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 "flags.h"
  41 #include "video_encoder.h"
  42 #include "pbo_frame_allocator.h"
  43 #include "ref_counted_gl_sync.h"
  44 #include "timebase.h"
  45
  46 class QOpenGLContext;
  47
  48 using namespace movit;
  49 using namespace std;
  50 using namespace std::chrono;
  51 using namespace std::placeholders;
  52 using namespace bmusb;
  53
  54 Mixer *global_mixer = nullptr;
  55 bool uses_mlock = false;
  56
  57 namespace {
  58
  59 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  60 {
  61         assert(in_channels >= out_channels);
  62         for (size_t i = 0; i < num_samples; ++i) {
  63                 for (size_t j = 0; j < out_channels; ++j) {
  64                         uint32_t s1 = *src++;
  65                         uint32_t s2 = *src++;
  66                         uint32_t s3 = *src++;
  67                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  68                         dst[i * out_channels + j] = int(s) * (1.0f / 2147483648.0f);
  69                 }
  70                 src += 3 * (in_channels - out_channels);
  71         }
  72 }
  73
  74 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  75 {
  76         assert(in_channels >= out_channels);
  77         for (size_t i = 0; i < num_samples; ++i) {
  78                 for (size_t j = 0; j < out_channels; ++j) {
  79                         int32_t s = le32toh(*(int32_t *)src);
  80                         dst[i * out_channels + j] = s * (1.0f / 2147483648.0f);
  81                         src += 4;
  82                 }
  83                 src += 4 * (in_channels - out_channels);
  84         }
  85 }
  86
  87 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  88 {
  89         if (interlaced) {
  90                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  91                         input_state->buffered_frames[card_index][frame_num] =
  92                                 input_state->buffered_frames[card_index][frame_num - 1];
  93                 }
  94                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  95         } else {
  96                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  97                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  98                 }
  99         }
 100 }
 101
 102 }  // namespace
 103
 104 void QueueLengthPolicy::update_policy(int queue_length)
 105 {
 106         if (queue_length < 0) {  // Starvation.
 107                 if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
 108                         ++safe_queue_length;
 109                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
 110                                 card_index, safe_queue_length);
 111                 }
 112                 frames_with_at_least_one = 0;
 113                 been_at_safe_point_since_last_starvation = false;
 114                 return;
 115         }
 116         if (queue_length > 0) {
 117                 if (queue_length >= int(safe_queue_length)) {
 118                         been_at_safe_point_since_last_starvation = true;
 119                 }
 120                 if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
 121                         --safe_queue_length;
 122                         fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
 123                                 card_index, safe_queue_length);
 124                         frames_with_at_least_one = 0;
 125                 }
 126         } else {
 127                 frames_with_at_least_one = 0;
 128         }
 129 }
 130
 131 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 132         : httpd(),
 133           num_cards(num_cards),
 134           mixer_surface(create_surface(format)),
 135           h264_encoder_surface(create_surface(format)),
 136           correlation(OUTPUT_FREQUENCY),
 137           level_compressor(OUTPUT_FREQUENCY),
 138           limiter(OUTPUT_FREQUENCY),
 139           compressor(OUTPUT_FREQUENCY)
 140 {
 141         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 142         check_error();
 143
 144         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 145         // will be halved when sampling them, and we need to compensate here.
 146         movit_texel_subpixel_precision /= 2.0;
 147
 148         resource_pool.reset(new ResourcePool);
 149         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 150         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 151                 output_channel[i].parent = this;
 152                 output_channel[i].channel = i;
 153         }
 154
 155         ImageFormat inout_format;
 156         inout_format.color_space = COLORSPACE_sRGB;
 157         inout_format.gamma_curve = GAMMA_sRGB;
 158
 159         // Display chain; shows the live output produced by the main chain (its RGBA version).
 160         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 161         check_error();
 162         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 163         display_chain->add_input(display_input);
 164         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 165         display_chain->set_dither_bits(0);  // Don't bother.
 166         display_chain->finalize();
 167
 168         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 169
 170         // Start listening for clients only once VideoEncoder has written its header, if any.
 171         httpd.start(9095);
 172
 173         // First try initializing the then PCI devices, then USB, then
 174         // fill up with fake cards until we have the desired number of cards.
 175         unsigned num_pci_devices = 0;
 176         unsigned card_index = 0;
 177
 178         {
 179                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 180                 if (decklink_iterator != nullptr) {
 181                         for ( ; card_index < num_cards; ++card_index) {
 182                                 IDeckLink *decklink;
 183                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 184                                         break;
 185                                 }
 186
 187                                 configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
 188                                 ++num_pci_devices;
 189                         }
 190                         decklink_iterator->Release();
 191                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 192                 } else {
 193                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 194                 }
 195         }
 196         unsigned num_usb_devices = BMUSBCapture::num_cards();
 197         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 198                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 199                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 200                 configure_card(card_index, capture, /*is_fake_capture=*/false);
 201         }
 202         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 203
 204         unsigned num_fake_cards = 0;
 205         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 206                 FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 207                 configure_card(card_index, capture, /*is_fake_capture=*/true);
 208         }
 209
 210         if (num_fake_cards > 0) {
 211                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 212         }
 213
 214         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 215         BMUSBCapture::start_bm_thread();
 216
 217         for (card_index = 0; card_index < num_cards; ++card_index) {
 218                 cards[card_index].queue_length_policy.reset(card_index);
 219                 cards[card_index].capture->start_bm_capture();
 220         }
 221
 222         // Set up stuff for NV12 conversion.
