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 "video_encoder.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 };
97 void QueueLengthPolicy::update_policy(int queue_length)
99 if (queue_length < 0) { // Starvation.
100 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
102 fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
103 card_index, safe_queue_length);
105 frames_with_at_least_one = 0;
106 been_at_safe_point_since_last_starvation = false;
109 if (queue_length > 0) {
110 if (queue_length >= int(safe_queue_length)) {
111 been_at_safe_point_since_last_starvation = true;
113 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
115 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
116 card_index, safe_queue_length);
117 frames_with_at_least_one = 0;
120 frames_with_at_least_one = 0;
124 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
126 num_cards(num_cards),
127 mixer_surface(create_surface(format)),
128 h264_encoder_surface(create_surface(format)),
129 correlation(OUTPUT_FREQUENCY),
130 level_compressor(OUTPUT_FREQUENCY),
131 limiter(OUTPUT_FREQUENCY),
132 compressor(OUTPUT_FREQUENCY)
134 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
137 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
138 // will be halved when sampling them, and we need to compensate here.
139 movit_texel_subpixel_precision /= 2.0;
141 resource_pool.reset(new ResourcePool);
142 theme.reset(new Theme(global_flags.theme_filename.c_str(), resource_pool.get(), num_cards));
143 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
144 output_channel[i].parent = this;
147 ImageFormat inout_format;
148 inout_format.color_space = COLORSPACE_sRGB;
149 inout_format.gamma_curve = GAMMA_sRGB;
151 // Display chain; shows the live output produced by the main chain (its RGBA version).
152 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
154 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
155 display_chain->add_input(display_input);
156 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
157 display_chain->set_dither_bits(0); // Don't bother.
158 display_chain->finalize();
160 video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
162 // Start listening for clients only once VideoEncoder has written its header, if any.
165 // First try initializing the PCI devices, then USB, until we have the desired number of cards.
166 unsigned num_pci_devices = 0, num_usb_devices = 0;
167 unsigned card_index = 0;
169 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
170 if (decklink_iterator != nullptr) {
171 for ( ; card_index < num_cards; ++card_index) {
173 if (decklink_iterator->Next(&decklink) != S_OK) {
177 configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
180 decklink_iterator->Release();
181 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
183 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
185 for ( ; card_index < num_cards; ++card_index) {
186 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
190 if (num_usb_devices > 0) {
191 BMUSBCapture::start_bm_thread();
194 for (card_index = 0; card_index < num_cards; ++card_index) {
195 cards[card_index].queue_length_policy.reset(card_index);
196 cards[card_index].capture->start_bm_capture();
199 // Set up stuff for NV12 conversion.
202 string cbcr_vert_shader =
205 "in vec2 position; \n"
206 "in vec2 texcoord; \n"
208 "uniform vec2 foo_chroma_offset_0; \n"
212 " // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
214 " // 2.000 0.000 0.000 -1.000 \n"
215 " // 0.000 2.000 0.000 -1.000 \n"
216 " // 0.000 0.000 -2.000 -1.000 \n"
217 " // 0.000 0.000 0.000 1.000 \n"
218 " gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
219 " vec2 flipped_tc = texcoord; \n"
220 " tc0 = flipped_tc + foo_chroma_offset_0; \n"
222 string cbcr_frag_shader =
225 "uniform sampler2D cbcr_tex; \n"
226 "out vec4 FragColor; \n"
228 " FragColor = texture(cbcr_tex, tc0); \n"
230 vector<string> frag_shader_outputs;
231 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
238 cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
239 cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
240 cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
242 r128.init(2, OUTPUT_FREQUENCY);
245 locut.init(FILTER_HPF, 2);
247 // If --flat-audio is given, turn off everything that messes with the sound,
248 // except the final makeup gain.
249 if (global_flags.flat_audio) {
250 set_locut_enabled(false);
251 set_limiter_enabled(false);
252 set_compressor_enabled(false);
255 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
256 // and there's a limit to how important the peak meter is.
