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>
17 #include <sys/resource.h>
20 #include <condition_variable>
31 #include "DeckLinkAPI.h"
33 #include "alsa_output.h"
34 #include "bmusb/bmusb.h"
35 #include "bmusb/fake_capture.h"
37 #include "decklink_capture.h"
39 #include "disk_space_estimator.h"
41 #include "input_mapping.h"
42 #include "pbo_frame_allocator.h"
43 #include "ref_counted_gl_sync.h"
44 #include "resampling_queue.h"
46 #include "video_encoder.h"
51 using namespace movit;
53 using namespace std::chrono;
54 using namespace std::placeholders;
55 using namespace bmusb;
57 Mixer *global_mixer = nullptr;
58 bool uses_mlock = false;
62 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
65 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
66 input_state->buffered_frames[card_index][frame_num] =
67 input_state->buffered_frames[card_index][frame_num - 1];
69 input_state->buffered_frames[card_index][0] = { frame, field_num };
71 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
72 input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
79 void QueueLengthPolicy::update_policy(int queue_length)
81 if (queue_length < 0) { // Starvation.
82 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
84 fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
85 card_index, safe_queue_length);
87 frames_with_at_least_one = 0;
88 been_at_safe_point_since_last_starvation = false;
91 if (queue_length > 0) {
92 if (queue_length >= int(safe_queue_length)) {
93 been_at_safe_point_since_last_starvation = true;
95 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
97 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
98 card_index, safe_queue_length);
99 frames_with_at_least_one = 0;
102 frames_with_at_least_one = 0;
106 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
108 num_cards(num_cards),
109 mixer_surface(create_surface(format)),
110 h264_encoder_surface(create_surface(format)),
111 audio_mixer(num_cards)
113 CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
116 // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
117 // will be halved when sampling them, and we need to compensate here.
118 movit_texel_subpixel_precision /= 2.0;
120 resource_pool.reset(new ResourcePool);
121 theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
122 for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
123 output_channel[i].parent = this;
124 output_channel[i].channel = i;
127 ImageFormat inout_format;
128 inout_format.color_space = COLORSPACE_sRGB;
129 inout_format.gamma_curve = GAMMA_sRGB;
131 // Display chain; shows the live output produced by the main chain (its RGBA version).
132 display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
134 display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, global_flags.width, global_flags.height); // FIXME: GL_UNSIGNED_BYTE is really wrong.
135 display_chain->add_input(display_input);
136 display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
137 display_chain->set_dither_bits(0); // Don't bother.
138 display_chain->finalize();
140 video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
142 // Start listening for clients only once VideoEncoder has written its header, if any.
145 // First try initializing the then PCI devices, then USB, then
146 // fill up with fake cards until we have the desired number of cards.
147 unsigned num_pci_devices = 0;
148 unsigned card_index = 0;
151 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
152 if (decklink_iterator != nullptr) {
153 for ( ; card_index < num_cards; ++card_index) {
155 if (decklink_iterator->Next(&decklink) != S_OK) {
159 configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
162 decklink_iterator->Release();
163 fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
165 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
168 unsigned num_usb_devices = BMUSBCapture::num_cards();
169 for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
170 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
171 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
172 configure_card(card_index, capture, /*is_fake_capture=*/false);
174 fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
176 unsigned num_fake_cards = 0;
177 for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
178 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
179 configure_card(card_index, capture, /*is_fake_capture=*/true);
182 if (num_fake_cards > 0) {
183 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
186 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
187 BMUSBCapture::start_bm_thread();
189 for (card_index = 0; card_index < num_cards; ++card_index) {
190 cards[card_index].queue_length_policy.reset(card_index);
191 cards[card_index].capture->start_bm_capture();
194 // Set up stuff for NV12 conversion.
