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

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <endian.h>
   7 #include <epoxy/egl.h>
   8 #include <movit/effect_chain.h>
   9 #include <movit/effect_util.h>
  10 #include <movit/flat_input.h>
  11 #include <movit/image_format.h>
  12 #include <movit/init.h>
  13 #include <movit/resource_pool.h>
  14 #include <movit/util.h>
  15 #include <stdint.h>
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <sys/time.h>
  19 #include <time.h>
  20 #include <algorithm>
  21 #include <chrono>
  22 #include <cmath>
  23 #include <condition_variable>
  24 #include <cstddef>
  25 #include <memory>
  26 #include <mutex>
  27 #include <string>
  28 #include <thread>
  29 #include <utility>
  30 #include <vector>
  31 #include <arpa/inet.h>
  32 #include <sys/time.h>
  33 #include <sys/resource.h>
  34
  35 #include "bmusb/bmusb.h"
  36 #include "bmusb/fake_capture.h"
  37 #include "context.h"
  38 #include "decklink_capture.h"
  39 #include "defs.h"
  40 #include "disk_space_estimator.h"
  41 #include "flags.h"
  42 #include "video_encoder.h"
  43 #include "pbo_frame_allocator.h"
  44 #include "ref_counted_gl_sync.h"
  45 #include "timebase.h"
  46
  47 class QOpenGLContext;
  48
  49 using namespace movit;
  50 using namespace std;
  51 using namespace std::chrono;
  52 using namespace std::placeholders;
  53 using namespace bmusb;
  54
  55 Mixer *global_mixer = nullptr;
  56 bool uses_mlock = false;
  57
  58 namespace {
  59
  60 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  61 {
  62         assert(in_channels >= out_channels);
  63         for (size_t i = 0; i < num_samples; ++i) {
  64                 for (size_t j = 0; j < out_channels; ++j) {
  65                         uint32_t s1 = *src++;
  66                         uint32_t s2 = *src++;
  67                         uint32_t s3 = *src++;
  68                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  69                         dst[i * out_channels + j] = int(s) * (1.0f / 2147483648.0f);
  70                 }
  71                 src += 3 * (in_channels - out_channels);
  72         }
  73 }
  74
  75 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  76 {
  77         assert(in_channels >= out_channels);
  78         for (size_t i = 0; i < num_samples; ++i) {
  79                 for (size_t j = 0; j < out_channels; ++j) {
  80                         int32_t s = le32toh(*(int32_t *)src);
  81                         dst[i * out_channels + j] = s * (1.0f / 2147483648.0f);
  82                         src += 4;
  83                 }
  84                 src += 4 * (in_channels - out_channels);
  85         }
  86 }
  87
  88 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  89 {
  90         if (interlaced) {
  91                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  92                         input_state->buffered_frames[card_index][frame_num] =
  93                                 input_state->buffered_frames[card_index][frame_num - 1];
  94                 }
  95                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  96         } else {
  97                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  98                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  99                 }
 100         }
 101 }
 102
 103 }  // namespace
 104
 105 void QueueLengthPolicy::update_policy(int queue_length)
 106 {
 107         if (queue_length < 0) {  // Starvation.
 108                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
 109                         ++safe_queue_length;
 110                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
 111                                 card_index, safe_queue_length);
 112                 }
 113                 frames_with_at_least_one = 0;
 114                 been_at_safe_point_since_last_starvation = false;
 115                 return;
 116         }
 117         if (queue_length > 0) {
 118                 if (queue_length >= int(safe_queue_length)) {
 119                         been_at_safe_point_since_last_starvation = true;
 120                 }
 121                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
 122                         --safe_queue_length;
 123                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
 124                                 card_index, safe_queue_length);
 125                         frames_with_at_least_one = 0;
 126                 }
 127         } else {
 128                 frames_with_at_least_one = 0;
 129         }
 130 }
 131
 132 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 133         : httpd(),
 134           num_cards(num_cards),
 135           mixer_surface(create_surface(format)),
 136           h264_encoder_surface(create_surface(format)),
 137           correlation(OUTPUT_FREQUENCY),
 138           level_compressor(OUTPUT_FREQUENCY),
 139           limiter(OUTPUT_FREQUENCY),
 140           compressor(OUTPUT_FREQUENCY)
 141 {
 142         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 143         check_error();
 144
 145         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 146         // will be halved when sampling them, and we need to compensate here.
 147         movit_texel_subpixel_precision /= 2.0;
 148
 149         resource_pool.reset(new ResourcePool);
 150         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 151         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 152                 output_channel[i].parent = this;
 153                 output_channel[i].channel = i;
 154         }
 155
 156         ImageFormat inout_format;
 157         inout_format.color_space = COLORSPACE_sRGB;
 158         inout_format.gamma_curve = GAMMA_sRGB;
 159
 160         // Display chain; shows the live output produced by the main chain (its RGBA version).
 161         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 162         check_error();
 163         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 164         display_chain->add_input(display_input);
 165         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 166         display_chain->set_dither_bits(0);  // Don't bother.
 167         display_chain->finalize();
 168
 169         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd, global_disk_space_estimator));
 170
 171         // Start listening for clients only once VideoEncoder has written its header, if any.
