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 void QueueLengthPolicy::update_policy(int queue_length)
116 if (queue_length < 0) { // Starvation.
117 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
119 fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
120 card_index, safe_queue_length);
122 frames_with_at_least_one = 0;
123 been_at_safe_point_since_last_starvation = false;
126 if (queue_length > 0) {
127 if (queue_length >= int(safe_queue_length)) {
128 been_at_safe_point_since_last_starvation = true;
130 if (++frames_with_at_least_one >= 50 && safe_queue_length > 0) {
132 fprintf(stderr, "Card %u: Spare frames for more than 50 frames, reducing safe limit to %u frames\n",
133 card_index, safe_queue_length);
134 frames_with_at_least_one = 0;
137 frames_with_at_least_one = 0;
141 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
142 : httpd(WIDTH, HEIGHT),
143 num_cards(num_cards),
144 mixer_surface(create_surface(format)),
145 h264_encoder_surface(create_surface(format)),
146 correlation(OUTPUT_FREQUENCY),
147 level_compressor(OUTPUT_FREQUENCY),
148 limiter(OUTPUT_FREQUENCY),
149 compressor(OUTPUT_FREQUENCY)
151 httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
154 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
157 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
158 // will be halved when sampling them, and we need to compensate here.
159 movit_texel_subpixel_precision /= 2.0;
161 resource_pool.reset(new ResourcePool);
162 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
163 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
164 output_channel[i].parent = this;
167 ImageFormat inout_format;
168 inout_format.color_space = COLORSPACE_sRGB;
169 inout_format.gamma_curve = GAMMA_sRGB;
171 // Display chain; shows the live output produced by the main chain (its RGBA version).
172 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
174 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
175 display_chain->add_input(display_input);
176 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
177 display_chain->set_dither_bits(0); // Don't bother.
178 display_chain->finalize();
180 h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
182 // First try initializing the PCI devices, then USB, until we have the desired number of cards.
183 unsigned num_pci_devices = 0, num_usb_devices = 0;
184 unsigned card_index = 0;
186 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
187 if (decklink_iterator != nullptr) {
188 for ( ; card_index < num_cards; ++card_index) {
190 if (decklink_iterator->Next(&decklink) != S_OK) {
194 configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
197 decklink_iterator->Release();
198 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
200 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
202 for ( ; card_index < num_cards; ++card_index) {
203 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
207 if (num_usb_devices > 0) {
208 BMUSBCapture::start_bm_thread();
211 for (card_index = 0; card_index < num_cards; ++card_index) {
212 cards[card_index].queue_length_policy.reset(card_index);
213 cards[card_index].capture->start_bm_capture();
216 // Set up stuff for NV12 conversion.
219 string cbcr_vert_shader =
222 "in vec2 position; \n"
223 "in vec2 texcoord; \n"
225 "uniform vec2 foo_chroma_offset_0; \n"
229 " // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
231 " // 2.000 0.000 0.000 -1.000 \n"
232 " // 0.000 2.000 0.000 -1.000 \n"
233 " // 0.000 0.000 -2.000 -1.000 \n"
234 " // 0.000 0.000 0.000 1.000 \n"
235 " gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
236 " vec2 flipped_tc = texcoord; \n"
237 " tc0 = flipped_tc + foo_chroma_offset_0; \n"
239 string cbcr_frag_shader =
242 "uniform sampler2D cbcr_tex; \n"
243 "out vec4 FragColor; \n"
245 " FragColor = texture(cbcr_tex, tc0); \n"
247 vector<string> frag_shader_outputs;
248 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
255 cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
256 cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
257 cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
259 r128.init(2, OUTPUT_FREQUENCY);
262 locut.init(FILTER_HPF, 2);
264 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
265 // and there's a limit to how important the peak meter is.
