10 #include <movit/effect_chain.h>
11 #include <movit/effect_util.h>
12 #include <movit/flat_input.h>
13 #include <movit/image_format.h>
14 #include <movit/resource_pool.h>
23 #include <condition_variable>
32 #include "bmusb/bmusb.h"
35 #include "h264encode.h"
36 #include "pbo_frame_allocator.h"
37 #include "ref_counted_gl_sync.h"
42 using namespace movit;
44 using namespace std::placeholders;
46 Mixer *global_mixer = nullptr;
50 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
52 for (size_t i = 0; i < num_samples; ++i) {
53 for (size_t j = 0; j < out_channels; ++j) {
57 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
58 dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
60 src += 3 * (in_channels - out_channels);
66 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
67 : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
69 mixer_surface(create_surface(format)),
70 h264_encoder_surface(create_surface(format)),
71 level_compressor(OUTPUT_FREQUENCY),
72 limiter(OUTPUT_FREQUENCY),
73 compressor(OUTPUT_FREQUENCY)
77 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
80 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
81 // will be halved when sampling them, and we need to compensate here.
82 movit_texel_subpixel_precision /= 2.0;
84 resource_pool.reset(new ResourcePool);
85 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
86 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
87 output_channel[i].parent = this;
90 ImageFormat inout_format;
91 inout_format.color_space = COLORSPACE_sRGB;
92 inout_format.gamma_curve = GAMMA_sRGB;
94 // Display chain; shows the live output produced by the main chain (its RGBA version).
95 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
97 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
98 display_chain->add_input(display_input);
99 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
100 display_chain->set_dither_bits(0); // Don't bother.
101 display_chain->finalize();
103 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
105 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
106 printf("Configuring card %d...\n", card_index);
107 CaptureCard *card = &cards[card_index];
108 card->usb = new BMUSBCapture(card_index);
109 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
110 card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44 + 1, WIDTH, HEIGHT));
111 card->usb->set_video_frame_allocator(card->frame_allocator.get());
112 card->surface = create_surface(format);
113 card->usb->set_dequeue_thread_callbacks(
115 eglBindAPI(EGL_OPENGL_API);
116 card->context = create_context(card->surface);
117 if (!make_current(card->context, card->surface)) {
118 printf("failed to create bmusb context\n");
123 resource_pool->clean_context();
125 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
126 card->usb->configure_card();
129 BMUSBCapture::start_bm_thread();
131 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
132 cards[card_index].usb->start_bm_capture();
135 //chain->enable_phase_timing(true);
137 // Set up stuff for NV12 conversion.
140 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
141 string cbcr_frag_shader =
144 "uniform sampler2D cbcr_tex; \n"
146 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
148 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
150 r128.init(2, OUTPUT_FREQUENCY);
153 locut.init(FILTER_HPF, 2);
155 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
156 // and there's a limit to how important the peak meter is.
157 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
159 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
164 resource_pool->release_glsl_program(cbcr_program_num);
165 BMUSBCapture::stop_bm_thread();
167 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
169 unique_lock<mutex> lock(bmusb_mutex);
170 cards[card_index].should_quit = true; // Unblock thread.
171 cards[card_index].new_data_ready_changed.notify_all();
173 cards[card_index].usb->stop_dequeue_thread();
176 h264_encoder.reset(nullptr);
181 int unwrap_timecode(uint16_t current_wrapped, int last)
183 uint16_t last_wrapped = last & 0xffff;
184 if (current_wrapped > last_wrapped) {
185 return (last & ~0xffff) | current_wrapped;
187 return 0x10000 + ((last & ~0xffff) | current_wrapped);
191 float find_peak(const float *samples, size_t num_samples)
193 float m = fabs(samples[0]);
194 for (size_t i = 1; i < num_samples; ++i) {
195 m = std::max(m, fabs(samples[i]));
200 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
202 size_t num_samples = in.size() / 2;
203 out_l->resize(num_samples);
204 out_r->resize(num_samples);
206 const float *inptr = in.data();
207 float *lptr = &(*out_l)[0];
208 float *rptr = &(*out_r)[0];
209 for (size_t i = 0; i < num_samples; ++i) {
217 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
218 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
219 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
221 CaptureCard *card = &cards[card_index];
223 int width, height, frame_rate_nom, frame_rate_den;
226 decode_video_format(video_format, &width, &height, &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now.
227 int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
229 size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
230 if (num_samples > OUTPUT_FREQUENCY / 10) {
231 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",
232 card_index, int(audio_frame.len), int(audio_offset),
233 timecode, int(video_frame.len), int(video_offset), video_format);
234 if (video_frame.owner) {
235 video_frame.owner->release_frame(video_frame);
237 if (audio_frame.owner) {
238 audio_frame.owner->release_frame(audio_frame);
243 int64_t local_pts = card->next_local_pts;
244 int dropped_frames = 0;
245 if (card->last_timecode != -1) {
246 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
249 // Convert the audio to stereo fp32 and add it.
