4 // The actual video mixer, running in its own separate background thread.
15 #include <condition_variable>
26 #include <movit/image_format.h>
28 #include "audio_mixer.h"
29 #include "bmusb/bmusb.h"
32 #include "input_state.h"
34 #include "pbo_frame_allocator.h"
35 #include "ref_counted_frame.h"
36 #include "ref_counted_gl_sync.h"
39 #include "video_encoder.h"
40 #include "ycbcr_interpretation.h"
43 class ChromaSubsampler;
47 class TimecodeRenderer;
57 // A class to estimate the future jitter. Used in QueueLengthPolicy (see below).
59 // There are many ways to estimate jitter; I've tested a few ones (and also
60 // some algorithms that don't explicitly model jitter) with different
61 // parameters on some real-life data in experiments/queue_drop_policy.cpp.
62 // This is one based on simple order statistics where I've added some margin in
63 // the number of starvation events; I believe that about one every hour would
64 // probably be acceptable, but this one typically goes lower than that, at the
65 // cost of 2–3 ms extra latency. (If the queue is hard-limited to one frame, it's
66 // possible to get ~10 ms further down, but this would mean framedrops every
67 // second or so.) The general strategy is: Take the 99.9-percentile jitter over
68 // last 5000 frames, multiply by two, and that's our worst-case jitter
69 // estimate. The fact that we're not using the max value means that we could
70 // actually even throw away very late frames immediately, which means we only
71 // get one user-visible event instead of seeing something both when the frame
72 // arrives late (duplicate frame) and then again when we drop.
75 static constexpr size_t history_length = 5000;
76 static constexpr double percentile = 0.999;
77 static constexpr double multiplier = 2.0;
80 void register_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
81 void unregister_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
87 void frame_arrived(std::chrono::steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames);
88 std::chrono::steady_clock::time_point get_expected_next_frame() const { return expected_timestamp; }
89 double estimate_max_jitter() const;
92 // A simple O(k) based algorithm for getting the k-th largest or
93 // smallest element from our window; we simply keep the multiset
94 // ordered (insertions and deletions are O(n) as always) and then
95 // iterate from one of the sides. If we had larger values of k,
96 // we could go for a more complicated setup with two sets or heaps
97 // (one increasing and one decreasing) that we keep balanced around
98 // the point, or it is possible to reimplement std::set with
99 // counts in each node. However, since k=5, we don't need this.
100 std::multiset<double> orders;
101 std::deque<std::multiset<double>::iterator> history;
103 std::chrono::steady_clock::time_point expected_timestamp = std::chrono::steady_clock::time_point::min();
105 // Metrics. There are no direct summaries for jitter, since we already have latency summaries.
106 std::atomic<int64_t> metric_input_underestimated_jitter_frames{0};
107 std::atomic<double> metric_input_estimated_max_jitter_seconds{0.0 / 0.0};
110 // For any card that's not the master (where we pick out the frames as they
111 // come, as fast as we can process), there's going to be a queue. The question
112 // is when we should drop frames from that queue (apart from the obvious
113 // dropping if the 16-frame queue should become full), especially given that
114 // the frame rate could be lower or higher than the master (either subtly or
115 // dramatically). We have two (conflicting) demands:
117 // 1. We want to avoid starving the queue.
118 // 2. We don't want to add more delay than is needed.
120 // Our general strategy is to drop as many frames as we can (helping for #2)
121 // that we think is safe for #1 given jitter. To this end, we measure the
122 // deviation from the expected arrival time for all cards, and use that for
123 // continuous jitter estimation.
125 // We then drop everything from the queue that we're sure we won't need to
126 // serve the output in the time before the next frame arrives. Typically,
127 // this means the queue will contain 0 or 1 frames, although more is also
128 // possible if the jitter is very high.
129 class QueueLengthPolicy {
131 QueueLengthPolicy() {}
132 void reset(unsigned card_index) {
133 this->card_index = card_index;
136 void register_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
137 void unregister_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
139 // Call after picking out a frame, so 0 means starvation.
140 void update_policy(std::chrono::steady_clock::time_point now,
141 std::chrono::steady_clock::time_point expected_next_frame,
142 int64_t input_frame_duration,
143 int64_t master_frame_duration,
144 double max_input_card_jitter_seconds,
145 double max_master_card_jitter_seconds);
146 unsigned get_safe_queue_length() const { return safe_queue_length; }
149 unsigned card_index; // For debugging and metrics only.