266 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
268 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
273 resource_pool->release_glsl_program(cbcr_program_num);
274 glDeleteBuffers(1, &cbcr_vbo);
275 BMUSBCapture::stop_bm_thread();
277 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
279 unique_lock<mutex> lock(bmusb_mutex);
280 cards[card_index].should_quit = true; // Unblock thread.
281 cards[card_index].new_frames_changed.notify_all();
283 cards[card_index].capture->stop_dequeue_thread();
286 h264_encoder.reset(nullptr);
289 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
291 printf("Configuring card %d...\n", card_index);
293 CaptureCard *card = &cards[card_index];
294 card->capture = capture;
295 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
296 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
297 card->capture->set_video_frame_allocator(card->frame_allocator.get());
298 card->surface = create_surface(format);
299 card->capture->set_dequeue_thread_callbacks(
301 eglBindAPI(EGL_OPENGL_API);
302 card->context = create_context(card->surface);
303 if (!make_current(card->context, card->surface)) {
304 printf("failed to create bmusb context\n");
309 resource_pool->clean_context();
311 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
312 card->capture->configure_card();
318 int unwrap_timecode(uint16_t current_wrapped, int last)
320 uint16_t last_wrapped = last & 0xffff;
321 if (current_wrapped > last_wrapped) {
322 return (last & ~0xffff) | current_wrapped;
324 return 0x10000 + ((last & ~0xffff) | current_wrapped);
328 float find_peak(const float *samples, size_t num_samples)
330 float m = fabs(samples[0]);
331 for (size_t i = 1; i < num_samples; ++i) {
332 m = max(m, fabs(samples[i]));
337 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
339 size_t num_samples = in.size() / 2;
340 out_l->resize(num_samples);
341 out_r->resize(num_samples);
343 const float *inptr = in.data();
344 float *lptr = &(*out_l)[0];
345 float *rptr = &(*out_r)[0];
346 for (size_t i = 0; i < num_samples; ++i) {
354 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
355 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
356 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
358 CaptureCard *card = &cards[card_index];
360 if (is_mode_scanning[card_index]) {
361 if (video_format.has_signal) {
362 // Found a stable signal, so stop scanning.
363 is_mode_scanning[card_index] = false;
365 static constexpr double switch_time_s = 0.5; // Should be enough time for the signal to stabilize.
367 clock_gettime(CLOCK_MONOTONIC, &now);
368 double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
369 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
370 if (sec_since_last_switch > switch_time_s) {
371 // It isn't this mode; try the next one.
372 mode_scanlist_index[card_index]++;
373 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
374 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
375 last_mode_scan_change[card_index] = now;
380 int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
382 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;
383 if (num_samples > OUTPUT_FREQUENCY / 10) {
384 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",
385 card_index, int(audio_frame.len), int(audio_offset),
386 timecode, int(video_frame.len), int(video_offset), video_format.id);
387 if (video_frame.owner) {
388 video_frame.owner->release_frame(video_frame);
390 if (audio_frame.owner) {
391 audio_frame.owner->release_frame(audio_frame);
396 int64_t local_pts = card->next_local_pts;
397 int dropped_frames = 0;
398 if (card->last_timecode != -1) {
399 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
402 // Convert the audio to stereo fp32 and add it.
404 audio.resize(num_samples * 2);
405 switch (audio_format.bits_per_sample) {
407 assert(num_samples == 0);
410 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
413 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
416 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
422 unique_lock<mutex> lock(card->audio_mutex);
424 // Number of samples per frame if we need to insert silence.
425 // (Could be nonintegral, but resampling will save us then.)
426 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
428 if (dropped_frames > MAX_FPS * 2) {
429 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
430 card_index, card->last_timecode, timecode);
431 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
433 } else if (dropped_frames > 0) {
434 // Insert silence as needed.
435 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
436 card_index, dropped_frames, timecode);
437 vector<float> silence(silence_samples * 2, 0.0f);
438 for (int i = 0; i < dropped_frames; ++i) {
439 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
440 // Note that if the format changed in the meantime, we have
441 // no way of detecting that; we just have to assume the frame length
442 // is always the same.
