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 "quicksync_encode.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)
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 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
154 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
155 // will be halved when sampling them, and we need to compensate here.
156 movit_texel_subpixel_precision /= 2.0;
158 resource_pool.reset(new ResourcePool);
159 theme.reset(new Theme(global_flags.theme_filename.c_str(), resource_pool.get(), num_cards));
160 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
161 output_channel[i].parent = this;
164 ImageFormat inout_format;
165 inout_format.color_space = COLORSPACE_sRGB;
166 inout_format.gamma_curve = GAMMA_sRGB;
168 // Display chain; shows the live output produced by the main chain (its RGBA version).
169 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
171 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
172 display_chain->add_input(display_input);
173 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
174 display_chain->set_dither_bits(0); // Don't bother.
175 display_chain->finalize();
177 quicksync_encoder.reset(new QuickSyncEncoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
178 quicksync_encoder->open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
180 // Start listening for clients only once H264Encoder has written its header, if any.
183 // First try initializing the PCI devices, then USB, until we have the desired number of cards.
184 unsigned num_pci_devices = 0, num_usb_devices = 0;
185 unsigned card_index = 0;
187 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
188 if (decklink_iterator != nullptr) {
189 for ( ; card_index < num_cards; ++card_index) {
191 if (decklink_iterator->Next(&decklink) != S_OK) {
195 configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
198 decklink_iterator->Release();
199 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
201 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
203 for ( ; card_index < num_cards; ++card_index) {
204 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
208 if (num_usb_devices > 0) {
209 BMUSBCapture::start_bm_thread();
212 for (card_index = 0; card_index < num_cards; ++card_index) {
213 cards[card_index].queue_length_policy.reset(card_index);
214 cards[card_index].capture->start_bm_capture();
217 // Set up stuff for NV12 conversion.
220 string cbcr_vert_shader =
223 "in vec2 position; \n"
224 "in vec2 texcoord; \n"
226 "uniform vec2 foo_chroma_offset_0; \n"
230 " // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
232 " // 2.000 0.000 0.000 -1.000 \n"
233 " // 0.000 2.000 0.000 -1.000 \n"
234 " // 0.000 0.000 -2.000 -1.000 \n"
235 " // 0.000 0.000 0.000 1.000 \n"
236 " gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
237 " vec2 flipped_tc = texcoord; \n"
238 " tc0 = flipped_tc + foo_chroma_offset_0; \n"
240 string cbcr_frag_shader =
243 "uniform sampler2D cbcr_tex; \n"
244 "out vec4 FragColor; \n"
246 " FragColor = texture(cbcr_tex, tc0); \n"
248 vector<string> frag_shader_outputs;
249 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
256 cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
257 cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
258 cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
260 r128.init(2, OUTPUT_FREQUENCY);
263 locut.init(FILTER_HPF, 2);
265 // If --flat-audio is given, turn off everything that messes with the sound,
266 // except the final makeup gain.
267 if (global_flags.flat_audio) {
268 set_locut_enabled(false);
269 set_limiter_enabled(false);
270 set_compressor_enabled(false);
273 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
274 // and there's a limit to how important the peak meter is.
275 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
277 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
282 resource_pool->release_glsl_program(cbcr_program_num);
283 glDeleteBuffers(1, &cbcr_vbo);
284 BMUSBCapture::stop_bm_thread();
286 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
288 unique_lock<mutex> lock(bmusb_mutex);
289 cards[card_index].should_quit = true; // Unblock thread.
290 cards[card_index].new_frames_changed.notify_all();
292 cards[card_index].capture->stop_dequeue_thread();
295 quicksync_encoder.reset(nullptr);
298 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
300 printf("Configuring card %d...\n", card_index);
302 CaptureCard *card = &cards[card_index];
303 card->capture = capture;
304 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
305 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
306 card->capture->set_video_frame_allocator(card->frame_allocator.get());
307 card->surface = create_surface(format);
308 card->capture->set_dequeue_thread_callbacks(
310 eglBindAPI(EGL_OPENGL_API);
311 card->context = create_context(card->surface);
312 if (!make_current(card->context, card->surface)) {
313 printf("failed to create bmusb context\n");
318 resource_pool->clean_context();
320 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
321 card->capture->configure_card();
327 int unwrap_timecode(uint16_t current_wrapped, int last)
329 uint16_t last_wrapped = last & 0xffff;
330 if (current_wrapped > last_wrapped) {
331 return (last & ~0xffff) | current_wrapped;
333 return 0x10000 + ((last & ~0xffff) | current_wrapped);
337 float find_peak(const float *samples, size_t num_samples)
339 float m = fabs(samples[0]);
340 for (size_t i = 1; i < num_samples; ++i) {
341 m = max(m, fabs(samples[i]));
346 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
348 size_t num_samples = in.size() / 2;
349 out_l->resize(num_samples);
350 out_r->resize(num_samples);
352 const float *inptr = in.data();
353 float *lptr = &(*out_l)[0];
354 float *rptr = &(*out_r)[0];
355 for (size_t i = 0; i < num_samples; ++i) {
363 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
364 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
365 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
367 CaptureCard *card = &cards[card_index];
369 if (is_mode_scanning[card_index]) {
370 if (video_format.has_signal) {
371 // Found a stable signal, so stop scanning.
