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 "context.h"
  35 #include "decklink_capture.h"
  36 #include "defs.h"
  37 #include "fake_capture.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
  50 Mixer *global_mixer = nullptr;
  51 bool uses_mlock = false;
  52
  53 namespace {
  54
  55 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  56 {
  57         assert(in_channels >= out_channels);
  58         for (size_t i = 0; i < num_samples; ++i) {
  59                 for (size_t j = 0; j < out_channels; ++j) {
  60                         uint32_t s1 = *src++;
  61                         uint32_t s2 = *src++;
  62                         uint32_t s3 = *src++;
  63                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  64                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  65                 }
  66                 src += 3 * (in_channels - out_channels);
  67         }
  68 }
  69
  70 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  71 {
  72         assert(in_channels >= out_channels);
  73         for (size_t i = 0; i < num_samples; ++i) {
  74                 for (size_t j = 0; j < out_channels; ++j) {
  75                         // Note: Assumes little-endian.
  76                         int32_t s = *(int32_t *)src;
  77                         dst[i * out_channels + j] = s * (1.0f / 4294967296.0f);
  78                         src += 4;
  79                 }
  80                 src += 4 * (in_channels - out_channels);
  81         }
  82 }
  83
  84 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  85 {
  86         if (interlaced) {
  87                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  88                         input_state->buffered_frames[card_index][frame_num] =
  89                                 input_state->buffered_frames[card_index][frame_num - 1];
  90                 }
  91                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  92         } else {
  93                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  94                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  95                 }
  96         }
  97 }
  98
  99 }  // namespace
 100
 101 void QueueLengthPolicy::update_policy(int queue_length)
 102 {
 103         if (queue_length < 0) {  // Starvation.
 104                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
 105                         ++safe_queue_length;
 106                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
 107                                 card_index, safe_queue_length);
 108                 }
 109                 frames_with_at_least_one = 0;
 110                 been_at_safe_point_since_last_starvation = false;
 111                 return;
 112         }
 113         if (queue_length > 0) {
 114                 if (queue_length >= int(safe_queue_length)) {
 115                         been_at_safe_point_since_last_starvation = true;
 116                 }
 117                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
 118                         --safe_queue_length;
 119                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
 120                                 card_index, safe_queue_length);
 121                         frames_with_at_least_one = 0;
 122                 }
 123         } else {
 124                 frames_with_at_least_one = 0;
 125         }
 126 }
 127
 128 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 129         : httpd(),
 130           num_cards(num_cards),
 131           mixer_surface(create_surface(format)),
 132           h264_encoder_surface(create_surface(format)),
 133           correlation(OUTPUT_FREQUENCY),
 134           level_compressor(OUTPUT_FREQUENCY),
 135           limiter(OUTPUT_FREQUENCY),
 136           compressor(OUTPUT_FREQUENCY)
 137 {
 138         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 139         check_error();
 140
 141         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 142         // will be halved when sampling them, and we need to compensate here.
 143         movit_texel_subpixel_precision /= 2.0;
 144
 145         resource_pool.reset(new ResourcePool);
 146         theme.reset(new Theme(global_flags.theme_filename.c_str(), resource_pool.get(), num_cards));
 147         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 148                 output_channel[i].parent = this;
 149                 output_channel[i].channel = i;
 150         }
 151
 152         ImageFormat inout_format;
 153         inout_format.color_space = COLORSPACE_sRGB;
 154         inout_format.gamma_curve = GAMMA_sRGB;
 155
 156         // Display chain; shows the live output produced by the main chain (its RGBA version).
 157         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 158         check_error();
 159         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 160         display_chain->add_input(display_input);
 161         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 162         display_chain->set_dither_bits(0);  // Don't bother.
 163         display_chain->finalize();
 164
 165         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 166
 167         // Start listening for clients only once VideoEncoder has written its header, if any.
 168         httpd.start(9095);
 169
 170         // First try initializing the fake devices, then PCI devices, then USB,
 171         // until we have the desired number of cards.
 172         unsigned num_pci_devices = 0, num_usb_devices = 0;
 173         unsigned card_index = 0;
 174
 175         assert(global_flags.num_fake_cards >= 0);  // Enforced in flags.cpp.
 176         unsigned num_fake_cards = global_flags.num_fake_cards;
 177
 178         assert(num_fake_cards <= num_cards);  // Enforced in flags.cpp.
 179         for ( ; card_index < num_fake_cards; ++card_index) {
 180                 configure_card(card_index, new FakeCapture(card_index), /*is_fake_capture=*/true);
 181         }
 182
 183         if (global_flags.num_fake_cards > 0) {
 184                 fprintf(stderr, "Initialized %d fake cards.\n", global_flags.num_fake_cards);
 185         }
 186
 187         if (card_index < num_cards) {
 188                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 189                 if (decklink_iterator != nullptr) {
 190                         for ( ; card_index < num_cards; ++card_index) {
 191                                 IDeckLink *decklink;
 192                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 193                                         break;
 194                                 }
 195
 196                                 configure_card(card_index, new DeckLinkCapture(decklink, card_index - num_fake_cards), /*is_fake_capture=*/false);
 197                                 ++num_pci_devices;
 198                         }
 199                         decklink_iterator->Release();
 200                         fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
 201                 } else {
 202                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 203                 }
 204         }
 205         for ( ; card_index < num_cards; ++card_index) {
 206                 BMUSBCapture *capture = new BMUSBCapture(card_index - num_pci_devices - num_fake_cards);
 207                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 208                 configure_card(card_index, capture, /*is_fake_capture=*/false);
 209                 ++num_usb_devices;
 210         }
 211
 212         if (num_usb_devices > 0) {
 213                 has_bmusb_thread = true;
 214                 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 215                 BMUSBCapture::start_bm_thread();
 216         }
 217
 218         for (card_index = 0; card_index < num_cards; ++card_index) {
 219                 cards[card_index].queue_length_policy.reset(card_index);
 220                 cards[card_index].capture->start_bm_capture();
 221         }
 222
 223         // Set up stuff for NV12 conversion.