 172         httpd.start(9095);
 173
 174         // First try initializing the then PCI devices, then USB, then
 175         // fill up with fake cards until we have the desired number of cards.
 176         unsigned num_pci_devices = 0;
 177         unsigned card_index = 0;
 178
 179         {
 180                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 181                 if (decklink_iterator != nullptr) {
 182                         for ( ; card_index < num_cards; ++card_index) {
 183                                 IDeckLink *decklink;
 184                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 185                                         break;
 186                                 }
 187
 188                                 configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
 189                                 ++num_pci_devices;
 190                         }
 191                         decklink_iterator->Release();
 192                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 193                 } else {
 194                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 195                 }
 196         }
 197         unsigned num_usb_devices = BMUSBCapture::num_cards();
 198         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 199                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 200                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 201                 configure_card(card_index, capture, /*is_fake_capture=*/false);
 202         }
 203         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 204
 205         unsigned num_fake_cards = 0;
 206         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 207                 FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 208                 configure_card(card_index, capture, /*is_fake_capture=*/true);
 209         }
 210
 211         if (num_fake_cards > 0) {
 212                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 213         }
 214
 215         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 216         BMUSBCapture::start_bm_thread();
 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         BMUSBCapture::stop_bm_thread();
 292
 293         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 294                 {
 295                         unique_lock<mutex> lock(bmusb_mutex);
 296                         cards[card_index].should_quit = true;  // Unblock thread.
 297                         cards[card_index].new_frames_changed.notify_all();
 298                 }
 299                 cards[card_index].capture->stop_dequeue_thread();
 300         }
 301
 302         video_encoder.reset(nullptr);
 303 }
 304
 305 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
 306 {
 307         printf("Configuring card %d...\n", card_index);
 308
 309         CaptureCard *card = &cards[card_index];
 310         if (card->capture != nullptr) {
 311                 card->capture->stop_dequeue_thread();
 312                 delete card->capture;
 313         }
 314         card->capture = capture;
 315         card->is_fake_capture = is_fake_capture;
 316         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 317         if (card->frame_allocator == nullptr) {
 318                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 319         }
 320         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 321         if (card->surface == nullptr) {
 322                 card->surface = create_surface_with_same_format(mixer_surface);
 323         }
 324         {
 325                 unique_lock<mutex> lock(cards[card_index].audio_mutex);
 326                 card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 327         }
 328         while (!card->new_frames.empty()) card->new_frames.pop();
 329         card->fractional_samples = 0;
 330         card->last_timecode = -1;
 331         card->next_local_pts = 0;
 332         card->capture->configure_card();
 333 }
 334
 335
 336 namespace {
 337
 338 int unwrap_timecode(uint16_t current_wrapped, int last)
 339 {
 340         uint16_t last_wrapped = last & 0xffff;
 341         if (current_wrapped > last_wrapped) {
 342                 return (last & ~0xffff) | current_wrapped;
 343         } else {
 344                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 345         }
 346 }
 347
 348 float find_peak(const float *samples, size_t num_samples)
 349 {
 350         float m = fabs(samples[0]);
 351         for (size_t i = 1; i < num_samples; ++i) {
 352                 m = max(m, fabs(samples[i]));
 353         }
 354         return m;
 355 }
 356
 357 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 358 {
 359         size_t num_samples = in.size() / 2;
 360         out_l->resize(num_samples);
 361         out_r->resize(num_samples);
 362
 363         const float *inptr = in.data();
 364         float *lptr = &(*out_l)[0];
 365         float *rptr = &(*out_r)[0];
 366         for (size_t i = 0; i < num_samples; ++i) {
 367                 *lptr++ = *inptr++;
 368                 *rptr++ = *inptr++;
 369         }
 370 }
 371
 372 }  // namespace
 373
 374 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 375                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 376                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 377 {
 378         CaptureCard *card = &cards[card_index];
 379
 380         if (is_mode_scanning[card_index]) {
 381                 if (video_format.has_signal) {
 382                         // Found a stable signal, so stop scanning.
 383                         is_mode_scanning[card_index] = false;
 384                 } else {
 385                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 386                         steady_clock::time_point now = steady_clock::now();
 387                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 388                         if (sec_since_last_switch > switch_time_s) {
 389                                 // It isn't this mode; try the next one.
 390                                 mode_scanlist_index[card_index]++;
 391                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 392                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 393                                 last_mode_scan_change[card_index] = now;
 394                         }
 395                 }
 396         }
 397
 398         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 399         assert(frame_length > 0);
 400
 401         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;
 402         if (num_samples > OUTPUT_FREQUENCY / 10) {
 403                 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",
 404                         card_index, int(audio_frame.len), int(audio_offset),
 405                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 406                 if (video_frame.owner) {
 407                         video_frame.owner->release_frame(video_frame);
 408                 }
 409                 if (audio_frame.owner) {
 410                         audio_frame.owner->release_frame(audio_frame);
 411                 }
 412                 return;
 413         }
 414
 415         int64_t local_pts = card->next_local_pts;
 416         int dropped_frames = 0;
 417         if (card->last_timecode != -1) {
 418                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 419         }
 420
 421         // Convert the audio to stereo fp32 and add it.
