git.sesse.net Git - nageru/blob - mixer.cpp

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   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 <stdint.h>
  14 #include <stdio.h>
  15 #include <stdlib.h>
  16 #include <sys/resource.h>
  17 #include <algorithm>
  18 #include <chrono>
  19 #include <condition_variable>
  20 #include <cstddef>
  21 #include <cstdint>
  22 #include <memory>
  23 #include <mutex>
  24 #include <ratio>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29
  30 #include "DeckLinkAPI.h"
  31 #include "LinuxCOM.h"
  32 #include "alsa_output.h"
  33 #include "bmusb/bmusb.h"
  34 #include "bmusb/fake_capture.h"
  35 #include "chroma_subsampler.h"
  36 #include "context.h"
  37 #include "decklink_capture.h"
  38 #include "defs.h"
  39 #include "disk_space_estimator.h"
  40 #include "flags.h"
  41 #include "input_mapping.h"
  42 #include "pbo_frame_allocator.h"
  43 #include "ref_counted_gl_sync.h"
  44 #include "resampling_queue.h"
  45 #include "timebase.h"
  46 #include "video_encoder.h"
  47
  48 class IDeckLink;
  49 class QOpenGLContext;
  50
  51 using namespace movit;
  52 using namespace std;
  53 using namespace std::chrono;
  54 using namespace std::placeholders;
  55 using namespace bmusb;
  56
  57 Mixer *global_mixer = nullptr;
  58 bool uses_mlock = false;
  59
  60 namespace {
  61
  62 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  63 {
  64         if (interlaced) {
  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];
  68                 }
  69                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  70         } else {
  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 };
  73                 }
  74         }
  75 }
  76
  77 }  // namespace
  78
  79 void QueueLengthPolicy::update_policy(int queue_length)
  80 {
  81         if (queue_length < 0) {  // Starvation.
  82                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
  83                         ++safe_queue_length;
  84                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
  85                                 card_index, safe_queue_length);
  86                 }
  87                 frames_with_at_least_one = 0;
  88                 been_at_safe_point_since_last_starvation = false;
  89                 return;
  90         }
  91         if (queue_length > 0) {
  92                 if (queue_length >= int(safe_queue_length)) {
  93                         been_at_safe_point_since_last_starvation = true;
  94                 }
  95                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
  96                         --safe_queue_length;
  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;
 100                 }
 101         } else {
 102                 frames_with_at_least_one = 0;
 103         }
 104 }
 105
 106 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 107         : httpd(),
 108           num_cards(num_cards),
 109           mixer_surface(create_surface(format)),
 110           h264_encoder_surface(create_surface(format)),
 111           audio_mixer(num_cards)
 112 {
 113         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 114         check_error();
 115
 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;
 119
 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;
 125         }
 126
 127         ImageFormat inout_format;
 128         inout_format.color_space = COLORSPACE_sRGB;
 129         inout_format.gamma_curve = GAMMA_sRGB;
 130
 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()));
 133         check_error();
 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();
 139
 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));
 141
 142         // Start listening for clients only once VideoEncoder has written its header, if any.
 143         httpd.start(9095);
 144
 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;
 149
 150         {
 151                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 152                 if (decklink_iterator != nullptr) {
 153                         for ( ; card_index < num_cards; ++card_index) {
 154                                 IDeckLink *decklink;
 155                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 156                                         break;
 157                                 }
 158
 159                                 configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
 160                                 ++num_pci_devices;
 161                         }
 162                         decklink_iterator->Release();
 163                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 164                 } else {
 165                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 166                 }
 167         }
 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);
 173         }
 174         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 175
 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);
 180         }
 181
 182         if (num_fake_cards > 0) {
 183                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 184         }
 185
 186         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 187         BMUSBCapture::start_bm_thread();
 188
 189         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 190                 cards[card_index].queue_length_policy.reset(card_index);
 191         }
 192
 193         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 194
 195         if (global_flags.enable_alsa_output) {
 196                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 197         }
 198 }
 199
 200 Mixer::~Mixer()
 201 {
 202         BMUSBCapture::stop_bm_thread();
 203
 204         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 205                 {
 206                         unique_lock<mutex> lock(card_mutex);
 207                         cards[card_index].should_quit = true;  // Unblock thread.