 224
 225         // Cb/Cr shader.
 226         string cbcr_vert_shader =
 227                 "#version 130 \n"
 228                 " \n"
 229                 "in vec2 position; \n"
 230                 "in vec2 texcoord; \n"
 231                 "out vec2 tc0; \n"
 232                 "uniform vec2 foo_chroma_offset_0; \n"
 233                 " \n"
 234                 "void main() \n"
 235                 "{ \n"
 236                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 237                 "    // \n"
 238                 "    //   2.000  0.000  0.000 -1.000 \n"
 239                 "    //   0.000  2.000  0.000 -1.000 \n"
 240                 "    //   0.000  0.000 -2.000 -1.000 \n"
 241                 "    //   0.000  0.000  0.000  1.000 \n"
 242                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 243                 "    vec2 flipped_tc = texcoord; \n"
 244                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 245                 "} \n";
 246         string cbcr_frag_shader =
 247                 "#version 130 \n"
 248                 "in vec2 tc0; \n"
 249                 "uniform sampler2D cbcr_tex; \n"
 250                 "out vec4 FragColor; \n"
 251                 "void main() { \n"
 252                 "    FragColor = texture(cbcr_tex, tc0); \n"
 253                 "} \n";
 254         vector<string> frag_shader_outputs;
 255         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 256
 257         float vertices[] = {
 258                 0.0f, 2.0f,
 259                 0.0f, 0.0f,
 260                 2.0f, 0.0f
 261         };
 262         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 263         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 264         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 265
 266         r128.init(2, OUTPUT_FREQUENCY);
 267         r128.integr_start();
 268
 269         locut.init(FILTER_HPF, 2);
 270
 271         // If --flat-audio is given, turn off everything that messes with the sound,
 272         // except the final makeup gain.
 273         if (global_flags.flat_audio) {
 274                 set_locut_enabled(false);
 275                 set_gain_staging_auto(false);
 276                 set_limiter_enabled(false);
 277                 set_compressor_enabled(false);
 278         }
 279
 280         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 281         // and there's a limit to how important the peak meter is.
 282         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 283
 284         if (global_flags.enable_alsa_output) {
 285                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 286         }
 287 }
 288
 289 Mixer::~Mixer()
 290 {
 291         resource_pool->release_glsl_program(cbcr_program_num);
 292         glDeleteBuffers(1, &cbcr_vbo);
 293         if (has_bmusb_thread) {
 294                 BMUSBCapture::stop_bm_thread();
 295         }
 296
 297         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 298                 {
 299                         unique_lock<mutex> lock(bmusb_mutex);
 300                         cards[card_index].should_quit = true;  // Unblock thread.
 301                         cards[card_index].new_frames_changed.notify_all();
 302                 }
 303                 cards[card_index].capture->stop_dequeue_thread();
 304         }
 305
 306         video_encoder.reset(nullptr);
 307 }
 308
 309 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
 310 {
 311         printf("Configuring card %d...\n", card_index);
 312
 313         CaptureCard *card = &cards[card_index];
 314         if (card->capture != nullptr) {
 315                 card->capture->stop_dequeue_thread();
 316                 delete card->capture;
 317         }
 318         card->capture = capture;
 319         card->is_fake_capture = is_fake_capture;
 320         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 321         if (card->frame_allocator == nullptr) {
 322                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 323         }
 324         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 325         if (card->surface == nullptr) {
 326                 card->surface = create_surface_with_same_format(mixer_surface);
 327         }
 328         {
 329                 unique_lock<mutex> lock(cards[card_index].audio_mutex);
 330                 card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 331         }
 332         while (!card->new_frames.empty()) card->new_frames.pop();
 333         card->fractional_samples = 0;
 334         card->last_timecode = -1;
 335         card->next_local_pts = 0;
 336         card->capture->configure_card();
 337 }
 338
 339
 340 namespace {
 341
 342 int unwrap_timecode(uint16_t current_wrapped, int last)
 343 {
 344         uint16_t last_wrapped = last & 0xffff;
 345         if (current_wrapped > last_wrapped) {
 346                 return (last & ~0xffff) | current_wrapped;
 347         } else {
 348                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 349         }
 350 }
 351
 352 float find_peak(const float *samples, size_t num_samples)
 353 {
 354         float m = fabs(samples[0]);
 355         for (size_t i = 1; i < num_samples; ++i) {
 356                 m = max(m, fabs(samples[i]));
 357         }
 358         return m;
 359 }
 360
 361 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 362 {
 363         size_t num_samples = in.size() / 2;
 364         out_l->resize(num_samples);
 365         out_r->resize(num_samples);
 366
 367         const float *inptr = in.data();
 368         float *lptr = &(*out_l)[0];
 369         float *rptr = &(*out_r)[0];
 370         for (size_t i = 0; i < num_samples; ++i) {
 371                 *lptr++ = *inptr++;
 372                 *rptr++ = *inptr++;
 373         }
 374 }
 375
 376 }  // namespace
 377
 378 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 379                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 380                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 381 {
 382         CaptureCard *card = &cards[card_index];
 383
 384         if (is_mode_scanning[card_index]) {
 385                 if (video_format.has_signal) {
 386                         // Found a stable signal, so stop scanning.
