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

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   7 #include <movit/effect_chain.h>
   8 #include <movit/effect_util.h>
   9 #include <movit/flat_input.h>
  10 #include <movit/image_format.h>
  11 #include <movit/init.h>
  12 #include <movit/resource_pool.h>
  13 #include <movit/util.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <sys/time.h>
  18 #include <time.h>
  19 #include <algorithm>
  20 #include <cmath>
  21 #include <condition_variable>
  22 #include <cstddef>
  23 #include <memory>
  24 #include <mutex>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29 #include <arpa/inet.h>
  30 #include <sys/time.h>
  31 #include <sys/resource.h>
  32
  33 #include "bmusb/bmusb.h"
  34 #include "context.h"
  35 #include "decklink_capture.h"
  36 #include "defs.h"
  37 #include "fake_capture.h"
  38 #include "flags.h"
  39 #include "video_encoder.h"
  40 #include "pbo_frame_allocator.h"
  41 #include "ref_counted_gl_sync.h"
  42 #include "timebase.h"
  43
  44 class QOpenGLContext;
  45
  46 using namespace movit;
  47 using namespace std;
  48 using namespace std::placeholders;
  49 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.c_str(), 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 fake devices, then PCI devices, then USB,
 172         // until we have the desired number of cards.
 173         unsigned num_pci_devices = 0, num_usb_devices = 0;
 174         unsigned card_index = 0;
 175
 176         assert(global_flags.num_fake_cards >= 0);  // Enforced in flags.cpp.
 177         unsigned num_fake_cards = global_flags.num_fake_cards;
 178
 179         assert(num_fake_cards <= num_cards);  // Enforced in flags.cpp.
 180         for ( ; card_index < num_fake_cards; ++card_index) {
 181                 configure_card(card_index, new FakeCapture(card_index), /*is_fake_capture=*/true);
 182         }
 183
 184         if (global_flags.num_fake_cards > 0) {
 185                 fprintf(stderr, "Initialized %d fake cards.\n", global_flags.num_fake_cards);
 186         }
 187
 188         if (card_index < num_cards) {
 189                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 190                 if (decklink_iterator != nullptr) {
 191                         for ( ; card_index < num_cards; ++card_index) {
 192                                 IDeckLink *decklink;
 193                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 194                                         break;
 195                                 }
 196
 197                                 configure_card(card_index, new DeckLinkCapture(decklink, card_index - num_fake_cards), /*is_fake_capture=*/false);
 198                                 ++num_pci_devices;
 199                         }
 200                         decklink_iterator->Release();
 201                         fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
 202                 } else {
 203                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 204                 }
 205         }
 206         for ( ; card_index < num_cards; ++card_index) {
 207                 BMUSBCapture *capture = new BMUSBCapture(card_index - num_pci_devices - num_fake_cards);
 208                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 209                 configure_card(card_index, capture, /*is_fake_capture=*/false);
 210                 ++num_usb_devices;
 211         }
 212
 213         if (num_usb_devices > 0) {
 214                 has_bmusb_thread = true;
 215                 BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 216                 BMUSBCapture::start_bm_thread();
 217         }
 218
 219         for (card_index = 0; card_index < num_cards; ++card_index) {
 220                 cards[card_index].queue_length_policy.reset(card_index);
 221                 cards[card_index].capture->start_bm_capture();
 222         }
 223
 224         // Set up stuff for NV12 conversion.
 225
 226         // Cb/Cr shader.
