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>
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);
64 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
65 : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
67 mixer_surface(create_surface(format)),
68 h264_encoder_surface(create_surface(format)),
69 level_compressor(OUTPUT_FREQUENCY),
70 limiter(OUTPUT_FREQUENCY),
71 compressor(OUTPUT_FREQUENCY)
75 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
78 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
79 // will be halved when sampling them, and we need to compensate here.
80 movit_texel_subpixel_precision /= 2.0;
82 resource_pool.reset(new ResourcePool);
83 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
84 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
85 output_channel[i].parent = this;
88 ImageFormat inout_format;
89 inout_format.color_space = COLORSPACE_sRGB;
90 inout_format.gamma_curve = GAMMA_sRGB;
92 // Display chain; shows the live output produced by the main chain (its RGBA version).
93 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
95 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
96 display_chain->add_input(display_input);
97 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
98 display_chain->set_dither_bits(0); // Don't bother.
99 display_chain->finalize();
101 h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
103 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
104 printf("Configuring card %d...\n", card_index);
105 CaptureCard *card = &cards[card_index];
106 card->usb = new BMUSBCapture(card_index);
107 card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
108 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
109 card->usb->set_video_frame_allocator(card->frame_allocator.get());
110 card->surface = create_surface(format);
111 card->usb->set_dequeue_thread_callbacks(
113 eglBindAPI(EGL_OPENGL_API);
114 card->context = create_context(card->surface);
115 if (!make_current(card->context, card->surface)) {
116 printf("failed to create bmusb context\n");
121 resource_pool->clean_context();
123 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
124 card->usb->configure_card();
127 BMUSBCapture::start_bm_thread();
129 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
130 cards[card_index].usb->start_bm_capture();
133 //chain->enable_phase_timing(true);
135 // Set up stuff for NV12 conversion.
138 string cbcr_vert_shader = read_file("vs-cbcr.130.vert");
139 string cbcr_frag_shader =
142 "uniform sampler2D cbcr_tex; \n"
144 " gl_FragColor = texture2D(cbcr_tex, tc0); \n"
146 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
148 r128.init(2, OUTPUT_FREQUENCY);
151 locut.init(FILTER_HPF, 2);
153 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
154 // and there's a limit to how important the peak meter is.
155 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
157 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
162 resource_pool->release_glsl_program(cbcr_program_num);
163 BMUSBCapture::stop_bm_thread();
165 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
167 unique_lock<mutex> lock(bmusb_mutex);
168 cards[card_index].should_quit = true; // Unblock thread.
169 cards[card_index].new_data_ready_changed.notify_all();
171 cards[card_index].usb->stop_dequeue_thread();
174 h264_encoder.reset(nullptr);
179 int unwrap_timecode(uint16_t current_wrapped, int last)
181 uint16_t last_wrapped = last & 0xffff;
182 if (current_wrapped > last_wrapped) {
183 return (last & ~0xffff) | current_wrapped;
185 return 0x10000 + ((last & ~0xffff) | current_wrapped);
189 float find_peak(const float *samples, size_t num_samples)
191 float m = fabs(samples[0]);
192 for (size_t i = 1; i < num_samples; ++i) {
193 m = std::max(m, fabs(samples[i]));
198 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
200 size_t num_samples = in.size() / 2;
201 out_l->resize(num_samples);
202 out_r->resize(num_samples);
204 const float *inptr = in.data();
205 float *lptr = &(*out_l)[0];
206 float *rptr = &(*out_r)[0];
207 for (size_t i = 0; i < num_samples; ++i) {
215 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
216 FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
217 FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
219 CaptureCard *card = &cards[card_index];
221 unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
224 decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
225 &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now.
226 int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
228 size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
229 if (num_samples > OUTPUT_FREQUENCY / 10) {
230 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",
231 card_index, int(audio_frame.len), int(audio_offset),
232 timecode, int(video_frame.len), int(video_offset), video_format);
233 if (video_frame.owner) {
234 video_frame.owner->release_frame(video_frame);
236 if (audio_frame.owner) {
237 audio_frame.owner->release_frame(audio_frame);
242 int64_t local_pts = card->next_local_pts;
243 int dropped_frames = 0;
244 if (card->last_timecode != -1) {
245 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
248 // Convert the audio to stereo fp32 and add it.
