X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.h;h=289ce45c9b5542da424942b4205ec825605d5848;hb=e284d1c7a2e18ee7e4aea082c4a57a3504a0f5e8;hp=58084817c70c9c78abef8ff9801ac71d9bd30778;hpb=138c12c13443d39f8b1168b5d54b36d9835c6405;p=nageru diff --git a/mixer.h b/mixer.h index 5808481..289ce45 100644 --- a/mixer.h +++ b/mixer.h @@ -23,6 +23,8 @@ #include #include +#include + #include "audio_mixer.h" #include "bmusb/bmusb.h" #include "defs.h" @@ -35,20 +37,76 @@ #include "theme.h" #include "timebase.h" #include "video_encoder.h" +#include "ycbcr_interpretation.h" class ALSAOutput; class ChromaSubsampler; class DeckLinkOutput; class QSurface; class QSurfaceFormat; +class TimecodeRenderer; +class v210Converter; namespace movit { class Effect; class EffectChain; -class FlatInput; class ResourcePool; +class YCbCrInput; } // namespace movit +// A class to estimate the future jitter. Used in QueueLengthPolicy (see below). +// +// There are many ways to estimate jitter; I've tested a few ones (and also +// some algorithms that don't explicitly model jitter) with different +// parameters on some real-life data in experiments/queue_drop_policy.cpp. +// This is one based on simple order statistics where I've added some margin in +// the number of starvation events; I believe that about one every hour would +// probably be acceptable, but this one typically goes lower than that, at the +// cost of 2–3 ms extra latency. (If the queue is hard-limited to one frame, it's +// possible to get ~10 ms further down, but this would mean framedrops every +// second or so.) The general strategy is: Take the 99.9-percentile jitter over +// last 5000 frames, multiply by two, and that's our worst-case jitter +// estimate. The fact that we're not using the max value means that we could +// actually even throw away very late frames immediately, which means we only +// get one user-visible event instead of seeing something both when the frame +// arrives late (duplicate frame) and then again when we drop. +class JitterHistory { +private: + static constexpr size_t history_length = 5000; + static constexpr double percentile = 0.999; + static constexpr double multiplier = 2.0; + +public: + void register_metrics(const std::vector> &labels); + void unregister_metrics(const std::vector> &labels); + + void clear() { + history.clear(); + orders.clear(); + } + void frame_arrived(std::chrono::steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames); + std::chrono::steady_clock::time_point get_expected_next_frame() const { return expected_timestamp; } + double estimate_max_jitter() const; + +private: + // A simple O(k) based algorithm for getting the k-th largest or + // smallest element from our window; we simply keep the multiset + // ordered (insertions and deletions are O(n) as always) and then + // iterate from one of the sides. If we had larger values of k, + // we could go for a more complicated setup with two sets or heaps + // (one increasing and one decreasing) that we keep balanced around + // the point, or it is possible to reimplement std::set with + // counts in each node. However, since k=5, we don't need this. + std::multiset orders; + std::deque::iterator> history; + + std::chrono::steady_clock::time_point expected_timestamp = std::chrono::steady_clock::time_point::min(); + + // Metrics. There are no direct summaries for jitter, since we already have latency summaries. + std::atomic metric_input_underestimated_jitter_frames{0}; + std::atomic metric_input_estimated_max_jitter_seconds{0.0 / 0.0}; +}; + // For any card that's not the master (where we pick out the frames as they // come, as fast as we can process), there's going to be a queue. The question // is when we should drop frames from that queue (apart from the obvious @@ -60,38 +118,39 @@ class ResourcePool; // 2. We don't want to add more delay than is needed. // // Our general strategy is to drop as many frames as we can (helping for #2) -// that we think is safe for #1 given jitter. To this end, we set a lower floor N, -// where we assume that if we have N frames in the queue, we're always safe from -// starvation. (Typically, N will be 0 or 1. It starts off at 0.) If we have -// more than N frames in the queue after reading out the one we need, we head-drop -// them to