+void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_CARDS], bool has_new_frame[MAX_CARDS], int num_samples[MAX_CARDS])
+{
+start:
+ // The first card is the master timer, so wait for it to have a new frame.
+ // TODO: Add a timeout.
+ unique_lock<mutex> lock(bmusb_mutex);
+ 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(); });
+
+ if (cards[master_card_index].new_frames.empty()) {
+ // We were woken up, but not due to a new frame. Deal with it
+ // and then restart.
+ assert(cards[master_card_index].capture->get_disconnected());
+ handle_hotplugged_cards();
+ goto start;
+ }
+
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
+ CaptureCard *card = &cards[card_index];
+ if (card->new_frames.empty()) {
+ assert(card_index != master_card_index);
+ card->queue_length_policy.update_policy(-1);
+ continue;
+ }
+ new_frames[card_index] = move(card->new_frames.front());
+ has_new_frame[card_index] = true;
+ card->new_frames.pop();
+ card->new_frames_changed.notify_all();
+
+ int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
+ num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
+ card->fractional_samples = num_samples_times_timebase % TIMEBASE;
+ assert(num_samples[card_index] >= 0);
+
+ if (card_index == master_card_index) {
+ // We don't use the queue length policy for the master card,
+ // but we will if it stops being the master. Thus, clear out
+ // the policy in case we switch in the future.
+ card->queue_length_policy.reset(card_index);
+ } else {
+ // If we have excess frames compared to the policy for this card,
+ // drop frames from the head.
+ card->queue_length_policy.update_policy(card->new_frames.size());
+ while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
+ card->new_frames.pop();
+ }
+ }
+ }
+}
+
+void Mixer::handle_hotplugged_cards()
+{
+ // Check for cards that have been disconnected since last frame.
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
+ CaptureCard *card = &cards[card_index];
+ if (card->capture->get_disconnected()) {
+ fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
+ FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
+ configure_card(card_index, capture, /*is_fake_capture=*/true);
+ card->queue_length_policy.reset(card_index);
+ card->capture->start_bm_capture();
+ }
+ }
+
+ // Check for cards that have been connected since last frame.
+ vector<libusb_device *> hotplugged_cards_copy;
+ {
+ lock_guard<mutex> lock(hotplug_mutex);
+ swap(hotplugged_cards, hotplugged_cards_copy);
+ }
+ for (libusb_device *new_dev : hotplugged_cards_copy) {
+ // Look for a fake capture card where we can stick this in.
+ int free_card_index = -1;
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
+ if (cards[card_index].is_fake_capture) {
+ free_card_index = int(card_index);
+ break;
+ }
+ }
+
+ if (free_card_index == -1) {
+ fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
+ libusb_unref_device(new_dev);
+ } else {
+ // BMUSBCapture takes ownership.
+ fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
+ CaptureCard *card = &cards[free_card_index];
+ BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
+ configure_card(free_card_index, capture, /*is_fake_capture=*/false);
+ card->queue_length_policy.reset(free_card_index);
+ capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
+ capture->start_bm_capture();
+ }
+ }
+}
+
+
+void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
+{
+ // Resample the audio as needed, including from previously dropped frames.
+ assert(num_cards > 0);
+ for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
+ {
+ // Signal to the audio thread to process this frame.
+ unique_lock<mutex> lock(audio_mutex);
+ audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
+ audio_task_queue_changed.notify_one();
+ }
+ if (frame_num != dropped_frames) {
+ // For dropped frames, increase the pts. Note that if the format changed
+ // in the meantime, we have no way of detecting that; we just have to
+ // assume the frame length is always the same.
+ pts_int += length_per_frame;
+ }
+ }
+}
+
+void Mixer::render_one_frame(int64_t duration)
+{
+ // Get the main chain from the theme, and set its state immediately.
+ Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
+ EffectChain *chain = theme_main_chain.chain;
+ theme_main_chain.setup_chain();
+ //theme_main_chain.chain->enable_phase_timing(true);
+
+ GLuint y_tex, cbcr_tex;
+ bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
+ assert(got_frame);
+
+ // Render main chain.
+ GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
+ GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
+ GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
+ check_error();
+ chain->render_to_fbo(fbo, WIDTH, HEIGHT);
+ resource_pool->release_fbo(fbo);
+
+ subsample_chroma(cbcr_full_tex, cbcr_tex);
+ resource_pool->release_2d_texture(cbcr_full_tex);
+
+ // Set the right state for rgba_tex.
+ glBindFramebuffer(GL_FRAMEBUFFER, 0);
+ glBindTexture(GL_TEXTURE_2D, rgba_tex);
+ glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+ glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+ glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
+ const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
+ RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
+
+ // The live frame just shows the RGBA texture we just rendered.
+ // It owns rgba_tex now.
+ DisplayFrame live_frame;
+ live_frame.chain = display_chain.get();
+ live_frame.setup_chain = [this, rgba_tex]{
+ display_input->set_texture_num(rgba_tex);
+ };
+ live_frame.ready_fence = fence;
+ live_frame.input_frames = {};
+ live_frame.temp_textures = { rgba_tex };
+ output_channel[OUTPUT_LIVE].output_frame(live_frame);
+
+ // Set up preview and any additional channels.
+ for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
+ DisplayFrame display_frame;
+ Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
+ display_frame.chain = chain.chain;
+ display_frame.setup_chain = chain.setup_chain;
+ display_frame.ready_fence = fence;
+ display_frame.input_frames = chain.input_frames;
+ display_frame.temp_textures = {};
+ output_channel[i].output_frame(display_frame);
+ }
+}
+
+void Mixer::send_audio_level_callback()
+{
+ if (audio_level_callback == nullptr) {
+ return;
+ }
+
+ unique_lock<mutex> lock(compressor_mutex);
+ double loudness_s = r128.loudness_S();
+ double loudness_i = r128.integrated();
+ double loudness_range_low = r128.range_min();
+ double loudness_range_high = r128.range_max();
+
+ audio_level_callback(loudness_s, 20.0 * log10(peak),
+ loudness_i, loudness_range_low, loudness_range_high,
+ gain_staging_db, 20.0 * log10(final_makeup_gain),
+ correlation.get_correlation());
+}
+
+void Mixer::audio_thread_func()
+{
+ while (!should_quit) {
+ AudioTask task;
+
+ {
+ unique_lock<mutex> lock(audio_mutex);
+ audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
+ if (should_quit) {
+ return;
+ }
+ task = audio_task_queue.front();
+ audio_task_queue.pop();
+ }
+
+ process_audio_one_frame(task.pts_int, task.num_samples);
+ }
+}
+
+void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
+{
+ vector<float> samples_card;
+ vector<float> samples_out;
+
+ // TODO: Allow mixing audio from several sources.
+ unsigned selected_audio_card = theme->map_signal(audio_source_channel);
+ assert(selected_audio_card < num_cards);
+
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
+ samples_card.resize(num_samples * 2);
+ {
+ unique_lock<mutex> lock(cards[card_index].audio_mutex);
+ cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples);
+ }
+ if (card_index == selected_audio_card) {
+ samples_out = move(samples_card);
+ }
+ }
+
+ // Cut away everything under 120 Hz (or whatever the cutoff is);
+ // we don't need it for voice, and it will reduce headroom
+ // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
+ // should be dampened.)
+ if (locut_enabled) {
+ locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
+ }
+
+ // Apply a level compressor to get the general level right.
+ // Basically, if it's over about -40 dBFS, we squeeze it down to that level
+ // (or more precisely, near it, since we don't use infinite ratio),
+ // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
+ // entirely arbitrary, but from practical tests with speech, it seems to
+ // put ut around -23 LUFS, so it's a reasonable starting point for later use.
+ {
+ unique_lock<mutex> lock(compressor_mutex);
+ if (level_compressor_enabled) {
+ float threshold = 0.01f; // -40 dBFS.