250 audio.resize(num_samples * 2);
251 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
255 unique_lock<mutex> lock(card->audio_mutex);
257 // Number of samples per frame if we need to insert silence.
258 // (Could be nonintegral, but resampling will save us then.)
259 int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
261 if (dropped_frames > MAX_FPS * 2) {
262 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
263 card_index, card->last_timecode, timecode);
264 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
266 } else if (dropped_frames > 0) {
267 // Insert silence as needed.
268 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
269 card_index, dropped_frames, timecode);
270 vector<float> silence;
271 silence.resize(silence_samples * 2);
272 for (int i = 0; i < dropped_frames; ++i) {
273 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
274 // Note that if the format changed in the meantime, we have
275 // no way of detecting that; we just have to assume the frame length
276 // is always the same.
277 local_pts += frame_length;
280 if (num_samples == 0) {
281 audio.resize(silence_samples * 2);
282 num_samples = silence_samples;
284 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
285 card->next_local_pts = local_pts + frame_length;
288 card->last_timecode = timecode;
290 // Done with the audio, so release it.
291 if (audio_frame.owner) {
292 audio_frame.owner->release_frame(audio_frame);
296 // Wait until the previous frame was consumed.
297 unique_lock<mutex> lock(bmusb_mutex);
298 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
299 if (card->should_quit) return;
302 if (video_frame.len - video_offset == 0 ||
303 video_frame.len - video_offset != size_t(width * (height + extra_lines_top + extra_lines_bottom) * 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 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
328 unsigned num_fields = interlaced ? 2 : 1;
329 timespec frame_upload_start;
331 // NOTE: This isn't deinterlacing. This is just sending the two fields along
332 // as separate frames without considering anything like the half-field offset.
333 // We'll need to add a proper deinterlacer on the receiving side to get this right.
334 assert(height % 2 == 0);
336 assert(frame_length % 2 == 0);
339 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
341 RefCountedFrame new_frame(video_frame);
343 // Upload the textures.
344 size_t cbcr_width = width / 2;
345 size_t cbcr_offset = video_offset / 2;
346 size_t y_offset = video_frame.size / 2 + video_offset / 2;
348 for (unsigned field = 0; field < num_fields; ++field) {
349 unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
351 if (userdata->tex_y[field] == 0 ||
352 userdata->tex_cbcr[field] == 0 ||
353 width != userdata->last_width[field] ||
354 height != userdata->last_height[field]) {
355 // We changed resolution since last use of this texture, so we need to create
356 // a new object. Note that this each card has its own PBOFrameAllocator,
357 // we don't need to worry about these flip-flopping between resolutions.
358 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
360 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
362 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
364 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
366 userdata->last_width[field] = width;
367 userdata->last_height[field] = height;
370 GLuint pbo = userdata->pbo;
372 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
374 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
376 //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
379 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
381 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)));
383 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
385 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
387 glBindTexture(GL_TEXTURE_2D, 0);
389 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
391 assert(fence != nullptr);
394 // Don't upload the second field as fast as we can; wait until
395 // the field time has approximately passed. (Otherwise, we could
396 // get timing jitter against the other sources, and possibly also
397 // against the video display, although the latter is not as critical.)
398 // This requires our system clock to be reasonably close to the
399 // video clock, but that's not an unreasonable assumption.
400 timespec second_field_start;
401 second_field_start.tv_nsec = frame_upload_start.tv_nsec +
402 frame_length * 1000000000 / TIMEBASE;
403 second_field_start.tv_sec = frame_upload_start.tv_sec +
404 second_field_start.tv_nsec / 1000000000;
405 second_field_start.tv_nsec %= 1000000000;
407 while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
408 &second_field_start, nullptr) == -1 &&
413 unique_lock<mutex> lock(bmusb_mutex);
414 card->new_data_ready = true;
415 card->new_frame = new_frame;
416 card->new_frame_length = frame_length;
417 card->new_frame_field = field;
418 card->new_data_ready_fence = fence;
419 card->dropped_frames = dropped_frames;
420 card->new_data_ready_changed.notify_all();
422 if (field != num_fields - 1) {
423 // Wait until the previous frame was consumed.