 227         string cbcr_vert_shader =
 228                 "#version 130 \n"
 229                 " \n"
 230                 "in vec2 position; \n"
 231                 "in vec2 texcoord; \n"
 232                 "out vec2 tc0; \n"
 233                 "uniform vec2 foo_chroma_offset_0; \n"
 234                 " \n"
 235                 "void main() \n"
 236                 "{ \n"
 237                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 238                 "    // \n"
 239                 "    //   2.000  0.000  0.000 -1.000 \n"
 240                 "    //   0.000  2.000  0.000 -1.000 \n"
 241                 "    //   0.000  0.000 -2.000 -1.000 \n"
 242                 "    //   0.000  0.000  0.000  1.000 \n"
 243                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 244                 "    vec2 flipped_tc = texcoord; \n"
 245                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 246                 "} \n";
 247         string cbcr_frag_shader =
 248                 "#version 130 \n"
 249                 "in vec2 tc0; \n"
 250                 "uniform sampler2D cbcr_tex; \n"
 251                 "out vec4 FragColor; \n"
 252                 "void main() { \n"
 253                 "    FragColor = texture(cbcr_tex, tc0); \n"
 254                 "} \n";
 255         vector<string> frag_shader_outputs;
 256         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 257
 258         float vertices[] = {
 259                 0.0f, 2.0f,
 260                 0.0f, 0.0f,
 261                 2.0f, 0.0f
 262         };
 263         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 264         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 265         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 266
 267         r128.init(2, OUTPUT_FREQUENCY);
 268         r128.integr_start();
 269
 270         locut.init(FILTER_HPF, 2);
 271
 272         set_locut_enabled(global_flags.locut_enabled);
 273         set_gain_staging_db(global_flags.initial_gain_staging_db);
 274         set_gain_staging_auto(global_flags.gain_staging_auto);
 275         set_compressor_enabled(global_flags.compressor_enabled);
 276         set_limiter_enabled(global_flags.limiter_enabled);
 277         set_final_makeup_gain_auto(global_flags.final_makeup_gain_auto);
 278
 279         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 280         // and there's a limit to how important the peak meter is.
 281         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 282
 283         if (global_flags.enable_alsa_output) {
 284                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 285         }
 286 }
 287
 288 Mixer::~Mixer()
 289 {
 290         resource_pool->release_glsl_program(cbcr_program_num);
 291         glDeleteBuffers(1, &cbcr_vbo);
 292         if (has_bmusb_thread) {
 293                 BMUSBCapture::stop_bm_thread();
 294         }
 295
 296         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 297                 {
 298                         unique_lock<mutex> lock(bmusb_mutex);
 299                         cards[card_index].should_quit = true;  // Unblock thread.
 300                         cards[card_index].new_frames_changed.notify_all();
 301                 }
 302                 cards[card_index].capture->stop_dequeue_thread();
 303         }
 304
 305         video_encoder.reset(nullptr);
 306 }
 307
 308 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
 309 {
 310         printf("Configuring card %d...\n", card_index);
 311
 312         CaptureCard *card = &cards[card_index];
 313         if (card->capture != nullptr) {
 314                 card->capture->stop_dequeue_thread();
 315                 delete card->capture;
 316         }
 317         card->capture = capture;
 318         card->is_fake_capture = is_fake_capture;
 319         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 320         if (card->frame_allocator == nullptr) {
 321                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 322         }
 323         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 324         if (card->surface == nullptr) {
 325                 card->surface = create_surface_with_same_format(mixer_surface);
 326         }
 327         {
 328                 unique_lock<mutex> lock(cards[card_index].audio_mutex);
 329                 card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 330         }
 331         while (!card->new_frames.empty()) card->new_frames.pop();
 332         card->fractional_samples = 0;
 333         card->last_timecode = -1;
 334         card->next_local_pts = 0;
 335         card->capture->configure_card();
 336 }
 337
 338
 339 namespace {
 340
 341 int unwrap_timecode(uint16_t current_wrapped, int last)
 342 {
 343         uint16_t last_wrapped = last & 0xffff;
 344         if (current_wrapped > last_wrapped) {
 345                 return (last & ~0xffff) | current_wrapped;
 346         } else {
 347                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 348         }
 349 }
 350
 351 float find_peak(const float *samples, size_t num_samples)
 352 {
 353         float m = fabs(samples[0]);
 354         for (size_t i = 1; i < num_samples; ++i) {
 355                 m = max(m, fabs(samples[i]));
 356         }
 357         return m;
 358 }
 359
 360 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 361 {
 362         size_t num_samples = in.size() / 2;
 363         out_l->resize(num_samples);
 364         out_r->resize(num_samples);
 365
 366         const float *inptr = in.data();
 367         float *lptr = &(*out_l)[0];
 368         float *rptr = &(*out_r)[0];
 369         for (size_t i = 0; i < num_samples; ++i) {
 370                 *lptr++ = *inptr++;
 371                 *rptr++ = *inptr++;
 372         }
 373 }
 374
 375 }  // namespace
 376
 377 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 378                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 379                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 380 {
 381         CaptureCard *card = &cards[card_index];
 382
 383         if (is_mode_scanning[card_index]) {
 384                 if (video_format.has_signal) {
 385                         // Found a stable signal, so stop scanning.