257 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
259 if (global_flags.enable_alsa_output) {
260 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
266 resource_pool->release_glsl_program(cbcr_program_num);
267 glDeleteBuffers(1, &cbcr_vbo);
268 BMUSBCapture::stop_bm_thread();
270 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
272 unique_lock<mutex> lock(bmusb_mutex);
273 cards[card_index].should_quit = true; // Unblock thread.
274 cards[card_index].new_frames_changed.notify_all();
276 cards[card_index].capture->stop_dequeue_thread();
279 video_encoder.reset(nullptr);
282 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
284 printf("Configuring card %d...\n", card_index);
286 CaptureCard *card = &cards[card_index];
287 card->capture = capture;
288 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
289 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
290 card->capture->set_video_frame_allocator(card->frame_allocator.get());
291 card->surface = create_surface(format);
292 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
293 card->capture->configure_card();
299 int unwrap_timecode(uint16_t current_wrapped, int last)
301 uint16_t last_wrapped = last & 0xffff;
302 if (current_wrapped > last_wrapped) {
303 return (last & ~0xffff) | current_wrapped;
305 return 0x10000 + ((last & ~0xffff) | current_wrapped);
309 float find_peak(const float *samples, size_t num_samples)
311 float m = fabs(samples[0]);
312 for (size_t i = 1; i < num_samples; ++i) {
313 m = max(m, fabs(samples[i]));
318 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
320 size_t num_samples = in.size() / 2;
321 out_l->resize(num_samples);
322 out_r->resize(num_samples);
324 const float *inptr = in.data();
325 float *lptr = &(*out_l)[0];
326 float *rptr = &(*out_r)[0];
327 for (size_t i = 0; i < num_samples; ++i) {
335 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
336 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
337 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
339 CaptureCard *card = &cards[card_index];
341 if (is_mode_scanning[card_index]) {
342 if (video_format.has_signal) {
343 // Found a stable signal, so stop scanning.
344 is_mode_scanning[card_index] = false;
346 static constexpr double switch_time_s = 0.5; // Should be enough time for the signal to stabilize.
348 clock_gettime(CLOCK_MONOTONIC, &now);
349 double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
350 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
351 if (sec_since_last_switch > switch_time_s) {
352 // It isn't this mode; try the next one.
353 mode_scanlist_index[card_index]++;
354 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
355 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
356 last_mode_scan_change[card_index] = now;
361 int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
363 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;
364 if (num_samples > OUTPUT_FREQUENCY / 10) {
365 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",
366 card_index, int(audio_frame.len), int(audio_offset),
367 timecode, int(video_frame.len), int(video_offset), video_format.id);
368 if (video_frame.owner) {
369 video_frame.owner->release_frame(video_frame);
371 if (audio_frame.owner) {
372 audio_frame.owner->release_frame(audio_frame);
377 int64_t local_pts = card->next_local_pts;
378 int dropped_frames = 0;
379 if (card->last_timecode != -1) {
380 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
383 // Convert the audio to stereo fp32 and add it.
385 audio.resize(num_samples * 2);
386 switch (audio_format.bits_per_sample) {
388 assert(num_samples == 0);
391 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
394 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
397 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
403 unique_lock<mutex> lock(card->audio_mutex);
405 // Number of samples per frame if we need to insert silence.
406 // (Could be nonintegral, but resampling will save us then.)
407 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
409 if (dropped_frames > MAX_FPS * 2) {
410 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
411 card_index, card->last_timecode, timecode);
412 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
414 } else if (dropped_frames > 0) {
415 // Insert silence as needed.
416 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
417 card_index, dropped_frames, timecode);
418 vector<float> silence(silence_samples * 2, 0.0f);
419 for (int i = 0; i < dropped_frames; ++i) {
420 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
421 // Note that if the format changed in the meantime, we have
422 // no way of detecting that; we just have to assume the frame length
423 // is always the same.