150 unsigned safe_queue_length = 0; // Can never go below zero.
153 std::atomic<int64_t> metric_input_queue_safe_length_frames{1};
158 // The surface format is used for offscreen destinations for OpenGL contexts we need.
159 Mixer(const QSurfaceFormat &format, unsigned num_cards);
164 void transition_clicked(int transition_num);
165 void channel_clicked(int preview_num);
170 OUTPUT_INPUT0, // 1, 2, 3, up to 15 follow numerically.
174 struct DisplayFrame {
175 // The chain for rendering this frame. To render a display frame,
176 // first wait for <ready_fence>, then call <setup_chain>
177 // to wire up all the inputs, and then finally call
178 // chain->render_to_screen() or similar.
179 movit::EffectChain *chain;
180 std::function<void()> setup_chain;
182 // Asserted when all the inputs are ready; you cannot render the chain
184 RefCountedGLsync ready_fence;
186 // Holds on to all the input frames needed for this display frame,
187 // so they are not released while still rendering.
188 std::vector<RefCountedFrame> input_frames;
190 // Textures that should be released back to the resource pool
191 // when this frame disappears, if any.
192 // TODO: Refcount these as well?
193 std::vector<GLuint> temp_textures;
195 // Implicitly frees the previous one if there's a new frame available.
196 bool get_display_frame(Output output, DisplayFrame *frame) {
197 return output_channel[output].get_display_frame(frame);
200 // NOTE: Callbacks will be called with a mutex held, so you should probably
201 // not do real work in them.
202 typedef std::function<void()> new_frame_ready_callback_t;
203 void add_frame_ready_callback(Output output, void *key, new_frame_ready_callback_t callback)
205 output_channel[output].add_frame_ready_callback(key, callback);
208 void remove_frame_ready_callback(Output output, void *key)
210 output_channel[output].remove_frame_ready_callback(key);
213 // TODO: Should this really be per-channel? Shouldn't it just be called for e.g. the live output?
214 typedef std::function<void(const std::vector<std::string> &)> transition_names_updated_callback_t;
215 void set_transition_names_updated_callback(Output output, transition_names_updated_callback_t callback)
217 output_channel[output].set_transition_names_updated_callback(callback);
220 typedef std::function<void(const std::string &)> name_updated_callback_t;
221 void set_name_updated_callback(Output output, name_updated_callback_t callback)
223 output_channel[output].set_name_updated_callback(callback);
226 typedef std::function<void(const std::string &)> color_updated_callback_t;
227 void set_color_updated_callback(Output output, color_updated_callback_t callback)
229 output_channel[output].set_color_updated_callback(callback);
232 std::vector<std::string> get_transition_names()
234 return theme->get_transition_names(pts());
237 unsigned get_num_channels() const
239 return theme->get_num_channels();
242 std::string get_channel_name(unsigned channel) const
244 return theme->get_channel_name(channel);
247 std::string get_channel_color(unsigned channel) const
249 return theme->get_channel_color(channel);
252 int get_channel_signal(unsigned channel) const
254 return theme->get_channel_signal(channel);
257 int map_signal(unsigned channel)
259 return theme->map_signal(channel);
262 unsigned get_master_clock() const
264 return master_clock_channel;
267 void set_master_clock(unsigned channel)
269 master_clock_channel = channel;
272 void set_signal_mapping(int signal, int card)
274 return theme->set_signal_mapping(signal, card);
277 YCbCrInterpretation get_input_ycbcr_interpretation(unsigned card_index) const;
278 void set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation);
280 bool get_supports_set_wb(unsigned channel) const
282 return theme->get_supports_set_wb(channel);
285 void set_wb(unsigned channel, double r, double g, double b) const
287 theme->set_wb(channel, r, g, b);
290 // Note: You can also get this through the global variable global_audio_mixer.
291 AudioMixer *get_audio_mixer() { return audio_mixer.get(); }
292 const AudioMixer *get_audio_mixer() const { return audio_mixer.get(); }
299 unsigned get_num_cards() const { return num_cards; }
301 std::string get_card_description(unsigned card_index) const {
302 assert(card_index < num_cards);
303 return cards[card_index].capture->get_description();
306 // The difference between this and the previous function is that if a card
307 // is used as the current output, get_card_description() will return the
308 // fake card that's replacing it for input, whereas this function will return
309 // the card's actual name.