 223
 224         // Cb/Cr shader.
 225         string cbcr_vert_shader =
 226                 "#version 130 \n"
 227                 " \n"
 228                 "in vec2 position; \n"
 229                 "in vec2 texcoord; \n"
 230                 "out vec2 tc0; \n"
 231                 "uniform vec2 foo_chroma_offset_0; \n"
 232                 " \n"
 233                 "void main() \n"
 234                 "{ \n"
 235                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 236                 "    // \n"
 237                 "    //   2.000  0.000  0.000 -1.000 \n"
 238                 "    //   0.000  2.000  0.000 -1.000 \n"
 239                 "    //   0.000  0.000 -2.000 -1.000 \n"
 240                 "    //   0.000  0.000  0.000  1.000 \n"
 241                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 242                 "    vec2 flipped_tc = texcoord; \n"
 243                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 244                 "} \n";
 245         string cbcr_frag_shader =
 246                 "#version 130 \n"
 247                 "in vec2 tc0; \n"
 248                 "uniform sampler2D cbcr_tex; \n"
 249                 "out vec4 FragColor; \n"
 250                 "void main() { \n"
 251                 "    FragColor = texture(cbcr_tex, tc0); \n"
 252                 "} \n";
 253         vector<string> frag_shader_outputs;
 254         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 255
 256         float vertices[] = {
 257                 0.0f, 2.0f,
 258                 0.0f, 0.0f,
 259                 2.0f, 0.0f
 260         };
 261         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 262         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 263         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 264
 265         r128.init(2, OUTPUT_FREQUENCY);
 266         r128.integr_start();
 267
 268         locut.init(FILTER_HPF, 2);
 269
 270         set_locut_enabled(global_flags.locut_enabled);
 271         set_gain_staging_db(global_flags.initial_gain_staging_db);
 272         set_gain_staging_auto(global_flags.gain_staging_auto);
 273         set_compressor_enabled(global_flags.compressor_enabled);
 274         set_limiter_enabled(global_flags.limiter_enabled);
 275         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 276
 277         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 278         // and there's a limit to how important the peak meter is.
 279         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 280
 281         if (global_flags.enable_alsa_output) {
 282                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 283         }
 284 }
 285
 286 Mixer::~Mixer()
 287 {
 288         resource_pool->release_glsl_program(cbcr_program_num);
 289         glDeleteBuffers(1, &cbcr_vbo);
 290         BMUSBCapture::stop_bm_thread();
 291
 292         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 293                 {
 294                         unique_lock<mutex> lock(bmusb_mutex);
 295                         cards[card_index].should_quit = true;  // Unblock thread.
 296                         cards[card_index].new_frames_changed.notify_all();
 297                 }
 298                 cards[card_index].capture->stop_dequeue_thread();
 299         }
 300
 301         video_encoder.reset(nullptr);
 302 }
 303
 304 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
 305 {
 306         printf("Configuring card %d...\n", card_index);
 307
 308         CaptureCard *card = &cards[card_index];
 309         if (card->capture != nullptr) {
 310                 card->capture->stop_dequeue_thread();
 311                 delete card->capture;
 312         }
 313         card->capture = capture;
 314         card->is_fake_capture = is_fake_capture;
 315         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 316         if (card->frame_allocator == nullptr) {
 317                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 318         }
 319         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 320         if (card->surface == nullptr) {
 321                 card->surface = create_surface_with_same_format(mixer_surface);
 322         }
 323         {
 324                 unique_lock<mutex> lock(cards[card_index].audio_mutex);
 325                 card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 326         }
 327         while (!card->new_frames.empty()) card->new_frames.pop();
 328         card->fractional_samples = 0;
 329         card->last_timecode = -1;
 330         card->next_local_pts = 0;
 331         card->capture->configure_card();
 332 }
 333
 334
 335 namespace {
 336
 337 int unwrap_timecode(uint16_t current_wrapped, int last)
 338 {
 339         uint16_t last_wrapped = last & 0xffff;
 340         if (current_wrapped > last_wrapped) {
 341                 return (last & ~0xffff) | current_wrapped;
 342         } else {
 343                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 344         }
 345 }
 346
 347 float find_peak(const float *samples, size_t num_samples)
 348 {
 349         float m = fabs(samples[0]);
 350         for (size_t i = 1; i < num_samples; ++i) {
 351                 m = max(m, fabs(samples[i]));
 352         }
 353         return m;
 354 }
 355
 356 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 357 {
 358         size_t num_samples = in.size() / 2;
 359         out_l->resize(num_samples);
 360         out_r->resize(num_samples);
 361
 362         const float *inptr = in.data();
 363         float *lptr = &(*out_l)[0];
 364         float *rptr = &(*out_r)[0];
 365         for (size_t i = 0; i < num_samples; ++i) {
 366                 *lptr++ = *inptr++;
 367                 *rptr++ = *inptr++;
 368         }
 369 }
 370
 371 }  // namespace
 372
 373 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 374                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 375                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 376 {
 377         CaptureCard *card = &cards[card_index];
 378
 379         if (is_mode_scanning[card_index]) {
 380                 if (video_format.has_signal) {
 381                         // Found a stable signal, so stop scanning.