372 is_mode_scanning[card_index] = false;
374 static constexpr double switch_time_s = 0.5; // Should be enough time for the signal to stabilize.
376 clock_gettime(CLOCK_MONOTONIC, &now);
377 double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
378 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
379 if (sec_since_last_switch > switch_time_s) {
380 // It isn't this mode; try the next one.
381 mode_scanlist_index[card_index]++;
382 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
383 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
384 last_mode_scan_change[card_index] = now;
389 int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
391 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;
392 if (num_samples > OUTPUT_FREQUENCY / 10) {
393 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",
394 card_index, int(audio_frame.len), int(audio_offset),
395 timecode, int(video_frame.len), int(video_offset), video_format.id);
396 if (video_frame.owner) {
397 video_frame.owner->release_frame(video_frame);
399 if (audio_frame.owner) {
400 audio_frame.owner->release_frame(audio_frame);
405 int64_t local_pts = card->next_local_pts;
406 int dropped_frames = 0;
407 if (card->last_timecode != -1) {
408 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
411 // Convert the audio to stereo fp32 and add it.
413 audio.resize(num_samples * 2);
414 switch (audio_format.bits_per_sample) {
416 assert(num_samples == 0);
419 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
422 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
425 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
431 unique_lock<mutex> lock(card->audio_mutex);
433 // Number of samples per frame if we need to insert silence.
434 // (Could be nonintegral, but resampling will save us then.)
435 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
437 if (dropped_frames > MAX_FPS * 2) {
438 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
439 card_index, card->last_timecode, timecode);
440 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
442 } else if (dropped_frames > 0) {
443 // Insert silence as needed.
444 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
445 card_index, dropped_frames, timecode);
446 vector<float> silence(silence_samples * 2, 0.0f);
447 for (int i = 0; i < dropped_frames; ++i) {
448 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
449 // Note that if the format changed in the meantime, we have
450 // no way of detecting that; we just have to assume the frame length
451 // is always the same.
452 local_pts += frame_length;
455 if (num_samples == 0) {
456 audio.resize(silence_samples * 2);
457 num_samples = silence_samples;
459 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
460 card->next_local_pts = local_pts + frame_length;
463 card->last_timecode = timecode;
465 // Done with the audio, so release it.
466 if (audio_frame.owner) {
467 audio_frame.owner->release_frame(audio_frame);
470 size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
471 if (video_frame.len - video_offset == 0 ||
472 video_frame.len - video_offset != expected_length) {
473 if (video_frame.len != 0) {
474 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
475 card_index, video_frame.len - video_offset, expected_length);
477 if (video_frame.owner) {
478 video_frame.owner->release_frame(video_frame);
481 // Still send on the information that we _had_ a frame, even though it's corrupted,
482 // so that pts can go up accordingly.
484 unique_lock<mutex> lock(bmusb_mutex);
485 CaptureCard::NewFrame new_frame;
486 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
487 new_frame.length = frame_length;
488 new_frame.interlaced = false;
489 new_frame.dropped_frames = dropped_frames;
490 card->new_frames.push(move(new_frame));
491 card->new_frames_changed.notify_all();
496 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
498 unsigned num_fields = video_format.interlaced ? 2 : 1;
499 timespec frame_upload_start;
500 if (video_format.interlaced) {
501 // Send the two fields along as separate frames; the other side will need to add
502 // a deinterlacer to actually get this right.