+ float ratio = 20.0f;
+ float attack_time = 0.5f;
+ float release_time = 20.0f;
+ float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
+ level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+ gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
+ } else {
+ // Just apply the gain we already had.
+ float g = pow(10.0f, gain_staging_db / 20.0f);
+ for (size_t i = 0; i < samples_out.size(); ++i) {
+ samples_out[i] *= g;
+ }
+ }
+ }
+
+#if 0
+ printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
+ level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
+ level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
+ 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
+#endif
+
+// float limiter_att, compressor_att;
+
+ // The real compressor.
+ if (compressor_enabled) {
+ float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
+ float ratio = 20.0f;
+ float attack_time = 0.005f;
+ float release_time = 0.040f;
+ float makeup_gain = 2.0f; // +6 dB.
+ compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+// compressor_att = compressor.get_attenuation();
+ }
+
+ // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
+ // Note that since ratio is not infinite, we could go slightly higher than this.
+ if (limiter_enabled) {
+ float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
+ float ratio = 30.0f;
+ float attack_time = 0.0f; // Instant.
+ float release_time = 0.020f;
+ float makeup_gain = 1.0f; // 0 dB.
+ limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+// limiter_att = limiter.get_attenuation();
+ }
+
+// printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
+
+ // At this point, we are most likely close to +0 LU, but all of our
+ // measurements have been on raw sample values, not R128 values.
+ // So we have a final makeup gain to get us to +0 LU; the gain
+ // adjustments required should be relatively small, and also, the
+ // offset shouldn't change much (only if the type of audio changes
+ // significantly). Thus, we shoot for updating this value basically
+ // “whenever we process buffers”, since the R128 calculation isn't exactly
+ // something we get out per-sample.
+ //
+ // Note that there's a feedback loop here, so we choose a very slow filter
+ // (half-time of 100 seconds).
+ double target_loudness_factor, alpha;
+ {
+ unique_lock<mutex> lock(compressor_mutex);
+ double loudness_lu = r128.loudness_M() - ref_level_lufs;
+ double current_makeup_lu = 20.0f * log10(final_makeup_gain);
+ target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
+
+ // If we're outside +/- 5 LU uncorrected, we don't count it as
+ // a normal signal (probably silence) and don't change the
+ // correction factor; just apply what we already have.
+ if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
+ alpha = 0.0;
+ } else {
+ // Formula adapted from
+ // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
+ const double half_time_s = 100.0;
+ const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
+ alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
+ }
+
+ double m = final_makeup_gain;
+ for (size_t i = 0; i < samples_out.size(); i += 2) {
+ samples_out[i + 0] *= m;
+ samples_out[i + 1] *= m;
+ m += (target_loudness_factor - m) * alpha;
+ }
+ final_makeup_gain = m;
+ }
+
+ // Upsample 4x to find interpolated peak.
+ peak_resampler.inp_data = samples_out.data();
+ peak_resampler.inp_count = samples_out.size() / 2;
+
+ vector<float> interpolated_samples_out;
+ interpolated_samples_out.resize(samples_out.size());
+ while (peak_resampler.inp_count > 0) { // About four iterations.
+ peak_resampler.out_data = &interpolated_samples_out[0];
+ peak_resampler.out_count = interpolated_samples_out.size() / 2;
+ peak_resampler.process();
+ size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
+ peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
+ peak_resampler.out_data = nullptr;
+ }
+
+ // Find R128 levels and L/R correlation.
+ vector<float> left, right;
+ deinterleave_samples(samples_out, &left, &right);
+ float *ptrs[] = { left.data(), right.data() };
+ {
+ unique_lock<mutex> lock(compressor_mutex);
+ r128.process(left.size(), ptrs);
+ correlation.process_samples(samples_out);
+ }
+
+ // Send the samples to the sound card.
+ if (alsa) {
+ alsa->write(samples_out);
+ }
+
+ // And finally add them to the output.
+ video_encoder->add_audio(frame_pts_int, move(samples_out));
+}
+