7 #include <movit/effect_chain.h>
8 #include <movit/effect_util.h>
9 #include <movit/flat_input.h>
10 #include <movit/image_format.h>
11 #include <movit/init.h>
12 #include <movit/resource_pool.h>
13 #include <movit/util.h>
21 #include <condition_variable>
29 #include <arpa/inet.h>
31 #include "bmusb/bmusb.h"
33 #include "decklink_capture.h"
36 #include "h264encode.h"
37 #include "pbo_frame_allocator.h"
38 #include "ref_counted_gl_sync.h"
43 using namespace movit;
45 using namespace std::placeholders;
47 Mixer *global_mixer = nullptr;
51 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
53 assert(in_channels >= out_channels);
54 for (size_t i = 0; i < num_samples; ++i) {
55 for (size_t j = 0; j < out_channels; ++j) {
59 uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
60 dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
62 src += 3 * (in_channels - out_channels);
66 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
68 assert(in_channels >= out_channels);
69 for (size_t i = 0; i < num_samples; ++i) {
70 for (size_t j = 0; j < out_channels; ++j) {
71 // Note: Assumes little-endian.
72 int32_t s = *(int32_t *)src;
73 dst[i * out_channels + j] = s * (1.0f / 4294967296.0f);
76 src += 4 * (in_channels - out_channels);
80 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
83 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
84 input_state->buffered_frames[card_index][frame_num] =
85 input_state->buffered_frames[card_index][frame_num - 1];
87 input_state->buffered_frames[card_index][0] = { frame, field_num };
89 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
90 input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
95 string generate_local_dump_filename(int frame)
97 time_t now = time(NULL);
99 localtime_r(&now, &now_tm);
102 strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm);
104 // Use the frame number to disambiguate between two cuts starting
105 // on the same second.
107 snprintf(filename, sizeof(filename), "%s%s-f%02d%s",
108 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
114 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
115 : httpd(WIDTH, HEIGHT),
116 num_cards(num_cards),
117 mixer_surface(create_surface(format)),
118 h264_encoder_surface(create_surface(format)),
119 correlation(OUTPUT_FREQUENCY),
120 level_compressor(OUTPUT_FREQUENCY),
121 limiter(OUTPUT_FREQUENCY),
122 compressor(OUTPUT_FREQUENCY)
124 httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
127 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
130 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
131 // will be halved when sampling them, and we need to compensate here.
132 movit_texel_subpixel_precision /= 2.0;
134 resource_pool.reset(new ResourcePool);
135 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
136 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
137 output_channel[i].parent = this;
140 ImageFormat inout_format;
141 inout_format.color_space = COLORSPACE_sRGB;
142 inout_format.gamma_curve = GAMMA_sRGB;
144 // Display chain; shows the live output produced by the main chain (its RGBA version).
145 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
147 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
148 display_chain->add_input(display_input);
149 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
150 display_chain->set_dither_bits(0); // Don't bother.
151 display_chain->finalize();
153 h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
155 // First try initializing the PCI devices, then USB, until we have the desired number of cards.
156 unsigned num_pci_devices = 0, num_usb_devices = 0;
157 unsigned card_index = 0;
159 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
160 if (decklink_iterator != nullptr) {
161 for ( ; card_index < num_cards; ++card_index) {
163 if (decklink_iterator->Next(&decklink) != S_OK) {
167 configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
170 decklink_iterator->Release();
171 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
173 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
175 for ( ; card_index < num_cards; ++card_index) {
176 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
180 if (num_usb_devices > 0) {
181 BMUSBCapture::start_bm_thread();
184 for (card_index = 0; card_index < num_cards; ++card_index) {
185 cards[card_index].capture->start_bm_capture();
188 // Set up stuff for NV12 conversion.
