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 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
164 r128.init(2, OUTPUT_FREQUENCY);
167 locut.init(FILTER_HPF, 2);
169 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
170 // and there's a limit to how important the peak meter is.
171 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
173 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
178 resource_pool->release_glsl_program(cbcr_program_num);
179 BMUSBCapture::stop_bm_thread();
181 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
183 unique_lock<mutex> lock(bmusb_mutex);
184 cards[card_index].should_quit = true; // Unblock thread.
185 cards[card_index].new_data_ready_changed.notify_all();
187 cards[card_index].usb->stop_dequeue_thread();
190 h264_encoder.reset(nullptr);
195 int unwrap_timecode(uint16_t current_wrapped, int last)
197 uint16_t last_wrapped = last & 0xffff;
198 if (current_wrapped > last_wrapped) {
199 return (last & ~0xffff) | current_wrapped;
201 return 0x10000 + ((last & ~0xffff) | current_wrapped);
205 float find_peak(const float *samples, size_t num_samples)
207 float m = fabs(samples[0]);
208 for (size_t i = 1; i < num_samples; ++i) {
209 m = std::max(m, fabs(samples[i]));
214 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
216 size_t num_samples = in.size() / 2;
217 out_l->resize(num_samples);
218 out_r->resize(num_samples);
220 const float *inptr = in.data();
221 float *lptr = &(*out_l)[0];
222 float *rptr = &(*out_r)[0];
223 for (size_t i = 0; i < num_samples; ++i) {
231 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
232 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
233 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
235 CaptureCard *card = &cards[card_index];
237 unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
240 decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
241 &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now.
242 int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
244 size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
245 if (num_samples > OUTPUT_FREQUENCY / 10) {
246 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",
247 card_index, int(audio_frame.len), int(audio_offset),
248 timecode, int(video_frame.len), int(video_offset), video_format);
249 if (video_frame.owner) {
250 video_frame.owner->release_frame(video_frame);
252 if (audio_frame.owner) {
253 audio_frame.owner->release_frame(audio_frame);
258 int64_t local_pts = card->next_local_pts;
259 int dropped_frames = 0;
260 if (card->last_timecode != -1) {
261 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
264 // Convert the audio to stereo fp32 and add it.
266 audio.resize(num_samples * 2);
267 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
271 unique_lock<mutex> lock(card->audio_mutex);
273 // Number of samples per frame if we need to insert silence.
274 // (Could be nonintegral, but resampling will save us then.)
275 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
277 if (dropped_frames > MAX_FPS * 2) {
278 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
279 card_index, card->last_timecode, timecode);
280 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
282 } else if (dropped_frames > 0) {
283 // Insert silence as needed.
284 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
285 card_index, dropped_frames, timecode);
286 vector<float> silence(silence_samples * 2, 0.0f);
287 for (int i = 0; i < dropped_frames; ++i) {
288 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
289 // Note that if the format changed in the meantime, we have
290 // no way of detecting that; we just have to assume the frame length
291 // is always the same.
292 local_pts += frame_length;
295 if (num_samples == 0) {
296 audio.resize(silence_samples * 2);
297 num_samples = silence_samples;
299 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
300 card->next_local_pts = local_pts + frame_length;
303 card->last_timecode = timecode;
305 // Done with the audio, so release it.
306 if (audio_frame.owner) {
307 audio_frame.owner->release_frame(audio_frame);
311 // Wait until the previous frame was consumed.
312 unique_lock<mutex> lock(bmusb_mutex);
313 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
314 if (card->should_quit) return;
317 if (video_frame.len - video_offset == 0 ||
318 video_frame.len - video_offset != size_t(width * (height + extra_lines_top + extra_lines_bottom) * 2)) {
319 if (video_frame.len != 0) {
320 printf("Card %d: Dropping video frame with wrong length (%ld)\n",
321 card_index, video_frame.len - video_offset);
323 if (video_frame.owner) {
324 video_frame.owner->release_frame(video_frame);
327 // Still send on the information that we _had_ a frame, even though it's corrupted,
328 // so that pts can go up accordingly.
330 unique_lock<mutex> lock(bmusb_mutex);
331 card->new_data_ready = true;
332 card->new_frame = RefCountedFrame(FrameAllocator::Frame());
333 card->new_frame_length = frame_length;
334 card->new_frame_interlaced = false;
335 card->new_data_ready_fence = nullptr;
336 card->dropped_frames = dropped_frames;
337 card->new_data_ready_changed.notify_all();
342 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
344 unsigned num_fields = interlaced ? 2 : 1;
345 timespec frame_upload_start;
347 // NOTE: This isn't deinterlacing. This is just sending the two fields along
348 // as separate frames without considering anything like the half-field offset.