503 assert(video_format.height % 2 == 0);
504 video_format.height /= 2;
505 assert(frame_length % 2 == 0);
508 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
510 userdata->last_interlaced = video_format.interlaced;
511 userdata->last_has_signal = video_format.has_signal;
512 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
513 userdata->last_frame_rate_den = video_format.frame_rate_den;
514 RefCountedFrame frame(video_frame);
516 // Upload the textures.
517 size_t cbcr_width = video_format.width / 2;
518 size_t cbcr_offset = video_offset / 2;
519 size_t y_offset = video_frame.size / 2 + video_offset / 2;
521 for (unsigned field = 0; field < num_fields; ++field) {
522 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
524 if (userdata->tex_y[field] == 0 ||
525 userdata->tex_cbcr[field] == 0 ||
526 video_format.width != userdata->last_width[field] ||
527 video_format.height != userdata->last_height[field]) {
528 // We changed resolution since last use of this texture, so we need to create
529 // a new object. Note that this each card has its own PBOFrameAllocator,
530 // we don't need to worry about these flip-flopping between resolutions.
531 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
533 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
535 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
537 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
539 userdata->last_width[field] = video_format.width;
540 userdata->last_height[field] = video_format.height;
543 GLuint pbo = userdata->pbo;
545 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
548 size_t field_y_start = y_offset + video_format.width * field_start_line;
549 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
551 if (global_flags.flush_pbos) {
552 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
554 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
558 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
560 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
562 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
564 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
566 glBindTexture(GL_TEXTURE_2D, 0);
568 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
570 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
572 assert(fence.get() != nullptr);
573 glFlush(); // Make sure the main thread doesn't have to wait until we push out enough frames to make a new command buffer.
577 // Don't upload the second field as fast as we can; wait until
578 // the field time has approximately passed. (Otherwise, we could
579 // get timing jitter against the other sources, and possibly also
580 // against the video display, although the latter is not as critical.)
581 // This requires our system clock to be reasonably close to the
582 // video clock, but that's not an unreasonable assumption.
583 timespec second_field_start;
584 second_field_start.tv_nsec = frame_upload_start.tv_nsec +
585 frame_length * 1000000000 / TIMEBASE;
586 second_field_start.tv_sec = frame_upload_start.tv_sec +
587 second_field_start.tv_nsec / 1000000000;
588 second_field_start.tv_nsec %= 1000000000;
590 while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
591 &second_field_start, nullptr) == -1 &&
596 unique_lock<mutex> lock(bmusb_mutex);
597 CaptureCard::NewFrame new_frame;
598 new_frame.frame = frame;
599 new_frame.length = frame_length;
600 new_frame.field = field;
601 new_frame.interlaced = video_format.interlaced;
602 new_frame.ready_fence = fence;
603 new_frame.dropped_frames = dropped_frames;
604 card->new_frames.push(move(new_frame));
605 card->new_frames_changed.notify_all();
610 void Mixer::thread_func()
612 eglBindAPI(EGL_OPENGL_API);
613 QOpenGLContext *context = create_context(mixer_surface);
614 if (!make_current(context, mixer_surface)) {
619 struct timespec start, now;
620 clock_gettime(CLOCK_MONOTONIC, &start);
623 int stats_dropped_frames = 0;
625 while (!should_quit) {
626 CaptureCard::NewFrame new_frames[MAX_CARDS];
627 bool has_new_frame[MAX_CARDS] = { false };
628 int num_samples[MAX_CARDS] = { 0 };
630 // TODO: Add a timeout.
631 unsigned master_card_index = theme->map_signal(master_clock_channel);
632 assert(master_card_index < num_cards);
634 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
635 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
636 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
637 send_audio_level_callback();
639 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
640 if (card_index == master_card_index || !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[master_card_index].frame->len == 0) {
655 ++stats_dropped_frames;
656 pts_int += new_frames[master_card_index].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.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
674 new_frame->ready_fence.reset();
679 int64_t duration = new_frames[master_card_index].length;
680 render_one_frame(duration);
684 clock_gettime(CLOCK_MONOTONIC, &now);
685 double elapsed = now.tv_sec - start.tv_sec +
686 1e-9 * (now.tv_nsec - start.tv_nsec);
687 if (frame % 100 == 0) {
688 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
689 frame, stats_dropped_frames, elapsed, frame / elapsed,
690 1e3 * elapsed / frame);
691 // chain->print_phase_timing();
694 if (should_cut.exchange(false)) { // Test and clear.