reduce the queue. -// -// N is reduced as follows: If the queue has had at least one spare frame for -// at least 50 (master) frames (ie., it's been too conservative for a second), -// we reduce N by 1 and reset the timers. TODO: Only do this if N ever actually -// touched the limit. +// that we think is safe for #1 given jitter. To this end, we measure the +// deviation from the expected arrival time for all cards, and use that for +// continuous jitter estimation. // -// Whenever the queue is starved (we needed a frame but there was none), -// and we've been at N since the last starvation, N was obviously too low, -// so we increment it. We will never set N above 5, though. +// We then drop everything from the queue that we're sure we won't need to +// serve the output in the time before the next frame arrives. Typically, +// this means the queue will contain 0 or 1 frames, although more is also +// possible if the jitter is very high. class QueueLengthPolicy { public: QueueLengthPolicy() {} void reset(unsigned card_index) { this->card_index = card_index; - safe_queue_length = 0; - frames_with_at_least_one = 0; - been_at_safe_point_since_last_starvation = false; } - void update_policy(int queue_length); // Give in -1 for starvation. + void register_metrics(const std::vector> &labels); + void unregister_metrics(const std::vector> &labels); + + // Call after picking out a frame, so 0 means starvation. + void update_policy(std::chrono::steady_clock::time_point now, + std::chrono::steady_clock::time_point expected_next_frame, + int64_t input_frame_duration, + int64_t master_frame_duration, + double max_input_card_jitter_seconds, + double max_master_card_jitter_seconds); unsigned get_safe_queue_length() const { return safe_queue_length; } private: - unsigned card_index; // For debugging only. - unsigned safe_queue_length = 0; // Called N in the comments. - unsigned frames_with_at_least_one = 0; - bool been_at_safe_point_since_last_starvation = false; + unsigned card_index; // For debugging and metrics only. + unsigned safe_queue_length = 0; // Can never go below zero. + + // Metrics. + std::atomic metric_input_queue_safe_length_frames{1}; }; class Mixer { @@ -138,10 +197,17 @@ public: return output_channel[output].get_display_frame(frame); } + // NOTE: Callbacks will be called with a mutex held, so you should probably + // not do real work in them. typedef std::function new_frame_ready_callback_t; - void set_frame_ready_callback(Output output, new_frame_ready_callback_t callback) + void add_frame_ready_callback(Output output, void *key, new_frame_ready_callback_t callback) + { + output_channel[output].add_frame_ready_callback(key, callback); + } + + void remove_frame_ready_callback(Output output, void *key) { - output_channel[output].set_frame_ready_callback(callback); + output_channel[output].remove_frame_ready_callback(key); } // TODO: Should this really be per-channel? Shouldn't it just be called for e.g. the live output? @@ -208,6 +274,9 @@ public: return theme->set_signal_mapping(signal, card); } + YCbCrInterpretation get_input_ycbcr_interpretation(unsigned card_index) const; + void set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation); + bool get_supports_set_wb(unsigned channel) const { return theme->get_supports_set_wb(channel); @@ -234,6 +303,25 @@ public: return cards[card_index].capture->get_description(); } + // The difference between this and the previous function is that if a card + // is used as the current output, get_card_description() will return the + // fake card that's replacing it for input, whereas this function will return + // the card's actual name. + std::string get_output_card_description(unsigned card_index) const { + assert(card_can_be_used_as_output(card_index)); + assert(card_index < num_cards); + if (cards[card_index].parked_capture) { + return cards[card_index].parked_capture->get_description(); + } else { + return cards[card_index].capture->get_description(); + } + } + + bool card_can_be_used_as_output(unsigned card_index) const { + assert(card_index < num_cards); + return cards[card_index].output != nullptr; + } + std::map get_available_video_modes(unsigned card_index) const { assert(card_index < num_cards); return cards[card_index].capture->get_available_video_modes(); @@ -285,9 +373,56 @@ public: video_encoder->change_x264_bitrate(rate_kbit); } + int get_output_card_index() const { // -1 = no output, just stream. + return desired_output_card_index; + } + + void set_output_card(int card_index) { // -1 = no output, just stream. + desired_output_card_index = card_index; + } + + std::map get_available_output_video_modes() const; + + uint32_t get_output_video_mode() const { + return desired_output_video_mode; + } + + void set_output_video_mode(uint32_t mode) { + desired_output_video_mode = mode; + } + + void set_display_timecode_in_stream(bool enable) { + display_timecode_in_stream = enable; + } + + void set_display_timecode_on_stdout(bool enable) { + display_timecode_on_stdout = enable; + } + + int64_t get_num_connected_clients() const { + return httpd.get_num_connected_clients(); + } + + std::vector get_theme_menu() { return theme->get_theme_menu(); } + + void theme_menu_entry_clicked(int lua_ref) { return theme->theme_menu_entry_clicked(lua_ref); } + + void set_theme_menu_callback(std::function callback) + { + theme->set_theme_menu_callback(callback); + } + private: - void configure_card(unsigned card_index, bmusb::CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output); - void set_output_card(int card_index); // -1 = no output, just stream. + struct CaptureCard; + + enum class CardType { + LIVE_CARD, + FAKE_CAPTURE, + FFMPEG_INPUT, + CEF_INPUT, + }; + void configure_card(unsigned card_index, bmusb::CaptureInterface *capture, CardType card_type, DeckLinkOutput *output); + void set_output_card_internal(int card_index); // Should only be called from the mixer thread. void bm_frame(unsigned card_index, uint16_t timecode, bmusb::FrameAllocator::Frame video_frame, size_t video_offset, bmusb::VideoFormat video_format, bmusb::FrameAllocator::Frame audio_frame, size_t audio_offset, bmusb::AudioFormat audio_format); @@ -296,47 +431,77 @@ private: void place_rectangle(movit::Effect *resample_effect, movit::Effect *padding_effect, float x0, float y0, float x1, float y1); void thread_func(); void handle_hotplugged_cards(); - void schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame); + 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); + std::string get_timecode_text() const; void render_one_frame(int64_t duration); void audio_thread_func(); void release_display_frame(DisplayFrame *frame); double pts() { return double(pts_int) / TIMEBASE; } + void trim_queue(CaptureCard *card, size_t safe_queue_length); + std::pair get_channels_json(); + std::pair get_channel_color_http(unsigned channel_idx); HTTPD httpd; - unsigned num_cards; + unsigned num_cards, num_video_inputs, num_html_inputs; QSurface *mixer_surface, *h264_encoder_surface, *decklink_output_surface; std::unique_ptr resource_pool; std::unique_ptr theme; std::atomic audio_source_channel{0}; std::atomic master_clock_channel{0}; // Gets overridden by if set. - std::atomic output_card_index{-1}; // -1 for none. + int output_card_index = -1; // -1 for none. + uint32_t output_video_mode = -1; + + // The mechanics of changing the output card and modes are so intricately connected + // with the work the mixer thread is doing. Thus, we don't change it directly, + // we just set this variable instead, which signals to the mixer thread that + // it should do the change before the next frame. This simplifies locking + // considerations immensely. + std::atomic desired_output_card_index{-1}; + std::atomic desired_output_video_mode{0}; + std::unique_ptr display_chain; std::unique_ptr chroma_subsampler; + std::unique_ptr v210_converter; std::unique_ptr video_encoder; + std::unique_ptr timecode_renderer; + std::atomic display_timecode_in_stream{false}; + std::atomic display_timecode_on_stdout{false}; + // Effects part of . Owned by . - movit::FlatInput *display_input; + movit::YCbCrInput *display_input; int64_t pts_int = 0; // In TIMEBASE units. + unsigned frame_num = 0; // Accumulated errors in number of 1/TIMEBASE audio samples. If OUTPUT_FREQUENCY divided by // frame rate is integer, will always stay zero. unsigned fractional_samples = 0; - std::mutex card_mutex; + mutable std::mutex card_mutex; bool has_bmusb_thread = false; struct CaptureCard { - bmusb::CaptureInterface *capture = nullptr; + std::unique_ptr capture; bool