443 local_pts += frame_length;
446 if (num_samples == 0) {
447 audio.resize(silence_samples * 2);
448 num_samples = silence_samples;
450 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
451 card->next_local_pts = local_pts + frame_length;
454 card->last_timecode = timecode;
456 // Done with the audio, so release it.
457 if (audio_frame.owner) {
458 audio_frame.owner->release_frame(audio_frame);
461 size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
462 if (video_frame.len - video_offset == 0 ||
463 video_frame.len - video_offset != expected_length) {
464 if (video_frame.len != 0) {
465 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
466 card_index, video_frame.len - video_offset, expected_length);
468 if (video_frame.owner) {
469 video_frame.owner->release_frame(video_frame);
472 // Still send on the information that we _had_ a frame, even though it's corrupted,
473 // so that pts can go up accordingly.
475 unique_lock<mutex> lock(bmusb_mutex);
476 CaptureCard::NewFrame new_frame;
477 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
478 new_frame.length = frame_length;
479 new_frame.interlaced = false;
480 new_frame.dropped_frames = dropped_frames;
481 card->new_frames.push(move(new_frame));
482 card->new_frames_changed.notify_all();
487 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
489 unsigned num_fields = video_format.interlaced ? 2 : 1;
490 timespec frame_upload_start;
491 if (video_format.interlaced) {
492 // Send the two fields along as separate frames; the other side will need to add
493 // a deinterlacer to actually get this right.
494 assert(video_format.height % 2 == 0);
495 video_format.height /= 2;
496 assert(frame_length % 2 == 0);
499 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
501 userdata->last_interlaced = video_format.interlaced;
502 userdata->last_has_signal = video_format.has_signal;
503 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
504 userdata->last_frame_rate_den = video_format.frame_rate_den;
505 RefCountedFrame frame(video_frame);
507 // Upload the textures.
508 size_t cbcr_width = video_format.width / 2;
509 size_t cbcr_offset = video_offset / 2;
510 size_t y_offset = video_frame.size / 2 + video_offset / 2;
512 for (unsigned field = 0; field < num_fields; ++field) {
513 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
515 if (userdata->tex_y[field] == 0 ||
516 userdata->tex_cbcr[field] == 0 ||
517 video_format.width != userdata->last_width[field] ||
518 video_format.height != userdata->last_height[field]) {
519 // We changed resolution since last use of this texture, so we need to create
520 // a new object. Note that this each card has its own PBOFrameAllocator,
521 // we don't need to worry about these flip-flopping between resolutions.
522 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
524 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
526 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
528 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
530 userdata->last_width[field] = video_format.width;
531 userdata->last_height[field] = video_format.height;
534 GLuint pbo = userdata->pbo;
536 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
538 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
541 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
543 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)));
545 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
547 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));
549 glBindTexture(GL_TEXTURE_2D, 0);
551 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
553 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
555 assert(fence.get() != nullptr);
558 // Don't upload the second field as fast as we can; wait until
559 // the field time has approximately passed. (Otherwise, we could
560 // get timing jitter against the other sources, and possibly also
561 // against the video display, although the latter is not as critical.)
562 // This requires our system clock to be reasonably close to the
563 // video clock, but that's not an unreasonable assumption.