251 audio.resize(num_samples * 2);
252 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
256 unique_lock<mutex> lock(card->audio_mutex);
258 // Number of samples per frame if we need to insert silence.
259 // (Could be nonintegral, but resampling will save us then.)
260 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
262 if (dropped_frames > MAX_FPS * 2) {
263 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
264 card_index, card->last_timecode, timecode);
265 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
267 } else if (dropped_frames > 0) {
268 // Insert silence as needed.
269 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
270 card_index, dropped_frames, timecode);
271 vector<float> silence;
272 silence.resize(silence_samples * 2);
273 for (int i = 0; i < dropped_frames; ++i) {
274 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
275 // Note that if the format changed in the meantime, we have
276 // no way of detecting that; we just have to assume the frame length
277 // is always the same.
278 local_pts += frame_length;
281 if (num_samples == 0) {
282 audio.resize(silence_samples * 2);
283 num_samples = silence_samples;
285 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
286 card->next_local_pts = local_pts + frame_length;
289 card->last_timecode = timecode;
291 // Done with the audio, so release it.
292 if (audio_frame.owner) {
293 audio_frame.owner->release_frame(audio_frame);
297 // Wait until the previous frame was consumed.
298 unique_lock<mutex> lock(bmusb_mutex);
299 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
300 if (card->should_quit) return;
303 if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
304 if (video_frame.len != 0) {
305 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
306 card_index, video_frame.len - video_offset);
308 if (video_frame.owner) {
309 video_frame.owner->release_frame(video_frame);
312 // Still send on the information that we _had_ a frame, even though it's corrupted,
313 // so that pts can go up accordingly.
315 unique_lock<mutex> lock(bmusb_mutex);
316 card->new_data_ready = true;
317 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
318 card->new_frame_length = frame_length;
319 card->new_data_ready_fence = nullptr;
320 card->dropped_frames = dropped_frames;
321 card->new_data_ready_changed.notify_all();
326 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
327 GLuint pbo = userdata->pbo;
329 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
331 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
333 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
336 // Upload the textures.
337 size_t skipped_lines = 25;
338 size_t cbcr_width = WIDTH / 2;
339 size_t cbcr_offset = video_offset / 2;
340 size_t y_offset = cbcr_offset + cbcr_width * (HEIGHT + EXTRAHEIGHT) * sizeof(uint16_t) + video_offset / 2;
342 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
344 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * skipped_lines * sizeof(uint16_t)));
346 glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
348 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH, HEIGHT, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + WIDTH * skipped_lines));
350 glBindTexture(GL_TEXTURE_2D, 0);
352 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
354 assert(fence != nullptr);
357 unique_lock<mutex> lock(bmusb_mutex);
358 card->new_data_ready = true;
359 card->new_frame = RefCountedFrame(video_frame);
360 card->new_frame_length = frame_length;
361 card->new_data_ready_fence = fence;
362 card->dropped_frames = dropped_frames;
363 card->new_data_ready_changed.notify_all();
367 void Mixer::thread_func()
369 eglBindAPI(EGL_OPENGL_API);
370 QOpenGLContext *context = create_context(mixer_surface);
371 if (!make_current(context, mixer_surface)) {
376 struct timespec start, now;
377 clock_gettime(CLOCK_MONOTONIC, &start);
380 int stats_dropped_frames = 0;
382 while (!should_quit) {
383 CaptureCard card_copy[MAX_CARDS];
384 int num_samples[MAX_CARDS];
387 unique_lock<mutex> lock(bmusb_mutex);
389 // The first card is the master timer, so wait for it to have a new frame.
390 // TODO: Make configurable, and with a timeout.
391 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
393 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
394 CaptureCard *card = &cards[card_index];
395 card_copy[card_index].usb = card->usb;
396 card_copy[card_index].new_data_ready = card->new_data_ready;
397 card_copy[card_index].new_frame = card->new_frame;
398 card_copy[card_index].new_frame_length = card->new_frame_length;
399 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
400 card_copy[card_index].dropped_frames = card->dropped_frames;
401 card->new_data_ready = false;
402 card->new_data_ready_changed.notify_all();
404 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
405 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
406 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
407 assert(num_samples[card_index] >= 0);
411 // Resample the audio as needed, including from previously dropped frames.
412 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
414 // Signal to the audio thread to process this frame.
415 unique_lock<mutex> lock(audio_mutex);
416 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
417 audio_task_queue_changed.notify_one();
419 if (frame_num != card_copy[0].dropped_frames) {
420 // For dropped frames, increase the pts. Note that if the format changed
421 // in the meantime, we have no way of detecting that; we just have to
422 // assume the frame length is always the same.
