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 (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;
125 if (queue_length > 0) {
126 if (++frames_with_at_least_one >= 50 && safe_queue_length > 0) {
128 fprintf(stderr, "Card %u: Spare frames for more than 50 frames, reducing safe limit to %u frames\n",
129 card_index, safe_queue_length);
130 frames_with_at_least_one = 0;
133 frames_with_at_least_one = 0;
137 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
138 : httpd(WIDTH, HEIGHT),
139 num_cards(num_cards),
140 mixer_surface(create_surface(format)),
141 h264_encoder_surface(create_surface(format)),
142 correlation(OUTPUT_FREQUENCY),
143 level_compressor(OUTPUT_FREQUENCY),
144 limiter(OUTPUT_FREQUENCY),
145 compressor(OUTPUT_FREQUENCY)
147 httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
150 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
153 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
154 // will be halved when sampling them, and we need to compensate here.
155 movit_texel_subpixel_precision /= 2.0;
157 resource_pool.reset(new ResourcePool);
158 theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
159 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
160 output_channel[i].parent = this;
163 ImageFormat inout_format;
164 inout_format.color_space = COLORSPACE_sRGB;
165 inout_format.gamma_curve = GAMMA_sRGB;
167 // Display chain; shows the live output produced by the main chain (its RGBA version).
168 display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
170 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong.
171 display_chain->add_input(display_input);
172 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
173 display_chain->set_dither_bits(0); // Don't bother.
174 display_chain->finalize();
176 h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
178 // First try initializing the PCI devices, then USB, until we have the desired number of cards.
179 unsigned num_pci_devices = 0, num_usb_devices = 0;
180 unsigned card_index = 0;
182 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
183 if (decklink_iterator != nullptr) {
184 for ( ; card_index < num_cards; ++card_index) {
186 if (decklink_iterator->Next(&decklink) != S_OK) {
190 configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
193 decklink_iterator->Release();
194 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
196 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
198 for ( ; card_index < num_cards; ++card_index) {
199 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
203 if (num_usb_devices > 0) {
204 BMUSBCapture::start_bm_thread();
207 for (card_index = 0; card_index < num_cards; ++card_index) {
208 cards[card_index].queue_length_policy.reset(card_index);
209 cards[card_index].capture->start_bm_capture();
212 // Set up stuff for NV12 conversion.
215 string cbcr_vert_shader =
218 "in vec2 position; \n"
219 "in vec2 texcoord; \n"
221 "uniform vec2 foo_chroma_offset_0; \n"
225 " // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
227 " // 2.000 0.000 0.000 -1.000 \n"
228 " // 0.000 2.000 0.000 -1.000 \n"
229 " // 0.000 0.000 -2.000 -1.000 \n"
230 " // 0.000 0.000 0.000 1.000 \n"
231 " gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
232 " vec2 flipped_tc = texcoord; \n"
233 " tc0 = flipped_tc + foo_chroma_offset_0; \n"
235 string cbcr_frag_shader =
238 "uniform sampler2D cbcr_tex; \n"
239 "out vec4 FragColor; \n"
241 " FragColor = texture(cbcr_tex, tc0); \n"
243 vector<string> frag_shader_outputs;
244 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
251 cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
252 cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
253 cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
255 r128.init(2, OUTPUT_FREQUENCY);
258 locut.init(FILTER_HPF, 2);
260 // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
261 // and there's a limit to how important the peak meter is.
262 peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
264 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
269 resource_pool->release_glsl_program(cbcr_program_num);
270 glDeleteBuffers(1, &cbcr_vbo);
271 BMUSBCapture::stop_bm_thread();
273 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
275 unique_lock<mutex> lock(bmusb_mutex);
276 cards[card_index].should_quit = true; // Unblock thread.