 382                         is_mode_scanning[card_index] = false;
 383                 } else {
 384                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 385                         steady_clock::time_point now = steady_clock::now();
 386                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 387                         if (sec_since_last_switch > switch_time_s) {
 388                                 // It isn't this mode; try the next one.
 389                                 mode_scanlist_index[card_index]++;
 390                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 391                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 392                                 last_mode_scan_change[card_index] = now;
 393                         }
 394                 }
 395         }
 396
 397         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 398         assert(frame_length > 0);
 399
 400         size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0;
 401         if (num_samples > OUTPUT_FREQUENCY / 10) {
 402                 printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
 403                         card_index, int(audio_frame.len), int(audio_offset),
 404                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 405                 if (video_frame.owner) {
 406                         video_frame.owner->release_frame(video_frame);
 407                 }
 408                 if (audio_frame.owner) {
 409                         audio_frame.owner->release_frame(audio_frame);
 410                 }
 411                 return;
 412         }
 413
 414         int64_t local_pts = card->next_local_pts;
 415         int dropped_frames = 0;
 416         if (card->last_timecode != -1) {
 417                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 418         }
 419
 420         // Convert the audio to stereo fp32 and add it.
 421         vector<float> audio;
 422         audio.resize(num_samples * 2);
 423         switch (audio_format.bits_per_sample) {
 424         case 0:
 425                 assert(num_samples == 0);
 426                 break;
 427         case 24:
 428                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 429                 break;
 430         case 32:
 431                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 432                 break;
 433         default:
 434                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 435                 assert(false);
 436         }
 437
 438         // Add the audio.
 439         {
 440                 unique_lock<mutex> lock(card->audio_mutex);
 441
 442                 // Number of samples per frame if we need to insert silence.
 443                 // (Could be nonintegral, but resampling will save us then.)
 444                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 445
 446                 if (dropped_frames > MAX_FPS * 2) {
 447                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 448                                 card_index, card->last_timecode, timecode);
 449                         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 450                         dropped_frames = 0;
 451                 } else if (dropped_frames > 0) {
 452                         // Insert silence as needed.
 453                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 454                                 card_index, dropped_frames, timecode);
 455                         vector<float> silence(silence_samples * 2, 0.0f);
 456                         for (int i = 0; i < dropped_frames; ++i) {
 457                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 458                                 // Note that if the format changed in the meantime, we have
 459                                 // no way of detecting that; we just have to assume the frame length
 460                                 // is always the same.
 461                                 local_pts += frame_length;
 462                         }
 463                 }
 464                 if (num_samples == 0) {
 465                         audio.resize(silence_samples * 2);
 466                         num_samples = silence_samples;
 467                 }
 468                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 469                 card->next_local_pts = local_pts + frame_length;
 470         }
 471
 472         card->last_timecode = timecode;
 473
 474         // Done with the audio, so release it.
 475         if (audio_frame.owner) {
 476                 audio_frame.owner->release_frame(audio_frame);
 477         }
 478
 479         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 480         if (video_frame.len - video_offset == 0 ||
 481             video_frame.len - video_offset != expected_length) {
 482                 if (video_frame.len != 0) {
 483                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 484                                 card_index, video_frame.len - video_offset, expected_length);
 485                 }
 486                 if (video_frame.owner) {
 487                         video_frame.owner->release_frame(video_frame);
 488                 }
 489
 490                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 491                 // so that pts can go up accordingly.
 492                 {
 493                         unique_lock<mutex> lock(bmusb_mutex);
 494                         CaptureCard::NewFrame new_frame;
 495                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 496                         new_frame.length = frame_length;
 497                         new_frame.interlaced = false;
 498                         new_frame.dropped_frames = dropped_frames;
 499                         card->new_frames.push(move(new_frame));
 500                         card->new_frames_changed.notify_all();
 501                 }
 502                 return;
 503         }
 504
 505         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 506
 507         unsigned num_fields = video_format.interlaced ? 2 : 1;
 508         steady_clock::time_point frame_upload_start;
 509         if (video_format.interlaced) {
 510                 // Send the two fields along as separate frames; the other side will need to add
 511                 // a deinterlacer to actually get this right.
 512                 assert(video_format.height % 2 == 0);
 513                 video_format.height /= 2;
 514                 assert(frame_length % 2 == 0);
 515                 frame_length /= 2;
 516                 num_fields = 2;
 517                 frame_upload_start = steady_clock::now();
 518         }
 519         userdata->last_interlaced = video_format.interlaced;
 520         userdata->last_has_signal = video_format.has_signal;
 521         userdata->last_is_connected = video_format.is_connected;
 522         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 523         userdata->last_frame_rate_den = video_format.frame_rate_den;
 524         RefCountedFrame frame(video_frame);
 525
 526         // Upload the textures.
 527         size_t cbcr_width = video_format.width / 2;
 528         size_t cbcr_offset = video_offset / 2;
 529         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 530
 531         for (unsigned field = 0; field < num_fields; ++field) {
 532                 // Put the actual texture upload in a lambda that is executed in the main thread.
 533                 // It is entirely possible to do this in the same thread (and it might even be
 534                 // faster, depending on the GPU and driver), but it appears to be trickling
 535                 // driver bugs very easily.