 422         vector<float> audio;
 423         audio.resize(num_samples * 2);
 424         switch (audio_format.bits_per_sample) {
 425         case 0:
 426                 assert(num_samples == 0);
 427                 break;
 428         case 24:
 429                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 430                 break;
 431         case 32:
 432                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 433                 break;
 434         default:
 435                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 436                 assert(false);
 437         }
 438
 439         // Add the audio.
 440         {
 441                 unique_lock<mutex> lock(card->audio_mutex);
 442
 443                 // Number of samples per frame if we need to insert silence.
 444                 // (Could be nonintegral, but resampling will save us then.)
 445                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 446
 447                 if (dropped_frames > MAX_FPS * 2) {
 448                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 449                                 card_index, card->last_timecode, timecode);
 450                         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 451                         dropped_frames = 0;
 452                 } else if (dropped_frames > 0) {
 453                         // Insert silence as needed.
 454                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 455                                 card_index, dropped_frames, timecode);
 456                         vector<float> silence(silence_samples * 2, 0.0f);
 457                         for (int i = 0; i < dropped_frames; ++i) {
 458                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 459                                 // Note that if the format changed in the meantime, we have
 460                                 // no way of detecting that; we just have to assume the frame length
 461                                 // is always the same.
 462                                 local_pts += frame_length;
 463                         }
 464                 }
 465                 if (num_samples == 0) {
 466                         audio.resize(silence_samples * 2);
 467                         num_samples = silence_samples;
 468                 }
 469                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 470                 card->next_local_pts = local_pts + frame_length;
 471         }
 472
 473         card->last_timecode = timecode;
 474
 475         // Done with the audio, so release it.
 476         if (audio_frame.owner) {
 477                 audio_frame.owner->release_frame(audio_frame);
 478         }
 479
 480         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 481         if (video_frame.len - video_offset == 0 ||
 482             video_frame.len - video_offset != expected_length) {
 483                 if (video_frame.len != 0) {
 484                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 485                                 card_index, video_frame.len - video_offset, expected_length);
 486                 }
 487                 if (video_frame.owner) {
 488                         video_frame.owner->release_frame(video_frame);
 489                 }
 490
 491                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 492                 // so that pts can go up accordingly.
 493                 {
 494                         unique_lock<mutex> lock(bmusb_mutex);
 495                         CaptureCard::NewFrame new_frame;
 496                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 497                         new_frame.length = frame_length;
 498                         new_frame.interlaced = false;
 499                         new_frame.dropped_frames = dropped_frames;
 500                         card->new_frames.push(move(new_frame));
 501                         card->new_frames_changed.notify_all();
 502                 }
 503                 return;
 504         }
 505
 506         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 507
 508         unsigned num_fields = video_format.interlaced ? 2 : 1;
 509         steady_clock::time_point frame_upload_start;
 510         if (video_format.interlaced) {
 511                 // Send the two fields along as separate frames; the other side will need to add
 512                 // a deinterlacer to actually get this right.
 513                 assert(video_format.height % 2 == 0);
 514                 video_format.height /= 2;
 515                 assert(frame_length % 2 == 0);
 516                 frame_length /= 2;
 517                 num_fields = 2;
 518                 frame_upload_start = steady_clock::now();
 519         }
 520         userdata->last_interlaced = video_format.interlaced;
 521         userdata->last_has_signal = video_format.has_signal;
 522         userdata->last_is_connected = video_format.is_connected;
 523         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 524         userdata->last_frame_rate_den = video_format.frame_rate_den;
 525         RefCountedFrame frame(video_frame);
 526
 527         // Upload the textures.
 528         size_t cbcr_width = video_format.width / 2;
 529         size_t cbcr_offset = video_offset / 2;
 530         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 531
 532         for (unsigned field = 0; field < num_fields; ++field) {
 533                 // Put the actual texture upload in a lambda that is executed in the main thread.
 534                 // It is entirely possible to do this in the same thread (and it might even be
 535                 // faster, depending on the GPU and driver), but it appears to be trickling
 536                 // driver bugs very easily.
 537                 //
 538                 // Note that this means we must hold on to the actual frame data in <userdata>
 539                 // until the upload command is run, but we hold on to <frame> much longer than that
 540                 // (in fact, all the way until we no longer use the texture in rendering).
 541                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
 542                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 543
 544                         if (userdata->tex_y[field] == 0 ||
 545                             userdata->tex_cbcr[field] == 0 ||
 546                             video_format.width != userdata->last_width[field] ||
 547                             video_format.height != userdata->last_height[field]) {
 548                                 // We changed resolution since last use of this texture, so we need to create
 549                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 550                                 // we don't need to worry about these flip-flopping between resolutions.