277 cards[card_index].new_frames_changed.notify_all();
279 cards[card_index].capture->stop_dequeue_thread();
282 h264_encoder.reset(nullptr);
285 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
287 printf("Configuring card %d...\n", card_index);
289 CaptureCard *card = &cards[card_index];
290 card->capture = capture;
291 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
292 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
293 card->capture->set_video_frame_allocator(card->frame_allocator.get());
294 card->surface = create_surface(format);
295 card->capture->set_dequeue_thread_callbacks(
297 eglBindAPI(EGL_OPENGL_API);
298 card->context = create_context(card->surface);
299 if (!make_current(card->context, card->surface)) {
300 printf("failed to create bmusb context\n");
305 resource_pool->clean_context();
307 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
308 card->capture->configure_card();
314 int unwrap_timecode(uint16_t current_wrapped, int last)
316 uint16_t last_wrapped = last & 0xffff;
317 if (current_wrapped > last_wrapped) {
318 return (last & ~0xffff) | current_wrapped;
320 return 0x10000 + ((last & ~0xffff) | current_wrapped);
324 float find_peak(const float *samples, size_t num_samples)
326 float m = fabs(samples[0]);
327 for (size_t i = 1; i < num_samples; ++i) {
328 m = max(m, fabs(samples[i]));
333 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
335 size_t num_samples = in.size() / 2;
336 out_l->resize(num_samples);
337 out_r->resize(num_samples);
339 const float *inptr = in.data();
340 float *lptr = &(*out_l)[0];
341 float *rptr = &(*out_r)[0];
342 for (size_t i = 0; i < num_samples; ++i) {
350 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
351 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
352 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
354 CaptureCard *card = &cards[card_index];
356 if (is_mode_scanning[card_index]) {
357 if (video_format.has_signal) {
358 // Found a stable signal, so stop scanning.
359 is_mode_scanning[card_index] = false;
361 static constexpr double switch_time_s = 0.5; // Should be enough time for the signal to stabilize.
363 clock_gettime(CLOCK_MONOTONIC, &now);
364 double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
365 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
366 if (sec_since_last_switch > switch_time_s) {
367 // It isn't this mode; try the next one.
368 mode_scanlist_index[card_index]++;
369 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
370 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
371 last_mode_scan_change[card_index] = now;
376 int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
378 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;
379 if (num_samples > OUTPUT_FREQUENCY / 10) {
380 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",
381 card_index, int(audio_frame.len), int(audio_offset),
382 timecode, int(video_frame.len), int(video_offset), video_format.id);
383 if (video_frame.owner) {
384 video_frame.owner->release_frame(video_frame);
386 if (audio_frame.owner) {
387 audio_frame.owner->release_frame(audio_frame);
392 int64_t local_pts = card->next_local_pts;
393 int dropped_frames = 0;
394 if (card->last_timecode != -1) {
395 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
398 // Convert the audio to stereo fp32 and add it.
400 audio.resize(num_samples * 2);
401 switch (audio_format.bits_per_sample) {
403 assert(num_samples == 0);
406 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
409 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
412 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
418 unique_lock<mutex> lock(card->audio_mutex);
420 // Number of samples per frame if we need to insert silence.
421 // (Could be nonintegral, but resampling will save us then.)
422 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
424 if (dropped_frames > MAX_FPS * 2) {
425 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
426 card_index, card->last_timecode, timecode);
427 card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
429 } else if (dropped_frames > 0) {
430 // Insert silence as needed.
431 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
432 card_index, dropped_frames, timecode);
433 vector<float> silence(silence_samples * 2, 0.0f);
434 for (int i = 0; i < dropped_frames; ++i) {
435 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
436 // Note that if the format changed in the meantime, we have
437 // no way of detecting that; we just have to assume the frame length
438 // is always the same.
439 local_pts += frame_length;
442 if (num_samples == 0) {
443 audio.resize(silence_samples * 2);
444 num_samples = silence_samples;
446 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
447 card->next_local_pts = local_pts + frame_length;
450 card->last_timecode = timecode;
452 // Done with the audio, so release it.
453 if (audio_frame.owner) {
454 audio_frame.owner->release_frame(audio_frame);
457 size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
458 if (video_frame.len - video_offset == 0 ||
459 video_frame.len - video_offset != expected_length) {
460 if (video_frame.len != 0) {
461 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
462 card_index, video_frame.len - video_offset, expected_length);
464 if (video_frame.owner) {
465 video_frame.owner->release_frame(video_frame);
468 // Still send on the information that we _had_ a frame, even though it's corrupted,
469 // so that pts can go up accordingly.