310 std::string get_output_card_description(unsigned card_index) const {
311 assert(card_can_be_used_as_output(card_index));
312 assert(card_index < num_cards);
313 if (cards[card_index].parked_capture) {
314 return cards[card_index].parked_capture->get_description();
316 return cards[card_index].capture->get_description();
320 bool card_can_be_used_as_output(unsigned card_index) const {
321 assert(card_index < num_cards);
322 return cards[card_index].output != nullptr;
325 bool card_is_ffmpeg(unsigned card_index) const {
326 assert(card_index < num_cards + num_video_inputs);
327 return cards[card_index].type == CardType::FFMPEG_INPUT;
330 std::map<uint32_t, bmusb::VideoMode> get_available_video_modes(unsigned card_index) const {
331 assert(card_index < num_cards);
332 return cards[card_index].capture->get_available_video_modes();
335 uint32_t get_current_video_mode(unsigned card_index) const {
336 assert(card_index < num_cards);
337 return cards[card_index].capture->get_current_video_mode();
340 void set_video_mode(unsigned card_index, uint32_t mode) {
341 assert(card_index < num_cards);
342 cards[card_index].capture->set_video_mode(mode);
345 void start_mode_scanning(unsigned card_index);
347 std::map<uint32_t, std::string> get_available_video_inputs(unsigned card_index) const {
348 assert(card_index < num_cards);
349 return cards[card_index].capture->get_available_video_inputs();
352 uint32_t get_current_video_input(unsigned card_index) const {
353 assert(card_index < num_cards);
354 return cards[card_index].capture->get_current_video_input();
357 void set_video_input(unsigned card_index, uint32_t input) {
358 assert(card_index < num_cards);
359 cards[card_index].capture->set_video_input(input);
362 std::map<uint32_t, std::string> get_available_audio_inputs(unsigned card_index) const {
363 assert(card_index < num_cards);
364 return cards[card_index].capture->get_available_audio_inputs();
367 uint32_t get_current_audio_input(unsigned card_index) const {
368 assert(card_index < num_cards);
369 return cards[card_index].capture->get_current_audio_input();
372 void set_audio_input(unsigned card_index, uint32_t input) {
373 assert(card_index < num_cards);
374 cards[card_index].capture->set_audio_input(input);
377 std::string get_ffmpeg_filename(unsigned card_index) const;
379 void set_ffmpeg_filename(unsigned card_index, const std::string &filename);
381 void change_x264_bitrate(unsigned rate_kbit) {
382 video_encoder->change_x264_bitrate(rate_kbit);
385 int get_output_card_index() const { // -1 = no output, just stream.
386 return desired_output_card_index;
389 void set_output_card(int card_index) { // -1 = no output, just stream.
390 desired_output_card_index = card_index;
393 std::map<uint32_t, bmusb::VideoMode> get_available_output_video_modes() const;
395 uint32_t get_output_video_mode() const {
396 return desired_output_video_mode;
399 void set_output_video_mode(uint32_t mode) {
400 desired_output_video_mode = mode;
403 void set_display_timecode_in_stream(bool enable) {
404 display_timecode_in_stream = enable;
407 void set_display_timecode_on_stdout(bool enable) {
408 display_timecode_on_stdout = enable;
411 int64_t get_num_connected_clients() const {
412 return httpd.get_num_connected_clients();
415 std::vector<Theme::MenuEntry> get_theme_menu() { return theme->get_theme_menu(); }
417 void theme_menu_entry_clicked(int lua_ref) { return theme->theme_menu_entry_clicked(lua_ref); }
419 void set_theme_menu_callback(std::function<void()> callback)
421 theme->set_theme_menu_callback(callback);
424 void wait_for_next_frame();
429 enum class CardType {
435 void configure_card(unsigned card_index, bmusb::CaptureInterface *capture, CardType card_type, DeckLinkOutput *output);
436 void set_output_card_internal(int card_index); // Should only be called from the mixer thread.