191 string cbcr_vert_shader =
194 "in vec2 position; \n"
195 "in vec2 texcoord; \n"
197 "uniform vec2 foo_chroma_offset_0; \n"
201 " // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
203 " // 2.000 0.000 0.000 -1.000 \n"
204 " // 0.000 2.000 0.000 -1.000 \n"
205 " // 0.000 0.000 -2.000 -1.000 \n"
206 " // 0.000 0.000 0.000 1.000 \n"
207 " gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
208 " vec2 flipped_tc = texcoord; \n"
209 " tc0 = flipped_tc + foo_chroma_offset_0; \n"
211 string cbcr_frag_shader =
214 "uniform sampler2D cbcr_tex; \n"
215 "out vec4 FragColor; \n"
217 " FragColor = texture(cbcr_tex, tc0); \n"
219 vector<string> frag_shader_outputs;
220 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
227 cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
228 cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
229 cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
231 r128.init(2, OUTPUT_FREQUENCY);
234 locut.init(FILTER_HPF, 2);
236 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
237 // and there's a limit to how important the peak meter is.
238 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
240 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
245 resource_pool->release_glsl_program(cbcr_program_num);
246 glDeleteBuffers(1, &cbcr_vbo);
247 BMUSBCapture::stop_bm_thread();
249 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
251 unique_lock<mutex> lock(bmusb_mutex);
252 cards[card_index].should_quit = true; // Unblock thread.
253 cards[card_index].new_frames_changed.notify_all();
255 cards[card_index].capture->stop_dequeue_thread();
258 h264_encoder.reset(nullptr);
261 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
263 printf("Configuring card %d...\n", card_index);
265 CaptureCard *card = &cards[card_index];
266 card->capture = capture;
267 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
268 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
269 card->capture->set_video_frame_allocator(card->frame_allocator.get());
270 card->surface = create_surface(format);
271 card->capture->set_dequeue_thread_callbacks(
273 eglBindAPI(EGL_OPENGL_API);
274 card->context = create_context(card->surface);
275 if (!make_current(card->context, card->surface)) {
276 printf("failed to create bmusb context\n");
281 resource_pool->clean_context();
283 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
284 card->capture->configure_card();
290 int unwrap_timecode(uint16_t current_wrapped, int last)
292 uint16_t last_wrapped = last & 0xffff;
293 if (current_wrapped > last_wrapped) {
294 return (last & ~0xffff) | current_wrapped;
296 return 0x10000 + ((last & ~0xffff) | current_wrapped);
300 float find_peak(const float *samples, size_t num_samples)
302 float m = fabs(samples[0]);
303 for (size_t i = 1; i < num_samples; ++i) {
304 m = max(m, fabs(samples[i]));
309 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
311 size_t num_samples = in.size() / 2;
312 out_l->resize(num_samples);
313 out_r->resize(num_samples);
315 const float *inptr = in.data();
316 float *lptr = &(*out_l)[0];
317 float *rptr = &(*out_r)[0];
318 for (size_t i = 0; i < num_samples; ++i) {
326 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
327 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
328 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
330 CaptureCard *card = &cards[card_index];
332 if (is_mode_scanning[card_index]) {
333 if (video_format.has_signal) {
334 // Found a stable signal, so stop scanning.
335 is_mode_scanning[card_index] = false;
337 static constexpr double switch_time_s = 0.5; // Should be enough time for the signal to stabilize.
339 clock_gettime(CLOCK_MONOTONIC, &now);
340 double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
341 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
342 if (sec_since_last_switch > switch_time_s) {
343 // It isn't this mode; try the next one.
344 mode_scanlist_index[card_index]++;
345 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
346 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
347 last_mode_scan_change[card_index] = now;
352 int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
354 size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
355 if (num_samples > OUTPUT_FREQUENCY / 10) {
356 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",
357 card_index, int(audio_frame.len), int(audio_offset),
358 timecode, int(video_frame.len), int(video_offset), video_format.id);
359 if (video_frame.owner) {
360 video_frame.owner->release_frame(video_frame);
362 if (audio_frame.owner) {
363 audio_frame.owner->release_frame(audio_frame);
368 int64_t local_pts = card->next_local_pts;
369 int dropped_frames = 0;
370 if (card->last_timecode != -1) {
371 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
374 // Convert the audio to stereo fp32 and add it.