471 unique_lock<mutex> lock(bmusb_mutex);
472 CaptureCard::NewFrame new_frame;
473 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
474 new_frame.length = frame_length;
475 new_frame.interlaced = false;
476 new_frame.dropped_frames = dropped_frames;
477 card->new_frames.push(move(new_frame));
478 card->new_frames_changed.notify_all();
483 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
485 unsigned num_fields = video_format.interlaced ? 2 : 1;
486 timespec frame_upload_start;
487 if (video_format.interlaced) {
488 // Send the two fields along as separate frames; the other side will need to add
489 // a deinterlacer to actually get this right.
490 assert(video_format.height % 2 == 0);
491 video_format.height /= 2;
492 assert(frame_length % 2 == 0);
495 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
497 userdata->last_interlaced = video_format.interlaced;
498 userdata->last_has_signal = video_format.has_signal;
499 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
500 userdata->last_frame_rate_den = video_format.frame_rate_den;
501 RefCountedFrame frame(video_frame);
503 // Upload the textures.
504 size_t cbcr_width = video_format.width / 2;
505 size_t cbcr_offset = video_offset / 2;
506 size_t y_offset = video_frame.size / 2 + video_offset / 2;
508 for (unsigned field = 0; field < num_fields; ++field) {
509 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
511 if (userdata->tex_y[field] == 0 ||
512 userdata->tex_cbcr[field] == 0 ||
513 video_format.width != userdata->last_width[field] ||
514 video_format.height != userdata->last_height[field]) {
515 // We changed resolution since last use of this texture, so we need to create
516 // a new object. Note that this each card has its own PBOFrameAllocator,
517 // we don't need to worry about these flip-flopping between resolutions.
518 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
520 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
522 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
524 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
526 userdata->last_width[field] = video_format.width;
527 userdata->last_height[field] = video_format.height;
530 GLuint pbo = userdata->pbo;
532 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
534 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
537 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
539 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)));
541 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
543 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));
545 glBindTexture(GL_TEXTURE_2D, 0);
547 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
549 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
551 assert(fence.get() != nullptr);
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.ready_fence = fence;
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 still be uploaded,
647 // tell the GPU to wait until it's there.
648 if (new_frame->ready_fence) {
649 glWaitSync(new_frame->ready_fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
651 new_frame->ready_fence.reset();
658 pts_int += new_frames[master_card_index].length;
660 clock_gettime(CLOCK_MONOTONIC, &now);
661 double elapsed = now.tv_sec - start.tv_sec +
662 1e-9 * (now.tv_nsec - start.tv_nsec);
663 if (frame % 100 == 0) {
664 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
665 frame, stats_dropped_frames, elapsed, frame / elapsed,
666 1e3 * elapsed / frame);
667 // chain->print_phase_timing();
670 if (should_cut.exchange(false)) { // Test and clear.
671 string filename = generate_local_dump_filename(frame);
672 printf("Starting new recording: %s\n", filename.c_str());
673 h264_encoder->shutdown();
674 httpd.close_output_file();
675 httpd.open_output_file(filename.c_str());
676 h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
680 // Reset every 100 frames, so that local variations in frame times
681 // (especially for the first few frames, when the shaders are
682 // compiled etc.) don't make it hard to measure for the entire
683 // remaining duration of the program.
684 if (frame == 10000) {
692 resource_pool->clean_context();
695 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])
697 // The first card is the master timer, so wait for it to have a new frame.
698 unique_lock<mutex> lock(bmusb_mutex);
699 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty(); });
701 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
702 CaptureCard *card = &cards[card_index];
703 if (card->new_frames.empty()) {
704 assert(card_index != master_card_index);
705 card->queue_length_policy.update_policy(-1);
708 new_frames[card_index] = move(card->new_frames.front());
709 has_new_frame[card_index] = true;
710 card->new_frames.pop();
711 card->new_frames_changed.notify_all();
713 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
714 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
715 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
716 assert(num_samples[card_index] >= 0);
718 if (card_index == master_card_index) {
719 // We don't use the queue length policy for the master card,
720 // but we will if it stops being the master. Thus, clear out
721 // the policy in case we switch in the future.