196 // Note: Due to the horizontally co-sited chroma/luma samples in H.264
197 // (chrome position is left for horizontal and center for vertical),
198 // we need to be a bit careful in our subsampling. A diagram will make
199 // this clearer, showing some luma and chroma samples:
214 // Clearly, the rightmost chroma sample here needs to be equivalent to
215 // b/4 + c/2 + d/4. (We could also implement more sophisticated filters,
216 // of course, but as long as the upsampling is not going to be equally
217 // sophisticated, it's probably not worth it.) If we sample once with
218 // no mipmapping, we get just c, ie., no actual filtering in the
219 // horizontal direction. (For the vertical direction, we can just
220 // sample in the middle to get the right filtering.) One could imagine
221 // we could use mipmapping (assuming we can create mipmaps cheaply),
222 // but then, what we'd get is this:
237 // which ends up sampling equally from a and b, which clearly isn't right. Instead,
238 // we need to do two (non-mipmapped) chroma samples, both hitting exactly in-between
241 // Sampling in-between b and c gives us the sample (b+c)/2, and similarly for c and d.
242 // Taking the average of these gives of (b+c)/4 + (c+d)/4 = b/4 + c/2 + d/4, which is
243 // exactly what we want.
245 // See also http://www.poynton.com/PDFs/Merging_RGB_and_422.pdf, pages 6–7.
248 string cbcr_vert_shader =
251 "in vec2 position; \n"
252 "in vec2 texcoord; \n"
253 "out vec2 tc0, tc1; \n"
254 "uniform vec2 foo_chroma_offset_0; \n"
255 "uniform vec2 foo_chroma_offset_1; \n"
259 " // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
261 " // 2.000 0.000 0.000 -1.000 \n"
262 " // 0.000 2.000 0.000 -1.000 \n"
263 " // 0.000 0.000 -2.000 -1.000 \n"
264 " // 0.000 0.000 0.000 1.000 \n"
265 " gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
266 " vec2 flipped_tc = texcoord; \n"
267 " tc0 = flipped_tc + foo_chroma_offset_0; \n"
268 " tc1 = flipped_tc + foo_chroma_offset_1; \n"
270 string cbcr_frag_shader =
272 "in vec2 tc0, tc1; \n"
273 "uniform sampler2D cbcr_tex; \n"
274 "out vec4 FragColor; \n"
276 " FragColor = 0.5 * (texture(cbcr_tex, tc0) + texture(cbcr_tex, tc1)); \n"
278 vector<string> frag_shader_outputs;
279 cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
286 cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
287 cbcr_texture_sampler_uniform = glGetUniformLocation(cbcr_program_num, "cbcr_tex");
288 cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
289 cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
291 if (global_flags.enable_alsa_output) {
292 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
298 resource_pool->release_glsl_program(cbcr_program_num);
299 glDeleteBuffers(1, &cbcr_vbo);
300 BMUSBCapture::stop_bm_thread();
302 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
304 unique_lock<mutex> lock(bmusb_mutex);
305 cards[card_index].should_quit = true; // Unblock thread.
306 cards[card_index].new_frames_changed.notify_all();
308 cards[card_index].capture->stop_dequeue_thread();
311 video_encoder.reset(nullptr);
314 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
316 printf("Configuring card %d...\n", card_index);
318 CaptureCard *card = &cards[card_index];
319 if (card->capture != nullptr) {
320 card->capture->stop_dequeue_thread();
321 delete card->capture;
323 card->capture = capture;
324 card->is_fake_capture = is_fake_capture;
325 card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
326 if (card->frame_allocator == nullptr) {
327 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height)); // 8 MB.
329 card->capture->set_video_frame_allocator(card->frame_allocator.get());
330 if (card->surface == nullptr) {
331 card->surface = create_surface_with_same_format(mixer_surface);
333 while (!card->new_frames.empty()) card->new_frames.pop();
334 card->fractional_samples = 0;
335 card->last_timecode = -1;
336 card->capture->configure_card();
338 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
339 audio_mixer.reset_resampler(device);
340 audio_mixer.set_display_name(device, card->capture->get_description());
341 audio_mixer.trigger_state_changed_callback();
347 int unwrap_timecode(uint16_t current_wrapped, int last)
349 uint16_t last_wrapped = last & 0xffff;
350 if (current_wrapped > last_wrapped) {
351 return (last & ~0xffff) | current_wrapped;
353 return 0x10000 + ((last & ~0xffff) | current_wrapped);
359 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
360 FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
361 FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
363 DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
364 CaptureCard *card = &cards[card_index];
366 if (is_mode_scanning[card_index]) {
367 if (video_format.has_signal) {
368 // Found a stable signal, so stop scanning.
