8 #include <movit/effect_chain.h>
9 #include <movit/effect_util.h>
10 #include <movit/flat_input.h>
11 #include <movit/image_format.h>
12 #include <movit/resource_pool.h>
13 #include <movit/util.h>
21 #include <condition_variable>
30 #include "bmusb/bmusb.h"
33 #include "h264encode.h"
34 #include "pbo_frame_allocator.h"
35 #include "ref_counted_gl_sync.h"
40 using namespace movit;
42 using namespace std::placeholders;
44 Mixer *global_mixer = nullptr;
48 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
50 for (size_t i = 0; i < num_samples; ++i) {
51 for (size_t j = 0; j < out_channels; ++j) {
55 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
56 dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
58 src += 3 * (in_channels - out_channels);
62 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
65 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
66 input_state->buffered_frames[card_index][frame_num] =
67 input_state->buffered_frames[card_index][frame_num - 1];
69 input_state->buffered_frames[card_index][0] = { frame, field_num };
71 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
72 input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
77 string generate_local_dump_filename(int frame)
79 time_t now = time(NULL);
81 localtime_r(&now, &now_tm);
84 strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm);
86 // Use the frame number to disambiguate between two cuts starting
87 // on the same second.
89 snprintf(filename, sizeof(filename), "%s%s-f%02d%s",
90 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
96 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
97 : httpd(WIDTH, HEIGHT),
99 mixer_surface(create_surface(format)),
100 h264_encoder_surface(create_surface(format)),
101 level_compressor(OUTPUT_FREQUENCY),
102 limiter(OUTPUT_FREQUENCY),
103 compressor(OUTPUT_FREQUENCY)
105 httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
108 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
111 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
112 // will be halved when sampling them, and we need to compensate here.
113 movit_texel_subpixel_precision /= 2.0;
115 resource_pool.reset(new ResourcePool);
116 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
117 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
118 output_channel[i].parent = this;
121 ImageFormat inout_format;
122 inout_format.color_space = COLORSPACE_sRGB;
123 inout_format.gamma_curve = GAMMA_sRGB;
125 // Display chain; shows the live output produced by the main chain (its RGBA version).
126 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
128 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
129 display_chain->add_input(display_input);
130 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
131 display_chain->set_dither_bits(0); // Don't bother.
132 display_chain->finalize();
134 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
136 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
137 printf("Configuring card %d...\n", card_index);
138 CaptureCard *card = &cards[card_index];
139 card->usb = new BMUSBCapture(card_index);
140 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
141 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
142 card->usb->set_video_frame_allocator(card->frame_allocator.get());
143 card->surface = create_surface(format);
144 card->usb->set_dequeue_thread_callbacks(
146 eglBindAPI(EGL_OPENGL_API);
147 card->context = create_context(card->surface);
148 if (!make_current(card->context, card->surface)) {
149 printf("failed to create bmusb context\n");
154 resource_pool->clean_context();
156 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
157 card->usb->configure_card();
160 BMUSBCapture::start_bm_thread();
162 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
163 cards[card_index].usb->start_bm_capture();
166 // Set up stuff for NV12 conversion.
169 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
170 string cbcr_frag_shader =
173 "uniform sampler2D cbcr_tex; \n"
175 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
177 vector<string> frag_shader_outputs;
178 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
180 r128.init(2, OUTPUT_FREQUENCY);
183 locut.init(FILTER_HPF, 2);
185 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
186 // and there's a limit to how important the peak meter is.
187 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
189 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
194 resource_pool->release_glsl_program(cbcr_program_num);
195 BMUSBCapture::stop_bm_thread();
197 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
199 unique_lock<mutex> lock(bmusb_mutex);
200 cards[card_index].should_quit = true; // Unblock thread.
