523 // by a FakeCapture. However, since reconstructing the real capture object
524 // with all its state can be annoying, it is not being deleted, just stopped
526 std::unique_ptr<bmusb::CaptureInterface> parked_capture;
528 std::unique_ptr<PBOFrameAllocator> frame_allocator;
530 // Stuff for the OpenGL context (for texture uploading).
531 QSurface *surface = nullptr;
534 RefCountedFrame frame;
535 int64_t length; // In TIMEBASE units.
537 unsigned field; // Which field (0 or 1) of the frame to use. Always 0 for progressive.
538 std::function<void()> upload_func; // Needs to be called to actually upload the texture to OpenGL.
539 unsigned dropped_frames = 0; // Number of dropped frames before this one.
540 std::chrono::steady_clock::time_point received_timestamp = std::chrono::steady_clock::time_point::min();
541 movit::RGBTriplet neutral_color{1.0f, 1.0f, 1.0f};
543 // Used for MJPEG encoding. (upload_func packs everything it needs
544 // into the functor, but would otherwise also use these.)
545 // width=0 or height=0 means a broken frame, ie., do not upload.
546 bmusb::VideoFormat video_format;
547 size_t y_offset, cbcr_offset;
549 std::deque<NewFrame> new_frames;
550 std::condition_variable new_frames_changed; // Set whenever new_frames is changed.
551 QueueLengthPolicy queue_length_policy; // Refers to the "new_frames" queue.
553 std::vector<int32_t> new_raw_audio;
555 int last_timecode = -1; // Unwrapped.
557 JitterHistory jitter_history;
560 std::vector<std::pair<std::string, std::string>> labels;
561 std::atomic<int64_t> metric_input_received_frames{0};
562 std::atomic<int64_t> metric_input_duped_frames{0};
563 std::atomic<int64_t> metric_input_dropped_frames_jitter{0};
564 std::atomic<int64_t> metric_input_dropped_frames_error{0};
565 std::atomic<int64_t> metric_input_resets{0};
566 std::atomic<int64_t> metric_input_queue_length_frames{0};
568 std::atomic<int64_t> metric_input_has_signal_bool{-1};
569 std::atomic<int64_t> metric_input_is_connected_bool{-1};
570 std::atomic<int64_t> metric_input_interlaced_bool{-1};
571 std::atomic<int64_t> metric_input_width_pixels{-1};
572 std::atomic<int64_t> metric_input_height_pixels{-1};
573 std::atomic<int64_t> metric_input_frame_rate_nom{-1};
574 std::atomic<int64_t> metric_input_frame_rate_den{-1};
575 std::atomic<int64_t> metric_input_sample_rate_hz{-1};
577 JitterHistory output_jitter_history;
578 CaptureCard cards[MAX_VIDEO_CARDS]; // Protected by <card_mutex>.
579 YCbCrInterpretation ycbcr_interpretation[MAX_VIDEO_CARDS]; // Protected by <card_mutex>.
580 movit::RGBTriplet last_received_neutral_color[MAX_VIDEO_CARDS]; // Used by the mixer thread only. Constructor-initialiezd.
581 std::unique_ptr<AudioMixer> audio_mixer; // Same as global_audio_mixer (see audio_mixer.h).
582 bool input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const;
583 struct OutputFrameInfo {
584 int dropped_frames; // Since last frame.
585 int num_samples; // Audio samples needed for this output frame.
586 int64_t frame_duration; // In TIMEBASE units.
588 std::chrono::steady_clock::time_point frame_timestamp;
590 OutputFrameInfo 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], std::vector<int32_t> raw_audio[MAX_VIDEO_CARDS]);
592 InputState input_state;
594 // Cards we have been noticed about being hotplugged, but haven't tried adding yet.
595 // Protected by its own mutex.
596 std::mutex hotplug_mutex;
597 std::vector<libusb_device *> hotplugged_cards;
599 class OutputChannel {
602 void output_frame(DisplayFrame &&frame);
603 bool get_display_frame(DisplayFrame *frame);
604 void add_frame_ready_callback(void *key, new_frame_ready_callback_t callback);
605 void remove_frame_ready_callback(void *key);
606 void set_transition_names_updated_callback(transition_names_updated_callback_t callback);
607 void set_name_updated_callback(name_updated_callback_t callback);
608 void set_color_updated_callback(color_updated_callback_t callback);
614 Mixer *parent = nullptr; // Not owned.
615 std::mutex frame_mutex;
616 DisplayFrame current_frame, ready_frame; // protected by <frame_mutex>
617 bool has_current_frame = false, has_ready_frame = false; // protected by <frame_mutex>
618 std::map<void *, new_frame_ready_callback_t> new_frame_ready_callbacks; // protected by <frame_mutex>
619 transition_names_updated_callback_t transition_names_updated_callback;
620 name_updated_callback_t name_updated_callback;
621 color_updated_callback_t color_updated_callback;
623 std::vector<std::string> last_transition_names;
624 std::string last_name, last_color;
626 OutputChannel output_channel[NUM_OUTPUTS];
628 std::thread mixer_thread;
629 std::thread audio_thread;
630 std::atomic<bool> should_quit{false};
631 std::atomic<bool> should_cut{false};
633 std::unique_ptr<ALSAOutput> alsa;
639 std::chrono::steady_clock::time_point frame_timestamp;
641 std::mutex audio_mutex;
642 std::condition_variable audio_task_queue_changed;
643 std::queue<AudioTask> audio_task_queue; // Under audio_mutex.
645 // For mode scanning.
646 bool is_mode_scanning[MAX_VIDEO_CARDS]{ false };
647 std::vector<uint32_t> mode_scanlist[MAX_VIDEO_CARDS];
648 unsigned mode_scanlist_index[MAX_VIDEO_CARDS]{ 0 };
649 std::chrono::steady_clock::time_point last_mode_scan_change[MAX_VIDEO_CARDS];
652 extern Mixer *global_mixer;
654 #endif // !defined(_MIXER_H)