 536                 //
 537                 // Note that this means we must hold on to the actual frame data in <userdata>
 538                 // until the upload command is run, but we hold on to <frame> much longer than that
 539                 // (in fact, all the way until we no longer use the texture in rendering).
 540                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
 541                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 542
 543                         if (userdata->tex_y[field] == 0 ||
 544                             userdata->tex_cbcr[field] == 0 ||
 545                             video_format.width != userdata->last_width[field] ||
 546                             video_format.height != userdata->last_height[field]) {
 547                                 // We changed resolution since last use of this texture, so we need to create
 548                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 549                                 // we don't need to worry about these flip-flopping between resolutions.
 550                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 551                                 check_error();
 552                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 553                                 check_error();
 554                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 555                                 check_error();
 556                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 557                                 check_error();
 558                                 userdata->last_width[field] = video_format.width;
 559                                 userdata->last_height[field] = video_format.height;
 560                         }
 561
 562                         GLuint pbo = userdata->pbo;
 563                         check_error();
 564                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 565                         check_error();
 566
 567                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 568                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 569
 570                         if (global_flags.flush_pbos) {
 571                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 572                                 check_error();
 573                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 574                                 check_error();
 575                         }
 576
 577                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 578                         check_error();
 579                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 580                         check_error();
 581                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 582                         check_error();
 583                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 584                         check_error();
 585                         glBindTexture(GL_TEXTURE_2D, 0);
 586                         check_error();
 587                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 588                         check_error();
 589                 };
 590
 591                 if (field == 1) {
 592                         // Don't upload the second field as fast as we can; wait until
 593                         // the field time has approximately passed. (Otherwise, we could
 594                         // get timing jitter against the other sources, and possibly also
 595                         // against the video display, although the latter is not as critical.)
 596                         // This requires our system clock to be reasonably close to the
 597                         // video clock, but that's not an unreasonable assumption.
 598                         steady_clock::time_point second_field_start = frame_upload_start +
 599                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 600                         this_thread::sleep_until(second_field_start);
 601                 }
 602
 603                 {
 604                         unique_lock<mutex> lock(bmusb_mutex);
 605                         CaptureCard::NewFrame new_frame;
 606                         new_frame.frame = frame;
 607                         new_frame.length = frame_length;
 608                         new_frame.field = field;
 609                         new_frame.interlaced = video_format.interlaced;
 610                         new_frame.upload_func = upload_func;
 611                         new_frame.dropped_frames = dropped_frames;
 612                         card->new_frames.push(move(new_frame));
 613                         card->new_frames_changed.notify_all();
 614                 }
 615         }
 616 }
 617
 618 void Mixer::bm_hotplug_add(libusb_device *dev)
 619 {
 620         lock_guard<mutex> lock(hotplug_mutex);
 621         hotplugged_cards.push_back(dev);
 622 }
 623
 624 void Mixer::bm_hotplug_remove(unsigned card_index)
 625 {
 626         cards[card_index].new_frames_changed.notify_all();
 627 }
 628
 629 void Mixer::thread_func()
 630 {
 631         eglBindAPI(EGL_OPENGL_API);
 632         QOpenGLContext *context = create_context(mixer_surface);
 633         if (!make_current(context, mixer_surface)) {
 634                 printf("oops\n");
 635                 exit(1);
 636         }
 637
 638         steady_clock::time_point start, now;
 639         start = steady_clock::now();
 640
 641         int frame = 0;
 642         int stats_dropped_frames = 0;
 643
 644         while (!should_quit) {
 645                 CaptureCard::NewFrame new_frames[MAX_CARDS];
 646                 bool has_new_frame[MAX_CARDS] = { false };
 647                 int num_samples[MAX_CARDS] = { 0 };
 648
 649                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 650                 assert(master_card_index < num_cards);
 651
 652                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 653                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 654                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 655                 send_audio_level_callback();
 656
 657                 handle_hotplugged_cards();
 658
 659                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 660                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 661                                 continue;
 662                         }
 663                         if (new_frames[card_index].frame->len == 0) {
 664                                 ++new_frames[card_index].dropped_frames;
 665                         }
 666                         if (new_frames[card_index].dropped_frames > 0) {
 667                                 printf("Card %u dropped %d frames before this\n",
 668                                         card_index, int(new_frames[card_index].dropped_frames));
 669                         }
 670                 }
 671
 672                 // If the first card is reporting a corrupted or otherwise dropped frame,
 673                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 674                 if (new_frames[master_card_index].frame->len == 0) {
 675                         ++stats_dropped_frames;
 676                         pts_int += new_frames[master_card_index].length;
 677                         continue;
 678                 }
 679
 680                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 681                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 682                                 continue;
 683
 684                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 685                         assert(new_frame->frame != nullptr);
 686                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 687                         check_error();
 688
 689                         // The new texture might need uploading before use.
 690                         if (new_frame->upload_func) {
 691                                 new_frame->upload_func();
 692                                 new_frame->upload_func = nullptr;
 693                         }
 694                 }
 695
 696                 int64_t frame_duration = new_frames[master_card_index].length;
 697                 render_one_frame(frame_duration);
 698                 ++frame;
 699                 pts_int += frame_duration;
 700
 701                 now = steady_clock::now();
 702                 double elapsed = duration<double>(now - start).count();
 703                 if (frame % 100 == 0) {
 704                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 705                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 706                                 1e3 * elapsed / frame);
 707                 //      chain->print_phase_timing();
 708
 709                         // Check our memory usage, to see if we are close to our mlockall()
 710                         // limit (if at all set).