695 string filename = generate_local_dump_filename(frame);
696 printf("Starting new recording: %s\n", filename.c_str());
697 quicksync_encoder->close_output_file();
698 quicksync_encoder->shutdown();
699 quicksync_encoder.reset(new QuickSyncEncoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
700 quicksync_encoder->open_output_file(filename.c_str());
704 // Reset every 100 frames, so that local variations in frame times
705 // (especially for the first few frames, when the shaders are
706 // compiled etc.) don't make it hard to measure for the entire
707 // remaining duration of the program.
708 if (frame == 10000) {
716 resource_pool->clean_context();
719 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])
721 // The first card is the master timer, so wait for it to have a new frame.
722 unique_lock<mutex> lock(bmusb_mutex);
723 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty(); });
725 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
726 CaptureCard *card = &cards[card_index];
727 if (card->new_frames.empty()) {
728 assert(card_index != master_card_index);
729 card->queue_length_policy.update_policy(-1);
732 new_frames[card_index] = move(card->new_frames.front());
733 has_new_frame[card_index] = true;
734 card->new_frames.pop();
735 card->new_frames_changed.notify_all();
737 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
738 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
739 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
740 assert(num_samples[card_index] >= 0);
742 if (card_index == master_card_index) {
743 // We don't use the queue length policy for the master card,
744 // but we will if it stops being the master. Thus, clear out
745 // the policy in case we switch in the future.
746 card->queue_length_policy.reset(card_index);
748 // If we have excess frames compared to the policy for this card,
749 // drop frames from the head.
750 card->queue_length_policy.update_policy(card->new_frames.size());
751 while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
752 card->new_frames.pop();
758 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
760 // Resample the audio as needed, including from previously dropped frames.
761 assert(num_cards > 0);
762 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
764 // Signal to the audio thread to process this frame.
765 unique_lock<mutex> lock(audio_mutex);
766 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
767 audio_task_queue_changed.notify_one();
769 if (frame_num != dropped_frames) {
770 // For dropped frames, increase the pts. Note that if the format changed
771 // in the meantime, we have no way of detecting that; we just have to
772 // assume the frame length is always the same.
773 pts_int += length_per_frame;
778 void Mixer::render_one_frame(int64_t duration)
780 // Get the main chain from the theme, and set its state immediately.
781 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
782 EffectChain *chain = theme_main_chain.chain;
783 theme_main_chain.setup_chain();
784 //theme_main_chain.chain->enable_phase_timing(true);
786 GLuint y_tex, cbcr_tex;
787 bool got_frame = quicksync_encoder->begin_frame(&y_tex, &cbcr_tex);
790 // Render main chain.
791 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
792 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
793 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
795 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
796 resource_pool->release_fbo(fbo);
798 subsample_chroma(cbcr_full_tex, cbcr_tex);
799 resource_pool->release_2d_texture(cbcr_full_tex);
801 // Set the right state for rgba_tex.
802 glBindFramebuffer(GL_FRAMEBUFFER, 0);
803 glBindTexture(GL_TEXTURE_2D, rgba_tex);
804 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
805 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
806 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
808 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
809 RefCountedGLsync fence = quicksync_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
811 // The live frame just shows the RGBA texture we just rendered.
812 // It owns rgba_tex now.
813 DisplayFrame live_frame;
814 live_frame.chain = display_chain.get();
815 live_frame.setup_chain = [this, rgba_tex]{
816 display_input->set_texture_num(rgba_tex);
818 live_frame.ready_fence = fence;
819 live_frame.input_frames = {};
820 live_frame.temp_textures = { rgba_tex };
821 output_channel[OUTPUT_LIVE].output_frame(live_frame);
823 // Set up preview and any additional channels.