 551                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 552                                 check_error();
 553                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 554                                 check_error();
 555                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 556                                 check_error();
 557                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 558                                 check_error();
 559                                 userdata->last_width[field] = video_format.width;
 560                                 userdata->last_height[field] = video_format.height;
 561                         }
 562
 563                         GLuint pbo = userdata->pbo;
 564                         check_error();
 565                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 566                         check_error();
 567
 568                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 569                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 570
 571                         if (global_flags.flush_pbos) {
 572                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 573                                 check_error();
 574                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 575                                 check_error();
 576                         }
 577
 578                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 579                         check_error();
 580                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 581                         check_error();
 582                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 583                         check_error();
 584                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 585                         check_error();
 586                         glBindTexture(GL_TEXTURE_2D, 0);
 587                         check_error();
 588                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 589                         check_error();
 590                 };
 591
 592                 if (field == 1) {
 593                         // Don't upload the second field as fast as we can; wait until
 594                         // the field time has approximately passed. (Otherwise, we could
 595                         // get timing jitter against the other sources, and possibly also
 596                         // against the video display, although the latter is not as critical.)
 597                         // This requires our system clock to be reasonably close to the
 598                         // video clock, but that's not an unreasonable assumption.
 599                         steady_clock::time_point second_field_start = frame_upload_start +
 600                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 601                         this_thread::sleep_until(second_field_start);
 602                 }
 603
 604                 {
 605                         unique_lock<mutex> lock(bmusb_mutex);
 606                         CaptureCard::NewFrame new_frame;
 607                         new_frame.frame = frame;
 608                         new_frame.length = frame_length;
 609                         new_frame.field = field;
 610                         new_frame.interlaced = video_format.interlaced;
 611                         new_frame.upload_func = upload_func;
 612                         new_frame.dropped_frames = dropped_frames;
 613                         card->new_frames.push(move(new_frame));
 614                         card->new_frames_changed.notify_all();
 615                 }
 616         }
 617 }
 618
 619 void Mixer::bm_hotplug_add(libusb_device *dev)
 620 {
 621         lock_guard<mutex> lock(hotplug_mutex);
 622         hotplugged_cards.push_back(dev);
 623 }
 624
 625 void Mixer::bm_hotplug_remove(unsigned card_index)
 626 {
 627         cards[card_index].new_frames_changed.notify_all();
 628 }
 629
 630 void Mixer::thread_func()
 631 {
 632         eglBindAPI(EGL_OPENGL_API);
 633         QOpenGLContext *context = create_context(mixer_surface);
 634         if (!make_current(context, mixer_surface)) {
 635                 printf("oops\n");
 636                 exit(1);
 637         }
 638
 639         steady_clock::time_point start, now;
 640         start = steady_clock::now();
 641
 642         int frame = 0;
 643         int stats_dropped_frames = 0;
 644
 645         while (!should_quit) {
 646                 CaptureCard::NewFrame new_frames[MAX_CARDS];
 647                 bool has_new_frame[MAX_CARDS] = { false };
 648                 int num_samples[MAX_CARDS] = { 0 };
 649
 650                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 651                 assert(master_card_index < num_cards);
 652
 653                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 654                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 655                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 656                 send_audio_level_callback();
 657
 658                 handle_hotplugged_cards();
 659
 660                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 661                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 662                                 continue;
 663                         }
 664                         if (new_frames[card_index].frame->len == 0) {
 665                                 ++new_frames[card_index].dropped_frames;
 666                         }
 667                         if (new_frames[card_index].dropped_frames > 0) {
 668                                 printf("Card %u dropped %d frames before this\n",
 669                                         card_index, int(new_frames[card_index].dropped_frames));
 670                         }
 671                 }
 672
 673                 // If the first card is reporting a corrupted or otherwise dropped frame,
 674                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 675                 if (new_frames[master_card_index].frame->len == 0) {
 676                         ++stats_dropped_frames;
 677                         pts_int += new_frames[master_card_index].length;
 678                         continue;
 679                 }
 680
 681                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 682                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 683                                 continue;
 684
 685                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 686                         assert(new_frame->frame != nullptr);
 687                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 688                         check_error();
 689
 690                         // The new texture might need uploading before use.
 691                         if (new_frame->upload_func) {
 692                                 new_frame->upload_func();
 693                                 new_frame->upload_func = nullptr;
 694                         }
 695                 }
 696
 697                 int64_t frame_duration = new_frames[master_card_index].length;
 698                 render_one_frame(frame_duration);
 699                 ++frame;
 700                 pts_int += frame_duration;
 701
 702                 now = steady_clock::now();
 703                 double elapsed = duration<double>(now - start).count();
 704                 if (frame % 100 == 0) {
 705                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 706                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 707                                 1e3 * elapsed / frame);
 708                 //      chain->print_phase_timing();
 709
 710                         // Check our memory usage, to see if we are close to our mlockall()
 711                         // limit (if at all set).
 712                         rusage used;
 713                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 714                                 perror("getrusage(RUSAGE_SELF)");
 715                                 assert(false);
 716                         }
 717
 718                         if (uses_mlock) {
 719                                 rlimit limit;
 720                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 721                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 722                                         assert(false);
 723                                 }
 724
 725                                 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 726                                         long(used.ru_maxrss / 1024),
 727                                         long(limit.rlim_cur / 1048576),
 728                                         float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 729                         } else {
 730                                 printf(", using %ld MB memory (not locked)",
 731                                         long(used.ru_maxrss / 1024));
 732                         }
 733
 734                         printf("\n");
 735                 }
 736
 737
 738                 if (should_cut.exchange(false)) {  // Test and clear.