201 cards[card_index].new_data_ready_changed.notify_all();
203 cards[card_index].usb->stop_dequeue_thread();
206 h264_encoder.reset(nullptr);
211 int unwrap_timecode(uint16_t current_wrapped, int last)
213 uint16_t last_wrapped = last & 0xffff;
214 if (current_wrapped > last_wrapped) {
215 return (last & ~0xffff) | current_wrapped;
217 return 0x10000 + ((last & ~0xffff) | current_wrapped);
221 float find_peak(const float *samples, size_t num_samples)
223 float m = fabs(samples[0]);
224 for (size_t i = 1; i < num_samples; ++i) {
225 m = max(m, fabs(samples[i]));
230 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
232 size_t num_samples = in.size() / 2;
233 out_l->resize(num_samples);
234 out_r->resize(num_samples);
236 const float *inptr = in.data();
237 float *lptr = &(*out_l)[0];
238 float *rptr = &(*out_r)[0];
239 for (size_t i = 0; i < num_samples; ++i) {
247 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
248 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
249 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
251 CaptureCard *card = &cards[card_index];
253 unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
256 decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
257 &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now.
258 int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
260 size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
261 if (num_samples > OUTPUT_FREQUENCY / 10) {
262 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",
263 card_index, int(audio_frame.len), int(audio_offset),
264 timecode, int(video_frame.len), int(video_offset), video_format);
265 if (video_frame.owner) {
266 video_frame.owner->release_frame(video_frame);
268 if (audio_frame.owner) {
269 audio_frame.owner->release_frame(audio_frame);
274 int64_t local_pts = card->next_local_pts;
275 int dropped_frames = 0;
276 if (card->last_timecode != -1) {
277 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
280 // Convert the audio to stereo fp32 and add it.
282 audio.resize(num_samples * 2);
283 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
287 unique_lock<mutex> lock(card->audio_mutex);
289 // Number of samples per frame if we need to insert silence.
290 // (Could be nonintegral, but resampling will save us then.)
291 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
293 if (dropped_frames > MAX_FPS * 2) {
294 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
295 card_index, card->last_timecode, timecode);
296 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
298 } else if (dropped_frames > 0) {
299 // Insert silence as needed.
300 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
301 card_index, dropped_frames, timecode);
302 vector<float> silence(silence_samples * 2, 0.0f);
303 for (int i = 0; i < dropped_frames; ++i) {
304 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
305 // Note that if the format changed in the meantime, we have
306 // no way of detecting that; we just have to assume the frame length
307 // is always the same.
308 local_pts += frame_length;
311 if (num_samples == 0) {
312 audio.resize(silence_samples * 2);
313 num_samples = silence_samples;
315 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
316 card->next_local_pts = local_pts + frame_length;
319 card->last_timecode = timecode;
321 // Done with the audio, so release it.
322 if (audio_frame.owner) {
323 audio_frame.owner->release_frame(audio_frame);
327 // Wait until the previous frame was consumed.
328 unique_lock<mutex> lock(bmusb_mutex);
329 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
330 if (card->should_quit) return;
333 size_t expected_length = width * (height + extra_lines_top + extra_lines_bottom) * 2;
334 if (video_frame.len - video_offset == 0 ||
335 video_frame.len - video_offset != expected_length) {
336 if (video_frame.len != 0) {
337 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
338 card_index, video_frame.len - video_offset, expected_length);
340 if (video_frame.owner) {
341 video_frame.owner->release_frame(video_frame);
344 // Still send on the information that we _had_ a frame, even though it's corrupted,
345 // so that pts can go up accordingly.
347 unique_lock<mutex> lock(bmusb_mutex);
348 card->new_data_ready = true;
349 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
350 card->new_frame_length = frame_length;
351 card->new_frame_interlaced = false;
352 card->new_data_ready_fence = nullptr;
353 card->dropped_frames = dropped_frames;
354 card->new_data_ready_changed.notify_all();
359 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
361 unsigned num_fields = interlaced ? 2 : 1;
362 timespec frame_upload_start;
364 // Send the two fields along as separate frames; the other side will need to add
365 // a deinterlacer to actually get this right.