369 is_mode_scanning[card_index] = false;
371 static constexpr double switch_time_s = 0.5; // Should be enough time for the signal to stabilize.
372 steady_clock::time_point now = steady_clock::now();
373 double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
374 if (sec_since_last_switch > switch_time_s) {
375 // It isn't this mode; try the next one.
376 mode_scanlist_index[card_index]++;
377 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
378 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
379 last_mode_scan_change[card_index] = now;
384 int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
385 assert(frame_length > 0);
387 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;
388 if (num_samples > OUTPUT_FREQUENCY / 10) {
389 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",
390 card_index, int(audio_frame.len), int(audio_offset),
391 timecode, int(video_frame.len), int(video_offset), video_format.id);
392 if (video_frame.owner) {
393 video_frame.owner->release_frame(video_frame);
395 if (audio_frame.owner) {
396 audio_frame.owner->release_frame(audio_frame);
401 int dropped_frames = 0;
402 if (card->last_timecode != -1) {
403 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
406 // Number of samples per frame if we need to insert silence.
407 // (Could be nonintegral, but resampling will save us then.)
408 const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
410 if (dropped_frames > MAX_FPS * 2) {
411 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
412 card_index, card->last_timecode, timecode);
413 audio_mixer.reset_resampler(device);
415 } else if (dropped_frames > 0) {
416 // Insert silence as needed.
417 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
418 card_index, dropped_frames, timecode);
422 success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
426 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length);
428 // Done with the audio, so release it.
429 if (audio_frame.owner) {
430 audio_frame.owner->release_frame(audio_frame);
433 card->last_timecode = timecode;
435 size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
436 if (video_frame.len - video_offset == 0 ||
437 video_frame.len - video_offset != expected_length) {
438 if (video_frame.len != 0) {
439 printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
440 card_index, video_frame.len - video_offset, expected_length);
442 if (video_frame.owner) {
443 video_frame.owner->release_frame(video_frame);
446 // Still send on the information that we _had_ a frame, even though it's corrupted,
447 // so that pts can go up accordingly.
449 unique_lock<mutex> lock(bmusb_mutex);
450 CaptureCard::NewFrame new_frame;
451 new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
452 new_frame.length = frame_length;
453 new_frame.interlaced = false;
454 new_frame.dropped_frames = dropped_frames;
455 card->new_frames.push(move(new_frame));
456 card->new_frames_changed.notify_all();
461 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
463 unsigned num_fields = video_format.interlaced ? 2 : 1;
464 steady_clock::time_point frame_upload_start;
465 if (video_format.interlaced) {
466 // Send the two fields along as separate frames; the other side will need to add
467 // a deinterlacer to actually get this right.
468 assert(video_format.height % 2 == 0);
469 video_format.height /= 2;
470 assert(frame_length % 2 == 0);
473 frame_upload_start = steady_clock::now();
475 userdata->last_interlaced = video_format.interlaced;
476 userdata->last_has_signal = video_format.has_signal;
477 userdata->last_is_connected = video_format.is_connected;
478 userdata->last_frame_rate_nom = video_format.frame_rate_nom;
479 userdata->last_frame_rate_den = video_format.frame_rate_den;
480 RefCountedFrame frame(video_frame);
482 // Upload the textures.
483 size_t cbcr_width = video_format.width / 2;
484 size_t cbcr_offset = video_offset / 2;
485 size_t y_offset = video_frame.size / 2 + video_offset / 2;
487 for (unsigned field = 0; field < num_fields; ++field) {
488 // Put the actual texture upload in a lambda that is executed in the main thread.
489 // It is entirely possible to do this in the same thread (and it might even be
490 // faster, depending on the GPU and driver), but it appears to be trickling
491 // driver bugs very easily.
493 // Note that this means we must hold on to the actual frame data in <userdata>
494 // until the upload command is run, but we hold on to <frame> much longer than that
495 // (in fact, all the way until we no longer use the texture in rendering).