349 // We'll need to add a proper deinterlacer on the receiving side to get this right.
350 assert(height % 2 == 0);
352 assert(frame_length % 2 == 0);
355 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
357 RefCountedFrame new_frame(video_frame);
359 // Upload the textures.
360 size_t cbcr_width = width / 2;
361 size_t cbcr_offset = video_offset / 2;
362 size_t y_offset = video_frame.size / 2 + video_offset / 2;
364 for (unsigned field = 0; field < num_fields; ++field) {
365 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
367 if (userdata->tex_y[field] == 0 ||
368 userdata->tex_cbcr[field] == 0 ||
369 width != userdata->last_width[field] ||
370 height != userdata->last_height[field]) {
371 // We changed resolution since last use of this texture, so we need to create
372 // a new object. Note that this each card has its own PBOFrameAllocator,
373 // we don't need to worry about these flip-flopping between resolutions.
374 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
376 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
378 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
380 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
382 userdata->last_width[field] = width;
383 userdata->last_height[field] = height;
386 GLuint pbo = userdata->pbo;
388 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
390 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
392 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
395 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
397 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)));
399 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
401 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
403 glBindTexture(GL_TEXTURE_2D, 0);
405 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
407 assert(fence != nullptr);
410 // Don't upload the second field as fast as we can; wait until
411 // the field time has approximately passed. (Otherwise, we could
412 // get timing jitter against the other sources, and possibly also
413 // against the video display, although the latter is not as critical.)
414 // This requires our system clock to be reasonably close to the
415 // video clock, but that's not an unreasonable assumption.
416 timespec second_field_start;
417 second_field_start.tv_nsec = frame_upload_start.tv_nsec +
418 frame_length * 1000000000 / TIMEBASE;
419 second_field_start.tv_sec = frame_upload_start.tv_sec +
420 second_field_start.tv_nsec / 1000000000;
421 second_field_start.tv_nsec %= 1000000000;
423 while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
424 &second_field_start, nullptr) == -1 &&
429 unique_lock<mutex> lock(bmusb_mutex);
430 card->new_data_ready = true;
431 card->new_frame = new_frame;
432 card->new_frame_length = frame_length;
433 card->new_frame_field = field;
434 card->new_frame_interlaced = interlaced;
435 card->new_data_ready_fence = fence;
436 card->dropped_frames = dropped_frames;
437 card->new_data_ready_changed.notify_all();
439 if (field != num_fields - 1) {
440 // Wait until the previous frame was consumed.
441 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
442 if (card->should_quit) return;
448 void Mixer::thread_func()
450 eglBindAPI(EGL_OPENGL_API);
451 QOpenGLContext *context = create_context(mixer_surface);
452 if (!make_current(context, mixer_surface)) {
457 struct timespec start, now;
458 clock_gettime(CLOCK_MONOTONIC, &start);
461 int stats_dropped_frames = 0;
463 while (!should_quit) {
464 CaptureCard card_copy[MAX_CARDS];
465 int num_samples[MAX_CARDS];
468 unique_lock<mutex> lock(bmusb_mutex);
470 // The first card is the master timer, so wait for it to have a new frame.
471 // TODO: Make configurable, and with a timeout.
472 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
474 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
475 CaptureCard *card = &cards[card_index];
476 card_copy[card_index].usb = card->usb;
477 card_copy[card_index].new_data_ready = card->new_data_ready;
478 card_copy[card_index].new_frame = card->new_frame;
479 card_copy[card_index].new_frame_length = card->new_frame_length;
480 card_copy[card_index].new_frame_field = card->new_frame_field;
481 card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
482 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
483 card_copy[card_index].dropped_frames = card->dropped_frames;
484 card->new_data_ready = false;
485 card->new_data_ready_changed.notify_all();
487 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
488 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
489 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
490 assert(num_samples[card_index] >= 0);
494 // Resample the audio as needed, including from previously dropped frames.
495 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
497 // Signal to the audio thread to process this frame.
498 unique_lock<mutex> lock(audio_mutex);
499 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
500 audio_task_queue_changed.notify_one();
502 if (frame_num != card_copy[0].dropped_frames) {
503 // For dropped frames, increase the pts. Note that if the format changed
504 // in the meantime, we have no way of detecting that; we just have to
505 // assume the frame length is always the same.