 739                         video_encoder->do_cut(frame);
 740                 }
 741
 742 #if 0
 743                 // Reset every 100 frames, so that local variations in frame times
 744                 // (especially for the first few frames, when the shaders are
 745                 // compiled etc.) don't make it hard to measure for the entire
 746                 // remaining duration of the program.
 747                 if (frame == 10000) {
 748                         frame = 0;
 749                         start = now;
 750                 }
 751 #endif
 752                 check_error();
 753         }
 754
 755         resource_pool->clean_context();
 756 }
 757
 758 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])
 759 {
 760 start:
 761         // The first card is the master timer, so wait for it to have a new frame.
 762         // TODO: Add a timeout.
 763         unique_lock<mutex> lock(bmusb_mutex);
 764         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(); });
 765
 766         if (cards[master_card_index].new_frames.empty()) {
 767                 // We were woken up, but not due to a new frame. Deal with it
 768                 // and then restart.
 769                 assert(cards[master_card_index].capture->get_disconnected());
 770                 handle_hotplugged_cards();
 771                 goto start;
 772         }
 773
 774         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 775                 CaptureCard *card = &cards[card_index];
 776                 if (card->new_frames.empty()) {
 777                         assert(card_index != master_card_index);
 778                         card->queue_length_policy.update_policy(-1);
 779                         continue;
 780                 }
 781                 new_frames[card_index] = move(card->new_frames.front());
 782                 has_new_frame[card_index] = true;
 783                 card->new_frames.pop();
 784                 card->new_frames_changed.notify_all();
 785
 786                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 787                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 788                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 789                 assert(num_samples[card_index] >= 0);
 790
 791                 if (card_index == master_card_index) {
 792                         // We don't use the queue length policy for the master card,
 793                         // but we will if it stops being the master. Thus, clear out
 794                         // the policy in case we switch in the future.
 795                         card->queue_length_policy.reset(card_index);
 796                 } else {
 797                         // If we have excess frames compared to the policy for this card,
 798                         // drop frames from the head.
 799                         card->queue_length_policy.update_policy(card->new_frames.size());
 800                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 801                                 card->new_frames.pop();
 802                         }
 803                 }
 804         }
 805 }
 806
 807 void Mixer::handle_hotplugged_cards()
 808 {
 809         // Check for cards that have been disconnected since last frame.
 810         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 811                 CaptureCard *card = &cards[card_index];
 812                 if (card->capture->get_disconnected()) {
 813                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 814                         FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 815                         configure_card(card_index, capture, /*is_fake_capture=*/true);
 816                         card->queue_length_policy.reset(card_index);
 817                         card->capture->start_bm_capture();
 818                 }
 819         }
 820
 821         // Check for cards that have been connected since last frame.
 822         vector<libusb_device *> hotplugged_cards_copy;
 823         {
 824                 lock_guard<mutex> lock(hotplug_mutex);
 825                 swap(hotplugged_cards, hotplugged_cards_copy);
 826         }
 827         for (libusb_device *new_dev : hotplugged_cards_copy) {
 828                 // Look for a fake capture card where we can stick this in.
 829                 int free_card_index = -1;
 830                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 831                         if (cards[card_index].is_fake_capture) {
 832                                 free_card_index = int(card_index);
 833                                 break;
 834                         }
 835                 }
 836
 837                 if (free_card_index == -1) {
 838                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 839                         libusb_unref_device(new_dev);
 840                 } else {
 841                         // BMUSBCapture takes ownership.
 842                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 843                         CaptureCard *card = &cards[free_card_index];
 844                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 845                         configure_card(free_card_index, capture, /*is_fake_capture=*/false);
 846                         card->queue_length_policy.reset(free_card_index);
 847                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 848                         capture->start_bm_capture();
 849                 }
 850         }
 851 }
 852
 853
 854 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 855 {
 856         // Resample the audio as needed, including from previously dropped frames.
 857         assert(num_cards > 0);
 858         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 859                 const bool dropped_frame = (frame_num != dropped_frames);
 860                 {
 861                         // Signal to the audio thread to process this frame.
 862                         // Note that if the frame is a dropped frame, we signal that
 863                         // we don't want to use this frame as base for adjusting
 864                         // the resampler rate. The reason for this is that the timing
 865                         // of these frames is often way too late; they typically don't
 866                         // “arrive” before we synthesize them. Thus, we could end up
 867                         // in a situation where we have inserted e.g. five audio frames
 868                         // into the queue before we then start pulling five of them
 869                         // back out. This makes ResamplingQueue overestimate the delay,
 870                         // causing undue resampler changes. (We _do_ use the last,
 871                         // non-dropped frame; perhaps we should just discard that as well,
 872                         // since dropped frames are expected to be rare, and it might be
 873                         // better to just wait until we have a slightly more normal situation).
 874                         unique_lock<mutex> lock(audio_mutex);
 875                         bool adjust_rate = !dropped_frame;
 876                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
 877                         audio_task_queue_changed.notify_one();
 878                 }
 879                 if (dropped_frame) {
 880                         // For dropped frames, increase the pts. Note that if the format changed
 881                         // in the meantime, we have no way of detecting that; we just have to
 882                         // assume the frame length is always the same.
