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 <movit/util.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <sys/time.h>
  18 #include <time.h>
  19 #include <algorithm>
  20 #include <cmath>
  21 #include <condition_variable>
  22 #include <cstddef>
  23 #include <memory>
  24 #include <mutex>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29 #include <arpa/inet.h>
  30 #include <sys/time.h>
  31 #include <sys/resource.h>
  32
  33 #include "bmusb/bmusb.h"
  34 #include "bmusb/fake_capture.h"
  35 #include "context.h"
  36 #include "decklink_capture.h"
  37 #include "defs.h"
  38 #include "flags.h"
  39 #include "video_encoder.h"
  40 #include "pbo_frame_allocator.h"
  41 #include "ref_counted_gl_sync.h"
  42 #include "timebase.h"
  43
  44 class QOpenGLContext;
  45
  46 using namespace movit;
  47 using namespace std;
  48 using namespace std::placeholders;
  49 using namespace bmusb;
  50
  51 Mixer *global_mixer = nullptr;
  52 bool uses_mlock = false;
  53
  54 namespace {
  55
  56 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  57 {
  58         assert(in_channels >= out_channels);
  59         for (size_t i = 0; i < num_samples; ++i) {
  60                 for (size_t j = 0; j < out_channels; ++j) {
  61                         uint32_t s1 = *src++;
  62                         uint32_t s2 = *src++;
  63                         uint32_t s3 = *src++;
  64                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  65                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  66                 }
  67                 src += 3 * (in_channels - out_channels);
  68         }
  69 }
  70
  71 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  72 {
  73         assert(in_channels >= out_channels);
  74         for (size_t i = 0; i < num_samples; ++i) {
  75                 for (size_t j = 0; j < out_channels; ++j) {
  76                         // Note: Assumes little-endian.
  77                         int32_t s = *(int32_t *)src;
  78                         dst[i * out_channels + j] = s * (1.0f / 4294967296.0f);
  79                         src += 4;
  80                 }
  81                 src += 4 * (in_channels - out_channels);
  82         }
  83 }
  84
  85 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  86 {
  87         if (interlaced) {
  88                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  89                         input_state->buffered_frames[card_index][frame_num] =
  90                                 input_state->buffered_frames[card_index][frame_num - 1];
  91                 }
  92                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  93         } else {
  94                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  95                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  96                 }
  97         }
  98 }
  99
 100 }  // namespace
 101
 102 void QueueLengthPolicy::update_policy(int queue_length)
 103 {
 104         if (queue_length < 0) {  // Starvation.
 105                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
 106                         ++safe_queue_length;
 107                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
 108                                 card_index, safe_queue_length);
 109                 }
 110                 frames_with_at_least_one = 0;
 111                 been_at_safe_point_since_last_starvation = false;
 112                 return;
 113         }
 114         if (queue_length > 0) {
 115                 if (queue_length >= int(safe_queue_length)) {
 116                         been_at_safe_point_since_last_starvation = true;
 117                 }
 118                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
 119                         --safe_queue_length;
 120                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
 121                                 card_index, safe_queue_length);
 122                         frames_with_at_least_one = 0;
 123                 }
 124         } else {
 125                 frames_with_at_least_one = 0;
 126         }
 127 }
 128
 129 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 130         : httpd(),
 131           num_cards(num_cards),
 132           mixer_surface(create_surface(format)),
 133           h264_encoder_surface(create_surface(format)),
 134           correlation(OUTPUT_FREQUENCY),
 135           level_compressor(OUTPUT_FREQUENCY),
 136           limiter(OUTPUT_FREQUENCY),
 137           compressor(OUTPUT_FREQUENCY)
 138 {
 139         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 140         check_error();
 141
 142         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 143         // will be halved when sampling them, and we need to compensate here.
 144         movit_texel_subpixel_precision /= 2.0;
 145
 146         resource_pool.reset(new ResourcePool);
 147         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 148         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 149                 output_channel[i].parent = this;
 150                 output_channel[i].channel = i;
 151         }
 152
 153         ImageFormat inout_format;
 154         inout_format.color_space = COLORSPACE_sRGB;
 155         inout_format.gamma_curve = GAMMA_sRGB;
 156
 157         // Display chain; shows the live output produced by the main chain (its RGBA version).
 158         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 159         check_error();
 160         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 161         display_chain->add_input(display_input);
 162         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 163         display_chain->set_dither_bits(0);  // Don't bother.
 164         display_chain->finalize();
 165
 166         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 167
 168         // Start listening for clients only once VideoEncoder has written its header, if any.
 169         httpd.start(9095);
 170
 171         // First try initializing the then PCI devices, then USB, then
 172         // fill up with fake cards until we have the desired number of cards.
