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

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