824 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
825 DisplayFrame display_frame;
826 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
827 display_frame.chain = chain.chain;
828 display_frame.setup_chain = chain.setup_chain;
829 display_frame.ready_fence = fence;
830 display_frame.input_frames = chain.input_frames;
831 display_frame.temp_textures = {};
832 output_channel[i].output_frame(display_frame);
836 void Mixer::send_audio_level_callback()
838 if (audio_level_callback == nullptr) {
842 unique_lock<mutex> lock(compressor_mutex);
843 double loudness_s = r128.loudness_S();
844 double loudness_i = r128.integrated();
845 double loudness_range_low = r128.range_min();
846 double loudness_range_high = r128.range_max();
848 audio_level_callback(loudness_s, 20.0 * log10(peak),
849 loudness_i, loudness_range_low, loudness_range_high,
850 gain_staging_db, 20.0 * log10(final_makeup_gain),
851 correlation.get_correlation());
854 void Mixer::audio_thread_func()
856 while (!should_quit) {
860 unique_lock<mutex> lock(audio_mutex);
861 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
862 task = audio_task_queue.front();
863 audio_task_queue.pop();
866 process_audio_one_frame(task.pts_int, task.num_samples);
870 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
872 vector<float> samples_card;
873 vector<float> samples_out;
875 // TODO: Allow mixing audio from several sources.
876 unsigned selected_audio_card = theme->map_signal(audio_source_channel);
877 assert(selected_audio_card < num_cards);
879 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
880 samples_card.resize(num_samples * 2);
882 unique_lock<mutex> lock(cards[card_index].audio_mutex);
883 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
884 printf("Card %d reported previous underrun.\n", card_index);
887 if (card_index == selected_audio_card) {
888 samples_out = move(samples_card);
892 // Cut away everything under 120 Hz (or whatever the cutoff is);
893 // we don't need it for voice, and it will reduce headroom
894 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
895 // should be dampened.)
897 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
900 // Apply a level compressor to get the general level right.
901 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
902 // (or more precisely, near it, since we don't use infinite ratio),
903 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
904 // entirely arbitrary, but from practical tests with speech, it seems to
905 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
907 unique_lock<mutex> lock(compressor_mutex);
908 if (level_compressor_enabled) {
909 float threshold = 0.01f; // -40 dBFS.
911 float attack_time = 0.5f;
912 float release_time = 20.0f;
913 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
914 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
915 gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
917 // Just apply the gain we already had.
918 float g = pow(10.0f, gain_staging_db / 20.0f);
919 for (size_t i = 0; i < samples_out.size(); ++i) {
926 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
927 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
928 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
929 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
932 // float limiter_att, compressor_att;
934 // The real compressor.
935 if (compressor_enabled) {
936 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
938 float attack_time = 0.005f;
939 float release_time = 0.040f;
940 float makeup_gain = 2.0f; // +6 dB.
941 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
942 // compressor_att = compressor.get_attenuation();
945 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
946 // Note that since ratio is not infinite, we could go slightly higher than this.
947 if (limiter_enabled) {
948 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
950 float attack_time = 0.0f; // Instant.
951 float release_time = 0.020f;
952 float makeup_gain = 1.0f; // 0 dB.
953 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
954 // limiter_att = limiter.get_attenuation();
957 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
959 // Upsample 4x to find interpolated peak.
960 peak_resampler.inp_data = samples_out.data();
961 peak_resampler.inp_count = samples_out.size() / 2;
963 vector<float> interpolated_samples_out;
964 interpolated_samples_out.resize(samples_out.size());
965 while (peak_resampler.inp_count > 0) { // About four iterations.
966 peak_resampler.out_data = &interpolated_samples_out[0];
967 peak_resampler.out_count = interpolated_samples_out.size() / 2;
968 peak_resampler.process();
969 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
970 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
971 peak_resampler.out_data = nullptr;
974 // At this point, we are most likely close to +0 LU, but all of our
975 // measurements have been on raw sample values, not R128 values.
976 // So we have a final makeup gain to get us to +0 LU; the gain
977 // adjustments required should be relatively small, and also, the
978 // offset shouldn't change much (only if the type of audio changes
979 // significantly). Thus, we shoot for updating this value basically
980 // “whenever we process buffers”, since the R128 calculation isn't exactly
981 // something we get out per-sample.
983 // Note that there's a feedback loop here, so we choose a very slow filter
984 // (half-time of 100 seconds).
985 double target_loudness_factor, alpha;
987 unique_lock<mutex> lock(compressor_mutex);
988 double loudness_lu = r128.loudness_M() - ref_level_lufs;
989 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
990 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
992 // If we're outside +/- 5 LU uncorrected, we don't count it as
993 // a normal signal (probably silence) and don't change the
994 // correction factor; just apply what we already have.
995 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
998 // Formula adapted from
999 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
1000 const double half_time_s = 100.0;
1001 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
1002 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
1005 double m = final_makeup_gain;
1006 for (size_t i = 0; i < samples_out.size(); i += 2) {
1007 samples_out[i + 0] *= m;
1008 samples_out[i + 1] *= m;
1009 m += (target_loudness_factor - m) * alpha;
1011 final_makeup_gain = m;
1014 // Find R128 levels and L/R correlation.