 387                         is_mode_scanning[card_index] = false;
 388                 } else {
 389                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 390                         timespec now;
 391                         clock_gettime(CLOCK_MONOTONIC, &now);
 392                         double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
 393                                 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
 394                         if (sec_since_last_switch > switch_time_s) {
 395                                 // It isn't this mode; try the next one.
 396                                 mode_scanlist_index[card_index]++;
 397                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 398                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 399                                 last_mode_scan_change[card_index] = now;
 400                         }
 401                 }
 402         }
 403
 404         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 405         assert(frame_length > 0);
 406
 407         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;
 408         if (num_samples > OUTPUT_FREQUENCY / 10) {
 409                 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",
 410                         card_index, int(audio_frame.len), int(audio_offset),
 411                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 412                 if (video_frame.owner) {
 413                         video_frame.owner->release_frame(video_frame);
 414                 }
 415                 if (audio_frame.owner) {
 416                         audio_frame.owner->release_frame(audio_frame);
 417                 }
 418                 return;
 419         }
 420
 421         int64_t local_pts = card->next_local_pts;
 422         int dropped_frames = 0;
 423         if (card->last_timecode != -1) {
 424                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 425         }
 426
 427         // Convert the audio to stereo fp32 and add it.
 428         vector<float> audio;
 429         audio.resize(num_samples * 2);
 430         switch (audio_format.bits_per_sample) {
 431         case 0:
 432                 assert(num_samples == 0);
 433                 break;
 434         case 24:
 435                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 436                 break;
 437         case 32:
 438                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 439                 break;
 440         default:
 441                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 442                 assert(false);
 443         }
 444
 445         // Add the audio.
 446         {
 447                 unique_lock<mutex> lock(card->audio_mutex);
 448
 449                 // Number of samples per frame if we need to insert silence.
 450                 // (Could be nonintegral, but resampling will save us then.)
 451                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 452
 453                 if (dropped_frames > MAX_FPS * 2) {
 454                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 455                                 card_index, card->last_timecode, timecode);
 456                         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 457                         dropped_frames = 0;
 458                 } else if (dropped_frames > 0) {
 459                         // Insert silence as needed.
 460                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 461                                 card_index, dropped_frames, timecode);
 462                         vector<float> silence(silence_samples * 2, 0.0f);
 463                         for (int i = 0; i < dropped_frames; ++i) {
 464                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 465                                 // Note that if the format changed in the meantime, we have
 466                                 // no way of detecting that; we just have to assume the frame length
 467                                 // is always the same.
 468                                 local_pts += frame_length;
 469                         }
 470                 }
 471                 if (num_samples == 0) {
 472                         audio.resize(silence_samples * 2);
 473                         num_samples = silence_samples;
 474                 }
 475                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 476                 card->next_local_pts = local_pts + frame_length;
 477         }
 478
 479         card->last_timecode = timecode;
 480
 481         // Done with the audio, so release it.
 482         if (audio_frame.owner) {
 483                 audio_frame.owner->release_frame(audio_frame);
 484         }
 485
 486         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 487         if (video_frame.len - video_offset == 0 ||
 488             video_frame.len - video_offset != expected_length) {
 489                 if (video_frame.len != 0) {
 490                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 491                                 card_index, video_frame.len - video_offset, expected_length);
 492                 }
 493                 if (video_frame.owner) {
 494                         video_frame.owner->release_frame(video_frame);
 495                 }
 496
 497                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 498                 // so that pts can go up accordingly.
 499                 {
 500                         unique_lock<mutex> lock(bmusb_mutex);
 501                         CaptureCard::NewFrame new_frame;
 502                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 503                         new_frame.length = frame_length;
 504                         new_frame.interlaced = false;
 505                         new_frame.dropped_frames = dropped_frames;
 506                         card->new_frames.push(move(new_frame));
 507                         card->new_frames_changed.notify_all();
 508                 }
 509                 return;
 510         }
 511
 512         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 513
 514         unsigned num_fields = video_format.interlaced ? 2 : 1;
 515         timespec frame_upload_start;
 516         if (video_format.interlaced) {
 517                 // Send the two fields along as separate frames; the other side will need to add
 518                 // a deinterlacer to actually get this right.
 519                 assert(video_format.height % 2 == 0);
 520                 video_format.height /= 2;
 521                 assert(frame_length % 2 == 0);
 522                 frame_length /= 2;
 523                 num_fields = 2;
 524                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 525         }
 526         userdata->last_interlaced = video_format.interlaced;
 527         userdata->last_has_signal = video_format.has_signal;
 528         userdata->last_is_connected = video_format.is_connected;
 529         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 530         userdata->last_frame_rate_den = video_format.frame_rate_den;
 531         RefCountedFrame frame(video_frame);
 532
 533         // Upload the textures.