424 card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
425 if (card->should_quit) return;
431 void Mixer::thread_func()
433 eglBindAPI(EGL_OPENGL_API);
434 QOpenGLContext *context = create_context(mixer_surface);
435 if (!make_current(context, mixer_surface)) {
440 struct timespec start, now;
441 clock_gettime(CLOCK_MONOTONIC, &start);
444 int stats_dropped_frames = 0;
446 while (!should_quit) {
447 CaptureCard card_copy[MAX_CARDS];
448 int num_samples[MAX_CARDS];
451 unique_lock<mutex> lock(bmusb_mutex);
453 // The first card is the master timer, so wait for it to have a new frame.
454 // TODO: Make configurable, and with a timeout.
455 cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
457 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
458 CaptureCard *card = &cards[card_index];
459 card_copy[card_index].usb = card->usb;
460 card_copy[card_index].new_data_ready = card->new_data_ready;
461 card_copy[card_index].new_frame = card->new_frame;
462 card_copy[card_index].new_frame_length = card->new_frame_length;
463 card_copy[card_index].new_frame_field = card->new_frame_field;
464 card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
465 card_copy[card_index].dropped_frames = card->dropped_frames;
466 card->new_data_ready = false;
467 card->new_data_ready_changed.notify_all();
469 int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
470 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
471 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
472 assert(num_samples[card_index] >= 0);
476 // Resample the audio as needed, including from previously dropped frames.
477 for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
479 // Signal to the audio thread to process this frame.
480 unique_lock<mutex> lock(audio_mutex);
481 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
482 audio_task_queue_changed.notify_one();
484 if (frame_num != card_copy[0].dropped_frames) {
485 // For dropped frames, increase the pts. Note that if the format changed
486 // in the meantime, we have no way of detecting that; we just have to
487 // assume the frame length is always the same.
488 ++stats_dropped_frames;
489 pts_int += card_copy[0].new_frame_length;
493 if (audio_level_callback != nullptr) {
494 double loudness_s = r128.loudness_S();
495 double loudness_i = r128.integrated();
496 double loudness_range_low = r128.range_min();
497 double loudness_range_high = r128.range_max();
499 audio_level_callback(loudness_s, 20.0 * log10(peak),
500 loudness_i, loudness_range_low, loudness_range_high,
501 last_gain_staging_db);
504 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
505 if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
506 ++card_copy[card_index].dropped_frames;
508 if (card_copy[card_index].dropped_frames > 0) {
509 printf("Card %u dropped %d frames before this\n",
510 card_index, int(card_copy[card_index].dropped_frames));
514 // If the first card is reporting a corrupted or otherwise dropped frame,
515 // just increase the pts (skipping over this frame) and don't try to compute anything new.
516 if (card_copy[0].new_frame->len == 0) {
517 ++stats_dropped_frames;
518 pts_int += card_copy[0].new_frame_length;
522 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
523 CaptureCard *card = &card_copy[card_index];
524 if (!card->new_data_ready || card->new_frame->len == 0)
527 assert(card->new_frame != nullptr);
528 bmusb_current_rendering_frame[card_index] = card->new_frame;
531 // The new texture might still be uploaded,
532 // tell the GPU to wait until it's there.
533 if (card->new_data_ready_fence) {
534 glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
536 glDeleteSync(card->new_data_ready_fence);
539 const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
540 theme->set_input_textures(
542 userdata->tex_y[card->new_frame_field],
543 userdata->tex_cbcr[card->new_frame_field],
544 userdata->last_width[card->new_frame_field],
545 userdata->last_height[card->new_frame_field]);
548 // Get the main chain from the theme, and set its state immediately.
549 pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
550 EffectChain *chain = theme_main_chain.first;
551 theme_main_chain.second();
553 GLuint y_tex, cbcr_tex;
554 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
557 // Render main chain.