496 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
497 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
499 if (userdata->tex_y[field] == 0 ||
500 userdata->tex_cbcr[field] == 0 ||
501 video_format.width != userdata->last_width[field] ||
502 video_format.height != userdata->last_height[field]) {
503 // We changed resolution since last use of this texture, so we need to create
504 // a new object. Note that this each card has its own PBOFrameAllocator,
505 // we don't need to worry about these flip-flopping between resolutions.
506 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
508 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
510 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
512 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
514 userdata->last_width[field] = video_format.width;
515 userdata->last_height[field] = video_format.height;
518 GLuint pbo = userdata->pbo;
520 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
523 size_t field_y_start = y_offset + video_format.width * field_start_line;
524 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
526 if (global_flags.flush_pbos) {
527 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
529 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
533 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
535 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
537 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
539 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
541 glBindTexture(GL_TEXTURE_2D, 0);
543 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
548 // Don't upload the second field as fast as we can; wait until
549 // the field time has approximately passed. (Otherwise, we could
550 // get timing jitter against the other sources, and possibly also
551 // against the video display, although the latter is not as critical.)
552 // This requires our system clock to be reasonably close to the
553 // video clock, but that's not an unreasonable assumption.
554 steady_clock::time_point second_field_start = frame_upload_start +
555 nanoseconds(frame_length * 1000000000 / TIMEBASE);
556 this_thread::sleep_until(second_field_start);
560 unique_lock<mutex> lock(bmusb_mutex);
561 CaptureCard::NewFrame new_frame;
562 new_frame.frame = frame;
563 new_frame.length = frame_length;
564 new_frame.field = field;
565 new_frame.interlaced = video_format.interlaced;
566 new_frame.upload_func = upload_func;
567 new_frame.dropped_frames = dropped_frames;
568 new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp.
569 card->new_frames.push(move(new_frame));
570 card->new_frames_changed.notify_all();
575 void Mixer::bm_hotplug_add(libusb_device *dev)
577 lock_guard<mutex> lock(hotplug_mutex);
578 hotplugged_cards.push_back(dev);
581 void Mixer::bm_hotplug_remove(unsigned card_index)
583 cards[card_index].new_frames_changed.notify_all();
586 void Mixer::thread_func()
588 eglBindAPI(EGL_OPENGL_API);
589 QOpenGLContext *context = create_context(mixer_surface);
590 if (!make_current(context, mixer_surface)) {
595 steady_clock::time_point start, now;
596 start = steady_clock::now();
599 int stats_dropped_frames = 0;
601 while (!should_quit) {
602 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
603 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
604 int num_samples[MAX_VIDEO_CARDS] = { 0 };
606 unsigned master_card_index = theme->map_signal(master_clock_channel);
607 assert(master_card_index < num_cards);
609 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
610 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
611 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
613 handle_hotplugged_cards();
615 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
616 if (card_index == master_card_index || !has_new_frame[card_index]) {
619 if (new_frames[card_index].frame->len == 0) {
620 ++new_frames[card_index].dropped_frames;
622 if (new_frames[card_index].dropped_frames > 0) {
623 printf("Card %u dropped %d frames before this\n",
624 card_index, int(new_frames[card_index].dropped_frames));
628 // If the first card is reporting a corrupted or otherwise dropped frame,
629 // just increase the pts (skipping over this frame) and don't try to compute anything new.
630 if (new_frames[master_card_index].frame->len == 0) {
631 ++stats_dropped_frames;
632 pts_int += new_frames[master_card_index].length;
636 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
637 if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
640 CaptureCard::NewFrame *new_frame = &new_frames[card_index];
641 assert(new_frame->frame != nullptr);
642 insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
645 // The new texture might need uploading before use.
646 if (new_frame->upload_func) {
647 new_frame->upload_func();
648 new_frame->upload_func = nullptr;
652 int64_t frame_duration = new_frames[master_card_index].length;
653 render_one_frame(frame_duration);
655 pts_int += frame_duration;
657 now = steady_clock::now();
658 double elapsed = duration<double>(now - start).count();
659 if (frame % 100 == 0) {
660 printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
661 frame, stats_dropped_frames, elapsed, frame / elapsed,
662 1e3 * elapsed / frame);
663 // chain->print_phase_timing();
665 // Check our memory usage, to see if we are close to our mlockall()
666 // limit (if at all set).