722 card->queue_length_policy.reset(card_index);
724 // If we have excess frames compared to the policy for this card,
725 // drop frames from the head.
726 card->queue_length_policy.update_policy(card->new_frames.size());
727 while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
728 card->new_frames.pop();
734 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
736 // Resample the audio as needed, including from previously dropped frames.
737 assert(num_cards > 0);
738 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
740 // Signal to the audio thread to process this frame.
741 unique_lock<mutex> lock(audio_mutex);
742 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
743 audio_task_queue_changed.notify_one();
745 if (frame_num != dropped_frames) {
746 // For dropped frames, increase the pts. Note that if the format changed
747 // in the meantime, we have no way of detecting that; we just have to
748 // assume the frame length is always the same.
749 pts_int += length_per_frame;
754 void Mixer::render_one_frame()
756 // Get the main chain from the theme, and set its state immediately.
757 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
758 EffectChain *chain = theme_main_chain.chain;
759 theme_main_chain.setup_chain();
760 //theme_main_chain.chain->enable_phase_timing(true);
762 GLuint y_tex, cbcr_tex;
763 bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
766 // Render main chain.
767 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
768 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up.
769 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
771 chain->render_to_fbo(fbo, WIDTH, HEIGHT);
772 resource_pool->release_fbo(fbo);
774 subsample_chroma(cbcr_full_tex, cbcr_tex);
775 resource_pool->release_2d_texture(cbcr_full_tex);
777 // Set the right state for rgba_tex.
778 glBindFramebuffer(GL_FRAMEBUFFER, 0);
779 glBindTexture(GL_TEXTURE_2D, rgba_tex);
780 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
781 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
782 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
784 RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
787 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
788 h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
790 // The live frame just shows the RGBA texture we just rendered.
791 // It owns rgba_tex now.
792 DisplayFrame live_frame;
793 live_frame.chain = display_chain.get();
794 live_frame.setup_chain = [this, rgba_tex]{
795 display_input->set_texture_num(rgba_tex);
797 live_frame.ready_fence = fence;
798 live_frame.input_frames = {};
799 live_frame.temp_textures = { rgba_tex };
800 output_channel[OUTPUT_LIVE].output_frame(live_frame);
802 // Set up preview and any additional channels.
803 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
804 DisplayFrame display_frame;
805 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions
806 display_frame.chain = chain.chain;
807 display_frame.setup_chain = chain.setup_chain;
808 display_frame.ready_fence = fence;
809 display_frame.input_frames = chain.input_frames;
810 display_frame.temp_textures = {};
811 output_channel[i].output_frame(display_frame);
815 void Mixer::send_audio_level_callback()
817 if (audio_level_callback == nullptr) {
821 unique_lock<mutex> lock(compressor_mutex);
822 double loudness_s = r128.loudness_S();
823 double loudness_i = r128.integrated();
824 double loudness_range_low = r128.range_min();
825 double loudness_range_high = r128.range_max();
827 audio_level_callback(loudness_s, 20.0 * log10(peak),
828 loudness_i, loudness_range_low, loudness_range_high,
829 gain_staging_db, 20.0 * log10(final_makeup_gain),
830 correlation.get_correlation());
833 void Mixer::audio_thread_func()
835 while (!should_quit) {
839 unique_lock<mutex> lock(audio_mutex);
840 audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
841 task = audio_task_queue.front();
842 audio_task_queue.pop();
845 process_audio_one_frame(task.pts_int, task.num_samples);
849 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
851 vector<float> samples_card;
852 vector<float> samples_out;
854 // TODO: Allow mixing audio from several sources.