 173         unsigned num_pci_devices = 0;
 174         unsigned card_index = 0;
 175
 176         {
 177                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 178                 if (decklink_iterator != nullptr) {
 179                         for ( ; card_index < num_cards; ++card_index) {
 180                                 IDeckLink *decklink;
 181                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 182                                         break;
 183                                 }
 184
 185                                 configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
 186                                 ++num_pci_devices;
 187                         }
 188                         decklink_iterator->Release();
 189                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 190                 } else {
 191                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 192                 }
 193         }
 194         unsigned num_usb_devices = BMUSBCapture::num_cards();
 195         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 196                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 197                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 198                 configure_card(card_index, capture, /*is_fake_capture=*/false);
 199         }
 200         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 201
 202         unsigned num_fake_cards = 0;
 203         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 204                 FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 205                 configure_card(card_index, capture, /*is_fake_capture=*/true);
 206         }
 207
 208         if (num_fake_cards > 0) {
 209                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 210         }
 211
 212         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 213         BMUSBCapture::start_bm_thread();
 214
 215         for (card_index = 0; card_index < num_cards; ++card_index) {
 216                 cards[card_index].queue_length_policy.reset(card_index);
 217                 cards[card_index].capture->start_bm_capture();
 218         }
 219
 220         // Set up stuff for NV12 conversion.
 221
 222         // Cb/Cr shader.
 223         string cbcr_vert_shader =
 224                 "#version 130 \n"
 225                 " \n"
 226                 "in vec2 position; \n"
 227                 "in vec2 texcoord; \n"
 228                 "out vec2 tc0; \n"
 229                 "uniform vec2 foo_chroma_offset_0; \n"
 230                 " \n"
 231                 "void main() \n"
 232                 "{ \n"
 233                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 234                 "    // \n"
 235                 "    //   2.000  0.000  0.000 -1.000 \n"
 236                 "    //   0.000  2.000  0.000 -1.000 \n"
 237                 "    //   0.000  0.000 -2.000 -1.000 \n"
 238                 "    //   0.000  0.000  0.000  1.000 \n"
 239                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 240                 "    vec2 flipped_tc = texcoord; \n"
 241                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 242                 "} \n";
 243         string cbcr_frag_shader =
 244                 "#version 130 \n"
 245                 "in vec2 tc0; \n"
 246                 "uniform sampler2D cbcr_tex; \n"
 247                 "out vec4 FragColor; \n"
 248                 "void main() { \n"
 249                 "    FragColor = texture(cbcr_tex, tc0); \n"
 250                 "} \n";
 251         vector<string> frag_shader_outputs;
 252         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 253
 254         float vertices[] = {
 255                 0.0f, 2.0f,
 256                 0.0f, 0.0f,
 257                 2.0f, 0.0f
 258         };
 259         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 260         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 261         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 262
 263         r128.init(2, OUTPUT_FREQUENCY);
 264         r128.integr_start();
 265
 266         locut.init(FILTER_HPF, 2);
 267
 268         set_locut_enabled(global_flags.locut_enabled);
 269         set_gain_staging_db(global_flags.initial_gain_staging_db);
 270         set_gain_staging_auto(global_flags.gain_staging_auto);
 271         set_compressor_enabled(global_flags.compressor_enabled);
 272         set_limiter_enabled(global_flags.limiter_enabled);
 273         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 274
 275         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 276         // and there's a limit to how important the peak meter is.
 277         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 278
 279         if (global_flags.enable_alsa_output) {
 280                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 281         }
 282 }
 283
 284 Mixer::~Mixer()
 285 {
 286         resource_pool->release_glsl_program(cbcr_program_num);
 287         glDeleteBuffers(1, &cbcr_vbo);
 288         BMUSBCapture::stop_bm_thread();
 289
 290         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 291                 {
 292                         unique_lock<mutex> lock(bmusb_mutex);
 293                         cards[card_index].should_quit = true;  // Unblock thread.
 294                         cards[card_index].new_frames_changed.notify_all();
 295                 }
 296                 cards[card_index].capture->stop_dequeue_thread();
 297         }
 298
 299         video_encoder.reset(nullptr);
 300 }
 301
 302 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
 303 {
 304         printf("Configuring card %d...\n", card_index);
 305
 306         CaptureCard *card = &cards[card_index];
 307         if (card->capture != nullptr) {
 308                 card->capture->stop_dequeue_thread();
 309                 delete card->capture;
 310         }
 311         card->capture = capture;
 312         card->is_fake_capture = is_fake_capture;
 313         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 314         if (card->frame_allocator == nullptr) {
 315                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 316         }
 317         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 318         if (card->surface == nullptr) {
 319                 card->surface = create_surface_with_same_format(mixer_surface);
 320         }
 321         {
 322                 unique_lock<mutex> lock(cards[card_index].audio_mutex);
 323                 card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 324         }
 325         while (!card->new_frames.empty()) card->new_frames.pop();
 326         card->fractional_samples = 0;
 327         card->last_timecode = -1;
 328         card->next_local_pts = 0;
 329         card->capture->configure_card();
 330 }
 331
 332
 333 namespace {
 334
 335 int unwrap_timecode(uint16_t current_wrapped, int last)
 336 {
 337         uint16_t last_wrapped = last & 0xffff;
 338         if (current_wrapped > last_wrapped) {
 339                 return (last & ~0xffff) | current_wrapped;
 340         } else {
 341                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 342         }
 343 }
 344
 345 float find_peak(const float *samples, size_t num_samples)
 346 {
 347         float m = fabs(samples[0]);
 348         for (size_t i = 1; i < num_samples; ++i) {
 349                 m = max(m, fabs(samples[i]));
 350         }
 351         return m;
 352 }
 353
 354 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 355 {
 356         size_t num_samples = in.size() / 2;
 357         out_l->resize(num_samples);
 358         out_r->resize(num_samples);
 359
 360         const float *inptr = in.data();
 361         float *lptr = &(*out_l)[0];
 362         float *rptr = &(*out_r)[0];
 363         for (size_t i = 0; i < num_samples; ++i) {
 364                 *lptr++ = *inptr++;
 365                 *rptr++ = *inptr++;
 366         }
 367 }
 368
 369 }  // namespace
 370
 371 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 372                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 373                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 374 {
 375         CaptureCard *card = &cards[card_index];
 376
 377         if (is_mode_scanning[card_index]) {
 378                 if (video_format.has_signal) {
 379                         // Found a stable signal, so stop scanning.