423 ++stats_dropped_frames;
424 pts_int += card_copy[0].new_frame_length;
428 if (audio_level_callback != nullptr) {
429 double loudness_s = r128.loudness_S();
430 double loudness_i = r128.integrated();
431 double loudness_range_low = r128.range_min();
432 double loudness_range_high = r128.range_max();
434 audio_level_callback(loudness_s, 20.0 * log10(peak),
435 loudness_i, loudness_range_low, loudness_range_high,
436 last_gain_staging_db);
439 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
440 if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
441 ++card_copy[card_index].dropped_frames;
443 if (card_copy[card_index].dropped_frames > 0) {
444 printf("Card %u dropped %d frames before this\n",
445 card_index, int(card_copy[card_index].dropped_frames));
449 // If the first card is reporting a corrupted or otherwise dropped frame,
450 // just increase the pts (skipping over this frame) and don't try to compute anything new.
451 if (card_copy[0].new_frame->len == 0) {
452 ++stats_dropped_frames;
453 pts_int += card_copy[0].new_frame_length;
457 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
458 CaptureCard *card = &card_copy[card_index];
459 if (!card->new_data_ready || card->new_frame->len == 0)
462 assert(card->new_frame != nullptr);
463 bmusb_current_rendering_frame[card_index] = card->new_frame;
466 // The new texture might still be uploaded,
467 // tell the GPU to wait until it's there.
468 if (card->new_data_ready_fence) {
469 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
471 glDeleteSync(card->new_data_ready_fence);
474 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
475 theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
478 // Get the main chain from the theme, and set its state immediately.
479 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
480 EffectChain *chain = theme_main_chain.first;
481 theme_main_chain.second();
483 GLuint y_tex, cbcr_tex;
484 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
487 // Render main chain.
488 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
489 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
490 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
492 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
493 resource_pool->release_fbo(fbo);
495 subsample_chroma(cbcr_full_tex, cbcr_tex);
496 resource_pool->release_2d_texture(cbcr_full_tex);
498 // Set the right state for rgba_tex.
499 glBindFramebuffer(GL_FRAMEBUFFER, 0);
500 glBindTexture(GL_TEXTURE_2D, rgba_tex);
501 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
502 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
503 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
505 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
508 // Make sure the H.264 gets a reference to all the
509 // input frames needed, so that they are not released back
510 // until the rendering is done.
511 vector<RefCountedFrame> input_frames;
512 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
513 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
515 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
516 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
518 pts_int += card_copy[0].new_frame_length;
520 // The live frame just shows the RGBA texture we just rendered.
521 // It owns rgba_tex now.
522 DisplayFrame live_frame;
523 live_frame.chain = display_chain.get();
524 live_frame.setup_chain = [this, rgba_tex]{
525 display_input->set_texture_num(rgba_tex);
527 live_frame.ready_fence = fence;
528 live_frame.input_frames = {};
529 live_frame.temp_textures = { rgba_tex };
530 output_channel[OUTPUT_LIVE].output_frame(live_frame);
532 // Set up preview and any additional channels.
533 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
534 DisplayFrame display_frame;
535 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
536 display_frame.chain = chain.first;
537 display_frame.setup_chain = chain.second;
538 display_frame.ready_fence = fence;
540 // FIXME: possible to do better?
541 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
542 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
544 display_frame.temp_textures = {};
545 output_channel[i].output_frame(display_frame);
548 clock_gettime(CLOCK_MONOTONIC, &now);
549 double elapsed = now.tv_sec - start.tv_sec +
550 1e-9 * (now.tv_nsec - start.tv_nsec);
551 if (frame % 100 == 0) {
552 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
553 frame, stats_dropped_frames, elapsed, frame / elapsed,
554 1e3 * elapsed / frame);
555 // chain->print_phase_timing();
559 // Reset every 100 frames, so that local variations in frame times
560 // (especially for the first few frames, when the shaders are
561 // compiled etc.) don't make it hard to measure for the entire
562 // remaining duration of the program.
563 if (frame == 10000) {
571 resource_pool->clean_context();
574 void Mixer::audio_thread_func()
576 while (!should_quit) {
580 unique_lock<mutex> lock(audio_mutex);
581 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
582 task = audio_task_queue.front();
583 audio_task_queue.pop();
586 process_audio_one_frame(task.pts_int, task.num_samples);
590 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
592 vector<float> samples_card;
593 vector<float> samples_out;
594 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
595 samples_card.resize(num_samples * 2);
597 unique_lock<mutex> lock(cards[card_index].audio_mutex);
598 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
599 printf("Card %d reported previous underrun.\n", card_index);
602 // TODO: Allow using audio from the other card(s) as well.
603 if (card_index == 0) {
604 samples_out = move(samples_card);
608 // Cut away everything under 120 Hz (or whatever the cutoff is);
609 // we don't need it for voice, and it will reduce headroom
610 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
611 // should be dampened.)