424 local_pts += frame_length;
427 if (num_samples == 0) {
428 audio.resize(silence_samples * 2);
429 num_samples = silence_samples;
431 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
432 card->next_local_pts = local_pts + frame_length;
435 card->last_timecode = timecode;
437 // Done with the audio, so release it.
438 if (audio_frame.owner) {
439 audio_frame.owner->release_frame(audio_frame);
442 size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
443 if (video_frame.len - video_offset == 0 ||
444 video_frame.len - video_offset != expected_length) {
445 if (video_frame.len != 0) {
446 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
447 card_index, video_frame.len - video_offset, expected_length);
449 if (video_frame.owner) {
450 video_frame.owner->release_frame(video_frame);
453 // Still send on the information that we _had_ a frame, even though it's corrupted,
454 // so that pts can go up accordingly.
456 unique_lock<mutex> lock(bmusb_mutex);
457 CaptureCard::NewFrame new_frame;
458 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
459 new_frame.length = frame_length;
460 new_frame.interlaced = false;
461 new_frame.dropped_frames = dropped_frames;
462 card->new_frames.push(move(new_frame));
463 card->new_frames_changed.notify_all();
468 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
470 unsigned num_fields = video_format.interlaced ? 2 : 1;
471 timespec frame_upload_start;
472 if (video_format.interlaced) {
473 // Send the two fields along as separate frames; the other side will need to add
474 // a deinterlacer to actually get this right.
475 assert(video_format.height % 2 == 0);
476 video_format.height /= 2;
477 assert(frame_length % 2 == 0);
480 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
482 userdata->last_interlaced = video_format.interlaced;
483 userdata->last_has_signal = video_format.has_signal;
484 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
485 userdata->last_frame_rate_den = video_format.frame_rate_den;
486 RefCountedFrame frame(video_frame);
488 // Upload the textures.
489 size_t cbcr_width = video_format.width / 2;
490 size_t cbcr_offset = video_offset / 2;
491 size_t y_offset = video_frame.size / 2 + video_offset / 2;
493 for (unsigned field = 0; field < num_fields; ++field) {
494 // Put the actual texture upload in a lambda that is executed in the main thread.
495 // It is entirely possible to do this in the same thread (and it might even be
496 // faster, depending on the GPU and driver), but it appears to be trickling
497 // driver bugs very easily.
499 // Note that this means we must hold on to the actual frame data in <userdata>
500 // until the upload command is run, but we hold on to <frame> much longer than that
501 // (in fact, all the way until we no longer use the texture in rendering).
502 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
503 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
505 if (userdata->tex_y[field] == 0 ||
506 userdata->tex_cbcr[field] == 0 ||
507 video_format.width != userdata->last_width[field] ||
508 video_format.height != userdata->last_height[field]) {
509 // We changed resolution since last use of this texture, so we need to create
510 // a new object. Note that this each card has its own PBOFrameAllocator,
511 // we don't need to worry about these flip-flopping between resolutions.
512 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
514 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
516 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
518 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
520 userdata->last_width[field] = video_format.width;
521 userdata->last_height[field] = video_format.height;
524 GLuint pbo = userdata->pbo;
526 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
529 size_t field_y_start = y_offset + video_format.width * field_start_line;
530 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
532 if (global_flags.flush_pbos) {
533 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
535 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
539 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
541 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
543 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
545 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
547 glBindTexture(GL_TEXTURE_2D, 0);
549 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
554 // Don't upload the second field as fast as we can; wait until
555 // the field time has approximately passed. (Otherwise, we could
556 // get timing jitter against the other sources, and possibly also
557 // against the video display, although the latter is not as critical.)
558 // This requires our system clock to be reasonably close to the
559 // video clock, but that's not an unreasonable assumption.