376 audio.resize(num_samples * 2);
377 switch (audio_format.bits_per_sample) {
379 assert(num_samples == 0);
382 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
385 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
388 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
394 unique_lock<mutex> lock(card->audio_mutex);
396 // Number of samples per frame if we need to insert silence.
397 // (Could be nonintegral, but resampling will save us then.)
398 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
400 if (dropped_frames > MAX_FPS * 2) {
401 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
402 card_index, card->last_timecode, timecode);
403 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
405 } else if (dropped_frames > 0) {
406 // Insert silence as needed.
407 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
408 card_index, dropped_frames, timecode);
409 vector<float> silence(silence_samples * 2, 0.0f);
410 for (int i = 0; i < dropped_frames; ++i) {
411 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
412 // Note that if the format changed in the meantime, we have
413 // no way of detecting that; we just have to assume the frame length
414 // is always the same.
415 local_pts += frame_length;
418 if (num_samples == 0) {
419 audio.resize(silence_samples * 2);
420 num_samples = silence_samples;
422 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
423 card->next_local_pts = local_pts + frame_length;
426 card->last_timecode = timecode;
428 // Done with the audio, so release it.
429 if (audio_frame.owner) {
430 audio_frame.owner->release_frame(audio_frame);
433 size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
434 if (video_frame.len - video_offset == 0 ||
435 video_frame.len - video_offset != expected_length) {
436 if (video_frame.len != 0) {
437 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
438 card_index, video_frame.len - video_offset, expected_length);
440 if (video_frame.owner) {
441 video_frame.owner->release_frame(video_frame);
444 // Still send on the information that we _had_ a frame, even though it's corrupted,
445 // so that pts can go up accordingly.
447 unique_lock<mutex> lock(bmusb_mutex);
448 CaptureCard::NewFrame new_frame;
449 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
450 new_frame.length = frame_length;
451 new_frame.interlaced = false;
452 new_frame.ready_fence = nullptr;
453 new_frame.dropped_frames = dropped_frames;
454 card->new_frames.push(move(new_frame));
455 card->new_frames_changed.notify_all();
460 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
462 unsigned num_fields = video_format.interlaced ? 2 : 1;
463 timespec frame_upload_start;
464 if (video_format.interlaced) {
465 // Send the two fields along as separate frames; the other side will need to add
466 // a deinterlacer to actually get this right.
467 assert(video_format.height % 2 == 0);
468 video_format.height /= 2;
469 assert(frame_length % 2 == 0);
472 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
474 userdata->last_interlaced = video_format.interlaced;
475 userdata->last_has_signal = video_format.has_signal;
476 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
477 userdata->last_frame_rate_den = video_format.frame_rate_den;
478 RefCountedFrame frame(video_frame);
480 // Upload the textures.
481 size_t cbcr_width = video_format.width / 2;
482 size_t cbcr_offset = video_offset / 2;
483 size_t y_offset = video_frame.size / 2 + video_offset / 2;
485 for (unsigned field = 0; field < num_fields; ++field) {
486 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
488 if (userdata->tex_y[field] == 0 ||
489 userdata->tex_cbcr[field] == 0 ||
490 video_format.width != userdata->last_width[field] ||
491 video_format.height != userdata->last_height[field]) {
492 // We changed resolution since last use of this texture, so we need to create
493 // a new object. Note that this each card has its own PBOFrameAllocator,
494 // we don't need to worry about these flip-flopping between resolutions.
495 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
497 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
499 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
501 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
503 userdata->last_width[field] = video_format.width;
504 userdata->last_height[field] = video_format.height;
507 GLuint pbo = userdata->pbo;
509 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
511 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
514 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
516 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
518 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
520 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + video_format.width * field_start_line));
522 glBindTexture(GL_TEXTURE_2D, 0);
524 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
526 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
528 assert(fence != nullptr);
531 // Don't upload the second field as fast as we can; wait until
532 // the field time has approximately passed. (Otherwise, we could
533 // get timing jitter against the other sources, and possibly also
534 // against the video display, although the latter is not as critical.)