 380                         is_mode_scanning[card_index] = false;
 381                 } else {
 382                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 383                         timespec now;
 384                         clock_gettime(CLOCK_MONOTONIC, &now);
 385                         double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
 386                                 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
 387                         if (sec_since_last_switch > switch_time_s) {
 388                                 // It isn't this mode; try the next one.
 389                                 mode_scanlist_index[card_index]++;
 390                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 391                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 392                                 last_mode_scan_change[card_index] = now;
 393                         }
 394                 }
 395         }
 396
 397         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 398         assert(frame_length > 0);
 399
 400         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;
 401         if (num_samples > OUTPUT_FREQUENCY / 10) {
 402                 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",
 403                         card_index, int(audio_frame.len), int(audio_offset),
 404                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 405                 if (video_frame.owner) {
 406                         video_frame.owner->release_frame(video_frame);
 407                 }
 408                 if (audio_frame.owner) {
 409                         audio_frame.owner->release_frame(audio_frame);
 410                 }
 411                 return;
 412         }
 413
 414         int64_t local_pts = card->next_local_pts;
 415         int dropped_frames = 0;
 416         if (card->last_timecode != -1) {
 417                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 418         }
 419
 420         // Convert the audio to stereo fp32 and add it.
 421         vector<float> audio;
 422         audio.resize(num_samples * 2);
 423         switch (audio_format.bits_per_sample) {
 424         case 0:
 425                 assert(num_samples == 0);
 426                 break;
 427         case 24:
 428                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 429                 break;
 430         case 32:
 431                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 432                 break;
 433         default:
 434                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 435                 assert(false);
 436         }
 437
 438         // Add the audio.
 439         {
 440                 unique_lock<mutex> lock(card->audio_mutex);
 441
 442                 // Number of samples per frame if we need to insert silence.
 443                 // (Could be nonintegral, but resampling will save us then.)
 444                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 445
 446                 if (dropped_frames > MAX_FPS * 2) {
 447                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 448                                 card_index, card->last_timecode, timecode);
 449                         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 450                         dropped_frames = 0;
 451                 } else if (dropped_frames > 0) {
 452                         // Insert silence as needed.
 453                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 454                                 card_index, dropped_frames, timecode);
 455                         vector<float> silence(silence_samples * 2, 0.0f);
 456                         for (int i = 0; i < dropped_frames; ++i) {
 457                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 458                                 // Note that if the format changed in the meantime, we have
 459                                 // no way of detecting that; we just have to assume the frame length
 460                                 // is always the same.
 461                                 local_pts += frame_length;
 462                         }
 463                 }
 464                 if (num_samples == 0) {
 465                         audio.resize(silence_samples * 2);
 466                         num_samples = silence_samples;
 467                 }
 468                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 469                 card->next_local_pts = local_pts + frame_length;
 470         }
 471
 472         card->last_timecode = timecode;
 473
 474         // Done with the audio, so release it.
 475         if (audio_frame.owner) {
 476                 audio_frame.owner->release_frame(audio_frame);
 477         }
 478
 479         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 480         if (video_frame.len - video_offset == 0 ||
 481             video_frame.len - video_offset != expected_length) {
 482                 if (video_frame.len != 0) {
 483                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 484                                 card_index, video_frame.len - video_offset, expected_length);
 485                 }
 486                 if (video_frame.owner) {
 487                         video_frame.owner->release_frame(video_frame);
 488                 }
 489
 490                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 491                 // so that pts can go up accordingly.
 492                 {
 493                         unique_lock<mutex> lock(bmusb_mutex);
 494                         CaptureCard::NewFrame new_frame;
 495                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 496                         new_frame.length = frame_length;
 497                         new_frame.interlaced = false;
 498                         new_frame.dropped_frames = dropped_frames;
 499                         card->new_frames.push(move(new_frame));
 500                         card->new_frames_changed.notify_all();
 501                 }
 502                 return;
 503         }
 504
 505         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 506
 507         unsigned num_fields = video_format.interlaced ? 2 : 1;
 508         timespec frame_upload_start;
 509         if (video_format.interlaced) {
 510                 // Send the two fields along as separate frames; the other side will need to add
 511                 // a deinterlacer to actually get this right.
 512                 assert(video_format.height % 2 == 0);
 513                 video_format.height /= 2;
 514                 assert(frame_length % 2 == 0);
 515                 frame_length /= 2;
 516                 num_fields = 2;
 517                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 518         }
 519         userdata->last_interlaced = video_format.interlaced;
 520         userdata->last_has_signal = video_format.has_signal;
 521         userdata->last_is_connected = video_format.is_connected;
 522         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 523         userdata->last_frame_rate_den = video_format.frame_rate_den;
 524         RefCountedFrame frame(video_frame);
 525
 526         // Upload the textures.