612 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
614 // Apply a level compressor to get the general level right.
615 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
616 // (or more precisely, near it, since we don't use infinite ratio),
617 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
618 // entirely arbitrary, but from practical tests with speech, it seems to
619 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
620 float ref_level_dbfs = -14.0f;
622 float threshold = 0.01f; // -40 dBFS.
624 float attack_time = 0.5f;
625 float release_time = 20.0f;
626 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
627 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
628 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
632 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
633 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
634 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
635 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
638 // float limiter_att, compressor_att;
640 // The real compressor.
641 if (compressor_enabled) {
642 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
644 float attack_time = 0.005f;
645 float release_time = 0.040f;
646 float makeup_gain = 2.0f; // +6 dB.
647 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
648 // compressor_att = compressor.get_attenuation();
651 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
652 // Note that since ratio is not infinite, we could go slightly higher than this.
653 if (limiter_enabled) {
654 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
656 float attack_time = 0.0f; // Instant.
657 float release_time = 0.020f;
658 float makeup_gain = 1.0f; // 0 dB.
659 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
660 // limiter_att = limiter.get_attenuation();
663 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
665 // Upsample 4x to find interpolated peak.
666 peak_resampler.inp_data = samples_out.data();
667 peak_resampler.inp_count = samples_out.size() / 2;
669 vector<float> interpolated_samples_out;
670 interpolated_samples_out.resize(samples_out.size());
671 while (peak_resampler.inp_count > 0) { // About four iterations.
672 peak_resampler.out_data = &interpolated_samples_out[0];
673 peak_resampler.out_count = interpolated_samples_out.size() / 2;
674 peak_resampler.process();
675 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
676 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
680 vector<float> left, right;
681 deinterleave_samples(samples_out, &left, &right);
682 float *ptrs[] = { left.data(), right.data() };
683 r128.process(left.size(), ptrs);
685 // Send the samples to the sound card.
687 alsa->write(samples_out);
690 // And finally add them to the output.
691 h264_encoder->add_audio(frame_pts_int, move(samples_out));
694 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
697 glGenVertexArrays(1, &vao);
706 glBindVertexArray(vao);
710 GLuint fbo = resource_pool->create_fbo(dst_tex);
711 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
712 glViewport(0, 0, WIDTH/2, HEIGHT/2);
715 glUseProgram(cbcr_program_num);
718 glActiveTexture(GL_TEXTURE0);
720 glBindTexture(GL_TEXTURE_2D, src_tex);
722 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
724 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
726 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
729 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
730 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
732 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
733 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
735 glDrawArrays(GL_TRIANGLES, 0, 3);
738 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
739 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
744 resource_pool->release_fbo(fbo);
745 glDeleteVertexArrays(1, &vao);
748 void Mixer::release_display_frame(DisplayFrame *frame)
750 for (GLuint texnum : frame->temp_textures) {
751 resource_pool->release_2d_texture(texnum);
753 frame->temp_textures.clear();
754 frame->ready_fence.reset();
755 frame->input_frames.clear();
760 mixer_thread = thread(&Mixer::thread_func, this);
761 audio_thread = thread(&Mixer::audio_thread_func, this);
771 void Mixer::transition_clicked(int transition_num)
773 theme->transition_clicked(transition_num, pts());
776 void Mixer::channel_clicked(int preview_num)
778 theme->channel_clicked(preview_num);
781 void Mixer::reset_meters()
783 peak_resampler.reset();
789 Mixer::OutputChannel::~OutputChannel()
791 if (has_current_frame) {
792 parent->release_display_frame(¤t_frame);
794 if (has_ready_frame) {
795 parent->release_display_frame(&ready_frame);
799 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
801 // Store this frame for display. Remove the ready frame if any
802 // (it was seemingly never used).
804 unique_lock<mutex> lock(frame_mutex);
805 if (has_ready_frame) {
806 parent->release_display_frame(&ready_frame);
809 has_ready_frame = true;
812 if (has_new_frame_ready_callback) {
813 new_frame_ready_callback();
817 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
819 unique_lock<mutex> lock(frame_mutex);
820 if (!has_current_frame && !has_ready_frame) {
824 if (has_current_frame && has_ready_frame) {
825 // We have a new ready frame. Toss the current one.
826 parent->release_display_frame(¤t_frame);
827 has_current_frame = false;
829 if (has_ready_frame) {
830 assert(!has_current_frame);
831 current_frame = ready_frame;
832 ready_frame.ready_fence.reset(); // Drop the refcount.
833 ready_frame.input_frames.clear(); // Drop the refcounts.
834 has_current_frame = true;
835 has_ready_frame = false;
838 *frame = current_frame;
842 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
844 new_frame_ready_callback = callback;
845 has_new_frame_ready_callback = true;