535 // This requires our system clock to be reasonably close to the
536 // video clock, but that's not an unreasonable assumption.
537 timespec second_field_start;
538 second_field_start.tv_nsec = frame_upload_start.tv_nsec +
539 frame_length * 1000000000 / TIMEBASE;
540 second_field_start.tv_sec = frame_upload_start.tv_sec +
541 second_field_start.tv_nsec / 1000000000;
542 second_field_start.tv_nsec %= 1000000000;
544 while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
545 &second_field_start, nullptr) == -1 &&
550 unique_lock<mutex> lock(bmusb_mutex);
551 CaptureCard::NewFrame new_frame;
552 new_frame.frame = frame;
553 new_frame.length = frame_length;
554 new_frame.field = field;
555 new_frame.interlaced = video_format.interlaced;
556 new_frame.ready_fence = fence;
557 new_frame.dropped_frames = dropped_frames;
558 card->new_frames.push(move(new_frame));
559 card->new_frames_changed.notify_all();
564 void Mixer::thread_func()
566 eglBindAPI(EGL_OPENGL_API);
567 QOpenGLContext *context = create_context(mixer_surface);
568 if (!make_current(context, mixer_surface)) {
573 struct timespec start, now;
574 clock_gettime(CLOCK_MONOTONIC, &start);
577 int stats_dropped_frames = 0;
579 while (!should_quit) {
580 CaptureCard::NewFrame new_frames[MAX_CARDS];
581 bool has_new_frame[MAX_CARDS] = { false };
582 int num_samples[MAX_CARDS] = { 0 };
585 unique_lock<mutex> lock(bmusb_mutex);
587 // The first card is the master timer, so wait for it to have a new frame.
588 // TODO: Make configurable, and with a timeout.
589 cards[0].new_frames_changed.wait(lock, [this]{ return !cards[0].new_frames.empty(); });
591 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
592 CaptureCard *card = &cards[card_index];
593 if (card->new_frames.empty()) {
596 new_frames[card_index] = move(card->new_frames.front());
597 has_new_frame[card_index] = true;
598 card->new_frames.pop();
599 card->new_frames_changed.notify_all();
601 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
602 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
603 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
604 assert(num_samples[card_index] >= 0);
608 // Resample the audio as needed, including from previously dropped frames.
609 assert(num_cards > 0);
610 for (unsigned frame_num = 0; frame_num < new_frames[0].dropped_frames + 1; ++frame_num) {
612 // Signal to the audio thread to process this frame.
613 unique_lock<mutex> lock(audio_mutex);
614 audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
615 audio_task_queue_changed.notify_one();
617 if (frame_num != new_frames[0].dropped_frames) {
618 // For dropped frames, increase the pts. Note that if the format changed
619 // in the meantime, we have no way of detecting that; we just have to
620 // assume the frame length is always the same.
621 ++stats_dropped_frames;
622 pts_int += new_frames[0].length;
626 if (audio_level_callback != nullptr) {
627 unique_lock<mutex> lock(compressor_mutex);
628 double loudness_s = r128.loudness_S();
629 double loudness_i = r128.integrated();
630 double loudness_range_low = r128.range_min();
631 double loudness_range_high = r128.range_max();
633 audio_level_callback(loudness_s, 20.0 * log10(peak),
634 loudness_i, loudness_range_low, loudness_range_high,
635 gain_staging_db, 20.0 * log10(final_makeup_gain),
636 correlation.get_correlation());
639 for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
640 if (!has_new_frame[card_index]) {
643 if (new_frames[card_index].frame->len == 0) {
644 ++new_frames[card_index].dropped_frames;
646 if (new_frames[card_index].dropped_frames > 0) {
647 printf("Card %u dropped %d frames before this\n",
648 card_index, int(new_frames[card_index].dropped_frames));
652 // If the first card is reporting a corrupted or otherwise dropped frame,
653 // just increase the pts (skipping over this frame) and don't try to compute anything new.