 883                         pts_int += length_per_frame;
 884                 }
 885         }
 886 }
 887
 888 void Mixer::render_one_frame(int64_t duration)
 889 {
 890         // Get the main chain from the theme, and set its state immediately.
 891         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 892         EffectChain *chain = theme_main_chain.chain;
 893         theme_main_chain.setup_chain();
 894         //theme_main_chain.chain->enable_phase_timing(true);
 895
 896         GLuint y_tex, cbcr_tex;
 897         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 898         assert(got_frame);
 899
 900         // Render main chain.
 901         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 902         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 903         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 904         check_error();
 905         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 906         resource_pool->release_fbo(fbo);
 907
 908         subsample_chroma(cbcr_full_tex, cbcr_tex);
 909         resource_pool->release_2d_texture(cbcr_full_tex);
 910
 911         // Set the right state for rgba_tex.
 912         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 913         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 914         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 915         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 916         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 917
 918         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 919         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 920
 921         // The live frame just shows the RGBA texture we just rendered.
 922         // It owns rgba_tex now.
 923         DisplayFrame live_frame;
 924         live_frame.chain = display_chain.get();
 925         live_frame.setup_chain = [this, rgba_tex]{
 926                 display_input->set_texture_num(rgba_tex);
 927         };
 928         live_frame.ready_fence = fence;
 929         live_frame.input_frames = {};
 930         live_frame.temp_textures = { rgba_tex };
 931         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 932
 933         // Set up preview and any additional channels.
 934         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 935                 DisplayFrame display_frame;
 936                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 937                 display_frame.chain = chain.chain;
 938                 display_frame.setup_chain = chain.setup_chain;
 939                 display_frame.ready_fence = fence;
 940                 display_frame.input_frames = chain.input_frames;
 941                 display_frame.temp_textures = {};
 942                 output_channel[i].output_frame(display_frame);
 943         }
 944 }
 945
 946 void Mixer::send_audio_level_callback()
 947 {
 948         if (audio_level_callback == nullptr) {
 949                 return;
 950         }
 951
 952         unique_lock<mutex> lock(compressor_mutex);
 953         double loudness_s = r128.loudness_S();
 954         double loudness_i = r128.integrated();
 955         double loudness_range_low = r128.range_min();
 956         double loudness_range_high = r128.range_max();
 957
 958         audio_level_callback(loudness_s, 20.0 * log10(peak),
 959                 loudness_i, loudness_range_low, loudness_range_high,
 960                 gain_staging_db, 20.0 * log10(final_makeup_gain),
 961                 correlation.get_correlation());
 962 }
 963
 964 void Mixer::audio_thread_func()
 965 {
 966         while (!should_quit) {
 967                 AudioTask task;
 968
 969                 {
 970                         unique_lock<mutex> lock(audio_mutex);
 971                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 972                         if (should_quit) {
 973                                 return;
 974                         }
 975                         task = audio_task_queue.front();
 976                         audio_task_queue.pop();
 977                 }
 978
 979                 process_audio_one_frame(task.pts_int, task.num_samples, task.adjust_rate);
 980         }
 981 }
 982
 983 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples, bool adjust_rate)
 984 {
 985         vector<float> samples_card;
 986         vector<float> samples_out;
 987         samples_out.resize(num_samples * 2);
 988
 989         // TODO: Allow more flexible input mapping.
 990         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 991         assert(selected_audio_card < num_cards);
 992
 993         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 994                 samples_card.resize(num_samples * 2);
 995                 {
 996                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 997                         ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
 998                                 adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
 999                         cards[card_index].resampling_queue->get_output_samples(
1000                                 double(frame_pts_int) / TIMEBASE,
1001                                 &samples_card[0],
1002                                 num_samples,
1003                                 rate_adjustment_policy);
1004                 }
1005                 if (card_index == 0) {
1006                         for (int i = 0; i < num_samples * 2; ++i) {
1007                                 samples_out[i] = samples_card[i];
1008                         }
1009                 } else {
1010                         for (int i = 0; i < num_samples * 2; ++i) {
1011                                 samples_out[i] += samples_card[i];
1012                         }
1013                 }
1014         }
1015
1016         // Cut away everything under 120 Hz (or whatever the cutoff is);
1017         // we don't need it for voice, and it will reduce headroom
1018         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
1019         // should be dampened.)
1020         if (locut_enabled) {
1021                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
1022         }
1023
1024         // Apply a level compressor to get the general level right.
1025         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
1026         // (or more precisely, near it, since we don't use infinite ratio),
1027         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
1028         // entirely arbitrary, but from practical tests with speech, it seems to
1029         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
1030         {
1031                 unique_lock<mutex> lock(compressor_mutex);
1032                 if (level_compressor_enabled) {
1033                         float threshold = 0.01f;   // -40 dBFS.
1034                         float ratio = 20.0f;
1035                         float attack_time = 0.5f;
1036                         float release_time = 20.0f;
1037                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
1038                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1039                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
1040                 } else {
1041                         // Just apply the gain we already had.
1042                         float g = pow(10.0f, gain_staging_db / 20.0f);
1043                         for (size_t i = 0; i < samples_out.size(); ++i) {
1044                                 samples_out[i] *= g;
1045                         }
1046                 }
1047         }
1048
1049 #if 0
1050         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
1051                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
1052                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
1053                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
1054 #endif
1055
1056 //      float limiter_att, compressor_att;
1057
1058         // The real compressor.