 534         size_t cbcr_width = video_format.width / 2;
 535         size_t cbcr_offset = video_offset / 2;
 536         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 537
 538         for (unsigned field = 0; field < num_fields; ++field) {
 539                 // Put the actual texture upload in a lambda that is executed in the main thread.
 540                 // It is entirely possible to do this in the same thread (and it might even be
 541                 // faster, depending on the GPU and driver), but it appears to be trickling
 542                 // driver bugs very easily.
 543                 //
 544                 // Note that this means we must hold on to the actual frame data in <userdata>
 545                 // until the upload command is run, but we hold on to <frame> much longer than that
 546                 // (in fact, all the way until we no longer use the texture in rendering).
 547                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
 548                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 549
 550                         if (userdata->tex_y[field] == 0 ||
 551                             userdata->tex_cbcr[field] == 0 ||
 552                             video_format.width != userdata->last_width[field] ||
 553                             video_format.height != userdata->last_height[field]) {
 554                                 // We changed resolution since last use of this texture, so we need to create
 555                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 556                                 // we don't need to worry about these flip-flopping between resolutions.
 557                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 558                                 check_error();
 559                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 560                                 check_error();
 561                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 562                                 check_error();
 563                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 564                                 check_error();
 565                                 userdata->last_width[field] = video_format.width;
 566                                 userdata->last_height[field] = video_format.height;
 567                         }
 568
 569                         GLuint pbo = userdata->pbo;
 570                         check_error();
 571                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 572                         check_error();
 573
 574                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 575                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 576
 577                         if (global_flags.flush_pbos) {
 578                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 579                                 check_error();
 580                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 581                                 check_error();
 582                         }
 583
 584                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 585                         check_error();
 586                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 587                         check_error();
 588                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 589                         check_error();
 590                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 591                         check_error();
 592                         glBindTexture(GL_TEXTURE_2D, 0);
 593                         check_error();
 594                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 595                         check_error();
 596                 };
 597
 598                 if (field == 1) {
 599                         // Don't upload the second field as fast as we can; wait until
 600                         // the field time has approximately passed. (Otherwise, we could
 601                         // get timing jitter against the other sources, and possibly also
 602                         // against the video display, although the latter is not as critical.)
 603                         // This requires our system clock to be reasonably close to the
 604                         // video clock, but that's not an unreasonable assumption.
 605                         timespec second_field_start;
 606                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 607                                 frame_length * 1000000000 / TIMEBASE;
 608                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 609                                 second_field_start.tv_nsec / 1000000000;
 610                         second_field_start.tv_nsec %= 1000000000;
 611
 612                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 613                                                &second_field_start, nullptr) == -1 &&
 614                                errno == EINTR) ;
 615                 }
 616
 617                 {
 618                         unique_lock<mutex> lock(bmusb_mutex);
 619                         CaptureCard::NewFrame new_frame;
 620                         new_frame.frame = frame;
 621                         new_frame.length = frame_length;
 622                         new_frame.field = field;
 623                         new_frame.interlaced = video_format.interlaced;
 624                         new_frame.upload_func = upload_func;
 625                         new_frame.dropped_frames = dropped_frames;
 626                         card->new_frames.push(move(new_frame));
 627                         card->new_frames_changed.notify_all();
 628                 }
 629         }
 630 }
 631
 632 void Mixer::bm_hotplug_add(libusb_device *dev)
 633 {
 634         lock_guard<mutex> lock(hotplug_mutex);
 635         hotplugged_cards.push_back(dev);
 636 }
 637
 638 void Mixer::bm_hotplug_remove(unsigned card_index)
 639 {
 640         cards[card_index].new_frames_changed.notify_all();
 641 }
 642
 643 void Mixer::thread_func()
 644 {
 645         eglBindAPI(EGL_OPENGL_API);
 646         QOpenGLContext *context = create_context(mixer_surface);
 647         if (!make_current(context, mixer_surface)) {
 648                 printf("oops\n");
 649                 exit(1);
 650         }
 651
 652         struct timespec start, now;
 653         clock_gettime(CLOCK_MONOTONIC, &start);
 654
 655         int frame = 0;
 656         int stats_dropped_frames = 0;
 657
 658         while (!should_quit) {
 659                 CaptureCard::NewFrame new_frames[MAX_CARDS];
 660                 bool has_new_frame[MAX_CARDS] = { false };
 661                 int num_samples[MAX_CARDS] = { 0 };
 662
 663                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 664                 assert(master_card_index < num_cards);
 665
 666                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 667                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 668                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 669                 send_audio_level_callback();
 670
 671                 handle_hotplugged_cards();
 672
 673                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 674                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 675                                 continue;
 676                         }
 677                         if (new_frames[card_index].frame->len == 0) {
 678                                 ++new_frames[card_index].dropped_frames;
 679                         }
 680                         if (new_frames[card_index].dropped_frames > 0) {
 681                                 printf("Card %u dropped %d frames before this\n",
 682                                         card_index, int(new_frames[card_index].dropped_frames));
 683                         }
 684                 }
 685
 686                 // If the first card is reporting a corrupted or otherwise dropped frame,
 687                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 688                 if (new_frames[master_card_index].frame->len == 0) {
 689                         ++stats_dropped_frames;
 690                         pts_int += new_frames[master_card_index].length;
 691                         continue;
 692                 }
 693
 694                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 695                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 696                                 continue;
 697
 698                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 699                         assert(new_frame->frame != nullptr);
 700                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 701                         check_error();
 702
 703                         // The new texture might need uploading before use.