 208                         cards[card_index].new_frames_changed.notify_all();
 209                 }
 210                 cards[card_index].capture->stop_dequeue_thread();
 211         }
 212
 213         video_encoder.reset(nullptr);
 214 }
 215
 216 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
 217 {
 218         printf("Configuring card %d...\n", card_index);
 219
 220         CaptureCard *card = &cards[card_index];
 221         if (card->capture != nullptr) {
 222                 card->capture->stop_dequeue_thread();
 223                 delete card->capture;
 224         }
 225         card->capture = capture;
 226         card->is_fake_capture = is_fake_capture;
 227         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 228         if (card->frame_allocator == nullptr) {
 229                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 230         }
 231         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 232         if (card->surface == nullptr) {
 233                 card->surface = create_surface_with_same_format(mixer_surface);
 234         }
 235         while (!card->new_frames.empty()) card->new_frames.pop();
 236         card->last_timecode = -1;
 237         card->capture->configure_card();
 238
 239         // NOTE: start_bm_capture() happens in thread_func().
 240
 241         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 242         audio_mixer.reset_resampler(device);
 243         audio_mixer.set_display_name(device, card->capture->get_description());
 244         audio_mixer.trigger_state_changed_callback();
 245 }
 246
 247
 248 namespace {
 249
 250 int unwrap_timecode(uint16_t current_wrapped, int last)
 251 {
 252         uint16_t last_wrapped = last & 0xffff;
 253         if (current_wrapped > last_wrapped) {
 254                 return (last & ~0xffff) | current_wrapped;
 255         } else {
 256                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 257         }
 258 }
 259
 260 }  // namespace
 261
 262 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 263                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 264                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 265 {
 266         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 267         CaptureCard *card = &cards[card_index];
 268
 269         if (is_mode_scanning[card_index]) {
 270                 if (video_format.has_signal) {
 271                         // Found a stable signal, so stop scanning.
 272                         is_mode_scanning[card_index] = false;
 273                 } else {
 274                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 275                         steady_clock::time_point now = steady_clock::now();
 276                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 277                         if (sec_since_last_switch > switch_time_s) {
 278                                 // It isn't this mode; try the next one.
 279                                 mode_scanlist_index[card_index]++;
 280                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 281                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 282                                 last_mode_scan_change[card_index] = now;
 283                         }
 284                 }
 285         }
 286
 287         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 288         assert(frame_length > 0);
 289
 290         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;
 291         if (num_samples > OUTPUT_FREQUENCY / 10) {
 292                 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",
 293                         card_index, int(audio_frame.len), int(audio_offset),
 294                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 295                 if (video_frame.owner) {
 296                         video_frame.owner->release_frame(video_frame);
 297                 }
 298                 if (audio_frame.owner) {
 299                         audio_frame.owner->release_frame(audio_frame);
 300                 }
 301                 return;
 302         }
 303
 304         int dropped_frames = 0;
 305         if (card->last_timecode != -1) {
 306                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 307         }
 308
 309         // Number of samples per frame if we need to insert silence.
 310         // (Could be nonintegral, but resampling will save us then.)
 311         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 312
 313         if (dropped_frames > MAX_FPS * 2) {
 314                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 315                         card_index, card->last_timecode, timecode);
 316                 audio_mixer.reset_resampler(device);
 317                 dropped_frames = 0;
 318         } else if (dropped_frames > 0) {
 319                 // Insert silence as needed.
 320                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 321                         card_index, dropped_frames, timecode);
 322
 323                 bool success;
 324                 do {
 325                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 326                 } while (!success);
 327         }
 328
 329         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length);
 330
 331         // Done with the audio, so release it.