1059         if (compressor_enabled) {
1060                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
1061                 float ratio = 20.0f;
1062                 float attack_time = 0.005f;
1063                 float release_time = 0.040f;
1064                 float makeup_gain = 2.0f;  // +6 dB.
1065                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1066 //              compressor_att = compressor.get_attenuation();
1067         }
1068
1069         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
1070         // Note that since ratio is not infinite, we could go slightly higher than this.
1071         if (limiter_enabled) {
1072                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
1073                 float ratio = 30.0f;
1074                 float attack_time = 0.0f;  // Instant.
1075                 float release_time = 0.020f;
1076                 float makeup_gain = 1.0f;  // 0 dB.
1077                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1078 //              limiter_att = limiter.get_attenuation();
1079         }
1080
1081 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
1082
1083         // At this point, we are most likely close to +0 LU, but all of our
1084         // measurements have been on raw sample values, not R128 values.
1085         // So we have a final makeup gain to get us to +0 LU; the gain
1086         // adjustments required should be relatively small, and also, the
1087         // offset shouldn't change much (only if the type of audio changes
1088         // significantly). Thus, we shoot for updating this value basically
1089         // “whenever we process buffers”, since the R128 calculation isn't exactly
1090         // something we get out per-sample.
1091         //
1092         // Note that there's a feedback loop here, so we choose a very slow filter
1093         // (half-time of 100 seconds).
1094         double target_loudness_factor, alpha;
1095         {
1096                 unique_lock<mutex> lock(compressor_mutex);
1097                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
1098                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
1099                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
1100
1101                 // If we're outside +/- 5 LU uncorrected, we don't count it as
1102                 // a normal signal (probably silence) and don't change the
1103                 // correction factor; just apply what we already have.
1104                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
1105                         alpha = 0.0;
1106                 } else {
1107                         // Formula adapted from
1108                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
1109                         const double half_time_s = 100.0;
1110                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
1111                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
1112                 }
1113
1114                 double m = final_makeup_gain;
1115                 for (size_t i = 0; i < samples_out.size(); i += 2) {
1116                         samples_out[i + 0] *= m;
1117                         samples_out[i + 1] *= m;
1118                         m += (target_loudness_factor - m) * alpha;
1119                 }
1120                 final_makeup_gain = m;
1121         }
1122
1123         // Upsample 4x to find interpolated peak.
1124         peak_resampler.inp_data = samples_out.data();
1125         peak_resampler.inp_count = samples_out.size() / 2;
1126
1127         vector<float> interpolated_samples_out;
1128         interpolated_samples_out.resize(samples_out.size());
1129         while (peak_resampler.inp_count > 0) {  // About four iterations.
1130                 peak_resampler.out_data = &interpolated_samples_out[0];
1131                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
1132                 peak_resampler.process();
1133                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
1134                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
1135                 peak_resampler.out_data = nullptr;
1136         }
1137
1138         // Find R128 levels and L/R correlation.
1139         vector<float> left, right;
1140         deinterleave_samples(samples_out, &left, &right);
1141         float *ptrs[] = { left.data(), right.data() };
1142         {
1143                 unique_lock<mutex> lock(compressor_mutex);
1144                 r128.process(left.size(), ptrs);
1145                 correlation.process_samples(samples_out);
1146         }
1147
1148         // Send the samples to the sound card.
1149         if (alsa) {
1150                 alsa->write(samples_out);
1151         }
1152
1153         // And finally add them to the output.
1154         video_encoder->add_audio(frame_pts_int, move(samples_out));
1155 }
1156
1157 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1158 {
1159         GLuint vao;
1160         glGenVertexArrays(1, &vao);
1161         check_error();
1162
1163         glBindVertexArray(vao);
1164         check_error();
1165
1166         // Extract Cb/Cr.