is_fake_capture; - DeckLinkOutput *output = nullptr; + CardType type; + std::unique_ptr output; + + // CEF only delivers frames when it actually has a change. + // If we trim the queue for latency reasons, we could thus + // end up in a situation trimming a frame that was meant to + // be displayed for a long time, which is really suboptimal. + // Thus, if we drop the last frame we have, may_have_dropped_last_frame + // is set to true, and the next starvation event will trigger + // us requestin a CEF repaint. + bool is_cef_capture, may_have_dropped_last_frame = false; // If this card is used for output (ie., output_card_index points to it), // it cannot simultaneously be uesd for capture, so gets replaced // by a FakeCapture. However, since reconstructing the real capture object // with all its state can be annoying, it is not being deleted, just stopped // and moved here. - bmusb::CaptureInterface *parked_capture = nullptr; + std::unique_ptr parked_capture; std::unique_ptr frame_allocator; @@ -352,21 +517,45 @@ private: unsigned dropped_frames = 0; // Number of dropped frames before this one. std::chrono::steady_clock::time_point received_timestamp = std::chrono::steady_clock::time_point::min(); }; - std::queue new_frames; + std::deque new_frames; bool should_quit = false; std::condition_variable new_frames_changed; // Set whenever new_frames (or should_quit) is changed. QueueLengthPolicy queue_length_policy; // Refers to the "new_frames" queue. int last_timecode = -1; // Unwrapped. + + JitterHistory jitter_history; + + // Metrics. + std::vector> labels; + std::atomic metric_input_received_frames{0}; + std::atomic metric_input_duped_frames{0}; + std::atomic metric_input_dropped_frames_jitter{0}; + std::atomic metric_input_dropped_frames_error{0}; + std::atomic metric_input_resets{0}; + std::atomic metric_input_queue_length_frames{0}; + + std::atomic metric_input_has_signal_bool{-1}; + std::atomic metric_input_is_connected_bool{-1}; + std::atomic metric_input_interlaced_bool{-1}; + std::atomic metric_input_width_pixels{-1}; + std::atomic metric_input_height_pixels{-1}; + std::atomic metric_input_frame_rate_nom{-1}; + std::atomic metric_input_frame_rate_den{-1}; + std::atomic metric_input_sample_rate_hz{-1}; }; + JitterHistory output_jitter_history; CaptureCard cards[MAX_VIDEO_CARDS]; // Protected by . + YCbCrInterpretation ycbcr_interpretation[MAX_VIDEO_CARDS]; // Protected by . AudioMixer audio_mixer; // Same as global_audio_mixer (see audio_mixer.h). bool input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const; struct OutputFrameInfo { int dropped_frames; // Since last frame. int num_samples; // Audio samples needed for this output frame. int64_t frame_duration; // In TIMEBASE units. + bool is_preroll; + std::chrono::steady_clock::time_point frame_timestamp; }; 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]); @@ -380,9 +569,10 @@ private: class OutputChannel { public: ~OutputChannel(); - void output_frame(DisplayFrame frame); + void output_frame(DisplayFrame &&frame); bool get_display_frame(DisplayFrame *frame); - void set_frame_ready_callback(new_frame_ready_callback_t callback); + void add_frame_ready_callback(void *key, new_frame_ready_callback_t callback); + void remove_frame_ready_callback(void *key); void set_transition_names_updated_callback(transition_names_updated_callback_t callback); void set_name_updated_callback(name_updated_callback_t callback); void set_color_updated_callback(color_updated_callback_t callback); @@ -395,7 +585,7 @@ private: std::mutex frame_mutex; DisplayFrame current_frame, ready_frame; // protected by bool has_current_frame = false, has_ready_frame = false; // protected by - new_frame_ready_callback_t new_frame_ready_callback; + std::map new_frame_ready_callbacks; // protected by transition_names_updated_callback_t transition_names_updated_callback; name_updated_callback_t name_updated_callback; color_updated_callback_t color_updated_callback; @@ -416,6 +606,7 @@ private: int64_t pts_int; int num_samples; bool adjust_rate; + std::chrono::steady_clock::time_point frame_timestamp; }; std::mutex audio_mutex; std::condition_variable audio_task_queue_changed; @@ -429,6 +620,5 @@ private: }; extern Mixer *global_mixer; -extern bool uses_mlock; #endif // !defined(_MIXER_H)