560 timespec second_field_start;
561 second_field_start.tv_nsec = frame_upload_start.tv_nsec +
562 frame_length * 1000000000 / TIMEBASE;
563 second_field_start.tv_sec = frame_upload_start.tv_sec +
564 second_field_start.tv_nsec / 1000000000;
565 second_field_start.tv_nsec %= 1000000000;
567 while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
568 &second_field_start, nullptr) == -1 &&
573 unique_lock<mutex> lock(bmusb_mutex);
574 CaptureCard::NewFrame new_frame;
575 new_frame.frame = frame;
576 new_frame.length = frame_length;
577 new_frame.field = field;
578 new_frame.interlaced = video_format.interlaced;
579 new_frame.upload_func = upload_func;
580 new_frame.dropped_frames = dropped_frames;
581 card->new_frames.push(move(new_frame));
582 card->new_frames_changed.notify_all();
587 void Mixer::thread_func()
589 eglBindAPI(EGL_OPENGL_API);
590 QOpenGLContext *context = create_context(mixer_surface);
591 if (!make_current(context, mixer_surface)) {
596 struct timespec start, now;
597 clock_gettime(CLOCK_MONOTONIC, &start);
600 int stats_dropped_frames = 0;
602 while (!should_quit) {
603 CaptureCard::NewFrame new_frames[MAX_CARDS];
604 bool has_new_frame[MAX_CARDS] = { false };
605 int num_samples[MAX_CARDS] = { 0 };
607 // TODO: Add a timeout.
608 unsigned master_card_index = theme->map_signal(master_clock_channel);
609 assert(master_card_index < num_cards);
611 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
612 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
613 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
614 send_audio_level_callback();
616 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
617 if (card_index == master_card_index || !has_new_frame[card_index]) {
620 if (new_frames[card_index].frame->len == 0) {
621 ++new_frames[card_index].dropped_frames;
623 if (new_frames[card_index].dropped_frames > 0) {
624 printf("Card %u dropped %d frames before this\n",
625 card_index, int(new_frames[card_index].dropped_frames));
629 // If the first card is reporting a corrupted or otherwise dropped frame,
630 // just increase the pts (skipping over this frame) and don't try to compute anything new.
631 if (new_frames[master_card_index].frame->len == 0) {
632 ++stats_dropped_frames;
633 pts_int += new_frames[master_card_index].length;
637 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
638 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
641 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
642 assert(new_frame->frame != nullptr);
643 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
646 // The new texture might need uploading before use.
647 if (new_frame->upload_func) {
648 new_frame->upload_func();
649 new_frame->upload_func = nullptr;
653 int64_t duration = new_frames[master_card_index].length;
654 render_one_frame(duration);
658 clock_gettime(CLOCK_MONOTONIC, &now);
659 double elapsed = now.tv_sec - start.tv_sec +
660 1e-9 * (now.tv_nsec - start.tv_nsec);
661 if (frame % 100 == 0) {
662 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
663 frame, stats_dropped_frames, elapsed, frame / elapsed,
664 1e3 * elapsed / frame);
665 // chain->print_phase_timing();
668 if (should_cut.exchange(false)) { // Test and clear.
669 video_encoder->do_cut(frame);
673 // Reset every 100 frames, so that local variations in frame times
674 // (especially for the first few frames, when the shaders are
675 // compiled etc.) don't make it hard to measure for the entire
676 // remaining duration of the program.
677 if (frame == 10000) {
685 resource_pool->clean_context();
688 void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_CARDS], bool has_new_frame[MAX_CARDS], int num_samples[MAX_CARDS])
690 // The first card is the master timer, so wait for it to have a new frame.
691 unique_lock<mutex> lock(bmusb_mutex);
692 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty(); });
694 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
695 CaptureCard *card = &cards[card_index];
696 if (card->new_frames.empty()) {
697 assert(card_index != master_card_index);
698 card->queue_length_policy.update_policy(-1);
701 new_frames[card_index] = move(card->new_frames.front());
702 has_new_frame[card_index] = true;
703 card->new_frames.pop();
704 card->new_frames_changed.notify_all();
706 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
707 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
708 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
709 assert(num_samples[card_index] >= 0);
711 if (card_index == master_card_index) {
712 // We don't use the queue length policy for the master card,
713 // but we will if it stops being the master. Thus, clear out
714 // the policy in case we switch in the future.