 704                         if (new_frame->upload_func) {
 705                                 new_frame->upload_func();
 706                                 new_frame->upload_func = nullptr;
 707                         }
 708                 }
 709
 710                 int64_t duration = new_frames[master_card_index].length;
 711                 render_one_frame(duration);
 712                 ++frame;
 713                 pts_int += duration;
 714
 715                 clock_gettime(CLOCK_MONOTONIC, &now);
 716                 double elapsed = now.tv_sec - start.tv_sec +
 717                         1e-9 * (now.tv_nsec - start.tv_nsec);
 718                 if (frame % 100 == 0) {
 719                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 720                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 721                                 1e3 * elapsed / frame);
 722                 //      chain->print_phase_timing();
 723
 724                         // Check our memory usage, to see if we are close to our mlockall()
 725                         // limit (if at all set).
 726                         rusage used;
 727                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 728                                 perror("getrusage(RUSAGE_SELF)");
 729                                 assert(false);
 730                         }
 731
 732                         if (uses_mlock) {
 733                                 rlimit limit;
 734                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 735                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 736                                         assert(false);
 737                                 }
 738
 739                                 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 740                                         long(used.ru_maxrss / 1024),
 741                                         long(limit.rlim_cur / 1048576),
 742                                         float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 743                         } else {
 744                                 printf(", using %ld MB memory (not locked)",
 745                                         long(used.ru_maxrss / 1024));
 746                         }
 747
 748                         printf("\n");
 749                 }
 750
 751
 752                 if (should_cut.exchange(false)) {  // Test and clear.
 753                         video_encoder->do_cut(frame);
 754                 }
 755
 756 #if 0
 757                 // Reset every 100 frames, so that local variations in frame times
 758                 // (especially for the first few frames, when the shaders are
 759                 // compiled etc.) don't make it hard to measure for the entire
 760                 // remaining duration of the program.
 761                 if (frame == 10000) {
 762                         frame = 0;
 763                         start = now;
 764                 }
 765 #endif
 766                 check_error();
 767         }
 768
 769         resource_pool->clean_context();
 770 }
 771
 772 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])
 773 {
 774 start:
 775         // The first card is the master timer, so wait for it to have a new frame.
 776         // TODO: Add a timeout.
 777         unique_lock<mutex> lock(bmusb_mutex);
 778         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(); });
 779
 780         if (cards[master_card_index].new_frames.empty()) {
 781                 // We were woken up, but not due to a new frame. Deal with it
 782                 // and then restart.
 783                 assert(cards[master_card_index].capture->get_disconnected());
 784                 handle_hotplugged_cards();
 785                 goto start;
 786         }
 787
 788         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 789                 CaptureCard *card = &cards[card_index];
 790                 if (card->new_frames.empty()) {
 791                         assert(card_index != master_card_index);
 792                         card->queue_length_policy.update_policy(-1);
 793                         continue;
 794                 }
 795                 new_frames[card_index] = move(card->new_frames.front());
 796                 has_new_frame[card_index] = true;
 797                 card->new_frames.pop();
 798                 card->new_frames_changed.notify_all();
 799
 800                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 801                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 802                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 803                 assert(num_samples[card_index] >= 0);
 804
 805                 if (card_index == master_card_index) {
 806                         // We don't use the queue length policy for the master card,
 807                         // but we will if it stops being the master. Thus, clear out
 808                         // the policy in case we switch in the future.
 809                         card->queue_length_policy.reset(card_index);
 810                 } else {
 811                         // If we have excess frames compared to the policy for this card,
 812                         // drop frames from the head.
 813                         card->queue_length_policy.update_policy(card->new_frames.size());
 814                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 815                                 card->new_frames.pop();
 816                         }
 817                 }
 818         }
 819 }
 820
 821 void Mixer::handle_hotplugged_cards()
 822 {
 823         // Check for cards that have been disconnected since last frame.
 824         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 825                 CaptureCard *card = &cards[card_index];
 826                 if (card->capture->get_disconnected()) {
 827                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 828                         configure_card(card_index, new FakeCapture(card_index), /*is_fake_capture=*/true);
 829                         card->queue_length_policy.reset(card_index);
 830                         card->capture->start_bm_capture();
 831                 }
 832         }
 833
 834         // Check for cards that have been connected since last frame.
 835         vector<libusb_device *> hotplugged_cards_copy;
 836         {
 837                 lock_guard<mutex> lock(hotplug_mutex);
 838                 swap(hotplugged_cards, hotplugged_cards_copy);
 839         }
 840         for (libusb_device *new_dev : hotplugged_cards_copy) {
 841                 // Look for a fake capture card where we can stick this in.
 842                 int free_card_index = -1;
 843                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 844                         if (cards[card_index].is_fake_capture) {
 845                                 free_card_index = int(card_index);
 846                                 break;
 847                         }
 848                 }
 849
 850                 if (free_card_index == -1) {
 851                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 852                         libusb_unref_device(new_dev);
 853                 } else {
 854                         // BMUSBCapture takes ownership.