558 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
559 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
560 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
562 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
563 resource_pool->release_fbo(fbo);
565 subsample_chroma(cbcr_full_tex, cbcr_tex);
566 resource_pool->release_2d_texture(cbcr_full_tex);
568 // Set the right state for rgba_tex.
569 glBindFramebuffer(GL_FRAMEBUFFER, 0);
570 glBindTexture(GL_TEXTURE_2D, rgba_tex);
571 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
572 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
573 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
575 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
578 // Make sure the H.264 gets a reference to all the
579 // input frames needed, so that they are not released back
580 // until the rendering is done.
581 vector<RefCountedFrame> input_frames;
582 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
583 input_frames.push_back(bmusb_current_rendering_frame[card_index]);
585 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
586 h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
588 pts_int += card_copy[0].new_frame_length;
590 // The live frame just shows the RGBA texture we just rendered.
591 // It owns rgba_tex now.
592 DisplayFrame live_frame;
593 live_frame.chain = display_chain.get();
594 live_frame.setup_chain = [this, rgba_tex]{
595 display_input->set_texture_num(rgba_tex);
597 live_frame.ready_fence = fence;
598 live_frame.input_frames = {};
599 live_frame.temp_textures = { rgba_tex };
600 output_channel[OUTPUT_LIVE].output_frame(live_frame);
602 // Set up preview and any additional channels.
603 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
604 DisplayFrame display_frame;
605 pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
606 display_frame.chain = chain.first;
607 display_frame.setup_chain = chain.second;
608 display_frame.ready_fence = fence;
610 // FIXME: possible to do better?
611 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
612 display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
614 display_frame.temp_textures = {};
615 output_channel[i].output_frame(display_frame);
618 clock_gettime(CLOCK_MONOTONIC, &now);
619 double elapsed = now.tv_sec - start.tv_sec +
620 1e-9 * (now.tv_nsec - start.tv_nsec);
621 if (frame % 100 == 0) {
622 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
623 frame, stats_dropped_frames, elapsed, frame / elapsed,
624 1e3 * elapsed / frame);
625 // chain->print_phase_timing();
629 // Reset every 100 frames, so that local variations in frame times
630 // (especially for the first few frames, when the shaders are
631 // compiled etc.) don't make it hard to measure for the entire
632 // remaining duration of the program.
633 if (frame == 10000) {
641 resource_pool->clean_context();
644 void Mixer::audio_thread_func()
646 while (!should_quit) {
650 unique_lock<mutex> lock(audio_mutex);
651 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
652 task = audio_task_queue.front();
653 audio_task_queue.pop();
656 process_audio_one_frame(task.pts_int, task.num_samples);
660 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
662 vector<float> samples_card;
663 vector<float> samples_out;
664 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
665 samples_card.resize(num_samples * 2);
667 unique_lock<mutex> lock(cards[card_index].audio_mutex);
668 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
669 printf("Card %d reported previous underrun.\n", card_index);
672 // TODO: Allow using audio from the other card(s) as well.
673 if (card_index == 0) {
674 samples_out = move(samples_card);
678 // Cut away everything under 120 Hz (or whatever the cutoff is);
679 // we don't need it for voice, and it will reduce headroom
680 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
681 // should be dampened.)
682 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
684 // Apply a level compressor to get the general level right.
685 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
686 // (or more precisely, near it, since we don't use infinite ratio),
687 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
688 // entirely arbitrary, but from practical tests with speech, it seems to
689 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
690 float ref_level_dbfs = -14.0f;
692 float threshold = 0.01f; // -40 dBFS.
694 float attack_time = 0.5f;
695 float release_time = 20.0f;
696 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
697 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
698 last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
702 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
703 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
704 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
705 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
708 // float limiter_att, compressor_att;
710 // The real compressor.
711 if (compressor_enabled) {
712 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
714 float attack_time = 0.005f;
715 float release_time = 0.040f;
716 float makeup_gain = 2.0f; // +6 dB.