564 timespec second_field_start;
565 second_field_start.tv_nsec = frame_upload_start.tv_nsec +
566 frame_length * 1000000000 / TIMEBASE;
567 second_field_start.tv_sec = frame_upload_start.tv_sec +
568 second_field_start.tv_nsec / 1000000000;
569 second_field_start.tv_nsec %= 1000000000;
571 while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
572 &second_field_start, nullptr) == -1 &&
577 unique_lock<mutex> lock(bmusb_mutex);
578 CaptureCard::NewFrame new_frame;
579 new_frame.frame = frame;
580 new_frame.length = frame_length;
581 new_frame.field = field;
582 new_frame.interlaced = video_format.interlaced;
583 new_frame.ready_fence = fence;
584 new_frame.dropped_frames = dropped_frames;
585 card->new_frames.push(move(new_frame));
586 card->new_frames_changed.notify_all();
591 void Mixer::thread_func()
593 eglBindAPI(EGL_OPENGL_API);
594 QOpenGLContext *context = create_context(mixer_surface);
595 if (!make_current(context, mixer_surface)) {
600 struct timespec start, now;
601 clock_gettime(CLOCK_MONOTONIC, &start);
604 int stats_dropped_frames = 0;
606 while (!should_quit) {
607 CaptureCard::NewFrame new_frames[MAX_CARDS];
608 bool has_new_frame[MAX_CARDS] = { false };
609 int num_samples[MAX_CARDS] = { 0 };
611 // TODO: Add a timeout.
612 unsigned master_card_index = theme->map_signal(master_clock_channel);
613 assert(master_card_index < num_cards);
615 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
616 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
617 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
618 send_audio_level_callback();
620 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
621 if (card_index == master_card_index || !has_new_frame[card_index]) {
624 if (new_frames[card_index].frame->len == 0) {
625 ++new_frames[card_index].dropped_frames;
627 if (new_frames[card_index].dropped_frames > 0) {
628 printf("Card %u dropped %d frames before this\n",
629 card_index, int(new_frames[card_index].dropped_frames));
633 // If the first card is reporting a corrupted or otherwise dropped frame,
634 // just increase the pts (skipping over this frame) and don't try to compute anything new.
635 if (new_frames[master_card_index].frame->len == 0) {
636 ++stats_dropped_frames;
637 pts_int += new_frames[master_card_index].length;
641 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
642 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
645 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
646 assert(new_frame->frame != nullptr);
647 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
650 // The new texture might still be uploaded,
651 // tell the GPU to wait until it's there.
652 if (new_frame->ready_fence) {
653 glWaitSync(new_frame->ready_fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
655 new_frame->ready_fence.reset();
662 pts_int += new_frames[master_card_index].length;
664 clock_gettime(CLOCK_MONOTONIC, &now);
665 double elapsed = now.tv_sec - start.tv_sec +
666 1e-9 * (now.tv_nsec - start.tv_nsec);
667 if (frame % 100 == 0) {
668 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
669 frame, stats_dropped_frames, elapsed, frame / elapsed,
670 1e3 * elapsed / frame);
671 // chain->print_phase_timing();
674 if (should_cut.exchange(false)) { // Test and clear.
675 string filename = generate_local_dump_filename(frame);
676 printf("Starting new recording: %s\n", filename.c_str());
677 h264_encoder->shutdown();
678 httpd.close_output_file();
679 httpd.open_output_file(filename.c_str());
680 h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
684 // Reset every 100 frames, so that local variations in frame times
685 // (especially for the first few frames, when the shaders are
686 // compiled etc.) don't make it hard to measure for the entire
687 // remaining duration of the program.
688 if (frame == 10000) {
696 resource_pool->clean_context();
699 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])
701 // The first card is the master timer, so wait for it to have a new frame.
702 unique_lock<mutex> lock(bmusb_mutex);
703 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty(); });
705 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
706 CaptureCard *card = &cards[card_index];
707 if (card->new_frames.empty()) {
708 assert(card_index != master_card_index);
709 card->queue_length_policy.update_policy(-1);
712 new_frames[card_index] = move(card->new_frames.front());
713 has_new_frame[card_index] = true;
714 card->new_frames.pop();
715 card->new_frames_changed.notify_all();
717 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
718 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
719 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
720 assert(num_samples[card_index] >= 0);
722 if (card_index == master_card_index) {
723 // We don't use the queue length policy for the master card,
724 // but we will if it stops being the master. Thus, clear out
725 // the policy in case we switch in the future.