668 if (getrusage(RUSAGE_SELF, &used) == -1) {
669 perror("getrusage(RUSAGE_SELF)");
675 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
676 perror("getrlimit(RLIMIT_MEMLOCK)");
680 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
681 long(used.ru_maxrss / 1024),
682 long(limit.rlim_cur / 1048576),
683 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
685 printf(", using %ld MB memory (not locked)",
686 long(used.ru_maxrss / 1024));
693 if (should_cut.exchange(false)) { // Test and clear.
694 video_encoder->do_cut(frame);
698 // Reset every 100 frames, so that local variations in frame times
699 // (especially for the first few frames, when the shaders are
700 // compiled etc.) don't make it hard to measure for the entire
701 // remaining duration of the program.
702 if (frame == 10000) {
710 resource_pool->clean_context();
713 void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS], int num_samples[MAX_VIDEO_CARDS])
716 // The first card is the master timer, so wait for it to have a new frame.
717 // TODO: Add a timeout.
718 unique_lock<mutex> lock(bmusb_mutex);
719 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
721 if (cards[master_card_index].new_frames.empty()) {
722 // We were woken up, but not due to a new frame. Deal with it
724 assert(cards[master_card_index].capture->get_disconnected());
725 handle_hotplugged_cards();
729 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
730 CaptureCard *card = &cards[card_index];
731 if (card->new_frames.empty()) {
732 assert(card_index != master_card_index);
733 card->queue_length_policy.update_policy(-1);
736 new_frames[card_index] = move(card->new_frames.front());
737 has_new_frame[card_index] = true;
738 card->new_frames.pop();
739 card->new_frames_changed.notify_all();
741 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
742 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
743 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
744 assert(num_samples[card_index] >= 0);
746 if (card_index == master_card_index) {
747 // We don't use the queue length policy for the master card,
748 // but we will if it stops being the master. Thus, clear out
749 // the policy in case we switch in the future.
750 card->queue_length_policy.reset(card_index);
752 // If we have excess frames compared to the policy for this card,
753 // drop frames from the head.
754 card->queue_length_policy.update_policy(card->new_frames.size());
755 while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
756 card->new_frames.pop();
762 void Mixer::handle_hotplugged_cards()
764 // Check for cards that have been disconnected since last frame.
765 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
766 CaptureCard *card = &cards[card_index];
767 if (card->capture->get_disconnected()) {
768 fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
769 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
770 configure_card(card_index, capture, /*is_fake_capture=*/true);
771 card->queue_length_policy.reset(card_index);
772 card->capture->start_bm_capture();
776 // Check for cards that have been connected since last frame.
777 vector<libusb_device *> hotplugged_cards_copy;
779 lock_guard<mutex> lock(hotplug_mutex);
780 swap(hotplugged_cards, hotplugged_cards_copy);
782 for (libusb_device *new_dev : hotplugged_cards_copy) {
783 // Look for a fake capture card where we can stick this in.
784 int free_card_index = -1;
785 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
786 if (cards[card_index].is_fake_capture) {
787 free_card_index = int(card_index);
792 if (free_card_index == -1) {
793 fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
794 libusb_unref_device(new_dev);
796 // BMUSBCapture takes ownership.
797 fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
798 CaptureCard *card = &cards[free_card_index];
799 BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
800 configure_card(free_card_index, capture, /*is_fake_capture=*/false);
801 card->queue_length_policy.reset(free_card_index);
802 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
803 capture->start_bm_capture();
809 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
811 // Resample the audio as needed, including from previously dropped frames.
812 assert(num_cards > 0);
813 for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
814 const bool dropped_frame = (frame_num != dropped_frames);
816 // Signal to the audio thread to process this frame.
817 // Note that if the frame is a dropped frame, we signal that
818 // we don't want to use this frame as base for adjusting
819 // the resampler rate. The reason for this is that the timing
820 // of these frames is often way too late; they typically don't
821 // “arrive” before we synthesize them. Thus, we could end up
822 // in a situation where we have inserted e.g. five audio frames
823 // into the queue before we then start pulling five of them
824 // back out. This makes ResamplingQueue overestimate the delay,
825 // causing undue resampler changes. (We _do_ use the last,
826 // non-dropped frame; perhaps we should just discard that as well,
827 // since dropped frames are expected to be rare, and it might be
828 // better to just wait until we have a slightly more normal situation).