717 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
718 // compressor_att = compressor.get_attenuation();
721 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
722 // Note that since ratio is not infinite, we could go slightly higher than this.
723 if (limiter_enabled) {
724 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
726 float attack_time = 0.0f; // Instant.
727 float release_time = 0.020f;
728 float makeup_gain = 1.0f; // 0 dB.
729 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
730 // limiter_att = limiter.get_attenuation();
733 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
735 // Upsample 4x to find interpolated peak.
736 peak_resampler.inp_data = samples_out.data();
737 peak_resampler.inp_count = samples_out.size() / 2;
739 vector<float> interpolated_samples_out;
740 interpolated_samples_out.resize(samples_out.size());
741 while (peak_resampler.inp_count > 0) { // About four iterations.
742 peak_resampler.out_data = &interpolated_samples_out[0];
743 peak_resampler.out_count = interpolated_samples_out.size() / 2;
744 peak_resampler.process();
745 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
746 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
750 vector<float> left, right;
751 deinterleave_samples(samples_out, &left, &right);
752 float *ptrs[] = { left.data(), right.data() };
753 r128.process(left.size(), ptrs);
755 // Send the samples to the sound card.
757 alsa->write(samples_out);
760 // And finally add them to the output.
761 h264_encoder->add_audio(frame_pts_int, move(samples_out));
764 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
767 glGenVertexArrays(1, &vao);
776 glBindVertexArray(vao);
780 GLuint fbo = resource_pool->create_fbo(dst_tex);
781 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
782 glViewport(0, 0, WIDTH/2, HEIGHT/2);
785 glUseProgram(cbcr_program_num);
788 glActiveTexture(GL_TEXTURE0);
790 glBindTexture(GL_TEXTURE_2D, src_tex);
792 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
794 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
796 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
799 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
800 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
802 GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices);
803 GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices.
805 glDrawArrays(GL_TRIANGLES, 0, 3);
808 cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo);
809 cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo);
814 resource_pool->release_fbo(fbo);
815 glDeleteVertexArrays(1, &vao);
818 void Mixer::release_display_frame(DisplayFrame *frame)
820 for (GLuint texnum : frame->temp_textures) {
821 resource_pool->release_2d_texture(texnum);
823 frame->temp_textures.clear();
824 frame->ready_fence.reset();
825 frame->input_frames.clear();
830 mixer_thread = thread(&Mixer::thread_func, this);
831 audio_thread = thread(&Mixer::audio_thread_func, this);
841 void Mixer::transition_clicked(int transition_num)
843 theme->transition_clicked(transition_num, pts());
846 void Mixer::channel_clicked(int preview_num)
848 theme->channel_clicked(preview_num);
851 void Mixer::reset_meters()
853 peak_resampler.reset();
859 Mixer::OutputChannel::~OutputChannel()
861 if (has_current_frame) {
862 parent->release_display_frame(¤t_frame);
864 if (has_ready_frame) {
865 parent->release_display_frame(&ready_frame);
869 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
871 // Store this frame for display. Remove the ready frame if any
872 // (it was seemingly never used).
874 unique_lock<mutex> lock(frame_mutex);
875 if (has_ready_frame) {
876 parent->release_display_frame(&ready_frame);
879 has_ready_frame = true;
882 if (has_new_frame_ready_callback) {
883 new_frame_ready_callback();
887 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
889 unique_lock<mutex> lock(frame_mutex);
890 if (!has_current_frame && !has_ready_frame) {
894 if (has_current_frame && has_ready_frame) {
895 // We have a new ready frame. Toss the current one.
896 parent->release_display_frame(¤t_frame);
897 has_current_frame = false;
899 if (has_ready_frame) {
900 assert(!has_current_frame);
901 current_frame = ready_frame;
902 ready_frame.ready_fence.reset(); // Drop the refcount.
903 ready_frame.input_frames.clear(); // Drop the refcounts.
904 has_current_frame = true;
905 has_ready_frame = false;
908 *frame = current_frame;
912 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
914 new_frame_ready_callback = callback;
915 has_new_frame_ready_callback = true;