 711                         rusage used;
 712                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 713                                 perror("getrusage(RUSAGE_SELF)");
 714                                 assert(false);
 715                         }
 716
 717                         if (uses_mlock) {
 718                                 rlimit limit;
 719                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 720                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 721                                         assert(false);
 722                                 }
 723
 724                                 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 725                                         long(used.ru_maxrss / 1024),
 726                                         long(limit.rlim_cur / 1048576),
 727                                         float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 728                         } else {
 729                                 printf(", using %ld MB memory (not locked)",
 730                                         long(used.ru_maxrss / 1024));
 731                         }
 732
 733                         printf("\n");
 734                 }
 735
 736
 737                 if (should_cut.exchange(false)) {  // Test and clear.
 738                         video_encoder->do_cut(frame);
 739                 }
 740
 741 #if 0
 742                 // Reset every 100 frames, so that local variations in frame times
 743                 // (especially for the first few frames, when the shaders are
 744                 // compiled etc.) don't make it hard to measure for the entire
 745                 // remaining duration of the program.
 746                 if (frame == 10000) {
 747                         frame = 0;
 748                         start = now;
 749                 }
 750 #endif
 751                 check_error();
 752         }
 753
 754         resource_pool->clean_context();
 755 }
 756
 757 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])
 758 {
 759 start:
 760         // The first card is the master timer, so wait for it to have a new frame.
 761         // TODO: Add a timeout.
 762         unique_lock<mutex> lock(bmusb_mutex);
 763         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(); });
 764
 765         if (cards[master_card_index].new_frames.empty()) {
 766                 // We were woken up, but not due to a new frame. Deal with it
 767                 // and then restart.
 768                 assert(cards[master_card_index].capture->get_disconnected());
 769                 handle_hotplugged_cards();
 770                 goto start;
 771         }
 772
 773         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 774                 CaptureCard *card = &cards[card_index];
 775                 if (card->new_frames.empty()) {
 776                         assert(card_index != master_card_index);
 777                         card->queue_length_policy.update_policy(-1);
 778                         continue;
 779                 }
 780                 new_frames[card_index] = move(card->new_frames.front());
 781                 has_new_frame[card_index] = true;
 782                 card->new_frames.pop();
 783                 card->new_frames_changed.notify_all();
 784
 785                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 786                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 787                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 788                 assert(num_samples[card_index] >= 0);
 789
 790                 if (card_index == master_card_index) {
 791                         // We don't use the queue length policy for the master card,
 792                         // but we will if it stops being the master. Thus, clear out
 793                         // the policy in case we switch in the future.
 794                         card->queue_length_policy.reset(card_index);
 795                 } else {
 796                         // If we have excess frames compared to the policy for this card,
 797                         // drop frames from the head.
 798                         card->queue_length_policy.update_policy(card->new_frames.size());
 799                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 800                                 card->new_frames.pop();
 801                         }
 802                 }
 803         }
 804 }
 805
 806 void Mixer::handle_hotplugged_cards()
 807 {
 808         // Check for cards that have been disconnected since last frame.
 809         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 810                 CaptureCard *card = &cards[card_index];
 811                 if (card->capture->get_disconnected()) {
 812                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 813                         FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 814                         configure_card(card_index, capture, /*is_fake_capture=*/true);
 815                         card->queue_length_policy.reset(card_index);
 816                         card->capture->start_bm_capture();
 817                 }
 818         }
 819
 820         // Check for cards that have been connected since last frame.
 821         vector<libusb_device *> hotplugged_cards_copy;
 822         {
 823                 lock_guard<mutex> lock(hotplug_mutex);
 824                 swap(hotplugged_cards, hotplugged_cards_copy);
 825         }
 826         for (libusb_device *new_dev : hotplugged_cards_copy) {
 827                 // Look for a fake capture card where we can stick this in.
 828                 int free_card_index = -1;
 829                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 830                         if (cards[card_index].is_fake_capture) {
 831                                 free_card_index = int(card_index);
 832                                 break;
 833                         }
 834                 }
 835
 836                 if (free_card_index == -1) {
 837                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 838                         libusb_unref_device(new_dev);
 839                 } else {
 840                         // BMUSBCapture takes ownership.
 841                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 842                         CaptureCard *card = &cards[free_card_index];
 843                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 844                         configure_card(free_card_index, capture, /*is_fake_capture=*/false);
 845                         card->queue_length_policy.reset(free_card_index);
 846                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 847                         capture->start_bm_capture();
 848                 }
 849         }
 850 }
 851
 852
 853 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 854 {
 855         // Resample the audio as needed, including from previously dropped frames.
 856         assert(num_cards > 0);
 857         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 858                 const bool dropped_frame = (frame_num != dropped_frames);
 859                 {
 860                         // Signal to the audio thread to process this frame.
 861                         // Note that if the frame is a dropped frame, we signal that
 862                         // we don't want to use this frame as base for adjusting
 863                         // the resampler rate. The reason for this is that the timing
 864                         // of these frames is often way too late; they typically don't
 865                         // “arrive” before we synthesize them. Thus, we could end up
 866                         // in a situation where we have inserted e.g. five audio frames
 867                         // into the queue before we then start pulling five of them
 868                         // back out. This makes ResamplingQueue overestimate the delay,
 869                         // causing undue resampler changes. (We _do_ use the last,
 870                         // non-dropped frame; perhaps we should just discard that as well,
 871                         // since dropped frames are expected to be rare, and it might be
 872                         // better to just wait until we have a slightly more normal situation).