 855                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 856                         CaptureCard *card = &cards[free_card_index];
 857                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 858                         configure_card(free_card_index, capture, /*is_fake_capture=*/false);
 859                         card->queue_length_policy.reset(free_card_index);
 860                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 861                         capture->start_bm_capture();
 862                 }
 863         }
 864 }
 865
 866
 867 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 868 {
 869         // Resample the audio as needed, including from previously dropped frames.
 870         assert(num_cards > 0);
 871         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 872                 {
 873                         // Signal to the audio thread to process this frame.
 874                         unique_lock<mutex> lock(audio_mutex);
 875                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
 876                         audio_task_queue_changed.notify_one();
 877                 }
 878                 if (frame_num != dropped_frames) {
 879                         // For dropped frames, increase the pts. Note that if the format changed
 880                         // in the meantime, we have no way of detecting that; we just have to
 881                         // assume the frame length is always the same.
 882                         pts_int += length_per_frame;
 883                 }
 884         }
 885 }
 886
 887 void Mixer::render_one_frame(int64_t duration)
 888 {
 889         // Get the main chain from the theme, and set its state immediately.
 890         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 891         EffectChain *chain = theme_main_chain.chain;
 892         theme_main_chain.setup_chain();
 893         //theme_main_chain.chain->enable_phase_timing(true);
 894
 895         GLuint y_tex, cbcr_tex;
 896         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 897         assert(got_frame);
 898
 899         // Render main chain.
 900         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 901         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 902         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 903         check_error();
 904         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 905         resource_pool->release_fbo(fbo);
 906
 907         subsample_chroma(cbcr_full_tex, cbcr_tex);
 908         resource_pool->release_2d_texture(cbcr_full_tex);
 909
 910         // Set the right state for rgba_tex.
 911         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 912         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 913         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 914         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 915         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 916
 917         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 918         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 919
 920         // The live frame just shows the RGBA texture we just rendered.
 921         // It owns rgba_tex now.
 922         DisplayFrame live_frame;
 923         live_frame.chain = display_chain.get();
 924         live_frame.setup_chain = [this, rgba_tex]{
 925                 display_input->set_texture_num(rgba_tex);
 926         };
 927         live_frame.ready_fence = fence;
 928         live_frame.input_frames = {};
 929         live_frame.temp_textures = { rgba_tex };
 930         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 931
 932         // Set up preview and any additional channels.
 933         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 934                 DisplayFrame display_frame;
 935                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 936                 display_frame.chain = chain.chain;
 937                 display_frame.setup_chain = chain.setup_chain;
 938                 display_frame.ready_fence = fence;
 939                 display_frame.input_frames = chain.input_frames;
 940                 display_frame.temp_textures = {};
 941                 output_channel[i].output_frame(display_frame);
 942         }
 943 }
 944
 945 void Mixer::send_audio_level_callback()
 946 {
 947         if (audio_level_callback == nullptr) {
 948                 return;
 949         }
 950
 951         unique_lock<mutex> lock(compressor_mutex);
 952         double loudness_s = r128.loudness_S();
 953         double loudness_i = r128.integrated();
 954         double loudness_range_low = r128.range_min();
 955         double loudness_range_high = r128.range_max();
 956
 957         audio_level_callback(loudness_s, 20.0 * log10(peak),
 958                 loudness_i, loudness_range_low, loudness_range_high,
 959                 gain_staging_db, 20.0 * log10(final_makeup_gain),
 960                 correlation.get_correlation());
 961 }
 962
 963 void Mixer::audio_thread_func()
 964 {
 965         while (!should_quit) {
 966                 AudioTask task;
 967
 968                 {
 969                         unique_lock<mutex> lock(audio_mutex);
 970                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 971                         if (should_quit) {
 972                                 return;
 973                         }
 974                         task = audio_task_queue.front();
 975                         audio_task_queue.pop();
 976                 }
 977
 978                 process_audio_one_frame(task.pts_int, task.num_samples);
 979         }
 980 }
 981
 982 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 983 {
 984         vector<float> samples_card;
 985         vector<float> samples_out;
 986
 987         // TODO: Allow mixing audio from several sources.
 988         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 989         assert(selected_audio_card < num_cards);
 990
 991         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 992                 samples_card.resize(num_samples * 2);
 993                 {
 994                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 995                         cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples);
 996                 }
 997                 if (card_index == selected_audio_card) {
 998                         samples_out = move(samples_card);
 999                 }
1000         }
1001
1002         // Cut away everything under 120 Hz (or whatever the cutoff is);
1003         // we don't need it for voice, and it will reduce headroom
1004         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
1005         // should be dampened.)
1006         if (locut_enabled) {
1007                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
1008         }
1009
1010         // Apply a level compressor to get the general level right.
1011         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
1012         // (or more precisely, near it, since we don't use infinite ratio),
1013         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
1014         // entirely arbitrary, but from practical tests with speech, it seems to
1015         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
1016         {
1017                 unique_lock<mutex> lock(compressor_mutex);
1018                 if (level_compressor_enabled) {
1019                         float threshold = 0.01f;   // -40 dBFS.
1020                         float ratio = 20.0f;
1021                         float attack_time = 0.5f;
1022                         float release_time = 20.0f;
1023                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
1024                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1025                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
1026                 } else {
1027                         // Just apply the gain we already had.