 527         size_t cbcr_width = video_format.width / 2;
 528         size_t cbcr_offset = video_offset / 2;
 529         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 530
 531         for (unsigned field = 0; field < num_fields; ++field) {
 532                 // Put the actual texture upload in a lambda that is executed in the main thread.
 533                 // It is entirely possible to do this in the same thread (and it might even be
 534                 // faster, depending on the GPU and driver), but it appears to be trickling
 535                 // driver bugs very easily.
 536                 //
 537                 // Note that this means we must hold on to the actual frame data in <userdata>
 538                 // until the upload command is run, but we hold on to <frame> much longer than that
 539                 // (in fact, all the way until we no longer use the texture in rendering).
 540                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
 541                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 542
 543                         if (userdata->tex_y[field] == 0 ||
 544                             userdata->tex_cbcr[field] == 0 ||
 545                             video_format.width != userdata->last_width[field] ||
 546                             video_format.height != userdata->last_height[field]) {
 547                                 // We changed resolution since last use of this texture, so we need to create
 548                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 549                                 // we don't need to worry about these flip-flopping between resolutions.
 550                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 551                                 check_error();
 552                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 553                                 check_error();
 554                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 555                                 check_error();
 556                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 557                                 check_error();
 558                                 userdata->last_width[field] = video_format.width;
 559                                 userdata->last_height[field] = video_format.height;
 560                         }
 561
 562                         GLuint pbo = userdata->pbo;
 563                         check_error();
 564                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 565                         check_error();
 566
 567                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 568                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 569
 570                         if (global_flags.flush_pbos) {
 571                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 572                                 check_error();
 573                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 574                                 check_error();
 575                         }
 576
 577                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 578                         check_error();
 579                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 580                         check_error();
 581                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 582                         check_error();
 583                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 584                         check_error();
 585                         glBindTexture(GL_TEXTURE_2D, 0);
 586                         check_error();
 587                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 588                         check_error();
 589                 };
 590
 591                 if (field == 1) {
 592                         // Don't upload the second field as fast as we can; wait until
 593                         // the field time has approximately passed. (Otherwise, we could
 594                         // get timing jitter against the other sources, and possibly also
 595                         // against the video display, although the latter is not as critical.)
 596                         // This requires our system clock to be reasonably close to the
 597                         // video clock, but that's not an unreasonable assumption.
 598                         timespec second_field_start;
 599                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 600                                 frame_length * 1000000000 / TIMEBASE;
 601                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 602                                 second_field_start.tv_nsec / 1000000000;
 603                         second_field_start.tv_nsec %= 1000000000;
 604
 605                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 606                                                &second_field_start, nullptr) == -1 &&
 607                                errno == EINTR) ;
 608                 }
 609
 610                 {
 611                         unique_lock<mutex> lock(bmusb_mutex);
 612                         CaptureCard::NewFrame new_frame;
 613                         new_frame.frame = frame;
 614                         new_frame.length = frame_length;
 615                         new_frame.field = field;
 616                         new_frame.interlaced = video_format.interlaced;
 617                         new_frame.upload_func = upload_func;
 618                         new_frame.dropped_frames = dropped_frames;
 619                         card->new_frames.push(move(new_frame));
 620                         card->new_frames_changed.notify_all();
 621                 }
 622         }
 623 }
 624
 625 void Mixer::bm_hotplug_add(libusb_device *dev)
 626 {
 627         lock_guard<mutex> lock(hotplug_mutex);
 628         hotplugged_cards.push_back(dev);
 629 }
 630
 631 void Mixer::bm_hotplug_remove(unsigned card_index)
 632 {
 633         cards[card_index].new_frames_changed.notify_all();
 634 }
 635
 636 void Mixer::thread_func()
 637 {
 638         eglBindAPI(EGL_OPENGL_API);
 639         QOpenGLContext *context = create_context(mixer_surface);
 640         if (!make_current(context, mixer_surface)) {
 641                 printf("oops\n");
 642                 exit(1);
 643         }
 644
 645         struct timespec start, now;
 646         clock_gettime(CLOCK_MONOTONIC, &start);
 647
 648         int frame = 0;
 649         int stats_dropped_frames = 0;
 650
 651         while (!should_quit) {
 652                 CaptureCard::NewFrame new_frames[MAX_CARDS];
 653                 bool has_new_frame[MAX_CARDS] = { false };
 654                 int num_samples[MAX_CARDS] = { 0 };
 655
 656                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 657                 assert(master_card_index < num_cards);
 658
 659                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 660                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 661                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 662                 send_audio_level_callback();
 663
 664                 handle_hotplugged_cards();
 665
 666                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 667                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 668                                 continue;
 669                         }
 670                         if (new_frames[card_index].frame->len == 0) {
 671                                 ++new_frames[card_index].dropped_frames;
 672                         }
 673                         if (new_frames[card_index].dropped_frames > 0) {
 674                                 printf("Card %u dropped %d frames before this\n",
 675                                         card_index, int(new_frames[card_index].dropped_frames));
 676                         }
 677                 }
 678
 679                 // If the first card is reporting a corrupted or otherwise dropped frame,
 680                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 681                 if (new_frames[master_card_index].frame->len == 0) {
 682                         ++stats_dropped_frames;
 683                         pts_int += new_frames[master_card_index].length;
 684                         continue;
 685                 }
 686
 687                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 688                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 689                                 continue;
 690
 691                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 692                         assert(new_frame->frame != nullptr);
 693                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 694                         check_error();
 695
 696                         // The new texture might need uploading before use.