 873                         unique_lock<mutex> lock(audio_mutex);
 874                         bool adjust_rate = !dropped_frame;
 875                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
 876                         audio_task_queue_changed.notify_one();
 877                 }
 878                 if (dropped_frame) {
 879                         // For dropped frames, increase the pts. Note that if the format changed
 880                         // in the meantime, we have no way of detecting that; we just have to
 881                         // assume the frame length is always the same.
 882                         pts_int += length_per_frame;
 883                 }
 884         }
 885 }
 886
 887 void Mixer::render_one_frame(int64_t duration)
 888 {
 889         // Get the main chain from the theme, and set its state immediately.
 890         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 891         EffectChain *chain = theme_main_chain.chain;
 892         theme_main_chain.setup_chain();
 893         //theme_main_chain.chain->enable_phase_timing(true);
 894
 895         GLuint y_tex, cbcr_tex;
 896         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 897         assert(got_frame);
 898
 899         // Render main chain.
 900         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 901         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 902         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 903         check_error();
 904         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 905         resource_pool->release_fbo(fbo);
 906
 907         subsample_chroma(cbcr_full_tex, cbcr_tex);
 908         resource_pool->release_2d_texture(cbcr_full_tex);
 909
 910         // Set the right state for rgba_tex.
 911         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 912         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 913         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 914         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 915         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 916
 917         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 918         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 919
 920         // The live frame just shows the RGBA texture we just rendered.
 921         // It owns rgba_tex now.
 922         DisplayFrame live_frame;
 923         live_frame.chain = display_chain.get();
 924         live_frame.setup_chain = [this, rgba_tex]{
 925                 display_input->set_texture_num(rgba_tex);
 926         };
 927         live_frame.ready_fence = fence;
 928         live_frame.input_frames = {};
 929         live_frame.temp_textures = { rgba_tex };
 930         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 931
 932         // Set up preview and any additional channels.
 933         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 934                 DisplayFrame display_frame;
 935                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 936                 display_frame.chain = chain.chain;
 937                 display_frame.setup_chain = chain.setup_chain;
 938                 display_frame.ready_fence = fence;
 939                 display_frame.input_frames = chain.input_frames;
 940                 display_frame.temp_textures = {};
 941                 output_channel[i].output_frame(display_frame);
 942         }
 943 }
 944
 945 void Mixer::send_audio_level_callback()
 946 {
 947         if (audio_level_callback == nullptr) {
 948                 return;
 949         }
 950
 951         unique_lock<mutex> lock(compressor_mutex);
 952         double loudness_s = r128.loudness_S();
 953         double loudness_i = r128.integrated();
 954         double loudness_range_low = r128.range_min();
 955         double loudness_range_high = r128.range_max();
 956
 957         audio_level_callback(loudness_s, 20.0 * log10(peak),
 958                 loudness_i, loudness_range_low, loudness_range_high,
 959                 gain_staging_db, 20.0 * log10(final_makeup_gain),
 960                 correlation.get_correlation());
 961 }
 962
 963 void Mixer::audio_thread_func()
 964 {
 965         while (!should_quit) {
 966                 AudioTask task;
 967
 968                 {
 969                         unique_lock<mutex> lock(audio_mutex);
 970                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 971                         if (should_quit) {
 972                                 return;
 973                         }
 974                         task = audio_task_queue.front();
 975                         audio_task_queue.pop();
 976                 }
 977
 978                 process_audio_one_frame(task.pts_int, task.num_samples, task.adjust_rate);
 979         }
 980 }
 981
 982 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples, bool adjust_rate)
 983 {
 984         vector<float> samples_card;
 985         vector<float> samples_out;
 986
 987         // TODO: Allow mixing audio from several sources.
 988         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 989         assert(selected_audio_card < num_cards);
 990
 991         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 992                 samples_card.resize(num_samples * 2);
 993                 {
 994                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 995                         ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
 996                                 adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
 997                         cards[card_index].resampling_queue->get_output_samples(
 998                                 double(frame_pts_int) / TIMEBASE,
 999                                 &samples_card[0],
1000                                 num_samples,
1001                                 rate_adjustment_policy);
1002                 }
1003                 if (card_index == selected_audio_card) {
1004                         samples_out = move(samples_card);
1005                 }
1006         }
1007
1008         // Cut away everything under 120 Hz (or whatever the cutoff is);
1009         // we don't need it for voice, and it will reduce headroom
1010         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
1011         // should be dampened.)
1012         if (locut_enabled) {
1013                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
1014         }
1015
1016         // Apply a level compressor to get the general level right.
1017         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
1018         // (or more precisely, near it, since we don't use infinite ratio),
1019         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
1020         // entirely arbitrary, but from practical tests with speech, it seems to
1021         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
1022         {
1023                 unique_lock<mutex> lock(compressor_mutex);
1024                 if (level_compressor_enabled) {
1025                         float threshold = 0.01f;   // -40 dBFS.
1026                         float ratio = 20.0f;
1027                         float attack_time = 0.5f;
1028                         float release_time = 20.0f;
1029                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
1030                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1031                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
1032                 } else {
1033                         // Just apply the gain we already had.