726 card->queue_length_policy.reset(card_index);
728 // If we have excess frames compared to the policy for this card,
729 // drop frames from the head.
730 card->queue_length_policy.update_policy(card->new_frames.size());
731 while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
732 card->new_frames.pop();
738 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
740 // Resample the audio as needed, including from previously dropped frames.
741 assert(num_cards > 0);
742 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
744 // Signal to the audio thread to process this frame.
745 unique_lock<mutex> lock(audio_mutex);
746 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
747 audio_task_queue_changed.notify_one();
749 if (frame_num != dropped_frames) {
750 // For dropped frames, increase the pts. Note that if the format changed
751 // in the meantime, we have no way of detecting that; we just have to
752 // assume the frame length is always the same.
753 pts_int += length_per_frame;
758 void Mixer::render_one_frame()
760 // Get the main chain from the theme, and set its state immediately.
761 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
762 EffectChain *chain = theme_main_chain.chain;
763 theme_main_chain.setup_chain();
764 //theme_main_chain.chain->enable_phase_timing(true);
766 GLuint y_tex, cbcr_tex;
767 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
770 // Render main chain.
771 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
772 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
773 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
775 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
776 resource_pool->release_fbo(fbo);
778 subsample_chroma(cbcr_full_tex, cbcr_tex);
779 resource_pool->release_2d_texture(cbcr_full_tex);
781 // Set the right state for rgba_tex.
782 glBindFramebuffer(GL_FRAMEBUFFER, 0);
783 glBindTexture(GL_TEXTURE_2D, rgba_tex);
784 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
785 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
786 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
788 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
791 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
792 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
794 // The live frame just shows the RGBA texture we just rendered.
795 // It owns rgba_tex now.
796 DisplayFrame live_frame;
797 live_frame.chain = display_chain.get();
798 live_frame.setup_chain = [this, rgba_tex]{
799 display_input->set_texture_num(rgba_tex);
801 live_frame.ready_fence = fence;
802 live_frame.input_frames = {};
803 live_frame.temp_textures = { rgba_tex };
804 output_channel[OUTPUT_LIVE].output_frame(live_frame);
806 // Set up preview and any additional channels.
807 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
808 DisplayFrame display_frame;
809 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
810 display_frame.chain = chain.chain;
811 display_frame.setup_chain = chain.setup_chain;
812 display_frame.ready_fence = fence;
813 display_frame.input_frames = chain.input_frames;
814 display_frame.temp_textures = {};
815 output_channel[i].output_frame(display_frame);
819 void Mixer::send_audio_level_callback()
821 if (audio_level_callback == nullptr) {
825 unique_lock<mutex> lock(compressor_mutex);
826 double loudness_s = r128.loudness_S();
827 double loudness_i = r128.integrated();
828 double loudness_range_low = r128.range_min();
829 double loudness_range_high = r128.range_max();
831 audio_level_callback(loudness_s, 20.0 * log10(peak),
832 loudness_i, loudness_range_low, loudness_range_high,
833 gain_staging_db, 20.0 * log10(final_makeup_gain),
834 correlation.get_correlation());
837 void Mixer::audio_thread_func()
839 while (!should_quit) {
843 unique_lock<mutex> lock(audio_mutex);
844 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
845 task = audio_task_queue.front();
846 audio_task_queue.pop();
849 process_audio_one_frame(task.pts_int, task.num_samples);
853 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
855 vector<float> samples_card;
856 vector<float> samples_out;
858 // TODO: Allow mixing audio from several sources.