 697                         if (new_frame->upload_func) {
 698                                 new_frame->upload_func();
 699                                 new_frame->upload_func = nullptr;
 700                         }
 701                 }
 702
 703                 int64_t duration = new_frames[master_card_index].length;
 704                 render_one_frame(duration);
 705                 ++frame;
 706                 pts_int += duration;
 707
 708                 clock_gettime(CLOCK_MONOTONIC, &now);
 709                 double elapsed = now.tv_sec - start.tv_sec +
 710                         1e-9 * (now.tv_nsec - start.tv_nsec);
 711                 if (frame % 100 == 0) {
 712                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 713                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 714                                 1e3 * elapsed / frame);
 715                 //      chain->print_phase_timing();
 716
 717                         // Check our memory usage, to see if we are close to our mlockall()
 718                         // limit (if at all set).
 719                         rusage used;
 720                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 721                                 perror("getrusage(RUSAGE_SELF)");
 722                                 assert(false);
 723                         }
 724
 725                         if (uses_mlock) {
 726                                 rlimit limit;
 727                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 728                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 729                                         assert(false);
 730                                 }
 731
 732                                 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 733                                         long(used.ru_maxrss / 1024),
 734                                         long(limit.rlim_cur / 1048576),
 735                                         float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 736                         } else {
 737                                 printf(", using %ld MB memory (not locked)",
 738                                         long(used.ru_maxrss / 1024));
 739                         }
 740
 741                         printf("\n");
 742                 }
 743
 744
 745                 if (should_cut.exchange(false)) {  // Test and clear.
 746                         video_encoder->do_cut(frame);
 747                 }
 748
 749 #if 0
 750                 // Reset every 100 frames, so that local variations in frame times
 751                 // (especially for the first few frames, when the shaders are
 752                 // compiled etc.) don't make it hard to measure for the entire
 753                 // remaining duration of the program.
 754                 if (frame == 10000) {
 755                         frame = 0;
 756                         start = now;
 757                 }
 758 #endif
 759                 check_error();
 760         }
 761
 762         resource_pool->clean_context();
 763 }
 764
 765 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])
 766 {
 767 start:
 768         // The first card is the master timer, so wait for it to have a new frame.
 769         // TODO: Add a timeout.
 770         unique_lock<mutex> lock(bmusb_mutex);
 771         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(); });
 772
 773         if (cards[master_card_index].new_frames.empty()) {
 774                 // We were woken up, but not due to a new frame. Deal with it
 775                 // and then restart.
 776                 assert(cards[master_card_index].capture->get_disconnected());
 777                 handle_hotplugged_cards();
 778                 goto start;
 779         }
 780
 781         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 782                 CaptureCard *card = &cards[card_index];
 783                 if (card->new_frames.empty()) {
 784                         assert(card_index != master_card_index);
 785                         card->queue_length_policy.update_policy(-1);
 786                         continue;
 787                 }
 788                 new_frames[card_index] = move(card->new_frames.front());
 789                 has_new_frame[card_index] = true;
 790                 card->new_frames.pop();
 791                 card->new_frames_changed.notify_all();
 792
 793                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 794                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 795                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 796                 assert(num_samples[card_index] >= 0);
 797
 798                 if (card_index == master_card_index) {
 799                         // We don't use the queue length policy for the master card,
 800                         // but we will if it stops being the master. Thus, clear out
 801                         // the policy in case we switch in the future.
 802                         card->queue_length_policy.reset(card_index);
 803                 } else {
 804                         // If we have excess frames compared to the policy for this card,
 805                         // drop frames from the head.
 806                         card->queue_length_policy.update_policy(card->new_frames.size());
 807                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 808                                 card->new_frames.pop();
 809                         }
 810                 }
 811         }
 812 }
 813
 814 void Mixer::handle_hotplugged_cards()
 815 {
 816         // Check for cards that have been disconnected since last frame.
 817         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 818                 CaptureCard *card = &cards[card_index];
 819                 if (card->capture->get_disconnected()) {
 820                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 821                         FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 822                         configure_card(card_index, capture, /*is_fake_capture=*/true);
 823                         card->queue_length_policy.reset(card_index);
 824                         card->capture->start_bm_capture();
 825                 }
 826         }
 827
 828         // Check for cards that have been connected since last frame.
 829         vector<libusb_device *> hotplugged_cards_copy;
 830         {
 831                 lock_guard<mutex> lock(hotplug_mutex);
 832                 swap(hotplugged_cards, hotplugged_cards_copy);
 833         }
 834         for (libusb_device *new_dev : hotplugged_cards_copy) {
 835                 // Look for a fake capture card where we can stick this in.
 836                 int free_card_index = -1;
 837                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 838                         if (cards[card_index].is_fake_capture) {
 839                                 free_card_index = int(card_index);
 840                                 break;
 841                         }
 842                 }
 843
 844                 if (free_card_index == -1) {
 845                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 846                         libusb_unref_device(new_dev);
 847                 } else {
 848                         // BMUSBCapture takes ownership.