366 assert(height % 2 == 0);
368 assert(frame_length % 2 == 0);
371 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
373 userdata->last_interlaced = interlaced;
374 userdata->last_frame_rate_nom = frame_rate_nom;
375 userdata->last_frame_rate_den = frame_rate_den;
376 RefCountedFrame new_frame(video_frame);
378 // Upload the textures.
379 size_t cbcr_width = width / 2;
380 size_t cbcr_offset = video_offset / 2;
381 size_t y_offset = video_frame.size / 2 + video_offset / 2;
383 for (unsigned field = 0; field < num_fields; ++field) {
384 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
386 if (userdata->tex_y[field] == 0 ||
387 userdata->tex_cbcr[field] == 0 ||
388 width != userdata->last_width[field] ||
389 height != userdata->last_height[field]) {
390 // We changed resolution since last use of this texture, so we need to create
391 // a new object. Note that this each card has its own PBOFrameAllocator,
392 // we don't need to worry about these flip-flopping between resolutions.
393 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
395 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
397 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
399 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
401 userdata->last_width[field] = width;
402 userdata->last_height[field] = height;
405 GLuint pbo = userdata->pbo;
407 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
409 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
412 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
414 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
416 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
418 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
420 glBindTexture(GL_TEXTURE_2D, 0);
422 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
424 assert(fence != nullptr);
427 // Don't upload the second field as fast as we can; wait until
428 // the field time has approximately passed. (Otherwise, we could
429 // get timing jitter against the other sources, and possibly also
430 // against the video display, although the latter is not as critical.)
431 // This requires our system clock to be reasonably close to the
432 // video clock, but that's not an unreasonable assumption.
433 timespec second_field_start;
434 second_field_start.tv_nsec = frame_upload_start.tv_nsec +
435 frame_length * 1000000000 / TIMEBASE;
436 second_field_start.tv_sec = frame_upload_start.tv_sec +
437 second_field_start.tv_nsec / 1000000000;
438 second_field_start.tv_nsec %= 1000000000;
440 while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
441 &second_field_start, nullptr) == -1 &&
446 unique_lock<mutex> lock(bmusb_mutex);
447 card->new_data_ready = true;
448 card->new_frame = new_frame;
449 card->new_frame_length = frame_length;
450 card->new_frame_field = field;
451 card->new_frame_interlaced = interlaced;
452 card->new_data_ready_fence = fence;
453 card->dropped_frames = dropped_frames;
454 card->new_data_ready_changed.notify_all();
456 if (field != num_fields - 1) {
457 // Wait until the previous frame was consumed.
458 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
459 if (card->should_quit) return;
465 void Mixer::thread_func()
467 eglBindAPI(EGL_OPENGL_API);
468 QOpenGLContext *context = create_context(mixer_surface);
469 if (!make_current(context, mixer_surface)) {
474 struct timespec start, now;
475 clock_gettime(CLOCK_MONOTONIC, &start);
478 int stats_dropped_frames = 0;
480 while (!should_quit) {
481 CaptureCard card_copy[MAX_CARDS];
482 int num_samples[MAX_CARDS];
485 unique_lock<mutex> lock(bmusb_mutex);
487 // The first card is the master timer, so wait for it to have a new frame.
488 // TODO: Make configurable, and with a timeout.
489 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
491 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
492 CaptureCard *card = &cards[card_index];
493 card_copy[card_index].usb = card->usb;
494 card_copy[card_index].new_data_ready = card->new_data_ready;
495 card_copy[card_index].new_frame = card->new_frame;
496 card_copy[card_index].new_frame_length = card->new_frame_length;
497 card_copy[card_index].new_frame_field = card->new_frame_field;
498 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
499 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
500 card_copy[card_index].dropped_frames = card->dropped_frames;
501 card->new_data_ready = false;
502 card->new_data_ready_changed.notify_all();
504 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
505 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
506 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
507 assert(num_samples[card_index] >= 0);
511 // Resample the audio as needed, including from previously dropped frames.