506 ++stats_dropped_frames;
507 pts_int += card_copy[0].new_frame_length;
511 if (audio_level_callback != nullptr) {
512 unique_lock<mutex> lock(r128_mutex);
513 double loudness_s = r128.loudness_S();
514 double loudness_i = r128.integrated();
515 double loudness_range_low = r128.range_min();
516 double loudness_range_high = r128.range_max();
518 audio_level_callback(loudness_s, 20.0 * log10(peak),
519 loudness_i, loudness_range_low, loudness_range_high,
520 last_gain_staging_db);
523 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
524 if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
525 ++card_copy[card_index].dropped_frames;
527 if (card_copy[card_index].dropped_frames > 0) {
528 printf("Card %u dropped %d frames before this\n",
529 card_index, int(card_copy[card_index].dropped_frames));
533 // If the first card is reporting a corrupted or otherwise dropped frame,
534 // just increase the pts (skipping over this frame) and don't try to compute anything new.
535 if (card_copy[0].new_frame->len == 0) {
536 ++stats_dropped_frames;
537 pts_int += card_copy[0].new_frame_length;
541 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
542 CaptureCard *card = &card_copy[card_index];
543 if (!card->new_data_ready || card->new_frame->len == 0)
546 assert(card->new_frame != nullptr);
547 insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
550 // The new texture might still be uploaded,
551 // tell the GPU to wait until it's there.
552 if (card->new_data_ready_fence) {
553 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
555 glDeleteSync(card->new_data_ready_fence);
560 // Get the main chain from the theme, and set its state immediately.
561 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
562 EffectChain *chain = theme_main_chain.chain;
563 theme_main_chain.setup_chain();
565 GLuint y_tex, cbcr_tex;
566 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
569 // Render main chain.
570 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
571 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
572 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
574 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
575 resource_pool->release_fbo(fbo);
577 subsample_chroma(cbcr_full_tex, cbcr_tex);
578 resource_pool->release_2d_texture(cbcr_full_tex);
580 // Set the right state for rgba_tex.
581 glBindFramebuffer(GL_FRAMEBUFFER, 0);
582 glBindTexture(GL_TEXTURE_2D, rgba_tex);
583 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
584 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
585 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
587 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
590 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
591 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
593 pts_int += card_copy[0].new_frame_length;
595 // The live frame just shows the RGBA texture we just rendered.
596 // It owns rgba_tex now.
597 DisplayFrame live_frame;
598 live_frame.chain = display_chain.get();
599 live_frame.setup_chain = [this, rgba_tex]{
600 display_input->set_texture_num(rgba_tex);
602 live_frame.ready_fence = fence;
603 live_frame.input_frames = {};
604 live_frame.temp_textures = { rgba_tex };
605 output_channel[OUTPUT_LIVE].output_frame(live_frame);
607 // Set up preview and any additional channels.
608 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
609 DisplayFrame display_frame;
610 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
611 display_frame.chain = chain.chain;
612 display_frame.setup_chain = chain.setup_chain;
613 display_frame.ready_fence = fence;
614 display_frame.input_frames = chain.input_frames;
615 display_frame.temp_textures = {};
616 output_channel[i].output_frame(display_frame);
619 clock_gettime(CLOCK_MONOTONIC, &now);
620 double elapsed = now.tv_sec - start.tv_sec +
621 1e-9 * (now.tv_nsec - start.tv_nsec);
622 if (frame % 100 == 0) {
623 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
624 frame, stats_dropped_frames, elapsed, frame / elapsed,
625 1e3 * elapsed / frame);
626 // chain->print_phase_timing();
630 // Reset every 100 frames, so that local variations in frame times
631 // (especially for the first few frames, when the shaders are
632 // compiled etc.) don't make it hard to measure for the entire
633 // remaining duration of the program.
634 if (frame == 10000) {
642 resource_pool->clean_context();
645 void Mixer::audio_thread_func()
647 while (!should_quit) {
651 unique_lock<mutex> lock(audio_mutex);
652 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
653 task = audio_task_queue.front();
654 audio_task_queue.pop();
657 process_audio_one_frame(task.pts_int, task.num_samples);
661 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
663 vector<float> samples_card;
664 vector<float> samples_out;
665 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
666 samples_card.resize(num_samples * 2);
668 unique_lock<mutex> lock(cards[card_index].audio_mutex);
669 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
670 printf("Card %d reported previous underrun.\n", card_index);
673 // TODO: Allow using audio from the other card(s) as well.
674 if (card_index == 0) {
675 samples_out = move(samples_card);
679 // Cut away everything under 120 Hz (or whatever the cutoff is);
680 // we don't need it for voice, and it will reduce headroom
681 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
682 // should be dampened.)
683 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
685 // Apply a level compressor to get the general level right.
