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 };
80 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
81 : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
83 mixer_surface(create_surface(format)),
84 h264_encoder_surface(create_surface(format)),
85 level_compressor(OUTPUT_FREQUENCY),
86 limiter(OUTPUT_FREQUENCY),
87 compressor(OUTPUT_FREQUENCY)
91 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
94 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
95 // will be halved when sampling them, and we need to compensate here.
96 movit_texel_subpixel_precision /= 2.0;
98 resource_pool.reset(new ResourcePool);
99 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
100 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
101 output_channel[i].parent = this;
104 ImageFormat inout_format;
105 inout_format.color_space = COLORSPACE_sRGB;
106 inout_format.gamma_curve = GAMMA_sRGB;
108 // Display chain; shows the live output produced by the main chain (its RGBA version).
109 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
111 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
112 display_chain->add_input(display_input);
113 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
114 display_chain->set_dither_bits(0); // Don't bother.
115 display_chain->finalize();
117 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
119 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
120 printf("Configuring card %d...\n", card_index);
121 CaptureCard *card = &cards[card_index];
122 card->usb = new BMUSBCapture(card_index);
123 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
124 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
125 card->usb->set_video_frame_allocator(card->frame_allocator.get());
126 card->surface = create_surface(format);
127 card->usb->set_dequeue_thread_callbacks(
129 eglBindAPI(EGL_OPENGL_API);
130 card->context = create_context(card->surface);
131 if (!make_current(card->context, card->surface)) {
132 printf("failed to create bmusb context\n");
137 resource_pool->clean_context();
139 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
140 card->usb->configure_card();
143 BMUSBCapture::start_bm_thread();
145 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
146 cards[card_index].usb->start_bm_capture();
149 //chain->enable_phase_timing(true);
151 // Set up stuff for NV12 conversion.
154 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
155 string cbcr_frag_shader =
158 "uniform sampler2D cbcr_tex; \n"
160 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
162 vector<string> frag_shader_outputs;
163 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
165 r128.init(2, OUTPUT_FREQUENCY);
168 locut.init(FILTER_HPF, 2);
170 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
171 // and there's a limit to how important the peak meter is.
172 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
174 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
179 resource_pool->release_glsl_program(cbcr_program_num);
180 BMUSBCapture::stop_bm_thread();
182 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
184 unique_lock<mutex> lock(bmusb_mutex);
185 cards[card_index].should_quit = true; // Unblock thread.
186 cards[card_index].new_data_ready_changed.notify_all();
188 cards[card_index].usb->stop_dequeue_thread();
191 h264_encoder.reset(nullptr);
196 int unwrap_timecode(uint16_t current_wrapped, int last)
198 uint16_t last_wrapped = last & 0xffff;
199 if (current_wrapped > last_wrapped) {
200 return (last & ~0xffff) | current_wrapped;
202 return 0x10000 + ((last & ~0xffff) | current_wrapped);
206 float find_peak(const float *samples, size_t num_samples)
208 float m = fabs(samples[0]);
209 for (size_t i = 1; i < num_samples; ++i) {
210 m = std::max(m, fabs(samples[i]));
215 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
217 size_t num_samples = in.size() / 2;
218 out_l->resize(num_samples);
219 out_r->resize(num_samples);
221 const float *inptr = in.data();
222 float *lptr = &(*out_l)[0];
223 float *rptr = &(*out_r)[0];
224 for (size_t i = 0; i < num_samples; ++i) {
232 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
233 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
234 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
236 CaptureCard *card = &cards[card_index];
238 unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
241 decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
242 &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now.
243 int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
245 size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
246 if (num_samples > OUTPUT_FREQUENCY / 10) {
247 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",
248 card_index, int(audio_frame.len), int(audio_offset),
249 timecode, int(video_frame.len), int(video_offset), video_format);
250 if (video_frame.owner) {
251 video_frame.owner->release_frame(video_frame);
253 if (audio_frame.owner) {
254 audio_frame.owner->release_frame(audio_frame);
259 int64_t local_pts = card->next_local_pts;
260 int dropped_frames = 0;
261 if (card->last_timecode != -1) {
262 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
265 // Convert the audio to stereo fp32 and add it.