654 if (new_frames[0].frame->len == 0) {
655 ++stats_dropped_frames;
656 pts_int += new_frames[0].length;
660 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
661 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
664 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
665 assert(new_frame->frame != nullptr);
666 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
669 // The new texture might still be uploaded,
670 // tell the GPU to wait until it's there.
671 if (new_frame->ready_fence) {
672 glWaitSync(new_frame->ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
674 glDeleteSync(new_frame->ready_fence);
679 // Get the main chain from the theme, and set its state immediately.
680 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
681 EffectChain *chain = theme_main_chain.chain;
682 theme_main_chain.setup_chain();
683 //theme_main_chain.chain->enable_phase_timing(true);
685 GLuint y_tex, cbcr_tex;
686 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
689 // Render main chain.
690 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
691 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
692 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
694 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
695 resource_pool->release_fbo(fbo);
697 subsample_chroma(cbcr_full_tex, cbcr_tex);
698 resource_pool->release_2d_texture(cbcr_full_tex);
700 // Set the right state for rgba_tex.
701 glBindFramebuffer(GL_FRAMEBUFFER, 0);
702 glBindTexture(GL_TEXTURE_2D, rgba_tex);
703 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
704 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
705 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
707 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
710 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
711 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
713 pts_int += new_frames[0].length;
715 // The live frame just shows the RGBA texture we just rendered.
716 // It owns rgba_tex now.
717 DisplayFrame live_frame;
718 live_frame.chain = display_chain.get();
719 live_frame.setup_chain = [this, rgba_tex]{
720 display_input->set_texture_num(rgba_tex);
722 live_frame.ready_fence = fence;
723 live_frame.input_frames = {};
724 live_frame.temp_textures = { rgba_tex };
725 output_channel[OUTPUT_LIVE].output_frame(live_frame);
727 // Set up preview and any additional channels.
728 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
729 DisplayFrame display_frame;
730 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
731 display_frame.chain = chain.chain;
732 display_frame.setup_chain = chain.setup_chain;
733 display_frame.ready_fence = fence;
734 display_frame.input_frames = chain.input_frames;
735 display_frame.temp_textures = {};
736 output_channel[i].output_frame(display_frame);
739 clock_gettime(CLOCK_MONOTONIC, &now);
740 double elapsed = now.tv_sec - start.tv_sec +
741 1e-9 * (now.tv_nsec - start.tv_nsec);
742 if (frame % 100 == 0) {
743 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
744 frame, stats_dropped_frames, elapsed, frame / elapsed,
745 1e3 * elapsed / frame);
746 // chain->print_phase_timing();
749 if (should_cut.exchange(false)) { // Test and clear.
750 string filename = generate_local_dump_filename(frame);
751 printf("Starting new recording: %s\n", filename.c_str());
752 h264_encoder->shutdown();
753 httpd.close_output_file();
754 httpd.open_output_file(filename.c_str());
755 h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
759 // Reset every 100 frames, so that local variations in frame times
760 // (especially for the first few frames, when the shaders are
761 // compiled etc.) don't make it hard to measure for the entire
762 // remaining duration of the program.
763 if (frame == 10000) {
771 resource_pool->clean_context();
774 void Mixer::audio_thread_func()
776 while (!should_quit) {
780 unique_lock<mutex> lock(audio_mutex);
781 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
782 task = audio_task_queue.front();
783 audio_task_queue.pop();
786 process_audio_one_frame(task.pts_int, task.num_samples);
790 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
792 vector<float> samples_card;
793 vector<float> samples_out;
795 // TODO: Allow mixing audio from several sources.