 332         if (audio_frame.owner) {
 333                 audio_frame.owner->release_frame(audio_frame);
 334         }
 335
 336         card->last_timecode = timecode;
 337
 338         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 339         if (video_frame.len - video_offset == 0 ||
 340             video_frame.len - video_offset != expected_length) {
 341                 if (video_frame.len != 0) {
 342                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 343                                 card_index, video_frame.len - video_offset, expected_length);
 344                 }
 345                 if (video_frame.owner) {
 346                         video_frame.owner->release_frame(video_frame);
 347                 }
 348
 349                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 350                 // so that pts can go up accordingly.
 351                 {
 352                         unique_lock<mutex> lock(card_mutex);
 353                         CaptureCard::NewFrame new_frame;
 354                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 355                         new_frame.length = frame_length;
 356                         new_frame.interlaced = false;
 357                         new_frame.dropped_frames = dropped_frames;
 358                         card->new_frames.push(move(new_frame));
 359                         card->new_frames_changed.notify_all();
 360                 }
 361                 return;
 362         }
 363
 364         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 365
 366         unsigned num_fields = video_format.interlaced ? 2 : 1;
 367         steady_clock::time_point frame_upload_start;
 368         bool interlaced_stride = false;
 369         if (video_format.interlaced) {
 370                 // Send the two fields along as separate frames; the other side will need to add
 371                 // a deinterlacer to actually get this right.
 372                 assert(video_format.height % 2 == 0);
 373                 video_format.height /= 2;
 374                 assert(frame_length % 2 == 0);
 375                 frame_length /= 2;
 376                 num_fields = 2;
 377                 if (video_format.second_field_start == 1) {
 378                         interlaced_stride = true;
 379                 }
 380                 frame_upload_start = steady_clock::now();
 381         }
 382         userdata->last_interlaced = video_format.interlaced;
 383         userdata->last_has_signal = video_format.has_signal;
 384         userdata->last_is_connected = video_format.is_connected;
 385         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 386         userdata->last_frame_rate_den = video_format.frame_rate_den;
 387         RefCountedFrame frame(video_frame);
 388
 389         // Upload the textures.
 390         size_t cbcr_width = video_format.width / 2;
 391         size_t cbcr_offset = video_offset / 2;
 392         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 393
 394         for (unsigned field = 0; field < num_fields; ++field) {
 395                 // Put the actual texture upload in a lambda that is executed in the main thread.
 396                 // It is entirely possible to do this in the same thread (and it might even be
 397                 // faster, depending on the GPU and driver), but it appears to be trickling
 398                 // driver bugs very easily.
 399                 //
 400                 // Note that this means we must hold on to the actual frame data in <userdata>
 401                 // until the upload command is run, but we hold on to <frame> much longer than that
 402                 // (in fact, all the way until we no longer use the texture in rendering).
 403                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 404                         unsigned field_start_line;
 405                         if (field == 1) {
 406                                 field_start_line = video_format.second_field_start;
 407                         } else {
 408                                 field_start_line = video_format.extra_lines_top;
 409                         }
 410
 411                         if (userdata->tex_y[field] == 0 ||
 412                             userdata->tex_cbcr[field] == 0 ||
 413                             video_format.width != userdata->last_width[field] ||
 414                             video_format.height != userdata->last_height[field]) {
 415                                 // We changed resolution since last use of this texture, so we need to create
 416                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 417                                 // we don't need to worry about these flip-flopping between resolutions.