267 audio.resize(num_samples * 2);
268 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
272 unique_lock<mutex> lock(card->audio_mutex);
274 // Number of samples per frame if we need to insert silence.
275 // (Could be nonintegral, but resampling will save us then.)
276 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
278 if (dropped_frames > MAX_FPS * 2) {
279 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
280 card_index, card->last_timecode, timecode);
281 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
283 } else if (dropped_frames > 0) {
284 // Insert silence as needed.
285 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
286 card_index, dropped_frames, timecode);
287 vector<float> silence(silence_samples * 2, 0.0f);
288 for (int i = 0; i < dropped_frames; ++i) {
289 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
290 // Note that if the format changed in the meantime, we have
291 // no way of detecting that; we just have to assume the frame length
292 // is always the same.
293 local_pts += frame_length;
296 if (num_samples == 0) {
297 audio.resize(silence_samples * 2);
298 num_samples = silence_samples;
300 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
301 card->next_local_pts = local_pts + frame_length;
304 card->last_timecode = timecode;
306 // Done with the audio, so release it.
307 if (audio_frame.owner) {
308 audio_frame.owner->release_frame(audio_frame);
312 // Wait until the previous frame was consumed.
313 unique_lock<mutex> lock(bmusb_mutex);
314 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
315 if (card->should_quit) return;
318 if (video_frame.len - video_offset == 0 ||
319 video_frame.len - video_offset != size_t(width * (height + extra_lines_top + extra_lines_bottom) * 2)) {
320 if (video_frame.len != 0) {
321 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
322 card_index, video_frame.len - video_offset);
324 if (video_frame.owner) {
325 video_frame.owner->release_frame(video_frame);
328 // Still send on the information that we _had_ a frame, even though it's corrupted,
329 // so that pts can go up accordingly.
331 unique_lock<mutex> lock(bmusb_mutex);
332 card->new_data_ready = true;
333 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
334 card->new_frame_length = frame_length;
335 card->new_frame_interlaced = false;
336 card->new_data_ready_fence = nullptr;
337 card->dropped_frames = dropped_frames;
338 card->new_data_ready_changed.notify_all();
343 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
345 unsigned num_fields = interlaced ? 2 : 1;
346 timespec frame_upload_start;
348 // NOTE: This isn't deinterlacing. This is just sending the two fields along
349 // as separate frames without considering anything like the half-field offset.
350 // We'll need to add a proper deinterlacer on the receiving side to get this right.
351 assert(height % 2 == 0);
353 assert(frame_length % 2 == 0);
356 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
358 RefCountedFrame new_frame(video_frame);
360 // Upload the textures.
361 size_t cbcr_width = width / 2;
362 size_t cbcr_offset = video_offset / 2;
363 size_t y_offset = video_frame.size / 2 + video_offset / 2;
365 for (unsigned field = 0; field < num_fields; ++field) {
366 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
368 if (userdata->tex_y[field] == 0 ||
369 userdata->tex_cbcr[field] == 0 ||
370 width != userdata->last_width[field] ||
371 height != userdata->last_height[field]) {
372 // We changed resolution since last use of this texture, so we need to create
373 // a new object. Note that this each card has its own PBOFrameAllocator,
374 // we don't need to worry about these flip-flopping between resolutions.
375 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
377 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
379 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
381 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
383 userdata->last_width[field] = width;
384 userdata->last_height[field] = height;
387 GLuint pbo = userdata->pbo;
389 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
391 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
393 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
396 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
398 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)));
400 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
402 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
404 glBindTexture(GL_TEXTURE_2D, 0);
406 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
408 assert(fence != nullptr);
411 // Don't upload the second field as fast as we can; wait until
412 // the field time has approximately passed. (Otherwise, we could
413 // get timing jitter against the other sources, and possibly also
414 // against the video display, although the latter is not as critical.)
415 // This requires our system clock to be reasonably close to the
416 // video clock, but that's not an unreasonable assumption.