437 void bm_frame(unsigned card_index, uint16_t timecode,
438 bmusb::FrameAllocator::Frame video_frame, size_t video_offset, bmusb::VideoFormat video_format,
439 bmusb::FrameAllocator::Frame audio_frame, size_t audio_offset, bmusb::AudioFormat audio_format);
440 void bm_hotplug_add(libusb_device *dev);
441 void bm_hotplug_remove(unsigned card_index);
442 void place_rectangle(movit::Effect *resample_effect, movit::Effect *padding_effect, float x0, float y0, float x1, float y1);
444 void handle_hotplugged_cards();
445 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);
446 std::string get_timecode_text() const;
447 void render_one_frame(int64_t duration);
448 void audio_thread_func();
449 void release_display_frame(DisplayFrame *frame);
450 double pts() { return double(pts_int) / TIMEBASE; }
451 void trim_queue(CaptureCard *card, size_t safe_queue_length);
452 std::pair<std::string, std::string> get_channels_json();
453 std::pair<std::string, std::string> get_channel_color_http(unsigned channel_idx);
456 unsigned num_cards, num_video_inputs, num_html_inputs = 0;
458 QSurface *mixer_surface, *h264_encoder_surface, *decklink_output_surface;
459 std::unique_ptr<movit::ResourcePool> resource_pool;
460 std::unique_ptr<Theme> theme;
461 std::atomic<unsigned> audio_source_channel{0};
462 std::atomic<int> master_clock_channel{0}; // Gets overridden by <output_card_index> if set.
463 int output_card_index = -1; // -1 for none.
464 uint32_t output_video_mode = -1;
466 // The mechanics of changing the output card and modes are so intricately connected
467 // with the work the mixer thread is doing. Thus, we don't change it directly,
468 // we just set this variable instead, which signals to the mixer thread that
469 // it should do the change before the next frame. This simplifies locking
470 // considerations immensely.
471 std::atomic<int> desired_output_card_index{-1};
472 std::atomic<uint32_t> desired_output_video_mode{0};
474 std::unique_ptr<movit::EffectChain> display_chain;
475 std::unique_ptr<ChromaSubsampler> chroma_subsampler;
476 std::unique_ptr<v210Converter> v210_converter;
477 std::unique_ptr<VideoEncoder> video_encoder;
479 std::unique_ptr<TimecodeRenderer> timecode_renderer;
480 std::atomic<bool> display_timecode_in_stream{false};
481 std::atomic<bool> display_timecode_on_stdout{false};
483 // Effects part of <display_chain>. Owned by <display_chain>.
484 movit::YCbCrInput *display_input;
486 int64_t pts_int = 0; // In TIMEBASE units.
488 mutable std::mutex frame_num_mutex;
489 std::condition_variable frame_num_updated;
490 unsigned frame_num = 0; // Under <frame_num_mutex>.
492 // Accumulated errors in number of 1/TIMEBASE audio samples. If OUTPUT_FREQUENCY divided by
493 // frame rate is integer, will always stay zero.
494 unsigned fractional_samples = 0;
496 mutable std::mutex card_mutex;
497 bool has_bmusb_thread = false;
499 std::unique_ptr<bmusb::CaptureInterface> capture;
500 bool is_fake_capture;
502 std::unique_ptr<DeckLinkOutput> output;
504 // CEF only delivers frames when it actually has a change.
505 // If we trim the queue for latency reasons, we could thus
506 // end up in a situation trimming a frame that was meant to
507 // be displayed for a long time, which is really suboptimal.
508 // Thus, if we drop the last frame we have, may_have_dropped_last_frame
509 // is set to true, and the next starvation event will trigger
510 // us requestin a CEF repaint.
511 bool is_cef_capture, may_have_dropped_last_frame = false;
513 // If this card is used for output (ie., output_card_index points to it),
514 // it cannot simultaneously be uesd for capture, so <capture> gets replaced
515 // by a FakeCapture. However, since reconstructing the real capture object
516 // with all its state can be annoying, it is not being deleted, just stopped
518 std::unique_ptr<bmusb::CaptureInterface> parked_capture;
520 std::unique_ptr<PBOFrameAllocator> frame_allocator;
522 // Stuff for the OpenGL context (for texture uploading).
523 QSurface *surface = nullptr;
526 RefCountedFrame frame;
527 int64_t length; // In TIMEBASE units.
529 unsigned field; // Which field (0 or 1) of the frame to use. Always 0 for progressive.
530 std::function<void()> upload_func; // Needs to be called to actually upload the texture to OpenGL.
531 unsigned dropped_frames = 0; // Number of dropped frames before this one.
532 std::chrono::steady_clock::time_point received_timestamp = std::chrono::steady_clock::time_point::min();
534 std::deque<NewFrame> new_frames;
535 bool should_quit = false;
536 std::condition_variable new_frames_changed; // Set whenever new_frames (or should_quit) is changed.