 849                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 850                         CaptureCard *card = &cards[free_card_index];
 851                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 852                         configure_card(free_card_index, capture, /*is_fake_capture=*/false);
 853                         card->queue_length_policy.reset(free_card_index);
 854                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 855                         capture->start_bm_capture();
 856                 }
 857         }
 858 }
 859
 860
 861 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 862 {
 863         // Resample the audio as needed, including from previously dropped frames.
 864         assert(num_cards > 0);
 865         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 866                 {
 867                         // Signal to the audio thread to process this frame.
 868                         unique_lock<mutex> lock(audio_mutex);
 869                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
 870                         audio_task_queue_changed.notify_one();
 871                 }
 872                 if (frame_num != dropped_frames) {
 873                         // For dropped frames, increase the pts. Note that if the format changed
 874                         // in the meantime, we have no way of detecting that; we just have to
 875                         // assume the frame length is always the same.
 876                         pts_int += length_per_frame;
 877                 }
 878         }
 879 }
 880
 881 void Mixer::render_one_frame(int64_t duration)
 882 {
 883         // Get the main chain from the theme, and set its state immediately.
 884         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 885         EffectChain *chain = theme_main_chain.chain;
 886         theme_main_chain.setup_chain();
 887         //theme_main_chain.chain->enable_phase_timing(true);
 888
 889         GLuint y_tex, cbcr_tex;
 890         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 891         assert(got_frame);
 892
 893         // Render main chain.
 894         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 895         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 896         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 897         check_error();
 898         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 899         resource_pool->release_fbo(fbo);
 900
 901         subsample_chroma(cbcr_full_tex, cbcr_tex);
 902         resource_pool->release_2d_texture(cbcr_full_tex);
 903
 904         // Set the right state for rgba_tex.
 905         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 906         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 907         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 908         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 909         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 910
 911         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 912         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 913
 914         // The live frame just shows the RGBA texture we just rendered.
 915         // It owns rgba_tex now.
 916         DisplayFrame live_frame;
 917         live_frame.chain = display_chain.get();
 918         live_frame.setup_chain = [this, rgba_tex]{
 919                 display_input->set_texture_num(rgba_tex);
 920         };
 921         live_frame.ready_fence = fence;
 922         live_frame.input_frames = {};
 923         live_frame.temp_textures = { rgba_tex };
 924         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 925
 926         // Set up preview and any additional channels.
 927         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 928                 DisplayFrame display_frame;
 929                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 930                 display_frame.chain = chain.chain;
 931                 display_frame.setup_chain = chain.setup_chain;
 932                 display_frame.ready_fence = fence;
 933                 display_frame.input_frames = chain.input_frames;
 934                 display_frame.temp_textures = {};
 935                 output_channel[i].output_frame(display_frame);
 936         }
 937 }
 938
 939 void Mixer::send_audio_level_callback()
 940 {
 941         if (audio_level_callback == nullptr) {
 942                 return;
 943         }
 944
 945         unique_lock<mutex> lock(compressor_mutex);
 946         double loudness_s = r128.loudness_S();
 947         double loudness_i = r128.integrated();
 948         double loudness_range_low = r128.range_min();
 949         double loudness_range_high = r128.range_max();
 950
 951         audio_level_callback(loudness_s, 20.0 * log10(peak),
 952                 loudness_i, loudness_range_low, loudness_range_high,
 953                 gain_staging_db, 20.0 * log10(final_makeup_gain),
 954                 correlation.get_correlation());
 955 }
 956
 957 void Mixer::audio_thread_func()
 958 {
 959         while (!should_quit) {
 960                 AudioTask task;
 961
 962                 {
 963                         unique_lock<mutex> lock(audio_mutex);
 964                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 965                         if (should_quit) {
 966                                 return;
 967                         }
 968                         task = audio_task_queue.front();
 969                         audio_task_queue.pop();
 970                 }
 971
 972                 process_audio_one_frame(task.pts_int, task.num_samples);
 973         }
 974 }
 975
 976 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 977 {
 978         vector<float> samples_card;
 979         vector<float> samples_out;
 980
 981         // TODO: Allow mixing audio from several sources.
 982         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 983         assert(selected_audio_card < num_cards);
 984
 985         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 986                 samples_card.resize(num_samples * 2);
 987                 {
 988                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 989                         cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples);
 990                 }
 991                 if (card_index == selected_audio_card) {
 992                         samples_out = move(samples_card);
 993                 }
 994         }
 995
 996         // Cut away everything under 120 Hz (or whatever the cutoff is);
 997         // we don't need it for voice, and it will reduce headroom
 998         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 999         // should be dampened.)
1000         if (locut_enabled) {
1001                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
1002         }
1003
1004         // Apply a level compressor to get the general level right.
1005         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
1006         // (or more precisely, near it, since we don't use infinite ratio),
1007         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
1008         // entirely arbitrary, but from practical tests with speech, it seems to
1009         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
1010         {
1011                 unique_lock<mutex> lock(compressor_mutex);
1012                 if (level_compressor_enabled) {
1013                         float threshold = 0.01f;   // -40 dBFS.
1014                         float ratio = 20.0f;
1015                         float attack_time = 0.5f;
1016                         float release_time = 20.0f;
1017                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
1018                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1019                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
1020                 } else {
1021                         // Just apply the gain we already had.