715 card->queue_length_policy.reset(card_index);
717 // If we have excess frames compared to the policy for this card,
718 // drop frames from the head.
719 card->queue_length_policy.update_policy(card->new_frames.size());
720 while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
721 card->new_frames.pop();
727 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
729 // Resample the audio as needed, including from previously dropped frames.
730 assert(num_cards > 0);
731 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
733 // Signal to the audio thread to process this frame.
734 unique_lock<mutex> lock(audio_mutex);
735 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
736 audio_task_queue_changed.notify_one();
738 if (frame_num != dropped_frames) {
739 // For dropped frames, increase the pts. Note that if the format changed
740 // in the meantime, we have no way of detecting that; we just have to
741 // assume the frame length is always the same.
742 pts_int += length_per_frame;
747 void Mixer::render_one_frame(int64_t duration)
749 // Get the main chain from the theme, and set its state immediately.
750 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
751 EffectChain *chain = theme_main_chain.chain;
752 theme_main_chain.setup_chain();
753 //theme_main_chain.chain->enable_phase_timing(true);
755 GLuint y_tex, cbcr_tex;
756 bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
759 // Render main chain.
760 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
761 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
762 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
764 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
765 resource_pool->release_fbo(fbo);
767 subsample_chroma(cbcr_full_tex, cbcr_tex);
768 resource_pool->release_2d_texture(cbcr_full_tex);
770 // Set the right state for rgba_tex.
771 glBindFramebuffer(GL_FRAMEBUFFER, 0);
772 glBindTexture(GL_TEXTURE_2D, rgba_tex);
773 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
774 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
775 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
777 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
778 RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
780 // The live frame just shows the RGBA texture we just rendered.
781 // It owns rgba_tex now.
782 DisplayFrame live_frame;
783 live_frame.chain = display_chain.get();
784 live_frame.setup_chain = [this, rgba_tex]{
785 display_input->set_texture_num(rgba_tex);
787 live_frame.ready_fence = fence;
788 live_frame.input_frames = {};
789 live_frame.temp_textures = { rgba_tex };
790 output_channel[OUTPUT_LIVE].output_frame(live_frame);
792 // Set up preview and any additional channels.
793 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
794 DisplayFrame display_frame;
795 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
796 display_frame.chain = chain.chain;
797 display_frame.setup_chain = chain.setup_chain;
798 display_frame.ready_fence = fence;
799 display_frame.input_frames = chain.input_frames;
800 display_frame.temp_textures = {};
801 output_channel[i].output_frame(display_frame);
805 void Mixer::send_audio_level_callback()
807 if (audio_level_callback == nullptr) {
811 unique_lock<mutex> lock(compressor_mutex);
812 double loudness_s = r128.loudness_S();
813 double loudness_i = r128.integrated();
814 double loudness_range_low = r128.range_min();
815 double loudness_range_high = r128.range_max();
817 audio_level_callback(loudness_s, 20.0 * log10(peak),
818 loudness_i, loudness_range_low, loudness_range_high,
819 gain_staging_db, 20.0 * log10(final_makeup_gain),
820 correlation.get_correlation());
823 void Mixer::audio_thread_func()
825 while (!should_quit) {
829 unique_lock<mutex> lock(audio_mutex);
830 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
831 task = audio_task_queue.front();
832 audio_task_queue.pop();
835 process_audio_one_frame(task.pts_int, task.num_samples);
839 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
841 vector<float> samples_card;
842 vector<float> samples_out;
844 // TODO: Allow mixing audio from several sources.