796 unsigned selected_audio_card = theme->map_signal(audio_source_channel);
797 assert(selected_audio_card < num_cards);
799 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
800 samples_card.resize(num_samples * 2);
802 unique_lock<mutex> lock(cards[card_index].audio_mutex);
803 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
804 printf("Card %d reported previous underrun.\n", card_index);
807 if (card_index == selected_audio_card) {
808 samples_out = move(samples_card);
812 // Cut away everything under 120 Hz (or whatever the cutoff is);
813 // we don't need it for voice, and it will reduce headroom
814 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
815 // should be dampened.)
817 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
820 // Apply a level compressor to get the general level right.
821 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
822 // (or more precisely, near it, since we don't use infinite ratio),
823 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
824 // entirely arbitrary, but from practical tests with speech, it seems to
825 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
827 unique_lock<mutex> lock(compressor_mutex);
828 if (level_compressor_enabled) {
829 float threshold = 0.01f; // -40 dBFS.
831 float attack_time = 0.5f;
832 float release_time = 20.0f;
833 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
834 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
835 gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
837 // Just apply the gain we already had.
838 float g = pow(10.0f, gain_staging_db / 20.0f);
839 for (size_t i = 0; i < samples_out.size(); ++i) {
846 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
847 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
848 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
849 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
852 // float limiter_att, compressor_att;
854 // The real compressor.
855 if (compressor_enabled) {
856 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
858 float attack_time = 0.005f;
859 float release_time = 0.040f;
860 float makeup_gain = 2.0f; // +6 dB.
861 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
862 // compressor_att = compressor.get_attenuation();
865 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
866 // Note that since ratio is not infinite, we could go slightly higher than this.
867 if (limiter_enabled) {
868 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
870 float attack_time = 0.0f; // Instant.
871 float release_time = 0.020f;
872 float makeup_gain = 1.0f; // 0 dB.
873 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
874 // limiter_att = limiter.get_attenuation();
877 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
879 // Upsample 4x to find interpolated peak.
880 peak_resampler.inp_data = samples_out.data();
881 peak_resampler.inp_count = samples_out.size() / 2;
883 vector<float> interpolated_samples_out;
884 interpolated_samples_out.resize(samples_out.size());
885 while (peak_resampler.inp_count > 0) { // About four iterations.
886 peak_resampler.out_data = &interpolated_samples_out[0];
887 peak_resampler.out_count = interpolated_samples_out.size() / 2;
888 peak_resampler.process();
889 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
890 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
891 peak_resampler.out_data = nullptr;
894 // At this point, we are most likely close to +0 LU, but all of our
895 // measurements have been on raw sample values, not R128 values.
896 // So we have a final makeup gain to get us to +0 LU; the gain
897 // adjustments required should be relatively small, and also, the
898 // offset shouldn't change much (only if the type of audio changes
899 // significantly). Thus, we shoot for updating this value basically
900 // “whenever we process buffers”, since the R128 calculation isn't exactly
901 // something we get out per-sample.
903 // Note that there's a feedback loop here, so we choose a very slow filter
904 // (half-time of 100 seconds).
905 double target_loudness_factor, alpha;
907 unique_lock<mutex> lock(compressor_mutex);
908 double loudness_lu = r128.loudness_M() - ref_level_lufs;
909 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
910 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
912 // If we're outside +/- 5 LU uncorrected, we don't count it as
913 // a normal signal (probably silence) and don't change the
914 // correction factor; just apply what we already have.
915 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
918 // Formula adapted from
919 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
920 const double half_time_s = 100.0;
921 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
922 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
925 double m = final_makeup_gain;
926 for (size_t i = 0; i < samples_out.size(); i += 2) {
927 samples_out[i + 0] *= m;
928 samples_out[i + 1] *= m;
929 m += (target_loudness_factor - m) * alpha;
931 final_makeup_gain = m;
934 // Find R128 levels and L/R correlation.