686 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
687 // (or more precisely, near it, since we don't use infinite ratio),
688 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
689 // entirely arbitrary, but from practical tests with speech, it seems to
690 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
691 float ref_level_dbfs = -14.0f;
693 float threshold = 0.01f; // -40 dBFS.
695 float attack_time = 0.5f;
696 float release_time = 20.0f;
697 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
698 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
699 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
703 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
704 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
705 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
706 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
709 // float limiter_att, compressor_att;
711 // The real compressor.
712 if (compressor_enabled) {
713 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
715 float attack_time = 0.005f;
716 float release_time = 0.040f;
717 float makeup_gain = 2.0f; // +6 dB.
718 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
719 // compressor_att = compressor.get_attenuation();
722 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
723 // Note that since ratio is not infinite, we could go slightly higher than this.
724 if (limiter_enabled) {
725 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
727 float attack_time = 0.0f; // Instant.
728 float release_time = 0.020f;
729 float makeup_gain = 1.0f; // 0 dB.
730 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
731 // limiter_att = limiter.get_attenuation();
734 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
736 // Upsample 4x to find interpolated peak.
737 peak_resampler.inp_data = samples_out.data();
738 peak_resampler.inp_count = samples_out.size() / 2;
740 vector<float> interpolated_samples_out;
741 interpolated_samples_out.resize(samples_out.size());
742 while (peak_resampler.inp_count > 0) { // About four iterations.
743 peak_resampler.out_data = &interpolated_samples_out[0];
744 peak_resampler.out_count = interpolated_samples_out.size() / 2;
745 peak_resampler.process();
746 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
747 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
751 vector<float> left, right;
752 deinterleave_samples(samples_out, &left, &right);
753 float *ptrs[] = { left.data(), right.data() };
755 unique_lock<mutex> lock(r128_mutex);
756 r128.process(left.size(), ptrs);
759 // Send the samples to the sound card.
761 alsa->write(samples_out);
764 // And finally add them to the output.
765 h264_encoder->add_audio(frame_pts_int, move(samples_out));
768 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
771 glGenVertexArrays(1, &vao);
780 glBindVertexArray(vao);
784 GLuint fbo = resource_pool->create_fbo(dst_tex);
785 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
786 glViewport(0, 0, WIDTH/2, HEIGHT/2);
789 glUseProgram(cbcr_program_num);
792 glActiveTexture(GL_TEXTURE0);
794 glBindTexture(GL_TEXTURE_2D, src_tex);
796 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
798 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
800 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
803 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
804 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
806 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
807 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
809 glDrawArrays(GL_TRIANGLES, 0, 3);
812 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
813 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
818 resource_pool->release_fbo(fbo);
819 glDeleteVertexArrays(1, &vao);
822 void Mixer::release_display_frame(DisplayFrame *frame)
824 for (GLuint texnum : frame->temp_textures) {
825 resource_pool->release_2d_texture(texnum);
827 frame->temp_textures.clear();
828 frame->ready_fence.reset();
829 frame->input_frames.clear();
834 mixer_thread = thread(&Mixer::thread_func, this);
835 audio_thread = thread(&Mixer::audio_thread_func, this);
845 void Mixer::transition_clicked(int transition_num)
847 theme->transition_clicked(transition_num, pts());
850 void Mixer::channel_clicked(int preview_num)
852 theme->channel_clicked(preview_num);
855 void Mixer::reset_meters()
857 peak_resampler.reset();
863 Mixer::OutputChannel::~OutputChannel()
865 if (has_current_frame) {
866 parent->release_display_frame(¤t_frame);
868 if (has_ready_frame) {
869 parent->release_display_frame(&ready_frame);
873 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
875 // Store this frame for display. Remove the ready frame if any
876 // (it was seemingly never used).
878 unique_lock<mutex> lock(frame_mutex);
879 if (has_ready_frame) {
880 parent->release_display_frame(&ready_frame);
883 has_ready_frame = true;
886 if (has_new_frame_ready_callback) {
887 new_frame_ready_callback();
891 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
893 unique_lock<mutex> lock(frame_mutex);
894 if (!has_current_frame && !has_ready_frame) {
898 if (has_current_frame && has_ready_frame) {
899 // We have a new ready frame. Toss the current one.
900 parent->release_display_frame(¤t_frame);
901 has_current_frame = false;
903 if (has_ready_frame) {
904 assert(!has_current_frame);
905 current_frame = ready_frame;
906 ready_frame.ready_fence.reset(); // Drop the refcount.
907 ready_frame.input_frames.clear(); // Drop the refcounts.
908 has_current_frame = true;
909 has_ready_frame = false;
912 *frame = current_frame;
916 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
918 new_frame_ready_callback = callback;
919 has_new_frame_ready_callback = true;