512 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
514 // Signal to the audio thread to process this frame.
515 unique_lock<mutex> lock(audio_mutex);
516 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
517 audio_task_queue_changed.notify_one();
519 if (frame_num != card_copy[0].dropped_frames) {
520 // For dropped frames, increase the pts. Note that if the format changed
521 // in the meantime, we have no way of detecting that; we just have to
522 // assume the frame length is always the same.
523 ++stats_dropped_frames;
524 pts_int += card_copy[0].new_frame_length;
528 if (audio_level_callback != nullptr) {
529 unique_lock<mutex> lock(compressor_mutex);
530 double loudness_s = r128.loudness_S();
531 double loudness_i = r128.integrated();
532 double loudness_range_low = r128.range_min();
533 double loudness_range_high = r128.range_max();
535 audio_level_callback(loudness_s, 20.0 * log10(peak),
536 loudness_i, loudness_range_low, loudness_range_high,
537 gain_staging_db, 20.0 * log10(final_makeup_gain));
540 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
541 if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
542 ++card_copy[card_index].dropped_frames;
544 if (card_copy[card_index].dropped_frames > 0) {
545 printf("Card %u dropped %d frames before this\n",
546 card_index, int(card_copy[card_index].dropped_frames));
550 // If the first card is reporting a corrupted or otherwise dropped frame,
551 // just increase the pts (skipping over this frame) and don't try to compute anything new.
552 if (card_copy[0].new_frame->len == 0) {
553 ++stats_dropped_frames;
554 pts_int += card_copy[0].new_frame_length;
558 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
559 CaptureCard *card = &card_copy[card_index];
560 if (!card->new_data_ready || card->new_frame->len == 0)
563 assert(card->new_frame != nullptr);
564 insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
567 // The new texture might still be uploaded,
568 // tell the GPU to wait until it's there.
569 if (card->new_data_ready_fence) {
570 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
572 glDeleteSync(card->new_data_ready_fence);
577 // Get the main chain from the theme, and set its state immediately.
578 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
579 EffectChain *chain = theme_main_chain.chain;
580 theme_main_chain.setup_chain();
581 //theme_main_chain.chain->enable_phase_timing(true);
583 GLuint y_tex, cbcr_tex;
584 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
587 // Render main chain.
588 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
589 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
590 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
592 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
593 resource_pool->release_fbo(fbo);
595 subsample_chroma(cbcr_full_tex, cbcr_tex);
596 resource_pool->release_2d_texture(cbcr_full_tex);
598 // Set the right state for rgba_tex.
599 glBindFramebuffer(GL_FRAMEBUFFER, 0);
600 glBindTexture(GL_TEXTURE_2D, rgba_tex);
601 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
602 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
603 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
605 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
608 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
609 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
611 pts_int += card_copy[0].new_frame_length;
613 // The live frame just shows the RGBA texture we just rendered.
614 // It owns rgba_tex now.
615 DisplayFrame live_frame;
616 live_frame.chain = display_chain.get();
617 live_frame.setup_chain = [this, rgba_tex]{
618 display_input->set_texture_num(rgba_tex);
620 live_frame.ready_fence = fence;
621 live_frame.input_frames = {};
622 live_frame.temp_textures = { rgba_tex };
623 output_channel[OUTPUT_LIVE].output_frame(live_frame);
625 // Set up preview and any additional channels.
626 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
627 DisplayFrame display_frame;
628 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
629 display_frame.chain = chain.chain;
630 display_frame.setup_chain = chain.setup_chain;
631 display_frame.ready_fence = fence;
632 display_frame.input_frames = chain.input_frames;
633 display_frame.temp_textures = {};
634 output_channel[i].output_frame(display_frame);
637 clock_gettime(CLOCK_MONOTONIC, &now);
638 double elapsed = now.tv_sec - start.tv_sec +
639 1e-9 * (now.tv_nsec - start.tv_nsec);
640 if (frame % 100 == 0) {
641 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
642 frame, stats_dropped_frames, elapsed, frame / elapsed,
643 1e3 * elapsed / frame);
644 // chain->print_phase_timing();
647 if (should_cut.exchange(false)) { // Test and clear.