1034                         float g = pow(10.0f, gain_staging_db / 20.0f);
1035                         for (size_t i = 0; i < samples_out.size(); ++i) {
1036                                 samples_out[i] *= g;
1037                         }
1038                 }
1039         }
1040
1041 #if 0
1042         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
1043                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
1044                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
1045                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
1046 #endif
1047
1048 //      float limiter_att, compressor_att;
1049
1050         // The real compressor.
1051         if (compressor_enabled) {
1052                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
1053                 float ratio = 20.0f;
1054                 float attack_time = 0.005f;
1055                 float release_time = 0.040f;
1056                 float makeup_gain = 2.0f;  // +6 dB.
1057                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1058 //              compressor_att = compressor.get_attenuation();
1059         }
1060
1061         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
1062         // Note that since ratio is not infinite, we could go slightly higher than this.
1063         if (limiter_enabled) {
1064                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
1065                 float ratio = 30.0f;
1066                 float attack_time = 0.0f;  // Instant.
1067                 float release_time = 0.020f;
1068                 float makeup_gain = 1.0f;  // 0 dB.
1069                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1070 //              limiter_att = limiter.get_attenuation();
1071         }
1072
1073 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
1074
1075         // At this point, we are most likely close to +0 LU, but all of our
1076         // measurements have been on raw sample values, not R128 values.
1077         // So we have a final makeup gain to get us to +0 LU; the gain
1078         // adjustments required should be relatively small, and also, the
1079         // offset shouldn't change much (only if the type of audio changes
1080         // significantly). Thus, we shoot for updating this value basically
1081         // “whenever we process buffers”, since the R128 calculation isn't exactly
1082         // something we get out per-sample.
1083         //
1084         // Note that there's a feedback loop here, so we choose a very slow filter
1085         // (half-time of 100 seconds).
1086         double target_loudness_factor, alpha;
1087         {
1088                 unique_lock<mutex> lock(compressor_mutex);
1089                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
1090                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
1091                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
1092
1093                 // If we're outside +/- 5 LU uncorrected, we don't count it as
1094                 // a normal signal (probably silence) and don't change the
1095                 // correction factor; just apply what we already have.
1096                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
1097                         alpha = 0.0;
1098                 } else {
1099                         // Formula adapted from
1100                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
1101                         const double half_time_s = 100.0;
1102                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
1103                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
1104                 }
1105
1106                 double m = final_makeup_gain;
1107                 for (size_t i = 0; i < samples_out.size(); i += 2) {
1108                         samples_out[i + 0] *= m;
1109                         samples_out[i + 1] *= m;
1110                         m += (target_loudness_factor - m) * alpha;
1111                 }
1112                 final_makeup_gain = m;
1113         }
1114
1115         // Upsample 4x to find interpolated peak.
1116         peak_resampler.inp_data = samples_out.data();
1117         peak_resampler.inp_count = samples_out.size() / 2;
1118
1119         vector<float> interpolated_samples_out;
1120         interpolated_samples_out.resize(samples_out.size());
1121         while (peak_resampler.inp_count > 0) {  // About four iterations.
1122                 peak_resampler.out_data = &interpolated_samples_out[0];
1123                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
1124                 peak_resampler.process();
1125                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
1126                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
1127                 peak_resampler.out_data = nullptr;
1128         }
1129
1130         // Find R128 levels and L/R correlation.
1131         vector<float> left, right;
1132         deinterleave_samples(samples_out, &left, &right);
1133         float *ptrs[] = { left.data(), right.data() };
1134         {
1135                 unique_lock<mutex> lock(compressor_mutex);
1136                 r128.process(left.size(), ptrs);
1137                 correlation.process_samples(samples_out);
1138         }
1139
1140         // Send the samples to the sound card.
1141         if (alsa) {
1142                 alsa->write(samples_out);
1143         }
1144
1145         // And finally add them to the output.
1146         video_encoder->add_audio(frame_pts_int, move(samples_out));
1147 }
1148
1149 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1150 {
1151         GLuint vao;
1152         glGenVertexArrays(1, &vao);
1153         check_error();
1154
1155         glBindVertexArray(vao);
1156         check_error();
1157
1158         // Extract Cb/Cr.