417 timespec second_field_start;
418 second_field_start.tv_nsec = frame_upload_start.tv_nsec +
419 frame_length * 1000000000 / TIMEBASE;
420 second_field_start.tv_sec = frame_upload_start.tv_sec +
421 second_field_start.tv_nsec / 1000000000;
422 second_field_start.tv_nsec %= 1000000000;
424 while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
425 &second_field_start, nullptr) == -1 &&
430 unique_lock<mutex> lock(bmusb_mutex);
431 card->new_data_ready = true;
432 card->new_frame = new_frame;
433 card->new_frame_length = frame_length;
434 card->new_frame_field = field;
435 card->new_frame_interlaced = interlaced;
436 card->new_data_ready_fence = fence;
437 card->dropped_frames = dropped_frames;
438 card->new_data_ready_changed.notify_all();
440 if (field != num_fields - 1) {
441 // Wait until the previous frame was consumed.
442 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
443 if (card->should_quit) return;
449 void Mixer::thread_func()
451 eglBindAPI(EGL_OPENGL_API);
452 QOpenGLContext *context = create_context(mixer_surface);
453 if (!make_current(context, mixer_surface)) {
458 struct timespec start, now;
459 clock_gettime(CLOCK_MONOTONIC, &start);
462 int stats_dropped_frames = 0;
464 while (!should_quit) {
465 CaptureCard card_copy[MAX_CARDS];
466 int num_samples[MAX_CARDS];
469 unique_lock<mutex> lock(bmusb_mutex);
471 // The first card is the master timer, so wait for it to have a new frame.
472 // TODO: Make configurable, and with a timeout.
473 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
475 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
476 CaptureCard *card = &cards[card_index];
477 card_copy[card_index].usb = card->usb;
478 card_copy[card_index].new_data_ready = card->new_data_ready;
479 card_copy[card_index].new_frame = card->new_frame;
480 card_copy[card_index].new_frame_length = card->new_frame_length;
481 card_copy[card_index].new_frame_field = card->new_frame_field;
482 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
483 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
484 card_copy[card_index].dropped_frames = card->dropped_frames;
485 card->new_data_ready = false;
486 card->new_data_ready_changed.notify_all();
488 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
489 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
490 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
491 assert(num_samples[card_index] >= 0);
495 // Resample the audio as needed, including from previously dropped frames.
496 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
498 // Signal to the audio thread to process this frame.
499 unique_lock<mutex> lock(audio_mutex);
500 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
501 audio_task_queue_changed.notify_one();
503 if (frame_num != card_copy[0].dropped_frames) {
504 // For dropped frames, increase the pts. Note that if the format changed
505 // in the meantime, we have no way of detecting that; we just have to
506 // assume the frame length is always the same.
507 ++stats_dropped_frames;
508 pts_int += card_copy[0].new_frame_length;
512 if (audio_level_callback != nullptr) {
513 unique_lock<mutex> lock(r128_mutex);
514 double loudness_s = r128.loudness_S();
515 double loudness_i = r128.integrated();
516 double loudness_range_low = r128.range_min();
517 double loudness_range_high = r128.range_max();
519 audio_level_callback(loudness_s, 20.0 * log10(peak),
520 loudness_i, loudness_range_low, loudness_range_high,
521 last_gain_staging_db);
524 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
525 if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
526 ++card_copy[card_index].dropped_frames;
528 if (card_copy[card_index].dropped_frames > 0) {
529 printf("Card %u dropped %d frames before this\n",
530 card_index, int(card_copy[card_index].dropped_frames));
534 // If the first card is reporting a corrupted or otherwise dropped frame,
535 // just increase the pts (skipping over this frame) and don't try to compute anything new.
536 if (card_copy[0].new_frame->len == 0) {
537 ++stats_dropped_frames;
538 pts_int += card_copy[0].new_frame_length;
542 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
543 CaptureCard *card = &card_copy[card_index];
544 if (!card->new_data_ready || card->new_frame->len == 0)
547 assert(card->new_frame != nullptr);
548 insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
551 // The new texture might still be uploaded,
552 // tell the GPU to wait until it's there.