845 unsigned selected_audio_card = theme->map_signal(audio_source_channel);
846 assert(selected_audio_card < num_cards);
848 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
849 samples_card.resize(num_samples * 2);
851 unique_lock<mutex> lock(cards[card_index].audio_mutex);
852 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
853 printf("Card %d reported previous underrun.\n", card_index);
856 if (card_index == selected_audio_card) {
857 samples_out = move(samples_card);
861 // Cut away everything under 120 Hz (or whatever the cutoff is);
862 // we don't need it for voice, and it will reduce headroom
863 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
864 // should be dampened.)
866 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
869 // Apply a level compressor to get the general level right.
870 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
871 // (or more precisely, near it, since we don't use infinite ratio),
872 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
873 // entirely arbitrary, but from practical tests with speech, it seems to
874 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
876 unique_lock<mutex> lock(compressor_mutex);
877 if (level_compressor_enabled) {
878 float threshold = 0.01f; // -40 dBFS.
880 float attack_time = 0.5f;
881 float release_time = 20.0f;
882 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
883 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
884 gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
886 // Just apply the gain we already had.
887 float g = pow(10.0f, gain_staging_db / 20.0f);
888 for (size_t i = 0; i < samples_out.size(); ++i) {
895 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
896 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
897 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
898 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
901 // float limiter_att, compressor_att;
903 // The real compressor.
904 if (compressor_enabled) {
905 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
907 float attack_time = 0.005f;
908 float release_time = 0.040f;
909 float makeup_gain = 2.0f; // +6 dB.
910 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
911 // compressor_att = compressor.get_attenuation();
914 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
915 // Note that since ratio is not infinite, we could go slightly higher than this.
916 if (limiter_enabled) {
917 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
919 float attack_time = 0.0f; // Instant.
920 float release_time = 0.020f;
921 float makeup_gain = 1.0f; // 0 dB.
922 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
923 // limiter_att = limiter.get_attenuation();
926 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
928 // Upsample 4x to find interpolated peak.
929 peak_resampler.inp_data = samples_out.data();
930 peak_resampler.inp_count = samples_out.size() / 2;
932 vector<float> interpolated_samples_out;
933 interpolated_samples_out.resize(samples_out.size());
934 while (peak_resampler.inp_count > 0) { // About four iterations.
935 peak_resampler.out_data = &interpolated_samples_out[0];
936 peak_resampler.out_count = interpolated_samples_out.size() / 2;
937 peak_resampler.process();
938 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
939 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
940 peak_resampler.out_data = nullptr;
943 // At this point, we are most likely close to +0 LU, but all of our
944 // measurements have been on raw sample values, not R128 values.
945 // So we have a final makeup gain to get us to +0 LU; the gain
946 // adjustments required should be relatively small, and also, the
947 // offset shouldn't change much (only if the type of audio changes
948 // significantly). Thus, we shoot for updating this value basically
949 // “whenever we process buffers”, since the R128 calculation isn't exactly
950 // something we get out per-sample.
952 // Note that there's a feedback loop here, so we choose a very slow filter
953 // (half-time of 100 seconds).
954 double target_loudness_factor, alpha;
956 unique_lock<mutex> lock(compressor_mutex);
957 double loudness_lu = r128.loudness_M() - ref_level_lufs;
958 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
959 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
961 // If we're outside +/- 5 LU uncorrected, we don't count it as
962 // a normal signal (probably silence) and don't change the
963 // correction factor; just apply what we already have.
964 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
967 // Formula adapted from
968 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
969 const double half_time_s = 100.0;
970 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
971 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
974 double m = final_makeup_gain;
975 for (size_t i = 0; i < samples_out.size(); i += 2) {
976 samples_out[i + 0] *= m;
977 samples_out[i + 1] *= m;
978 m += (target_loudness_factor - m) * alpha;
980 final_makeup_gain = m;
983 // Find R128 levels and L/R correlation.
984 vector<float> left, right;
985 deinterleave_samples(samples_out, &left, &right);
986 float *ptrs[] = { left.data(), right.data() };
988 unique_lock<mutex> lock(compressor_mutex);
989 r128.process(left.size(), ptrs);
990 correlation.process_samples(samples_out);
993 // Send the samples to the sound card.