1167         GLuint fbo = resource_pool->create_fbo(dst_tex);
1168         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1169         glViewport(0, 0, WIDTH/2, HEIGHT/2);
1170         check_error();
1171
1172         glUseProgram(cbcr_program_num);
1173         check_error();
1174
1175         glActiveTexture(GL_TEXTURE0);
1176         check_error();
1177         glBindTexture(GL_TEXTURE_2D, src_tex);
1178         check_error();
1179         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1180         check_error();
1181         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1182         check_error();
1183         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1184         check_error();
1185
1186         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1187         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1188
1189         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1190         check_error();
1191
1192         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1193                 glEnableVertexAttribArray(attr_index);
1194                 check_error();
1195                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1196                 check_error();
1197         }
1198
1199         glDrawArrays(GL_TRIANGLES, 0, 3);
1200         check_error();
1201
1202         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1203                 glDisableVertexAttribArray(attr_index);
1204                 check_error();
1205         }
1206
1207         glUseProgram(0);
1208         check_error();
1209         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1210         check_error();
1211
1212         resource_pool->release_fbo(fbo);
1213         glDeleteVertexArrays(1, &vao);
1214 }
1215
1216 void Mixer::release_display_frame(DisplayFrame *frame)
1217 {
1218         for (GLuint texnum : frame->temp_textures) {
1219                 resource_pool->release_2d_texture(texnum);
1220         }
1221         frame->temp_textures.clear();
1222         frame->ready_fence.reset();
1223         frame->input_frames.clear();
1224 }
1225
1226 void Mixer::start()
1227 {
1228         mixer_thread = thread(&Mixer::thread_func, this);
1229         audio_thread = thread(&Mixer::audio_thread_func, this);
1230 }
1231
1232 void Mixer::quit()
1233 {
1234         should_quit = true;
1235         audio_task_queue_changed.notify_one();
1236         mixer_thread.join();
1237         audio_thread.join();
1238 }
1239
1240 void Mixer::transition_clicked(int transition_num)
1241 {
1242         theme->transition_clicked(transition_num, pts());
1243 }
1244
1245 void Mixer::channel_clicked(int preview_num)
1246 {
1247         theme->channel_clicked(preview_num);
1248 }
1249
1250 void Mixer::reset_meters()
1251 {
1252         peak_resampler.reset();
1253         peak = 0.0f;
1254         r128.reset();
1255         r128.integr_start();
1256         correlation.reset();
1257 }
1258
1259 void Mixer::start_mode_scanning(unsigned card_index)
1260 {
1261         assert(card_index < num_cards);
1262         if (is_mode_scanning[card_index]) {
1263                 return;
1264         }
1265         is_mode_scanning[card_index] = true;
1266         mode_scanlist[card_index].clear();
1267         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1268                 mode_scanlist[card_index].push_back(mode.first);
1269         }
1270         assert(!mode_scanlist[card_index].empty());
1271         mode_scanlist_index[card_index] = 0;
1272         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1273         last_mode_scan_change[card_index] = steady_clock::now();
1274 }
1275
1276 Mixer::OutputChannel::~OutputChannel()
1277 {
1278         if (has_current_frame) {
1279                 parent->release_display_frame(&current_frame);
1280         }
1281         if (has_ready_frame) {
1282                 parent->release_display_frame(&ready_frame);
1283         }
1284 }
1285
1286 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1287 {
1288         // Store this frame for display. Remove the ready frame if any
1289         // (it was seemingly never used).
1290         {
1291                 unique_lock<mutex> lock(frame_mutex);
1292                 if (has_ready_frame) {
1293                         parent->release_display_frame(&ready_frame);
1294                 }
1295                 ready_frame = frame;
1296                 has_ready_frame = true;
1297         }
1298
1299         if (new_frame_ready_callback) {
1300                 new_frame_ready_callback();
1301         }
1302
1303         // Reduce the number of callbacks by filtering duplicates. The reason
1304         // why we bother doing this is that Qt seemingly can get into a state
1305         // where its builds up an essentially unbounded queue of signals,
1306         // consuming more and more memory, and there's no good way of collapsing
1307         // user-defined signals or limiting the length of the queue.
1308         if (transition_names_updated_callback) {
1309                 vector<string> transition_names = global_mixer->get_transition_names();
1310                 bool changed = false;
1311                 if (transition_names.size() != last_transition_names.size()) {
1312                         changed = true;
1313                 } else {
1314                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1315                                 if (transition_names[i] != last_transition_names[i]) {
1316                                         changed = true;
1317                                         break;
1318                                 }
1319                         }
1320                 }
1321                 if (changed) {
1322                         transition_names_updated_callback(transition_names);
1323                         last_transition_names = transition_names;
1324                 }
1325         }
1326         if (name_updated_callback) {
1327                 string name = global_mixer->get_channel_name(channel);
1328                 if (name != last_name) {
1329                         name_updated_callback(name);
1330                         last_name = name;
1331                 }
1332         }
1333         if (color_updated_callback) {
1334                 string color = global_mixer->get_channel_color(channel);
1335                 if (color != last_color) {
1336                         color_updated_callback(color);
1337                         last_color = color;
1338                 }
1339         }
1340 }
1341
1342 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1343 {
1344         unique_lock<mutex> lock(frame_mutex);
1345         if (!has_current_frame && !has_ready_frame) {
1346                 return false;
1347         }
1348
1349         if (has_current_frame && has_ready_frame) {
1350                 // We have a new ready frame. Toss the current one.
1351                 parent->release_display_frame(&current_frame);
1352                 has_current_frame = false;
1353         }
1354         if (has_ready_frame) {
1355                 assert(!has_current_frame);
1356                 current_frame = ready_frame;
1357                 ready_frame.ready_fence.reset();  // Drop the refcount.
1358                 ready_frame.input_frames.clear();  // Drop the refcounts.
1359                 has_current_frame = true;
1360                 has_ready_frame = false;
1361         }
1362
1363         *frame = current_frame;
1364         return true;
1365 }
1366
1367 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1368 {
1369         new_frame_ready_callback = callback;
1370 }
1371
1372 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1373 {
1374         transition_names_updated_callback = callback;
1375 }
1376
1377 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1378 {
1379         name_updated_callback = callback;
1380 }
1381
1382 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1383 {
1384         color_updated_callback = callback;
1385 }
1386
1387 mutex RefCountedGLsync::fence_lock;