1028                         float g = pow(10.0f, gain_staging_db / 20.0f);
1029                         for (size_t i = 0; i < samples_out.size(); ++i) {
1030                                 samples_out[i] *= g;
1031                         }
1032                 }
1033         }
1034
1035 #if 0
1036         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
1037                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
1038                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
1039                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
1040 #endif
1041
1042 //      float limiter_att, compressor_att;
1043
1044         // The real compressor.
1045         if (compressor_enabled) {
1046                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
1047                 float ratio = 20.0f;
1048                 float attack_time = 0.005f;
1049                 float release_time = 0.040f;
1050                 float makeup_gain = 2.0f;  // +6 dB.
1051                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1052 //              compressor_att = compressor.get_attenuation();
1053         }
1054
1055         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
1056         // Note that since ratio is not infinite, we could go slightly higher than this.
1057         if (limiter_enabled) {
1058                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
1059                 float ratio = 30.0f;
1060                 float attack_time = 0.0f;  // Instant.
1061                 float release_time = 0.020f;
1062                 float makeup_gain = 1.0f;  // 0 dB.
1063                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1064 //              limiter_att = limiter.get_attenuation();
1065         }
1066
1067 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
1068
1069         // Upsample 4x to find interpolated peak.
1070         peak_resampler.inp_data = samples_out.data();
1071         peak_resampler.inp_count = samples_out.size() / 2;
1072
1073         vector<float> interpolated_samples_out;
1074         interpolated_samples_out.resize(samples_out.size());
1075         while (peak_resampler.inp_count > 0) {  // About four iterations.
1076                 peak_resampler.out_data = &interpolated_samples_out[0];
1077                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
1078                 peak_resampler.process();
1079                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
1080                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
1081                 peak_resampler.out_data = nullptr;
1082         }
1083
1084         // At this point, we are most likely close to +0 LU, but all of our
1085         // measurements have been on raw sample values, not R128 values.
1086         // So we have a final makeup gain to get us to +0 LU; the gain
1087         // adjustments required should be relatively small, and also, the
1088         // offset shouldn't change much (only if the type of audio changes
1089         // significantly). Thus, we shoot for updating this value basically
1090         // “whenever we process buffers”, since the R128 calculation isn't exactly
1091         // something we get out per-sample.
1092         //
1093         // Note that there's a feedback loop here, so we choose a very slow filter
1094         // (half-time of 100 seconds).
1095         double target_loudness_factor, alpha;
1096         {
1097                 unique_lock<mutex> lock(compressor_mutex);
1098                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
1099                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
1100                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
1101
1102                 // If we're outside +/- 5 LU uncorrected, we don't count it as
1103                 // a normal signal (probably silence) and don't change the
1104                 // correction factor; just apply what we already have.
1105                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
1106                         alpha = 0.0;
1107                 } else {
1108                         // Formula adapted from
1109                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
1110                         const double half_time_s = 100.0;
1111                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
1112                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
1113                 }
1114
1115                 double m = final_makeup_gain;
1116                 for (size_t i = 0; i < samples_out.size(); i += 2) {
1117                         samples_out[i + 0] *= m;
1118                         samples_out[i + 1] *= m;
1119                         m += (target_loudness_factor - m) * alpha;
1120                 }
1121                 final_makeup_gain = m;
1122         }
1123
1124         // Find R128 levels and L/R correlation.
1125         vector<float> left, right;
1126         deinterleave_samples(samples_out, &left, &right);
1127         float *ptrs[] = { left.data(), right.data() };
1128         {
1129                 unique_lock<mutex> lock(compressor_mutex);
1130                 r128.process(left.size(), ptrs);
1131                 correlation.process_samples(samples_out);
1132         }
1133
1134         // Send the samples to the sound card.
1135         if (alsa) {
1136                 alsa->write(samples_out);
1137         }
1138
1139         // And finally add them to the output.
1140         video_encoder->add_audio(frame_pts_int, move(samples_out));
1141 }
1142
1143 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1144 {
1145         GLuint vao;
1146         glGenVertexArrays(1, &vao);
1147         check_error();
1148
1149         glBindVertexArray(vao);
1150         check_error();
1151
1152         // Extract Cb/Cr.