859 unsigned selected_audio_card = theme->map_signal(audio_source_channel);
860 assert(selected_audio_card < num_cards);
862 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
863 samples_card.resize(num_samples * 2);
865 unique_lock<mutex> lock(cards[card_index].audio_mutex);
866 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
867 printf("Card %d reported previous underrun.\n", card_index);
870 if (card_index == selected_audio_card) {
871 samples_out = move(samples_card);
875 // Cut away everything under 120 Hz (or whatever the cutoff is);
876 // we don't need it for voice, and it will reduce headroom
877 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
878 // should be dampened.)
880 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
883 // Apply a level compressor to get the general level right.
884 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
885 // (or more precisely, near it, since we don't use infinite ratio),
886 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
887 // entirely arbitrary, but from practical tests with speech, it seems to
888 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
890 unique_lock<mutex> lock(compressor_mutex);
891 if (level_compressor_enabled) {
892 float threshold = 0.01f; // -40 dBFS.
894 float attack_time = 0.5f;
895 float release_time = 20.0f;
896 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
897 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
898 gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
900 // Just apply the gain we already had.
901 float g = pow(10.0f, gain_staging_db / 20.0f);
902 for (size_t i = 0; i < samples_out.size(); ++i) {
909 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
910 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
911 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
912 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
915 // float limiter_att, compressor_att;
917 // The real compressor.
918 if (compressor_enabled) {
919 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
921 float attack_time = 0.005f;
922 float release_time = 0.040f;
923 float makeup_gain = 2.0f; // +6 dB.
924 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
925 // compressor_att = compressor.get_attenuation();
928 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
929 // Note that since ratio is not infinite, we could go slightly higher than this.
930 if (limiter_enabled) {
931 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
933 float attack_time = 0.0f; // Instant.
934 float release_time = 0.020f;
935 float makeup_gain = 1.0f; // 0 dB.
936 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
937 // limiter_att = limiter.get_attenuation();
940 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
942 // Upsample 4x to find interpolated peak.
943 peak_resampler.inp_data = samples_out.data();
944 peak_resampler.inp_count = samples_out.size() / 2;
946 vector<float> interpolated_samples_out;
947 interpolated_samples_out.resize(samples_out.size());
948 while (peak_resampler.inp_count > 0) { // About four iterations.
949 peak_resampler.out_data = &interpolated_samples_out[0];
950 peak_resampler.out_count = interpolated_samples_out.size() / 2;
951 peak_resampler.process();
952 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
953 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
954 peak_resampler.out_data = nullptr;
957 // At this point, we are most likely close to +0 LU, but all of our
958 // measurements have been on raw sample values, not R128 values.
959 // So we have a final makeup gain to get us to +0 LU; the gain
960 // adjustments required should be relatively small, and also, the
961 // offset shouldn't change much (only if the type of audio changes
962 // significantly). Thus, we shoot for updating this value basically
963 // “whenever we process buffers”, since the R128 calculation isn't exactly
964 // something we get out per-sample.
966 // Note that there's a feedback loop here, so we choose a very slow filter
967 // (half-time of 100 seconds).
968 double target_loudness_factor, alpha;
970 unique_lock<mutex> lock(compressor_mutex);
971 double loudness_lu = r128.loudness_M() - ref_level_lufs;
972 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
973 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
975 // If we're outside +/- 5 LU uncorrected, we don't count it as
976 // a normal signal (probably silence) and don't change the
977 // correction factor; just apply what we already have.
978 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
981 // Formula adapted from
982 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
983 const double half_time_s = 100.0;
984 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
985 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
988 double m = final_makeup_gain;
989 for (size_t i = 0; i < samples_out.size(); i += 2) {
990 samples_out[i + 0] *= m;
991 samples_out[i + 1] *= m;
992 m += (target_loudness_factor - m) * alpha;
994 final_makeup_gain = m;
997 // Find R128 levels and L/R correlation.
998 vector<float> left, right;
999 deinterleave_samples(samples_out, &left, &right);
1000 float *ptrs[] = { left.data(), right.data() };
1002 unique_lock<mutex> lock(compressor_mutex);
1003 r128.process(left.size(), ptrs);
1004 correlation.process_samples(samples_out);
1007 // Send the samples to the sound card.