538 QueueLengthPolicy queue_length_policy; // Refers to the "new_frames" queue.
540 int last_timecode = -1; // Unwrapped.
542 JitterHistory jitter_history;
545 std::vector<std::pair<std::string, std::string>> labels;
546 std::atomic<int64_t> metric_input_received_frames{0};
547 std::atomic<int64_t> metric_input_duped_frames{0};
548 std::atomic<int64_t> metric_input_dropped_frames_jitter{0};
549 std::atomic<int64_t> metric_input_dropped_frames_error{0};
550 std::atomic<int64_t> metric_input_resets{0};
551 std::atomic<int64_t> metric_input_queue_length_frames{0};
553 std::atomic<int64_t> metric_input_has_signal_bool{-1};
554 std::atomic<int64_t> metric_input_is_connected_bool{-1};
555 std::atomic<int64_t> metric_input_interlaced_bool{-1};
556 std::atomic<int64_t> metric_input_width_pixels{-1};
557 std::atomic<int64_t> metric_input_height_pixels{-1};
558 std::atomic<int64_t> metric_input_frame_rate_nom{-1};
559 std::atomic<int64_t> metric_input_frame_rate_den{-1};
560 std::atomic<int64_t> metric_input_sample_rate_hz{-1};
562 JitterHistory output_jitter_history;
563 CaptureCard cards[MAX_VIDEO_CARDS]; // Protected by <card_mutex>.
564 YCbCrInterpretation ycbcr_interpretation[MAX_VIDEO_CARDS]; // Protected by <card_mutex>.
565 std::unique_ptr<AudioMixer> audio_mixer; // Same as global_audio_mixer (see audio_mixer.h).
566 bool input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const;
567 struct OutputFrameInfo {
568 int dropped_frames; // Since last frame.
569 int num_samples; // Audio samples needed for this output frame.
570 int64_t frame_duration; // In TIMEBASE units.
572 std::chrono::steady_clock::time_point frame_timestamp;
574 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]);
576 InputState input_state;
578 // Cards we have been noticed about being hotplugged, but haven't tried adding yet.
579 // Protected by its own mutex.
580 std::mutex hotplug_mutex;
581 std::vector<libusb_device *> hotplugged_cards;
583 class OutputChannel {
586 void output_frame(DisplayFrame &&frame);
587 bool get_display_frame(DisplayFrame *frame);
588 void add_frame_ready_callback(void *key, new_frame_ready_callback_t callback);
589 void remove_frame_ready_callback(void *key);
590 void set_transition_names_updated_callback(transition_names_updated_callback_t callback);
591 void set_name_updated_callback(name_updated_callback_t callback);
592 void set_color_updated_callback(color_updated_callback_t callback);
598 Mixer *parent = nullptr; // Not owned.
599 std::mutex frame_mutex;
600 DisplayFrame current_frame, ready_frame; // protected by <frame_mutex>
601 bool has_current_frame = false, has_ready_frame = false; // protected by <frame_mutex>
602 std::map<void *, new_frame_ready_callback_t> new_frame_ready_callbacks; // protected by <frame_mutex>
603 transition_names_updated_callback_t transition_names_updated_callback;
604 name_updated_callback_t name_updated_callback;
605 color_updated_callback_t color_updated_callback;
607 std::vector<std::string> last_transition_names;
608 std::string last_name, last_color;
610 OutputChannel output_channel[NUM_OUTPUTS];
612 std::thread mixer_thread;
613 std::thread audio_thread;
614 std::atomic<bool> should_quit{false};
615 std::atomic<bool> should_cut{false};
617 std::unique_ptr<ALSAOutput> alsa;
623 std::chrono::steady_clock::time_point frame_timestamp;
625 std::mutex audio_mutex;
626 std::condition_variable audio_task_queue_changed;
627 std::queue<AudioTask> audio_task_queue; // Under audio_mutex.
629 // For mode scanning.
630 bool is_mode_scanning[MAX_VIDEO_CARDS]{ false };
631 std::vector<uint32_t> mode_scanlist[MAX_VIDEO_CARDS];
632 unsigned mode_scanlist_index[MAX_VIDEO_CARDS]{ 0 };
633 std::chrono::steady_clock::time_point last_mode_scan_change[MAX_VIDEO_CARDS];
636 extern Mixer *global_mixer;
638 #endif // !defined(_MIXER_H)