648 string filename = generate_local_dump_filename(frame);
649 printf("Starting new recording: %s\n", filename.c_str());
650 h264_encoder->shutdown();
651 httpd.close_output_file();
652 httpd.open_output_file(filename.c_str());
653 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
657 // Reset every 100 frames, so that local variations in frame times
658 // (especially for the first few frames, when the shaders are
659 // compiled etc.) don't make it hard to measure for the entire
660 // remaining duration of the program.
661 if (frame == 10000) {
669 resource_pool->clean_context();
672 void Mixer::audio_thread_func()
674 while (!should_quit) {
678 unique_lock<mutex> lock(audio_mutex);
679 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
680 task = audio_task_queue.front();
681 audio_task_queue.pop();
684 process_audio_one_frame(task.pts_int, task.num_samples);
688 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
690 vector<float> samples_card;
691 vector<float> samples_out;
692 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
693 samples_card.resize(num_samples * 2);
695 unique_lock<mutex> lock(cards[card_index].audio_mutex);
696 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
697 printf("Card %d reported previous underrun.\n", card_index);
700 // TODO: Allow using audio from the other card(s) as well.
701 if (card_index == 0) {
702 samples_out = move(samples_card);
706 // Cut away everything under 120 Hz (or whatever the cutoff is);
707 // we don't need it for voice, and it will reduce headroom
708 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
709 // should be dampened.)
710 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
712 // Apply a level compressor to get the general level right.
713 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
714 // (or more precisely, near it, since we don't use infinite ratio),
715 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
716 // entirely arbitrary, but from practical tests with speech, it seems to
717 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
719 unique_lock<mutex> lock(compressor_mutex);
720 if (level_compressor_enabled) {
721 float threshold = 0.01f; // -40 dBFS.
723 float attack_time = 0.5f;
724 float release_time = 20.0f;
725 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
726 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
727 gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
729 // Just apply the gain we already had.
730 float g = pow(10.0f, gain_staging_db / 20.0f);
731 for (size_t i = 0; i < samples_out.size(); ++i) {
738 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
739 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
740 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
741 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
744 // float limiter_att, compressor_att;
746 // The real compressor.
747 if (compressor_enabled) {
748 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
750 float attack_time = 0.005f;
751 float release_time = 0.040f;
752 float makeup_gain = 2.0f; // +6 dB.
753 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
754 // compressor_att = compressor.get_attenuation();
757 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
758 // Note that since ratio is not infinite, we could go slightly higher than this.
759 if (limiter_enabled) {
760 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
762 float attack_time = 0.0f; // Instant.
763 float release_time = 0.020f;
764 float makeup_gain = 1.0f; // 0 dB.
765 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
766 // limiter_att = limiter.get_attenuation();
769 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
771 // Upsample 4x to find interpolated peak.
772 peak_resampler.inp_data = samples_out.data();
773 peak_resampler.inp_count = samples_out.size() / 2;
775 vector<float> interpolated_samples_out;
776 interpolated_samples_out.resize(samples_out.size());
777 while (peak_resampler.inp_count > 0) { // About four iterations.
778 peak_resampler.out_data = &interpolated_samples_out[0];
779 peak_resampler.out_count = interpolated_samples_out.size() / 2;
780 peak_resampler.process();
781 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
782 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
785 // At this point, we are most likely close to +0 LU, but all of our
786 // measurements have been on raw sample values, not R128 values.
787 // So we have a final makeup gain to get us to +0 LU; the gain
788 // adjustments required should be relatively small, and also, the
789 // offset shouldn't change much (only if the type of audio changes
790 // significantly). Thus, we shoot for updating this value basically
791 // “whenever we process buffers”, since the R128 calculation isn't exactly
792 // something we get out per-sample.
794 // Note that there's a feedback loop here, so we choose a very slow filter
795 // (half-time of 100 seconds).