1009 alsa->write(samples_out);
1012 // And finally add them to the output.
1013 h264_encoder->add_audio(frame_pts_int, move(samples_out));
1016 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1019 glGenVertexArrays(1, &vao);
1022 glBindVertexArray(vao);
1026 GLuint fbo = resource_pool->create_fbo(dst_tex);
1027 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1028 glViewport(0, 0, WIDTH/2, HEIGHT/2);
1031 glUseProgram(cbcr_program_num);
1034 glActiveTexture(GL_TEXTURE0);
1036 glBindTexture(GL_TEXTURE_2D, src_tex);
1038 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1040 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1042 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1045 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1046 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1048 glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1051 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1052 glEnableVertexAttribArray(attr_index);
1054 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1058 glDrawArrays(GL_TRIANGLES, 0, 3);
1061 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1062 glDisableVertexAttribArray(attr_index);
1068 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1071 resource_pool->release_fbo(fbo);
1072 glDeleteVertexArrays(1, &vao);
1075 void Mixer::release_display_frame(DisplayFrame *frame)
1077 for (GLuint texnum : frame->temp_textures) {
1078 resource_pool->release_2d_texture(texnum);
1080 frame->temp_textures.clear();
1081 frame->ready_fence.reset();
1082 frame->input_frames.clear();
1087 mixer_thread = thread(&Mixer::thread_func, this);
1088 audio_thread = thread(&Mixer::audio_thread_func, this);
1094 mixer_thread.join();
1095 audio_thread.join();
1098 void Mixer::transition_clicked(int transition_num)
1100 theme->transition_clicked(transition_num, pts());
1103 void Mixer::channel_clicked(int preview_num)
1105 theme->channel_clicked(preview_num);
1108 void Mixer::reset_meters()
1110 peak_resampler.reset();
1113 r128.integr_start();
1114 correlation.reset();
1117 void Mixer::start_mode_scanning(unsigned card_index)
1119 assert(card_index < num_cards);
1120 if (is_mode_scanning[card_index]) {
1123 is_mode_scanning[card_index] = true;
1124 mode_scanlist[card_index].clear();
1125 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1126 mode_scanlist[card_index].push_back(mode.first);
1128 assert(!mode_scanlist[card_index].empty());
1129 mode_scanlist_index[card_index] = 0;
1130 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1131 clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1134 Mixer::OutputChannel::~OutputChannel()
1136 if (has_current_frame) {
1137 parent->release_display_frame(¤t_frame);
1139 if (has_ready_frame) {
1140 parent->release_display_frame(&ready_frame);
1144 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1146 // Store this frame for display. Remove the ready frame if any
1147 // (it was seemingly never used).
1149 unique_lock<mutex> lock(frame_mutex);
1150 if (has_ready_frame) {
1151 parent->release_display_frame(&ready_frame);
1153 ready_frame = frame;
1154 has_ready_frame = true;
1157 if (has_new_frame_ready_callback) {
1158 new_frame_ready_callback();
1162 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1164 unique_lock<mutex> lock(frame_mutex);
1165 if (!has_current_frame && !has_ready_frame) {
1169 if (has_current_frame && has_ready_frame) {
1170 // We have a new ready frame. Toss the current one.
1171 parent->release_display_frame(¤t_frame);
1172 has_current_frame = false;
1174 if (has_ready_frame) {
1175 assert(!has_current_frame);
1176 current_frame = ready_frame;
1177 ready_frame.ready_fence.reset(); // Drop the refcount.
1178 ready_frame.input_frames.clear(); // Drop the refcounts.
1179 has_current_frame = true;
1180 has_ready_frame = false;
1183 *frame = current_frame;
1187 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1189 new_frame_ready_callback = callback;
1190 has_new_frame_ready_callback = true;