1015 vector<float> left, right;
1016 deinterleave_samples(samples_out, &left, &right);
1017 float *ptrs[] = { left.data(), right.data() };
1019 unique_lock<mutex> lock(compressor_mutex);
1020 r128.process(left.size(), ptrs);
1021 correlation.process_samples(samples_out);
1024 // Send the samples to the sound card.
1026 alsa->write(samples_out);
1029 // And finally add them to the output.
1030 quicksync_encoder->add_audio(frame_pts_int, move(samples_out));
1033 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1036 glGenVertexArrays(1, &vao);
1039 glBindVertexArray(vao);
1043 GLuint fbo = resource_pool->create_fbo(dst_tex);
1044 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1045 glViewport(0, 0, WIDTH/2, HEIGHT/2);
1048 glUseProgram(cbcr_program_num);
1051 glActiveTexture(GL_TEXTURE0);
1053 glBindTexture(GL_TEXTURE_2D, src_tex);
1055 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1057 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1059 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1062 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1063 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1065 glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1068 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1069 glEnableVertexAttribArray(attr_index);
1071 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1075 glDrawArrays(GL_TRIANGLES, 0, 3);
1078 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1079 glDisableVertexAttribArray(attr_index);
1085 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1088 resource_pool->release_fbo(fbo);
1089 glDeleteVertexArrays(1, &vao);
1092 void Mixer::release_display_frame(DisplayFrame *frame)
1094 for (GLuint texnum : frame->temp_textures) {
1095 resource_pool->release_2d_texture(texnum);
1097 frame->temp_textures.clear();
1098 frame->ready_fence.reset();
1099 frame->input_frames.clear();
1104 mixer_thread = thread(&Mixer::thread_func, this);
1105 audio_thread = thread(&Mixer::audio_thread_func, this);
1111 mixer_thread.join();
1112 audio_thread.join();
1115 void Mixer::transition_clicked(int transition_num)
1117 theme->transition_clicked(transition_num, pts());
1120 void Mixer::channel_clicked(int preview_num)
1122 theme->channel_clicked(preview_num);
1125 void Mixer::reset_meters()
1127 peak_resampler.reset();
1130 r128.integr_start();
1131 correlation.reset();
1134 void Mixer::start_mode_scanning(unsigned card_index)
1136 assert(card_index < num_cards);
1137 if (is_mode_scanning[card_index]) {
1140 is_mode_scanning[card_index] = true;
1141 mode_scanlist[card_index].clear();
1142 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1143 mode_scanlist[card_index].push_back(mode.first);
1145 assert(!mode_scanlist[card_index].empty());
1146 mode_scanlist_index[card_index] = 0;
1147 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1148 clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1151 Mixer::OutputChannel::~OutputChannel()
1153 if (has_current_frame) {
1154 parent->release_display_frame(¤t_frame);
1156 if (has_ready_frame) {
1157 parent->release_display_frame(&ready_frame);
1161 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1163 // Store this frame for display. Remove the ready frame if any
1164 // (it was seemingly never used).
1166 unique_lock<mutex> lock(frame_mutex);
1167 if (has_ready_frame) {
1168 parent->release_display_frame(&ready_frame);
1170 ready_frame = frame;
1171 has_ready_frame = true;
1174 if (has_new_frame_ready_callback) {
1175 new_frame_ready_callback();
1179 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1181 unique_lock<mutex> lock(frame_mutex);
1182 if (!has_current_frame && !has_ready_frame) {
1186 if (has_current_frame && has_ready_frame) {
1187 // We have a new ready frame. Toss the current one.
1188 parent->release_display_frame(¤t_frame);
1189 has_current_frame = false;
1191 if (has_ready_frame) {
1192 assert(!has_current_frame);
1193 current_frame = ready_frame;
1194 ready_frame.ready_fence.reset(); // Drop the refcount.
1195 ready_frame.input_frames.clear(); // Drop the refcounts.
1196 has_current_frame = true;
1197 has_ready_frame = false;
1200 *frame = current_frame;
1204 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1206 new_frame_ready_callback = callback;
1207 has_new_frame_ready_callback = true;
1210 mutex RefCountedGLsync::fence_lock;