796 double target_loudness_factor, alpha;
798 unique_lock<mutex> lock(compressor_mutex);
799 double loudness_lu = r128.loudness_M() - ref_level_lufs;
800 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
801 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
803 // If we're outside +/- 5 LU uncorrected, we don't count it as
804 // a normal signal (probably silence) and don't change the
805 // correction factor; just apply what we already have.
806 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
809 // Formula adapted from
810 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
811 const double half_time_s = 100.0;
812 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
813 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
816 double m = final_makeup_gain;
817 for (size_t i = 0; i < samples_out.size(); i += 2) {
818 samples_out[i + 0] *= m;
819 samples_out[i + 1] *= m;
820 m += (target_loudness_factor - m) * alpha;
822 final_makeup_gain = m;
826 vector<float> left, right;
827 deinterleave_samples(samples_out, &left, &right);
828 float *ptrs[] = { left.data(), right.data() };
830 unique_lock<mutex> lock(compressor_mutex);
831 r128.process(left.size(), ptrs);
834 // Send the samples to the sound card.
836 alsa->write(samples_out);
839 // And finally add them to the output.
840 h264_encoder->add_audio(frame_pts_int, move(samples_out));
843 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
846 glGenVertexArrays(1, &vao);
855 glBindVertexArray(vao);
859 GLuint fbo = resource_pool->create_fbo(dst_tex);
860 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
861 glViewport(0, 0, WIDTH/2, HEIGHT/2);
864 glUseProgram(cbcr_program_num);
867 glActiveTexture(GL_TEXTURE0);
869 glBindTexture(GL_TEXTURE_2D, src_tex);
871 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
873 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
875 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
878 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
879 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
881 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
882 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
884 glDrawArrays(GL_TRIANGLES, 0, 3);
887 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
888 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
893 resource_pool->release_fbo(fbo);
894 glDeleteVertexArrays(1, &vao);
897 void Mixer::release_display_frame(DisplayFrame *frame)
899 for (GLuint texnum : frame->temp_textures) {
900 resource_pool->release_2d_texture(texnum);
902 frame->temp_textures.clear();
903 frame->ready_fence.reset();
904 frame->input_frames.clear();
909 mixer_thread = thread(&Mixer::thread_func, this);
910 audio_thread = thread(&Mixer::audio_thread_func, this);
920 void Mixer::transition_clicked(int transition_num)
922 theme->transition_clicked(transition_num, pts());
925 void Mixer::channel_clicked(int preview_num)
927 theme->channel_clicked(preview_num);
930 void Mixer::reset_meters()
932 peak_resampler.reset();
938 Mixer::OutputChannel::~OutputChannel()
940 if (has_current_frame) {
941 parent->release_display_frame(¤t_frame);
943 if (has_ready_frame) {
944 parent->release_display_frame(&ready_frame);
948 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
950 // Store this frame for display. Remove the ready frame if any
951 // (it was seemingly never used).
953 unique_lock<mutex> lock(frame_mutex);
954 if (has_ready_frame) {
955 parent->release_display_frame(&ready_frame);
958 has_ready_frame = true;
961 if (has_new_frame_ready_callback) {
962 new_frame_ready_callback();
966 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
968 unique_lock<mutex> lock(frame_mutex);
969 if (!has_current_frame && !has_ready_frame) {
973 if (has_current_frame && has_ready_frame) {
974 // We have a new ready frame. Toss the current one.
975 parent->release_display_frame(¤t_frame);
976 has_current_frame = false;
978 if (has_ready_frame) {
979 assert(!has_current_frame);
980 current_frame = ready_frame;
981 ready_frame.ready_fence.reset(); // Drop the refcount.
982 ready_frame.input_frames.clear(); // Drop the refcounts.
983 has_current_frame = true;
984 has_ready_frame = false;
987 *frame = current_frame;
991 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
993 new_frame_ready_callback = callback;
994 has_new_frame_ready_callback = true;