1022                         float g = pow(10.0f, gain_staging_db / 20.0f);
1023                         for (size_t i = 0; i < samples_out.size(); ++i) {
1024                                 samples_out[i] *= g;
1025                         }
1026                 }
1027         }
1028
1029 #if 0
1030         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
1031                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
1032                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
1033                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
1034 #endif
1035
1036 //      float limiter_att, compressor_att;
1037
1038         // The real compressor.
1039         if (compressor_enabled) {
1040                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
1041                 float ratio = 20.0f;
1042                 float attack_time = 0.005f;
1043                 float release_time = 0.040f;
1044                 float makeup_gain = 2.0f;  // +6 dB.
1045                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1046 //              compressor_att = compressor.get_attenuation();
1047         }
1048
1049         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
1050         // Note that since ratio is not infinite, we could go slightly higher than this.
1051         if (limiter_enabled) {
1052                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
1053                 float ratio = 30.0f;
1054                 float attack_time = 0.0f;  // Instant.
1055                 float release_time = 0.020f;
1056                 float makeup_gain = 1.0f;  // 0 dB.
1057                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1058 //              limiter_att = limiter.get_attenuation();
1059         }
1060
1061 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
1062
1063         // Upsample 4x to find interpolated peak.
1064         peak_resampler.inp_data = samples_out.data();
1065         peak_resampler.inp_count = samples_out.size() / 2;
1066
1067         vector<float> interpolated_samples_out;
1068         interpolated_samples_out.resize(samples_out.size());
1069         while (peak_resampler.inp_count > 0) {  // About four iterations.
1070                 peak_resampler.out_data = &interpolated_samples_out[0];
1071                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
1072                 peak_resampler.process();
1073                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
1074                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
1075                 peak_resampler.out_data = nullptr;
1076         }
1077
1078         // At this point, we are most likely close to +0 LU, but all of our
1079         // measurements have been on raw sample values, not R128 values.
1080         // So we have a final makeup gain to get us to +0 LU; the gain
1081         // adjustments required should be relatively small, and also, the
1082         // offset shouldn't change much (only if the type of audio changes
1083         // significantly). Thus, we shoot for updating this value basically
1084         // “whenever we process buffers”, since the R128 calculation isn't exactly
1085         // something we get out per-sample.
1086         //
1087         // Note that there's a feedback loop here, so we choose a very slow filter
1088         // (half-time of 100 seconds).
1089         double target_loudness_factor, alpha;
1090         {
1091                 unique_lock<mutex> lock(compressor_mutex);
1092                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
1093                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
1094                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
1095
1096                 // If we're outside +/- 5 LU uncorrected, we don't count it as
1097                 // a normal signal (probably silence) and don't change the
1098                 // correction factor; just apply what we already have.
1099                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
1100                         alpha = 0.0;
1101                 } else {
1102                         // Formula adapted from
1103                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
1104                         const double half_time_s = 100.0;
1105                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
1106                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
1107                 }
1108
1109                 double m = final_makeup_gain;
1110                 for (size_t i = 0; i < samples_out.size(); i += 2) {
1111                         samples_out[i + 0] *= m;
1112                         samples_out[i + 1] *= m;
1113                         m += (target_loudness_factor - m) * alpha;
1114                 }
1115                 final_makeup_gain = m;
1116         }
1117
1118         // Find R128 levels and L/R correlation.
1119         vector<float> left, right;
1120         deinterleave_samples(samples_out, &left, &right);
1121         float *ptrs[] = { left.data(), right.data() };
1122         {
1123                 unique_lock<mutex> lock(compressor_mutex);
1124                 r128.process(left.size(), ptrs);
1125                 correlation.process_samples(samples_out);
1126         }
1127
1128         // Send the samples to the sound card.
1129         if (alsa) {
1130                 alsa->write(samples_out);
1131         }
1132
1133         // And finally add them to the output.
1134         video_encoder->add_audio(frame_pts_int, move(samples_out));
1135 }
1136
1137 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1138 {
1139         GLuint vao;
1140         glGenVertexArrays(1, &vao);
1141         check_error();
1142
1143         glBindVertexArray(vao);
1144         check_error();
1145
1146         // Extract Cb/Cr.