 386                         is_mode_scanning[card_index] = false;
 387                 } else {
 388                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 389                         timespec now;
 390                         clock_gettime(CLOCK_MONOTONIC, &now);
 391                         double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
 392                                 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
 393                         if (sec_since_last_switch > switch_time_s) {
 394                                 // It isn't this mode; try the next one.
 395                                 mode_scanlist_index[card_index]++;
 396                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 397                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 398                                 last_mode_scan_change[card_index] = now;
 399                         }
 400                 }
 401         }
 402
 403         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 404         assert(frame_length > 0);
 405
 406         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;
 407         if (num_samples > OUTPUT_FREQUENCY / 10) {
 408                 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",
 409                         card_index, int(audio_frame.len), int(audio_offset),
 410                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 411                 if (video_frame.owner) {
 412                         video_frame.owner->release_frame(video_frame);
 413                 }
 414                 if (audio_frame.owner) {
 415                         audio_frame.owner->release_frame(audio_frame);
 416                 }
 417                 return;
 418         }
 419
 420         int64_t local_pts = card->next_local_pts;
 421         int dropped_frames = 0;
 422         if (card->last_timecode != -1) {
 423                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 424         }
 425
 426         // Convert the audio to stereo fp32 and add it.
 427         vector<float> audio;
 428         audio.resize(num_samples * 2);
 429         switch (audio_format.bits_per_sample) {
 430         case 0:
 431                 assert(num_samples == 0);
 432                 break;
 433         case 24:
 434                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 435                 break;
 436         case 32:
 437                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 438                 break;
 439         default:
 440                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 441                 assert(false);
 442         }
 443
 444         // Add the audio.
 445         {
 446                 unique_lock<mutex> lock(card->audio_mutex);
 447
 448                 // Number of samples per frame if we need to insert silence.
 449                 // (Could be nonintegral, but resampling will save us then.)
 450                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 451
 452                 if (dropped_frames > MAX_FPS * 2) {
 453                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 454                                 card_index, card->last_timecode, timecode);
 455                         card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 456                         dropped_frames = 0;
 457                 } else if (dropped_frames > 0) {
 458                         // Insert silence as needed.
 459                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 460                                 card_index, dropped_frames, timecode);
 461                         vector<float> silence(silence_samples * 2, 0.0f);
 462                         for (int i = 0; i < dropped_frames; ++i) {
 463                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 464                                 // Note that if the format changed in the meantime, we have
 465                                 // no way of detecting that; we just have to assume the frame length
 466                                 // is always the same.
 467                                 local_pts += frame_length;
 468                         }
 469                 }
 470                 if (num_samples == 0) {
 471                         audio.resize(silence_samples * 2);
 472                         num_samples = silence_samples;
 473                 }
 474                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 475                 card->next_local_pts = local_pts + frame_length;
 476         }
 477
 478         card->last_timecode = timecode;
 479
 480         // Done with the audio, so release it.
 481         if (audio_frame.owner) {
 482                 audio_frame.owner->release_frame(audio_frame);
 483         }
 484
 485         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 486         if (video_frame.len - video_offset == 0 ||
 487             video_frame.len - video_offset != expected_length) {
 488                 if (video_frame.len != 0) {
 489                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 490                                 card_index, video_frame.len - video_offset, expected_length);
 491                 }
 492                 if (video_frame.owner) {
 493                         video_frame.owner->release_frame(video_frame);
 494                 }
 495
 496                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 497                 // so that pts can go up accordingly.
 498                 {
 499                         unique_lock<mutex> lock(bmusb_mutex);
 500                         CaptureCard::NewFrame new_frame;
 501                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 502                         new_frame.length = frame_length;
 503                         new_frame.interlaced = false;
 504                         new_frame.dropped_frames = dropped_frames;
 505                         card->new_frames.push(move(new_frame));
 506                         card->new_frames_changed.notify_all();
 507                 }
 508                 return;
 509         }
 510
 511         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 512
 513         unsigned num_fields = video_format.interlaced ? 2 : 1;
 514         timespec frame_upload_start;
 515         if (video_format.interlaced) {
 516                 // Send the two fields along as separate frames; the other side will need to add
 517                 // a deinterlacer to actually get this right.
 518                 assert(video_format.height % 2 == 0);
 519                 video_format.height /= 2;
 520                 assert(frame_length % 2 == 0);
 521                 frame_length /= 2;
 522                 num_fields = 2;
 523                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 524         }
 525         userdata->last_interlaced = video_format.interlaced;
 526         userdata->last_has_signal = video_format.has_signal;
 527         userdata->last_is_connected = video_format.is_connected;
 528         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 529         userdata->last_frame_rate_den = video_format.frame_rate_den;
 530         RefCountedFrame frame(video_frame);
 531
 532         // Upload the textures.