829 unique_lock<mutex> lock(audio_mutex);
830 bool adjust_rate = !dropped_frame;
831 audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
832 audio_task_queue_changed.notify_one();
835 // For dropped frames, increase the pts. Note that if the format changed
836 // in the meantime, we have no way of detecting that; we just have to
837 // assume the frame length is always the same.
838 pts_int += length_per_frame;
843 void Mixer::render_one_frame(int64_t duration)
845 // Get the main chain from the theme, and set its state immediately.
846 Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
847 EffectChain *chain = theme_main_chain.chain;
848 theme_main_chain.setup_chain();
849 //theme_main_chain.chain->enable_phase_timing(true);
851 GLuint y_tex, cbcr_tex;
852 bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
855 // Render main chain.
856 GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
857 GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, global_flags.width, global_flags.height); // Saves texture bandwidth, although dithering gets messed up.
858 GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
860 chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
861 resource_pool->release_fbo(fbo);
863 subsample_chroma(cbcr_full_tex, cbcr_tex);
864 resource_pool->release_2d_texture(cbcr_full_tex);
866 // Set the right state for rgba_tex.
867 glBindFramebuffer(GL_FRAMEBUFFER, 0);
868 glBindTexture(GL_TEXTURE_2D, rgba_tex);
869 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
870 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
871 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
873 const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
874 RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
876 // The live frame just shows the RGBA texture we just rendered.
877 // It owns rgba_tex now.
878 DisplayFrame live_frame;
879 live_frame.chain = display_chain.get();
880 live_frame.setup_chain = [this, rgba_tex]{
881 display_input->set_texture_num(rgba_tex);
883 live_frame.ready_fence = fence;
884 live_frame.input_frames = {};
885 live_frame.temp_textures = { rgba_tex };
886 output_channel[OUTPUT_LIVE].output_frame(live_frame);
888 // Set up preview and any additional channels.
889 for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
890 DisplayFrame display_frame;
891 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions
892 display_frame.chain = chain.chain;
893 display_frame.setup_chain = chain.setup_chain;
894 display_frame.ready_fence = fence;
895 display_frame.input_frames = chain.input_frames;
896 display_frame.temp_textures = {};
897 output_channel[i].output_frame(display_frame);
901 void Mixer::audio_thread_func()
903 while (!should_quit) {
907 unique_lock<mutex> lock(audio_mutex);
908 audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
912 task = audio_task_queue.front();
913 audio_task_queue.pop();
916 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
917 task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
918 vector<float> samples_out = audio_mixer.get_output(
919 double(task.pts_int) / TIMEBASE,
921 rate_adjustment_policy);
923 // Send the samples to the sound card, then add them to the output.