553 if (card->new_data_ready_fence) {
554 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
556 glDeleteSync(card->new_data_ready_fence);
561 // Get the main chain from the theme, and set its state immediately.
562 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
563 EffectChain *chain = theme_main_chain.chain;
564 theme_main_chain.setup_chain();
566 GLuint y_tex, cbcr_tex;
567 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
570 // Render main chain.
571 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
572 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
573 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
575 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
576 resource_pool->release_fbo(fbo);
578 subsample_chroma(cbcr_full_tex, cbcr_tex);
579 resource_pool->release_2d_texture(cbcr_full_tex);
581 // Set the right state for rgba_tex.
582 glBindFramebuffer(GL_FRAMEBUFFER, 0);
583 glBindTexture(GL_TEXTURE_2D, rgba_tex);
584 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
585 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
586 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
588 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
591 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
592 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
594 pts_int += card_copy[0].new_frame_length;
596 // The live frame just shows the RGBA texture we just rendered.
597 // It owns rgba_tex now.
598 DisplayFrame live_frame;
599 live_frame.chain = display_chain.get();
600 live_frame.setup_chain = [this, rgba_tex]{
601 display_input->set_texture_num(rgba_tex);
603 live_frame.ready_fence = fence;
604 live_frame.input_frames = {};
605 live_frame.temp_textures = { rgba_tex };
606 output_channel[OUTPUT_LIVE].output_frame(live_frame);
608 // Set up preview and any additional channels.
609 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
610 DisplayFrame display_frame;
611 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
612 display_frame.chain = chain.chain;
613 display_frame.setup_chain = chain.setup_chain;
614 display_frame.ready_fence = fence;
615 display_frame.input_frames = chain.input_frames;
616 display_frame.temp_textures = {};
617 output_channel[i].output_frame(display_frame);
620 clock_gettime(CLOCK_MONOTONIC, &now);
621 double elapsed = now.tv_sec - start.tv_sec +
622 1e-9 * (now.tv_nsec - start.tv_nsec);
623 if (frame % 100 == 0) {
624 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
625 frame, stats_dropped_frames, elapsed, frame / elapsed,
626 1e3 * elapsed / frame);
627 // chain->print_phase_timing();
631 // Reset every 100 frames, so that local variations in frame times
632 // (especially for the first few frames, when the shaders are
633 // compiled etc.) don't make it hard to measure for the entire
634 // remaining duration of the program.
635 if (frame == 10000) {
643 resource_pool->clean_context();
646 void Mixer::audio_thread_func()
648 while (!should_quit) {
652 unique_lock<mutex> lock(audio_mutex);
653 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
654 task = audio_task_queue.front();
655 audio_task_queue.pop();
658 process_audio_one_frame(task.pts_int, task.num_samples);
662 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
664 vector<float> samples_card;
665 vector<float> samples_out;
666 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
667 samples_card.resize(num_samples * 2);
669 unique_lock<mutex> lock(cards[card_index].audio_mutex);
670 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
671 printf("Card %d reported previous underrun.\n", card_index);
674 // TODO: Allow using audio from the other card(s) as well.
675 if (card_index == 0) {
676 samples_out = move(samples_card);
680 // Cut away everything under 120 Hz (or whatever the cutoff is);
681 // we don't need it for voice, and it will reduce headroom
682 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
683 // should be dampened.)
684 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
686 // Apply a level compressor to get the general level right.
687 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
688 // (or more precisely, near it, since we don't use infinite ratio),
689 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
690 // entirely arbitrary, but from practical tests with speech, it seems to
691 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
692 float ref_level_dbfs = -14.0f;
694 float threshold = 0.01f; // -40 dBFS.
696 float attack_time = 0.5f;
697 float release_time = 20.0f;
698 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
699 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
700 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
704 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
705 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
706 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
707 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
710 // float limiter_att, compressor_att;
712 // The real compressor.
713 if (compressor_enabled) {
714 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
716 float attack_time = 0.005f;
717 float release_time = 0.040f;
718 float makeup_gain = 2.0f; // +6 dB.