 418                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 419                                 check_error();
 420                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 421                                 check_error();
 422                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 423                                 check_error();
 424                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 425                                 check_error();
 426                                 userdata->last_width[field] = video_format.width;
 427                                 userdata->last_height[field] = video_format.height;
 428                         }
 429
 430                         GLuint pbo = userdata->pbo;
 431                         check_error();
 432                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 433                         check_error();
 434
 435                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 436                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 437
 438                         if (global_flags.flush_pbos) {
 439                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 440                                 check_error();
 441                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 442                                 check_error();
 443                         }
 444
 445                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 446                         check_error();
 447                         if (interlaced_stride) {
 448                                 glPixelStorei(GL_UNPACK_ROW_LENGTH, cbcr_width * 2);
 449                                 check_error();
 450                         } else {
 451                                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 452                                 check_error();
 453                         }
 454                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 455                         check_error();
 456                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 457                         check_error();
 458                         if (interlaced_stride) {
 459                                 glPixelStorei(GL_UNPACK_ROW_LENGTH, video_format.width * 2);
 460                                 check_error();
 461                         } else {
 462                                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 463                                 check_error();
 464                         }
 465                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 466                         check_error();
 467                         glBindTexture(GL_TEXTURE_2D, 0);
 468                         check_error();
 469                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 470                         check_error();
 471                         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 472                         check_error();
 473                 };
 474
 475                 if (field == 1) {
 476                         // Don't upload the second field as fast as we can; wait until
 477                         // the field time has approximately passed. (Otherwise, we could
 478                         // get timing jitter against the other sources, and possibly also
 479                         // against the video display, although the latter is not as critical.)
 480                         // This requires our system clock to be reasonably close to the
 481                         // video clock, but that's not an unreasonable assumption.
 482                         steady_clock::time_point second_field_start = frame_upload_start +
 483                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 484                         this_thread::sleep_until(second_field_start);
 485                 }
 486
 487                 {
 488                         unique_lock<mutex> lock(card_mutex);
 489                         CaptureCard::NewFrame new_frame;
 490                         new_frame.frame = frame;
 491                         new_frame.length = frame_length;
 492                         new_frame.field = field;
 493                         new_frame.interlaced = video_format.interlaced;
 494                         new_frame.upload_func = upload_func;
 495                         new_frame.dropped_frames = dropped_frames;
 496                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 497                         card->new_frames.push(move(new_frame));
 498                         card->new_frames_changed.notify_all();
 499                 }
 500         }
 501 }
 502
 503 void Mixer::bm_hotplug_add(libusb_device *dev)
 504 {
 505         lock_guard<mutex> lock(hotplug_mutex);
 506         hotplugged_cards.push_back(dev);
 507 }
 508
 509 void Mixer::bm_hotplug_remove(unsigned card_index)
 510 {
 511         cards[card_index].new_frames_changed.notify_all();
 512 }
 513
 514 void Mixer::thread_func()
 515 {
 516         eglBindAPI(EGL_OPENGL_API);
 517         QOpenGLContext *context = create_context(mixer_surface);
 518         if (!make_current(context, mixer_surface)) {
 519                 printf("oops\n");
 520                 exit(1);
 521         }
 522
 523         // Start the actual capture. (We don't want to do it before we're actually ready
 524         // to process output frames.)
 525         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 526                 cards[card_index].capture->start_bm_capture();
 527         }
 528
 529         steady_clock::time_point start, now;
 530         start = steady_clock::now();
 531
 532         int frame = 0;
 533         int stats_dropped_frames = 0;
 534
 535         while (!should_quit) {
 536                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 537                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 538
 539                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 540                 assert(master_card_index < num_cards);
 541
 542                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame);
 543                 schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration);
 544                 stats_dropped_frames += output_frame_info.dropped_frames;
 545
 546                 handle_hotplugged_cards();
 547
 548                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 549                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 550                                 continue;
 551                         }
 552                         if (new_frames[card_index].frame->len == 0) {
 553                                 ++new_frames[card_index].dropped_frames;
 554                         }
 555                         if (new_frames[card_index].dropped_frames > 0) {
 556                                 printf("Card %u dropped %d frames before this\n",
 557                                         card_index, int(new_frames[card_index].dropped_frames));
 558                         }
 559                 }
 560
 561                 // If the first card is reporting a corrupted or otherwise dropped frame,
 562                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 563                 if (new_frames[master_card_index].frame->len == 0) {
 564                         ++stats_dropped_frames;
 565                         pts_int += new_frames[master_card_index].length;
 566                         continue;
 567                 }
 568
 569                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 570                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 571                                 continue;
 572
 573                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 574                         assert(new_frame->frame != nullptr);
 575                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 576                         check_error();
 577
 578                         // The new texture might need uploading before use.