855 unsigned selected_audio_card = theme->map_signal(audio_source_channel);
856 assert(selected_audio_card < num_cards);
858 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
859 samples_card.resize(num_samples * 2);
861 unique_lock<mutex> lock(cards[card_index].audio_mutex);
862 if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
863 printf("Card %d reported previous underrun.\n", card_index);
866 if (card_index == selected_audio_card) {
867 samples_out = move(samples_card);
871 // Cut away everything under 120 Hz (or whatever the cutoff is);
872 // we don't need it for voice, and it will reduce headroom
873 // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
874 // should be dampened.)
876 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
879 // Apply a level compressor to get the general level right.
880 // Basically, if it's over about -40 dBFS, we squeeze it down to that level
881 // (or more precisely, near it, since we don't use infinite ratio),
882 // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
883 // entirely arbitrary, but from practical tests with speech, it seems to
884 // put ut around -23 LUFS, so it's a reasonable starting point for later use.
886 unique_lock<mutex> lock(compressor_mutex);
887 if (level_compressor_enabled) {
888 float threshold = 0.01f; // -40 dBFS.
890 float attack_time = 0.5f;
891 float release_time = 20.0f;
892 float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
893 level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
894 gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
896 // Just apply the gain we already had.
897 float g = pow(10.0f, gain_staging_db / 20.0f);
898 for (size_t i = 0; i < samples_out.size(); ++i) {
905 printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
906 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
907 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
908 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
911 // float limiter_att, compressor_att;
913 // The real compressor.
914 if (compressor_enabled) {
915 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
917 float attack_time = 0.005f;
918 float release_time = 0.040f;
919 float makeup_gain = 2.0f; // +6 dB.
920 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
921 // compressor_att = compressor.get_attenuation();
924 // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
925 // Note that since ratio is not infinite, we could go slightly higher than this.
926 if (limiter_enabled) {
927 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
929 float attack_time = 0.0f; // Instant.
930 float release_time = 0.020f;
931 float makeup_gain = 1.0f; // 0 dB.
932 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
933 // limiter_att = limiter.get_attenuation();
936 // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
938 // Upsample 4x to find interpolated peak.
939 peak_resampler.inp_data = samples_out.data();
940 peak_resampler.inp_count = samples_out.size() / 2;
942 vector<float> interpolated_samples_out;
943 interpolated_samples_out.resize(samples_out.size());
944 while (peak_resampler.inp_count > 0) { // About four iterations.
945 peak_resampler.out_data = &interpolated_samples_out[0];
946 peak_resampler.out_count = interpolated_samples_out.size() / 2;
947 peak_resampler.process();
948 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
949 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
950 peak_resampler.out_data = nullptr;
953 // At this point, we are most likely close to +0 LU, but all of our
954 // measurements have been on raw sample values, not R128 values.
955 // So we have a final makeup gain to get us to +0 LU; the gain
956 // adjustments required should be relatively small, and also, the
957 // offset shouldn't change much (only if the type of audio changes
958 // significantly). Thus, we shoot for updating this value basically
959 // “whenever we process buffers”, since the R128 calculation isn't exactly
960 // something we get out per-sample.
962 // Note that there's a feedback loop here, so we choose a very slow filter
963 // (half-time of 100 seconds).
964 double target_loudness_factor, alpha;
966 unique_lock<mutex> lock(compressor_mutex);
967 double loudness_lu = r128.loudness_M() - ref_level_lufs;
968 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
969 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
971 // If we're outside +/- 5 LU uncorrected, we don't count it as
972 // a normal signal (probably silence) and don't change the
973 // correction factor; just apply what we already have.
974 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
977 // Formula adapted from
978 // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
979 const double half_time_s = 100.0;
980 const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
981 alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
984 double m = final_makeup_gain;
985 for (size_t i = 0; i < samples_out.size(); i += 2) {
986 samples_out[i + 0] *= m;
987 samples_out[i + 1] *= m;
988 m += (target_loudness_factor - m) * alpha;
990 final_makeup_gain = m;
993 // Find R128 levels and L/R correlation.