719 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
720 // compressor_att = compressor.get_attenuation();
723 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
724 // Note that since ratio is not infinite, we could go slightly higher than this.
725 if (limiter_enabled) {
726 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
728 float attack_time = 0.0f; // Instant.
729 float release_time = 0.020f;
730 float makeup_gain = 1.0f; // 0 dB.
731 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
732 // limiter_att = limiter.get_attenuation();
735 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
737 // Upsample 4x to find interpolated peak.
738 peak_resampler.inp_data = samples_out.data();
739 peak_resampler.inp_count = samples_out.size() / 2;
741 vector<float> interpolated_samples_out;
742 interpolated_samples_out.resize(samples_out.size());
743 while (peak_resampler.inp_count > 0) { // About four iterations.
744 peak_resampler.out_data = &interpolated_samples_out[0];
745 peak_resampler.out_count = interpolated_samples_out.size() / 2;
746 peak_resampler.process();
747 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
748 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
752 vector<float> left, right;
753 deinterleave_samples(samples_out, &left, &right);
754 float *ptrs[] = { left.data(), right.data() };
756 unique_lock<mutex> lock(r128_mutex);
757 r128.process(left.size(), ptrs);
760 // Send the samples to the sound card.
762 alsa->write(samples_out);
765 // And finally add them to the output.
766 h264_encoder->add_audio(frame_pts_int, move(samples_out));
769 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
772 glGenVertexArrays(1, &vao);
781 glBindVertexArray(vao);
785 GLuint fbo = resource_pool->create_fbo(dst_tex);
786 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
787 glViewport(0, 0, WIDTH/2, HEIGHT/2);
790 glUseProgram(cbcr_program_num);
793 glActiveTexture(GL_TEXTURE0);
795 glBindTexture(GL_TEXTURE_2D, src_tex);
797 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
799 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
801 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
804 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
805 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
807 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
808 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
810 glDrawArrays(GL_TRIANGLES, 0, 3);
813 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
814 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
819 resource_pool->release_fbo(fbo);
820 glDeleteVertexArrays(1, &vao);
823 void Mixer::release_display_frame(DisplayFrame *frame)
825 for (GLuint texnum : frame->temp_textures) {
826 resource_pool->release_2d_texture(texnum);
828 frame->temp_textures.clear();
829 frame->ready_fence.reset();
830 frame->input_frames.clear();
835 mixer_thread = thread(&Mixer::thread_func, this);
836 audio_thread = thread(&Mixer::audio_thread_func, this);
846 void Mixer::transition_clicked(int transition_num)
848 theme->transition_clicked(transition_num, pts());
851 void Mixer::channel_clicked(int preview_num)
853 theme->channel_clicked(preview_num);
856 void Mixer::reset_meters()
858 peak_resampler.reset();
864 Mixer::OutputChannel::~OutputChannel()
866 if (has_current_frame) {
867 parent->release_display_frame(¤t_frame);
869 if (has_ready_frame) {
870 parent->release_display_frame(&ready_frame);
874 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
876 // Store this frame for display. Remove the ready frame if any
877 // (it was seemingly never used).
879 unique_lock<mutex> lock(frame_mutex);
880 if (has_ready_frame) {
881 parent->release_display_frame(&ready_frame);
884 has_ready_frame = true;
887 if (has_new_frame_ready_callback) {
888 new_frame_ready_callback();
892 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
894 unique_lock<mutex> lock(frame_mutex);
895 if (!has_current_frame && !has_ready_frame) {
899 if (has_current_frame && has_ready_frame) {
900 // We have a new ready frame. Toss the current one.
901 parent->release_display_frame(¤t_frame);
902 has_current_frame = false;
904 if (has_ready_frame) {
905 assert(!has_current_frame);
906 current_frame = ready_frame;
907 ready_frame.ready_fence.reset(); // Drop the refcount.
908 ready_frame.input_frames.clear(); // Drop the refcounts.
909 has_current_frame = true;
910 has_ready_frame = false;
913 *frame = current_frame;
917 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
919 new_frame_ready_callback = callback;
920 has_new_frame_ready_callback = true;