 579                         if (new_frame->upload_func) {
 580                                 new_frame->upload_func();
 581                                 new_frame->upload_func = nullptr;
 582                         }
 583                 }
 584
 585                 int64_t frame_duration = output_frame_info.frame_duration;
 586                 render_one_frame(frame_duration);
 587                 ++frame;
 588                 pts_int += frame_duration;
 589
 590                 now = steady_clock::now();
 591                 double elapsed = duration<double>(now - start).count();
 592                 if (frame % 100 == 0) {
 593                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 594                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 595                                 1e3 * elapsed / frame);
 596                 //      chain->print_phase_timing();
 597
 598                         // Check our memory usage, to see if we are close to our mlockall()
 599                         // limit (if at all set).
 600                         rusage used;
 601                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 602                                 perror("getrusage(RUSAGE_SELF)");
 603                                 assert(false);
 604                         }
 605
 606                         if (uses_mlock) {
 607                                 rlimit limit;
 608                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 609                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 610                                         assert(false);
 611                                 }
 612
 613                                 if (limit.rlim_cur == 0) {
 614                                         printf(", using %ld MB memory (locked)",
 615                                                 long(used.ru_maxrss / 1024));
 616                                 } else {
 617                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 618                                                 long(used.ru_maxrss / 1024),
 619                                                 long(limit.rlim_cur / 1048576),
 620                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 621                                 }
 622                         } else {
 623                                 printf(", using %ld MB memory (not locked)",
 624                                         long(used.ru_maxrss / 1024));
 625                         }
 626
 627                         printf("\n");
 628                 }
 629
 630
 631                 if (should_cut.exchange(false)) {  // Test and clear.
 632                         video_encoder->do_cut(frame);
 633                 }
 634
 635 #if 0
 636                 // Reset every 100 frames, so that local variations in frame times
 637                 // (especially for the first few frames, when the shaders are
 638                 // compiled etc.) don't make it hard to measure for the entire
 639                 // remaining duration of the program.
 640                 if (frame == 10000) {
 641                         frame = 0;
 642                         start = now;
 643                 }
 644 #endif
 645                 check_error();
 646         }
 647
 648         resource_pool->clean_context();
 649 }
 650
 651 Mixer::OutputFrameInfo 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])
 652 {
 653         OutputFrameInfo output_frame_info;
 654
 655 start:
 656         // The first card is the master timer, so wait for it to have a new frame.
 657         // TODO: Add a timeout.
 658         unique_lock<mutex> lock(card_mutex);
 659         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(); });
 660
 661         if (cards[master_card_index].new_frames.empty()) {
 662                 // We were woken up, but not due to a new frame. Deal with it
 663                 // and then restart.
 664                 assert(cards[master_card_index].capture->get_disconnected());
 665                 handle_hotplugged_cards();
 666                 goto start;
 667         }
 668
 669         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 670                 CaptureCard *card = &cards[card_index];
 671                 if (card->new_frames.empty()) {
 672                         assert(card_index != master_card_index);
 673                         card->queue_length_policy.update_policy(-1);
 674                         continue;
 675                 }
 676                 new_frames[card_index] = move(card->new_frames.front());
 677                 has_new_frame[card_index] = true;
 678                 card->new_frames.pop();
 679                 card->new_frames_changed.notify_all();
 680
 681                 if (card_index == master_card_index) {
 682                         // We don't use the queue length policy for the master card,
 683                         // but we will if it stops being the master. Thus, clear out
 684                         // the policy in case we switch in the future.