925 alsa->write(samples_out);
927 video_encoder->add_audio(task.pts_int, move(samples_out));
931 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
934 glGenVertexArrays(1, &vao);
937 glBindVertexArray(vao);
941 GLuint fbo = resource_pool->create_fbo(dst_tex);
942 glBindFramebuffer(GL_FRAMEBUFFER, fbo);
943 glViewport(0, 0, global_flags.width/2, global_flags.height/2);
946 glUseProgram(cbcr_program_num);
949 glActiveTexture(GL_TEXTURE0);
951 glBindTexture(GL_TEXTURE_2D, src_tex);
953 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
955 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
957 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
960 float chroma_offset_0[] = { -1.0f / global_flags.width, 0.0f };
961 float chroma_offset_1[] = { -0.0f / global_flags.width, 0.0f };
962 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
963 set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_1", chroma_offset_1);
965 glUniform1i(cbcr_texture_sampler_uniform, 0);
967 glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
970 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
971 glEnableVertexAttribArray(attr_index);
973 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
977 glDrawArrays(GL_TRIANGLES, 0, 3);
980 for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
981 glDisableVertexAttribArray(attr_index);
987 glBindFramebuffer(GL_FRAMEBUFFER, 0);
990 resource_pool->release_fbo(fbo);
991 glDeleteVertexArrays(1, &vao);
994 void Mixer::release_display_frame(DisplayFrame *frame)
996 for (GLuint texnum : frame->temp_textures) {
997 resource_pool->release_2d_texture(texnum);
999 frame->temp_textures.clear();
1000 frame->ready_fence.reset();
1001 frame->input_frames.clear();
1006 mixer_thread = thread(&Mixer::thread_func, this);
1007 audio_thread = thread(&Mixer::audio_thread_func, this);
1013 audio_task_queue_changed.notify_one();
1014 mixer_thread.join();
1015 audio_thread.join();
1018 void Mixer::transition_clicked(int transition_num)
1020 theme->transition_clicked(transition_num, pts());
1023 void Mixer::channel_clicked(int preview_num)
1025 theme->channel_clicked(preview_num);
1028 void Mixer::start_mode_scanning(unsigned card_index)
1030 assert(card_index < num_cards);
1031 if (is_mode_scanning[card_index]) {
1034 is_mode_scanning[card_index] = true;
1035 mode_scanlist[card_index].clear();
1036 for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1037 mode_scanlist[card_index].push_back(mode.first);
1039 assert(!mode_scanlist[card_index].empty());
1040 mode_scanlist_index[card_index] = 0;
1041 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1042 last_mode_scan_change[card_index] = steady_clock::now();
1045 Mixer::OutputChannel::~OutputChannel()
1047 if (has_current_frame) {
1048 parent->release_display_frame(¤t_frame);
1050 if (has_ready_frame) {
1051 parent->release_display_frame(&ready_frame);
1055 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1057 // Store this frame for display. Remove the ready frame if any
1058 // (it was seemingly never used).
1060 unique_lock<mutex> lock(frame_mutex);
1061 if (has_ready_frame) {
1062 parent->release_display_frame(&ready_frame);
1064 ready_frame = frame;
1065 has_ready_frame = true;
1068 if (new_frame_ready_callback) {
1069 new_frame_ready_callback();
1072 // Reduce the number of callbacks by filtering duplicates. The reason
1073 // why we bother doing this is that Qt seemingly can get into a state
1074 // where its builds up an essentially unbounded queue of signals,
1075 // consuming more and more memory, and there's no good way of collapsing
1076 // user-defined signals or limiting the length of the queue.
1077 if (transition_names_updated_callback) {
1078 vector<string> transition_names = global_mixer->get_transition_names();
1079 bool changed = false;
1080 if (transition_names.size() != last_transition_names.size()) {
1083 for (unsigned i = 0; i < transition_names.size(); ++i) {
1084 if (transition_names[i] != last_transition_names[i]) {
1091 transition_names_updated_callback(transition_names);
1092 last_transition_names = transition_names;
1095 if (name_updated_callback) {
1096 string name = global_mixer->get_channel_name(channel);
1097 if (name != last_name) {
1098 name_updated_callback(name);
1102 if (color_updated_callback) {
1103 string color = global_mixer->get_channel_color(channel);
1104 if (color != last_color) {
1105 color_updated_callback(color);
1111 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1113 unique_lock<mutex> lock(frame_mutex);
1114 if (!has_current_frame && !has_ready_frame) {
1118 if (has_current_frame && has_ready_frame) {
1119 // We have a new ready frame. Toss the current one.
1120 parent->release_display_frame(¤t_frame);
1121 has_current_frame = false;
1123 if (has_ready_frame) {
1124 assert(!has_current_frame);
1125 current_frame = ready_frame;
1126 ready_frame.ready_fence.reset(); // Drop the refcount.
1127 ready_frame.input_frames.clear(); // Drop the refcounts.
1128 has_current_frame = true;
1129 has_ready_frame = false;
1132 *frame = current_frame;
1136 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1138 new_frame_ready_callback = callback;
1141 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1143 transition_names_updated_callback = callback;
1146 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1148 name_updated_callback = callback;
1151 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1153 color_updated_callback = callback;
1156 mutex RefCountedGLsync::fence_lock;