935 vector<float> left, right;
936 deinterleave_samples(samples_out, &left, &right);
937 float *ptrs[] = { left.data(), right.data() };
939 unique_lock<mutex> lock(compressor_mutex);
940 r128.process(left.size(), ptrs);
941 correlation.process_samples(samples_out);
944 // Send the samples to the sound card.
946 alsa->write(samples_out);
949 // And finally add them to the output.
950 h264_encoder->add_audio(frame_pts_int, move(samples_out));
953 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
956 glGenVertexArrays(1, &vao);
959 glBindVertexArray(vao);
963 GLuint fbo = resource_pool->create_fbo(dst_tex);
964 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
965 glViewport(0, 0, WIDTH/2, HEIGHT/2);
968 glUseProgram(cbcr_program_num);
971 glActiveTexture(GL_TEXTURE0);
973 glBindTexture(GL_TEXTURE_2D, src_tex);
975 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
977 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
979 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
982 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
983 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
985 glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
988 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
989 glEnableVertexAttribArray(attr_index);
991 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
995 glDrawArrays(GL_TRIANGLES, 0, 3);
998 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
999 glDisableVertexAttribArray(attr_index);
1005 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1008 resource_pool->release_fbo(fbo);
1009 glDeleteVertexArrays(1, &vao);
1012 void Mixer::release_display_frame(DisplayFrame *frame)
1014 for (GLuint texnum : frame->temp_textures) {
1015 resource_pool->release_2d_texture(texnum);
1017 frame->temp_textures.clear();
1018 frame->ready_fence.reset();
1019 frame->input_frames.clear();
1024 mixer_thread = thread(&Mixer::thread_func, this);
1025 audio_thread = thread(&Mixer::audio_thread_func, this);
1031 mixer_thread.join();
1032 audio_thread.join();
1035 void Mixer::transition_clicked(int transition_num)
1037 theme->transition_clicked(transition_num, pts());
1040 void Mixer::channel_clicked(int preview_num)
1042 theme->channel_clicked(preview_num);
1045 void Mixer::reset_meters()
1047 peak_resampler.reset();
1050 r128.integr_start();
1051 correlation.reset();
1054 void Mixer::start_mode_scanning(unsigned card_index)
1056 assert(card_index < num_cards);
1057 if (is_mode_scanning[card_index]) {
1060 is_mode_scanning[card_index] = true;
1061 mode_scanlist[card_index].clear();
1062 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1063 mode_scanlist[card_index].push_back(mode.first);
1065 assert(!mode_scanlist[card_index].empty());
1066 mode_scanlist_index[card_index] = 0;
1067 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1068 clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1071 Mixer::OutputChannel::~OutputChannel()
1073 if (has_current_frame) {
1074 parent->release_display_frame(¤t_frame);
1076 if (has_ready_frame) {
1077 parent->release_display_frame(&ready_frame);
1081 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1083 // Store this frame for display. Remove the ready frame if any
1084 // (it was seemingly never used).
1086 unique_lock<mutex> lock(frame_mutex);
1087 if (has_ready_frame) {
1088 parent->release_display_frame(&ready_frame);
1090 ready_frame = frame;
1091 has_ready_frame = true;
1094 if (has_new_frame_ready_callback) {
1095 new_frame_ready_callback();
1099 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1101 unique_lock<mutex> lock(frame_mutex);
1102 if (!has_current_frame && !has_ready_frame) {
1106 if (has_current_frame && has_ready_frame) {
1107 // We have a new ready frame. Toss the current one.
1108 parent->release_display_frame(¤t_frame);
1109 has_current_frame = false;
1111 if (has_ready_frame) {
1112 assert(!has_current_frame);
1113 current_frame = ready_frame;
1114 ready_frame.ready_fence.reset(); // Drop the refcount.
1115 ready_frame.input_frames.clear(); // Drop the refcounts.
1116 has_current_frame = true;
1117 has_ready_frame = false;
1120 *frame = current_frame;
1124 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1126 new_frame_ready_callback = callback;
1127 has_new_frame_ready_callback = true;