995 alsa->write(samples_out);
998 // And finally add them to the output.
999 video_encoder->add_audio(frame_pts_int, move(samples_out));
1002 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1005 glGenVertexArrays(1, &vao);
1008 glBindVertexArray(vao);
1012 GLuint fbo = resource_pool->create_fbo(dst_tex);
1013 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1014 glViewport(0, 0, WIDTH/2, HEIGHT/2);
1017 glUseProgram(cbcr_program_num);
1020 glActiveTexture(GL_TEXTURE0);
1022 glBindTexture(GL_TEXTURE_2D, src_tex);
1024 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1026 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1028 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1031 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1032 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1034 glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1037 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1038 glEnableVertexAttribArray(attr_index);
1040 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1044 glDrawArrays(GL_TRIANGLES, 0, 3);
1047 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1048 glDisableVertexAttribArray(attr_index);
1054 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1057 resource_pool->release_fbo(fbo);
1058 glDeleteVertexArrays(1, &vao);
1061 void Mixer::release_display_frame(DisplayFrame *frame)
1063 for (GLuint texnum : frame->temp_textures) {
1064 resource_pool->release_2d_texture(texnum);
1066 frame->temp_textures.clear();
1067 frame->ready_fence.reset();
1068 frame->input_frames.clear();
1073 mixer_thread = thread(&Mixer::thread_func, this);
1074 audio_thread = thread(&Mixer::audio_thread_func, this);
1080 mixer_thread.join();
1081 audio_thread.join();
1084 void Mixer::transition_clicked(int transition_num)
1086 theme->transition_clicked(transition_num, pts());
1089 void Mixer::channel_clicked(int preview_num)
1091 theme->channel_clicked(preview_num);
1094 void Mixer::reset_meters()
1096 peak_resampler.reset();
1099 r128.integr_start();
1100 correlation.reset();
1103 void Mixer::start_mode_scanning(unsigned card_index)
1105 assert(card_index < num_cards);
1106 if (is_mode_scanning[card_index]) {
1109 is_mode_scanning[card_index] = true;
1110 mode_scanlist[card_index].clear();
1111 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1112 mode_scanlist[card_index].push_back(mode.first);
1114 assert(!mode_scanlist[card_index].empty());
1115 mode_scanlist_index[card_index] = 0;
1116 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1117 clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1120 Mixer::OutputChannel::~OutputChannel()
1122 if (has_current_frame) {
1123 parent->release_display_frame(¤t_frame);
1125 if (has_ready_frame) {
1126 parent->release_display_frame(&ready_frame);
1130 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1132 // Store this frame for display. Remove the ready frame if any
1133 // (it was seemingly never used).
1135 unique_lock<mutex> lock(frame_mutex);
1136 if (has_ready_frame) {
1137 parent->release_display_frame(&ready_frame);
1139 ready_frame = frame;
1140 has_ready_frame = true;
1143 if (has_new_frame_ready_callback) {
1144 new_frame_ready_callback();
1148 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1150 unique_lock<mutex> lock(frame_mutex);
1151 if (!has_current_frame && !has_ready_frame) {
1155 if (has_current_frame && has_ready_frame) {
1156 // We have a new ready frame. Toss the current one.
1157 parent->release_display_frame(¤t_frame);
1158 has_current_frame = false;
1160 if (has_ready_frame) {
1161 assert(!has_current_frame);
1162 current_frame = ready_frame;
1163 ready_frame.ready_fence.reset(); // Drop the refcount.
1164 ready_frame.input_frames.clear(); // Drop the refcounts.
1165 has_current_frame = true;
1166 has_ready_frame = false;
1169 *frame = current_frame;
1173 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1175 new_frame_ready_callback = callback;
1176 has_new_frame_ready_callback = true;
1179 mutex RefCountedGLsync::fence_lock;