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