1147         GLuint fbo = resource_pool->create_fbo(dst_tex);
1148         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1149         glViewport(0, 0, WIDTH/2, HEIGHT/2);
1150         check_error();
1151
1152         glUseProgram(cbcr_program_num);
1153         check_error();
1154
1155         glActiveTexture(GL_TEXTURE0);
1156         check_error();
1157         glBindTexture(GL_TEXTURE_2D, src_tex);
1158         check_error();
1159         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1160         check_error();
1161         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1162         check_error();
1163         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1164         check_error();
1165
1166         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1167         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1168
1169         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1170         check_error();
1171
1172         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1173                 glEnableVertexAttribArray(attr_index);
1174                 check_error();
1175                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1176                 check_error();
1177         }
1178
1179         glDrawArrays(GL_TRIANGLES, 0, 3);
1180         check_error();
1181
1182         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1183                 glDisableVertexAttribArray(attr_index);
1184                 check_error();
1185         }
1186
1187         glUseProgram(0);
1188         check_error();
1189         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1190         check_error();
1191
1192         resource_pool->release_fbo(fbo);
1193         glDeleteVertexArrays(1, &vao);
1194 }
1195
1196 void Mixer::release_display_frame(DisplayFrame *frame)
1197 {
1198         for (GLuint texnum : frame->temp_textures) {
1199                 resource_pool->release_2d_texture(texnum);
1200         }
1201         frame->temp_textures.clear();
1202         frame->ready_fence.reset();
1203         frame->input_frames.clear();
1204 }
1205
1206 void Mixer::start()
1207 {
1208         mixer_thread = thread(&Mixer::thread_func, this);
1209         audio_thread = thread(&Mixer::audio_thread_func, this);
1210 }
1211
1212 void Mixer::quit()
1213 {
1214         should_quit = true;
1215         audio_task_queue_changed.notify_one();
1216         mixer_thread.join();
1217         audio_thread.join();
1218 }
1219
1220 void Mixer::transition_clicked(int transition_num)
1221 {
1222         theme->transition_clicked(transition_num, pts());
1223 }
1224
1225 void Mixer::channel_clicked(int preview_num)
1226 {
1227         theme->channel_clicked(preview_num);
1228 }
1229
1230 void Mixer::reset_meters()
1231 {
1232         peak_resampler.reset();
1233         peak = 0.0f;
1234         r128.reset();
1235         r128.integr_start();
1236         correlation.reset();
1237 }
1238
1239 void Mixer::start_mode_scanning(unsigned card_index)
1240 {
1241         assert(card_index < num_cards);
1242         if (is_mode_scanning[card_index]) {
1243                 return;
1244         }
1245         is_mode_scanning[card_index] = true;
1246         mode_scanlist[card_index].clear();
1247         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1248                 mode_scanlist[card_index].push_back(mode.first);
1249         }
1250         assert(!mode_scanlist[card_index].empty());
1251         mode_scanlist_index[card_index] = 0;
1252         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1253         clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1254 }
1255
1256 Mixer::OutputChannel::~OutputChannel()
1257 {
1258         if (has_current_frame) {
1259                 parent->release_display_frame(&current_frame);
1260         }
1261         if (has_ready_frame) {
1262                 parent->release_display_frame(&ready_frame);
1263         }
1264 }
1265
1266 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1267 {
1268         // Store this frame for display. Remove the ready frame if any
1269         // (it was seemingly never used).
1270         {
1271                 unique_lock<mutex> lock(frame_mutex);
1272                 if (has_ready_frame) {
1273                         parent->release_display_frame(&ready_frame);
1274                 }
1275                 ready_frame = frame;
1276                 has_ready_frame = true;
1277         }
1278
1279         if (new_frame_ready_callback) {
1280                 new_frame_ready_callback();
1281         }
1282
1283         // Reduce the number of callbacks by filtering duplicates. The reason
1284         // why we bother doing this is that Qt seemingly can get into a state
1285         // where its builds up an essentially unbounded queue of signals,
1286         // consuming more and more memory, and there's no good way of collapsing
1287         // user-defined signals or limiting the length of the queue.
1288         if (transition_names_updated_callback) {
1289                 vector<string> transition_names = global_mixer->get_transition_names();
1290                 bool changed = false;
1291                 if (transition_names.size() != last_transition_names.size()) {
1292                         changed = true;
1293                 } else {
1294                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1295                                 if (transition_names[i] != last_transition_names[i]) {
1296                                         changed = true;
1297                                         break;
1298                                 }
1299                         }
1300                 }
1301                 if (changed) {
1302                         transition_names_updated_callback(transition_names);
1303                         last_transition_names = transition_names;
1304                 }
1305         }
1306         if (name_updated_callback) {
1307                 string name = global_mixer->get_channel_name(channel);
1308                 if (name != last_name) {
1309                         name_updated_callback(name);
1310                         last_name = name;
1311                 }
1312         }
1313         if (color_updated_callback) {
1314                 string color = global_mixer->get_channel_color(channel);
1315                 if (color != last_color) {
1316                         color_updated_callback(color);
1317                         last_color = color;
1318                 }
1319         }
1320 }
1321
1322 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1323 {
1324         unique_lock<mutex> lock(frame_mutex);
1325         if (!has_current_frame && !has_ready_frame) {
1326                 return false;
1327         }
1328
1329         if (has_current_frame && has_ready_frame) {
1330                 // We have a new ready frame. Toss the current one.
1331                 parent->release_display_frame(&current_frame);
1332                 has_current_frame = false;
1333         }
1334         if (has_ready_frame) {
1335                 assert(!has_current_frame);
1336                 current_frame = ready_frame;
1337                 ready_frame.ready_fence.reset();  // Drop the refcount.
1338                 ready_frame.input_frames.clear();  // Drop the refcounts.
1339                 has_current_frame = true;
1340                 has_ready_frame = false;
1341         }
1342
1343         *frame = current_frame;
1344         return true;
1345 }
1346
1347 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1348 {
1349         new_frame_ready_callback = callback;
1350 }
1351
1352 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1353 {
1354         transition_names_updated_callback = callback;
1355 }
1356
1357 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1358 {
1359         name_updated_callback = callback;
1360 }
1361
1362 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1363 {
1364         color_updated_callback = callback;
1365 }
1366
1367 mutex RefCountedGLsync::fence_lock;