 685                         card->queue_length_policy.reset(card_index);
 686                 } else {
 687                         // If we have excess frames compared to the policy for this card,
 688                         // drop frames from the head.
 689                         card->queue_length_policy.update_policy(card->new_frames.size());
 690                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 691                                 card->new_frames.pop();
 692                         }
 693                 }
 694         }
 695
 696         output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 697         output_frame_info.frame_duration = new_frames[master_card_index].length;
 698
 699         // This might get off by a fractional sample when changing master card
 700         // between ones with different frame rates, but that's fine.
 701         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 702         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 703         fractional_samples = num_samples_times_timebase % TIMEBASE;
 704         assert(output_frame_info.num_samples >= 0);
 705
 706         return output_frame_info;
 707 }
 708
 709 void Mixer::handle_hotplugged_cards()
 710 {
 711         // Check for cards that have been disconnected since last frame.
 712         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 713                 CaptureCard *card = &cards[card_index];
 714                 if (card->capture->get_disconnected()) {
 715                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 716                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 717                         configure_card(card_index, capture, /*is_fake_capture=*/true);
 718                         card->queue_length_policy.reset(card_index);
 719                         card->capture->start_bm_capture();
 720                 }
 721         }
 722
 723         // Check for cards that have been connected since last frame.
 724         vector<libusb_device *> hotplugged_cards_copy;
 725         {
 726                 lock_guard<mutex> lock(hotplug_mutex);
 727                 swap(hotplugged_cards, hotplugged_cards_copy);
 728         }
 729         for (libusb_device *new_dev : hotplugged_cards_copy) {
 730                 // Look for a fake capture card where we can stick this in.
 731                 int free_card_index = -1;
 732                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 733                         if (cards[card_index].is_fake_capture) {
 734                                 free_card_index = int(card_index);
 735                                 break;
 736                         }
 737                 }
 738
 739                 if (free_card_index == -1) {
 740                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 741                         libusb_unref_device(new_dev);
 742                 } else {
 743                         // BMUSBCapture takes ownership.
 744                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 745                         CaptureCard *card = &cards[free_card_index];
 746                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 747                         configure_card(free_card_index, capture, /*is_fake_capture=*/false);
 748                         card->queue_length_policy.reset(free_card_index);
 749                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 750                         capture->start_bm_capture();
 751                 }
 752         }
 753 }
 754
 755
 756 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 757 {
 758         // Resample the audio as needed, including from previously dropped frames.
 759         assert(num_cards > 0);
 760         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 761                 const bool dropped_frame = (frame_num != dropped_frames);
 762                 {
 763                         // Signal to the audio thread to process this frame.
 764                         // Note that if the frame is a dropped frame, we signal that
 765                         // we don't want to use this frame as base for adjusting
 766                         // the resampler rate. The reason for this is that the timing
 767                         // of these frames is often way too late; they typically don't
 768                         // “arrive” before we synthesize them. Thus, we could end up
 769                         // in a situation where we have inserted e.g. five audio frames
 770                         // into the queue before we then start pulling five of them
 771                         // back out. This makes ResamplingQueue overestimate the delay,
 772                         // causing undue resampler changes. (We _do_ use the last,
 773                         // non-dropped frame; perhaps we should just discard that as well,
 774                         // since dropped frames are expected to be rare, and it might be
 775                         // better to just wait until we have a slightly more normal situation).
 776                         unique_lock<mutex> lock(audio_mutex);
 777                         bool adjust_rate = !dropped_frame;
 778                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
 779                         audio_task_queue_changed.notify_one();
 780                 }
 781                 if (dropped_frame) {
 782                         // For dropped frames, increase the pts. Note that if the format changed
 783                         // in the meantime, we have no way of detecting that; we just have to
 784                         // assume the frame length is always the same.
 785                         pts_int += length_per_frame;
 786                 }
 787         }
 788 }
 789
 790 void Mixer::render_one_frame(int64_t duration)
 791 {
 792         // Get the main chain from the theme, and set its state immediately.