994 vector<float> left, right;
995 deinterleave_samples(samples_out, &left, &right);
996 float *ptrs[] = { left.data(), right.data() };
998 unique_lock<mutex> lock(compressor_mutex);
999 r128.process(left.size(), ptrs);
1000 correlation.process_samples(samples_out);
1003 // Send the samples to the sound card.
1005 alsa->write(samples_out);
1008 // And finally add them to the output.
1009 h264_encoder->add_audio(frame_pts_int, move(samples_out));
1012 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1015 glGenVertexArrays(1, &vao);
1018 glBindVertexArray(vao);
1022 GLuint fbo = resource_pool->create_fbo(dst_tex);
1023 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1024 glViewport(0, 0, WIDTH/2, HEIGHT/2);
1027 glUseProgram(cbcr_program_num);
1030 glActiveTexture(GL_TEXTURE0);
1032 glBindTexture(GL_TEXTURE_2D, src_tex);
1034 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1036 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1038 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1041 float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1042 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1044 glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1047 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1048 glEnableVertexAttribArray(attr_index);
1050 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1054 glDrawArrays(GL_TRIANGLES, 0, 3);
1057 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1058 glDisableVertexAttribArray(attr_index);
1064 glBindFramebuffer(GL_FRAMEBUFFER, 0);
1067 resource_pool->release_fbo(fbo);
1068 glDeleteVertexArrays(1, &vao);
1071 void Mixer::release_display_frame(DisplayFrame *frame)
1073 for (GLuint texnum : frame->temp_textures) {
1074 resource_pool->release_2d_texture(texnum);
1076 frame->temp_textures.clear();
1077 frame->ready_fence.reset();
1078 frame->input_frames.clear();
1083 mixer_thread = thread(&Mixer::thread_func, this);
1084 audio_thread = thread(&Mixer::audio_thread_func, this);
1090 mixer_thread.join();
1091 audio_thread.join();
1094 void Mixer::transition_clicked(int transition_num)
1096 theme->transition_clicked(transition_num, pts());
1099 void Mixer::channel_clicked(int preview_num)
1101 theme->channel_clicked(preview_num);
1104 void Mixer::reset_meters()
1106 peak_resampler.reset();
1109 r128.integr_start();
1110 correlation.reset();
1113 void Mixer::start_mode_scanning(unsigned card_index)
1115 assert(card_index < num_cards);
1116 if (is_mode_scanning[card_index]) {
1119 is_mode_scanning[card_index] = true;
1120 mode_scanlist[card_index].clear();
1121 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1122 mode_scanlist[card_index].push_back(mode.first);
1124 assert(!mode_scanlist[card_index].empty());
1125 mode_scanlist_index[card_index] = 0;
1126 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1127 clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1130 Mixer::OutputChannel::~OutputChannel()
1132 if (has_current_frame) {
1133 parent->release_display_frame(¤t_frame);
1135 if (has_ready_frame) {
1136 parent->release_display_frame(&ready_frame);
1140 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1142 // Store this frame for display. Remove the ready frame if any
1143 // (it was seemingly never used).
1145 unique_lock<mutex> lock(frame_mutex);
1146 if (has_ready_frame) {
1147 parent->release_display_frame(&ready_frame);
1149 ready_frame = frame;
1150 has_ready_frame = true;
1153 if (has_new_frame_ready_callback) {
1154 new_frame_ready_callback();
1158 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1160 unique_lock<mutex> lock(frame_mutex);
1161 if (!has_current_frame && !has_ready_frame) {
1165 if (has_current_frame && has_ready_frame) {
1166 // We have a new ready frame. Toss the current one.
1167 parent->release_display_frame(¤t_frame);
1168 has_current_frame = false;
1170 if (has_ready_frame) {
1171 assert(!has_current_frame);
1172 current_frame = ready_frame;
1173 ready_frame.ready_fence.reset(); // Drop the refcount.
1174 ready_frame.input_frames.clear(); // Drop the refcounts.
1175 has_current_frame = true;
1176 has_ready_frame = false;
1179 *frame = current_frame;
1183 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1185 new_frame_ready_callback = callback;
1186 has_new_frame_ready_callback = true;