1153         GLuint fbo = resource_pool->create_fbo(dst_tex);
1154         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1155         glViewport(0, 0, WIDTH/2, HEIGHT/2);
1156         check_error();
1157
1158         glUseProgram(cbcr_program_num);
1159         check_error();
1160
1161         glActiveTexture(GL_TEXTURE0);
1162         check_error();
1163         glBindTexture(GL_TEXTURE_2D, src_tex);
1164         check_error();
1165         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1166         check_error();
1167         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1168         check_error();
1169         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1170         check_error();
1171
1172         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1173         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1174
1175         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1176         check_error();
1177
1178         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1179                 glEnableVertexAttribArray(attr_index);
1180                 check_error();
1181                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1182                 check_error();
1183         }
1184
1185         glDrawArrays(GL_TRIANGLES, 0, 3);
1186         check_error();
1187
1188         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1189                 glDisableVertexAttribArray(attr_index);
1190                 check_error();
1191         }
1192
1193         glUseProgram(0);
1194         check_error();
1195         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1196         check_error();
1197
1198         resource_pool->release_fbo(fbo);
1199         glDeleteVertexArrays(1, &vao);
1200 }
1201
1202 void Mixer::release_display_frame(DisplayFrame *frame)
1203 {
1204         for (GLuint texnum : frame->temp_textures) {
1205                 resource_pool->release_2d_texture(texnum);
1206         }
1207         frame->temp_textures.clear();
1208         frame->ready_fence.reset();
1209         frame->input_frames.clear();
1210 }
1211
1212 void Mixer::start()
1213 {
1214         mixer_thread = thread(&Mixer::thread_func, this);
1215         audio_thread = thread(&Mixer::audio_thread_func, this);
1216 }
1217
1218 void Mixer::quit()
1219 {
1220         should_quit = true;
1221         audio_task_queue_changed.notify_one();
1222         mixer_thread.join();
1223         audio_thread.join();
1224 }
1225
1226 void Mixer::transition_clicked(int transition_num)
1227 {
1228         theme->transition_clicked(transition_num, pts());
1229 }
1230
1231 void Mixer::channel_clicked(int preview_num)
1232 {
1233         theme->channel_clicked(preview_num);
1234 }
1235
1236 void Mixer::reset_meters()
1237 {
1238         peak_resampler.reset();
1239         peak = 0.0f;
1240         r128.reset();
1241         r128.integr_start();
1242         correlation.reset();
1243 }
1244
1245 void Mixer::start_mode_scanning(unsigned card_index)
1246 {
1247         assert(card_index < num_cards);
1248         if (is_mode_scanning[card_index]) {
1249                 return;
1250         }
1251         is_mode_scanning[card_index] = true;
1252         mode_scanlist[card_index].clear();
1253         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1254                 mode_scanlist[card_index].push_back(mode.first);
1255         }
1256         assert(!mode_scanlist[card_index].empty());
1257         mode_scanlist_index[card_index] = 0;
1258         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1259         clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1260 }
1261
1262 Mixer::OutputChannel::~OutputChannel()
1263 {
1264         if (has_current_frame) {
1265                 parent->release_display_frame(&current_frame);
1266         }
1267         if (has_ready_frame) {
1268                 parent->release_display_frame(&ready_frame);
1269         }
1270 }
1271
1272 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1273 {
1274         // Store this frame for display. Remove the ready frame if any
1275         // (it was seemingly never used).
1276         {
1277                 unique_lock<mutex> lock(frame_mutex);
1278                 if (has_ready_frame) {
1279                         parent->release_display_frame(&ready_frame);
1280                 }
1281                 ready_frame = frame;
1282                 has_ready_frame = true;
1283         }
1284
1285         if (new_frame_ready_callback) {
1286                 new_frame_ready_callback();
1287         }
1288
1289         // Reduce the number of callbacks by filtering duplicates. The reason
1290         // why we bother doing this is that Qt seemingly can get into a state
1291         // where its builds up an essentially unbounded queue of signals,
1292         // consuming more and more memory, and there's no good way of collapsing
1293         // user-defined signals or limiting the length of the queue.
1294         if (transition_names_updated_callback) {
1295                 vector<string> transition_names = global_mixer->get_transition_names();
1296                 bool changed = false;
1297                 if (transition_names.size() != last_transition_names.size()) {
1298                         changed = true;
1299                 } else {
1300                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1301                                 if (transition_names[i] != last_transition_names[i]) {
1302                                         changed = true;
1303                                         break;
1304                                 }
1305                         }
1306                 }
1307                 if (changed) {
1308                         transition_names_updated_callback(transition_names);
1309                         last_transition_names = transition_names;
1310                 }
1311         }
1312         if (name_updated_callback) {
1313                 string name = global_mixer->get_channel_name(channel);
1314                 if (name != last_name) {
1315                         name_updated_callback(name);
1316                         last_name = name;
1317                 }
1318         }
1319         if (color_updated_callback) {
1320                 string color = global_mixer->get_channel_color(channel);
1321                 if (color != last_color) {
1322                         color_updated_callback(color);
1323                         last_color = color;
1324                 }
1325         }
1326 }
1327
1328 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1329 {
1330         unique_lock<mutex> lock(frame_mutex);
1331         if (!has_current_frame && !has_ready_frame) {
1332                 return false;
1333         }
1334
1335         if (has_current_frame && has_ready_frame) {
1336                 // We have a new ready frame. Toss the current one.
1337                 parent->release_display_frame(&current_frame);
1338                 has_current_frame = false;
1339         }
1340         if (has_ready_frame) {
1341                 assert(!has_current_frame);
1342                 current_frame = ready_frame;
1343                 ready_frame.ready_fence.reset();  // Drop the refcount.
1344                 ready_frame.input_frames.clear();  // Drop the refcounts.
1345                 has_current_frame = true;
1346                 has_ready_frame = false;
1347         }
1348
1349         *frame = current_frame;
1350         return true;
1351 }
1352
1353 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1354 {
1355         new_frame_ready_callback = callback;
1356 }
1357
1358 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1359 {
1360         transition_names_updated_callback = callback;
1361 }
1362
1363 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1364 {
1365         name_updated_callback = callback;
1366 }
1367
1368 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1369 {
1370         color_updated_callback = callback;
1371 }
1372
1373 mutex RefCountedGLsync::fence_lock;