 793         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
 794         EffectChain *chain = theme_main_chain.chain;
 795         theme_main_chain.setup_chain();
 796         //theme_main_chain.chain->enable_phase_timing(true);
 797
 798         GLuint y_tex, cbcr_tex;
 799         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 800         assert(got_frame);
 801
 802         // Render main chain.
 803         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
 804         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, global_flags.width, global_flags.height);  // Saves texture bandwidth, although dithering gets messed up.
 805         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 806         check_error();
 807         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
 808         resource_pool->release_fbo(fbo);
 809
 810         chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
 811         resource_pool->release_2d_texture(cbcr_full_tex);
 812
 813         // Set the right state for rgba_tex.
 814         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 815         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 816         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 817         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 818         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 819
 820         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
 821         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 822
 823         // The live frame just shows the RGBA texture we just rendered.
 824         // It owns rgba_tex now.
 825         DisplayFrame live_frame;
 826         live_frame.chain = display_chain.get();
 827         live_frame.setup_chain = [this, rgba_tex]{
 828                 display_input->set_texture_num(rgba_tex);
 829         };
 830         live_frame.ready_fence = fence;
 831         live_frame.input_frames = {};
 832         live_frame.temp_textures = { rgba_tex };
 833         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 834
 835         // Set up preview and any additional channels.
 836         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 837                 DisplayFrame display_frame;
 838                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
 839                 display_frame.chain = chain.chain;
 840                 display_frame.setup_chain = chain.setup_chain;
 841                 display_frame.ready_fence = fence;
 842                 display_frame.input_frames = chain.input_frames;
 843                 display_frame.temp_textures = {};
 844                 output_channel[i].output_frame(display_frame);
 845         }
 846 }
 847
 848 void Mixer::audio_thread_func()
 849 {
 850         while (!should_quit) {
 851                 AudioTask task;
 852
 853                 {
 854                         unique_lock<mutex> lock(audio_mutex);
 855                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 856                         if (should_quit) {
 857                                 return;
 858                         }
 859                         task = audio_task_queue.front();
 860                         audio_task_queue.pop();
 861                 }
 862
 863                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
 864                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
 865                 vector<float> samples_out = audio_mixer.get_output(
 866                         double(task.pts_int) / TIMEBASE,
 867                         task.num_samples,
 868                         rate_adjustment_policy);
 869
 870                 // Send the samples to the sound card, then add them to the output.
 871                 if (alsa) {
 872                         alsa->write(samples_out);
 873                 }
 874                 decklink_output->send_audio(task.pts_int, samples_out);
 875                 video_encoder->add_audio(task.pts_int, move(samples_out));
 876         }
 877 }
 878
 879 void Mixer::release_display_frame(DisplayFrame *frame)
 880 {
 881         for (GLuint texnum : frame->temp_textures) {
 882                 resource_pool->release_2d_texture(texnum);
 883         }
 884         frame->temp_textures.clear();
 885         frame->ready_fence.reset();
 886         frame->input_frames.clear();
 887 }
 888
 889 void Mixer::start()
 890 {
 891         mixer_thread = thread(&Mixer::thread_func, this);
 892         audio_thread = thread(&Mixer::audio_thread_func, this);
 893 }
 894
 895 void Mixer::quit()
 896 {
 897         should_quit = true;
 898         audio_task_queue_changed.notify_one();
 899         mixer_thread.join();
 900         audio_thread.join();
 901 }
 902
 903 void Mixer::transition_clicked(int transition_num)
 904 {
 905         theme->transition_clicked(transition_num, pts());
 906 }
 907
 908 void Mixer::channel_clicked(int preview_num)
 909 {
 910         theme->channel_clicked(preview_num);
 911 }
 912
 913 void Mixer::start_mode_scanning(unsigned card_index)
 914 {
 915         assert(card_index < num_cards);
 916         if (is_mode_scanning[card_index]) {
 917                 return;
 918         }
 919         is_mode_scanning[card_index] = true;
 920         mode_scanlist[card_index].clear();
 921         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
 922                 mode_scanlist[card_index].push_back(mode.first);
 923         }
 924         assert(!mode_scanlist[card_index].empty());
 925         mode_scanlist_index[card_index] = 0;
 926         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
 927         last_mode_scan_change[card_index] = steady_clock::now();
 928 }
 929
 930 Mixer::OutputChannel::~OutputChannel()
 931 {
 932         if (has_current_frame) {
 933                 parent->release_display_frame(&current_frame);
 934         }
 935         if (has_ready_frame) {
 936                 parent->release_display_frame(&ready_frame);
 937         }
 938 }
 939
 940 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
 941 {
 942         // Store this frame for display. Remove the ready frame if any
 943         // (it was seemingly never used).