1159         GLuint fbo = resource_pool->create_fbo(dst_tex);
1160         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1161         glViewport(0, 0, WIDTH/2, HEIGHT/2);
1162         check_error();
1163
1164         glUseProgram(cbcr_program_num);
1165         check_error();
1166
1167         glActiveTexture(GL_TEXTURE0);
1168         check_error();
1169         glBindTexture(GL_TEXTURE_2D, src_tex);
1170         check_error();
1171         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1172         check_error();
1173         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1174         check_error();
1175         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1176         check_error();
1177
1178         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1179         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1180
1181         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1182         check_error();
1183
1184         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1185                 glEnableVertexAttribArray(attr_index);
1186                 check_error();
1187                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1188                 check_error();
1189         }
1190
1191         glDrawArrays(GL_TRIANGLES, 0, 3);
1192         check_error();
1193
1194         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1195                 glDisableVertexAttribArray(attr_index);
1196                 check_error();
1197         }
1198
1199         glUseProgram(0);
1200         check_error();
1201         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1202         check_error();
1203
1204         resource_pool->release_fbo(fbo);
1205         glDeleteVertexArrays(1, &vao);
1206 }
1207
1208 void Mixer::release_display_frame(DisplayFrame *frame)
1209 {
1210         for (GLuint texnum : frame->temp_textures) {
1211                 resource_pool->release_2d_texture(texnum);
1212         }
1213         frame->temp_textures.clear();
1214         frame->ready_fence.reset();
1215         frame->input_frames.clear();
1216 }
1217
1218 void Mixer::start()
1219 {
1220         mixer_thread = thread(&Mixer::thread_func, this);
1221         audio_thread = thread(&Mixer::audio_thread_func, this);
1222 }
1223
1224 void Mixer::quit()
1225 {
1226         should_quit = true;
1227         audio_task_queue_changed.notify_one();
1228         mixer_thread.join();
1229         audio_thread.join();
1230 }
1231
1232 void Mixer::transition_clicked(int transition_num)
1233 {
1234         theme->transition_clicked(transition_num, pts());
1235 }
1236
1237 void Mixer::channel_clicked(int preview_num)
1238 {
1239         theme->channel_clicked(preview_num);
1240 }
1241
1242 void Mixer::reset_meters()
1243 {
1244         peak_resampler.reset();
1245         peak = 0.0f;
1246         r128.reset();
1247         r128.integr_start();
1248         correlation.reset();
1249 }
1250
1251 void Mixer::start_mode_scanning(unsigned card_index)
1252 {
1253         assert(card_index < num_cards);
1254         if (is_mode_scanning[card_index]) {
1255                 return;
1256         }
1257         is_mode_scanning[card_index] = true;
1258         mode_scanlist[card_index].clear();
1259         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1260                 mode_scanlist[card_index].push_back(mode.first);
1261         }
1262         assert(!mode_scanlist[card_index].empty());
1263         mode_scanlist_index[card_index] = 0;
1264         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1265         last_mode_scan_change[card_index] = steady_clock::now();
1266 }
1267
1268 Mixer::OutputChannel::~OutputChannel()
1269 {
1270         if (has_current_frame) {
1271                 parent->release_display_frame(&current_frame);
1272         }
1273         if (has_ready_frame) {
1274                 parent->release_display_frame(&ready_frame);
1275         }
1276 }
1277
1278 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1279 {
1280         // Store this frame for display. Remove the ready frame if any
1281         // (it was seemingly never used).
1282         {
1283                 unique_lock<mutex> lock(frame_mutex);
1284                 if (has_ready_frame) {
1285                         parent->release_display_frame(&ready_frame);
1286                 }
1287                 ready_frame = frame;
1288                 has_ready_frame = true;
1289         }
1290
1291         if (new_frame_ready_callback) {
1292                 new_frame_ready_callback();
1293         }
1294
1295         // Reduce the number of callbacks by filtering duplicates. The reason
1296         // why we bother doing this is that Qt seemingly can get into a state
1297         // where its builds up an essentially unbounded queue of signals,
1298         // consuming more and more memory, and there's no good way of collapsing
1299         // user-defined signals or limiting the length of the queue.
1300         if (transition_names_updated_callback) {
1301                 vector<string> transition_names = global_mixer->get_transition_names();
1302                 bool changed = false;
1303                 if (transition_names.size() != last_transition_names.size()) {
1304                         changed = true;
1305                 } else {
1306                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1307                                 if (transition_names[i] != last_transition_names[i]) {
1308                                         changed = true;
1309                                         break;
1310                                 }
1311                         }
1312                 }
1313                 if (changed) {
1314                         transition_names_updated_callback(transition_names);
1315                         last_transition_names = transition_names;
1316                 }
1317         }
1318         if (name_updated_callback) {
1319                 string name = global_mixer->get_channel_name(channel);
1320                 if (name != last_name) {
1321                         name_updated_callback(name);
1322                         last_name = name;
1323                 }
1324         }
1325         if (color_updated_callback) {
1326                 string color = global_mixer->get_channel_color(channel);
1327                 if (color != last_color) {
1328                         color_updated_callback(color);
1329                         last_color = color;
1330                 }
1331         }
1332 }
1333
1334 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1335 {
1336         unique_lock<mutex> lock(frame_mutex);
1337         if (!has_current_frame && !has_ready_frame) {
1338                 return false;
1339         }
1340
1341         if (has_current_frame && has_ready_frame) {
1342                 // We have a new ready frame. Toss the current one.
1343                 parent->release_display_frame(&current_frame);
1344                 has_current_frame = false;
1345         }
1346         if (has_ready_frame) {
1347                 assert(!has_current_frame);
1348                 current_frame = ready_frame;
1349                 ready_frame.ready_fence.reset();  // Drop the refcount.
1350                 ready_frame.input_frames.clear();  // Drop the refcounts.
1351                 has_current_frame = true;
1352                 has_ready_frame = false;
1353         }
1354
1355         *frame = current_frame;
1356         return true;
1357 }
1358
1359 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1360 {
1361         new_frame_ready_callback = callback;
1362 }
1363
1364 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1365 {
1366         transition_names_updated_callback = callback;
1367 }
1368
1369 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1370 {
1371         name_updated_callback = callback;
1372 }
1373
1374 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1375 {
1376         color_updated_callback = callback;
1377 }
1378
1379 mutex RefCountedGLsync::fence_lock;