 944         {
 945                 unique_lock<mutex> lock(frame_mutex);
 946                 if (has_ready_frame) {
 947                         parent->release_display_frame(&ready_frame);
 948                 }
 949                 ready_frame = frame;
 950                 has_ready_frame = true;
 951         }
 952
 953         if (new_frame_ready_callback) {
 954                 new_frame_ready_callback();
 955         }
 956
 957         // Reduce the number of callbacks by filtering duplicates. The reason
 958         // why we bother doing this is that Qt seemingly can get into a state
 959         // where its builds up an essentially unbounded queue of signals,
 960         // consuming more and more memory, and there's no good way of collapsing
 961         // user-defined signals or limiting the length of the queue.
 962         if (transition_names_updated_callback) {
 963                 vector<string> transition_names = global_mixer->get_transition_names();
 964                 bool changed = false;
 965                 if (transition_names.size() != last_transition_names.size()) {
 966                         changed = true;
 967                 } else {
 968                         for (unsigned i = 0; i < transition_names.size(); ++i) {
 969                                 if (transition_names[i] != last_transition_names[i]) {
 970                                         changed = true;
 971                                         break;
 972                                 }
 973                         }
 974                 }
 975                 if (changed) {
 976                         transition_names_updated_callback(transition_names);
 977                         last_transition_names = transition_names;
 978                 }
 979         }
 980         if (name_updated_callback) {
 981                 string name = global_mixer->get_channel_name(channel);
 982                 if (name != last_name) {
 983                         name_updated_callback(name);
 984                         last_name = name;
 985                 }
 986         }
 987         if (color_updated_callback) {
 988                 string color = global_mixer->get_channel_color(channel);
 989                 if (color != last_color) {
 990                         color_updated_callback(color);
 991                         last_color = color;
 992                 }
 993         }
 994 }
 995
 996 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
 997 {
 998         unique_lock<mutex> lock(frame_mutex);
 999         if (!has_current_frame && !has_ready_frame) {
1000                 return false;
1001         }
1002
1003         if (has_current_frame && has_ready_frame) {
1004                 // We have a new ready frame. Toss the current one.
1005                 parent->release_display_frame(&current_frame);
1006                 has_current_frame = false;
1007         }
1008         if (has_ready_frame) {
1009                 assert(!has_current_frame);
1010                 current_frame = ready_frame;
1011                 ready_frame.ready_fence.reset();  // Drop the refcount.
1012                 ready_frame.input_frames.clear();  // Drop the refcounts.
1013                 has_current_frame = true;
1014                 has_ready_frame = false;
1015         }
1016
1017         *frame = current_frame;
1018         return true;
1019 }
1020
1021 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1022 {
1023         new_frame_ready_callback = callback;
1024 }
1025
1026 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1027 {
1028         transition_names_updated_callback = callback;
1029 }
1030
1031 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1032 {
1033         name_updated_callback = callback;
1034 }
1035
1036 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1037 {
1038         color_updated_callback = callback;
1039 }
1040
1041 mutex RefCountedGLsync::fence_lock;