 533         size_t cbcr_width = video_format.width / 2;
 534         size_t cbcr_offset = video_offset / 2;
 535         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 536
 537         for (unsigned field = 0; field < num_fields; ++field) {
 538                 // Put the actual texture upload in a lambda that is executed in the main thread.
 539                 // It is entirely possible to do this in the same thread (and it might even be
 540                 // faster, depending on the GPU and driver), but it appears to be trickling
 541                 // driver bugs very easily.
 542                 //
 543                 // Note that this means we must hold on to the actual frame data in <userdata>
 544                 // until the upload command is run, but we hold on to <frame> much longer than that
 545                 // (in fact, all the way until we no longer use the texture in rendering).
 546                 auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() {
 547                         unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 548
 549                         if (userdata->tex_y[field] == 0 ||
 550                             userdata->tex_cbcr[field] == 0 ||
 551                             video_format.width != userdata->last_width[field] ||
 552                             video_format.height != userdata->last_height[field]) {
 553                                 // We changed resolution since last use of this texture, so we need to create
 554                                 // a new object. Note that this each card has its own PBOFrameAllocator,
 555                                 // we don't need to worry about these flip-flopping between resolutions.
 556                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 557                                 check_error();
 558                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 559                                 check_error();
 560                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 561                                 check_error();
 562                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 563                                 check_error();
 564                                 userdata->last_width[field] = video_format.width;
 565                                 userdata->last_height[field] = video_format.height;
 566                         }
 567
 568                         GLuint pbo = userdata->pbo;
 569                         check_error();
 570                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
 571                         check_error();
 572
 573                         size_t field_y_start = y_offset + video_format.width * field_start_line;
 574                         size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 575
 576                         if (global_flags.flush_pbos) {
 577                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
 578                                 check_error();
 579                                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
 580                                 check_error();
 581                         }
 582
 583                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 584                         check_error();
 585                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
 586                         check_error();
 587                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 588                         check_error();
 589                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
 590                         check_error();
 591                         glBindTexture(GL_TEXTURE_2D, 0);
 592                         check_error();
 593                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 594                         check_error();
 595                 };
 596
 597                 if (field == 1) {
 598                         // Don't upload the second field as fast as we can; wait until
 599                         // the field time has approximately passed. (Otherwise, we could
 600                         // get timing jitter against the other sources, and possibly also
 601                         // against the video display, although the latter is not as critical.)
 602                         // This requires our system clock to be reasonably close to the
 603                         // video clock, but that's not an unreasonable assumption.
 604                         timespec second_field_start;
 605                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 606                                 frame_length * 1000000000 / TIMEBASE;
 607                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 608                                 second_field_start.tv_nsec / 1000000000;
 609                         second_field_start.tv_nsec %= 1000000000;
 610
 611                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 612                                                &second_field_start, nullptr) == -1 &&
 613                                errno == EINTR) ;
 614                 }
 615
 616                 {
 617                         unique_lock<mutex> lock(bmusb_mutex);
 618                         CaptureCard::NewFrame new_frame;
 619                         new_frame.frame = frame;
 620                         new_frame.length = frame_length;
 621                         new_frame.field = field;
 622                         new_frame.interlaced = video_format.interlaced;
 623                         new_frame.upload_func = upload_func;
 624                         new_frame.dropped_frames = dropped_frames;
 625                         card->new_frames.push(move(new_frame));
 626                         card->new_frames_changed.notify_all();
 627                 }
 628         }
 629 }
 630
 631 void Mixer::bm_hotplug_add(libusb_device *dev)
 632 {
 633         lock_guard<mutex> lock(hotplug_mutex);
 634         hotplugged_cards.push_back(dev);
 635 }
 636
 637 void Mixer::bm_hotplug_remove(unsigned card_index)
 638 {
 639         cards[card_index].new_frames_changed.notify_all();
 640 }
 641
 642 void Mixer::thread_func()
 643 {
 644         eglBindAPI(EGL_OPENGL_API);
 645         QOpenGLContext *context = create_context(mixer_surface);
 646         if (!make_current(context, mixer_surface)) {
 647                 printf("oops\n");
 648                 exit(1);
 649         }
 650
 651         struct timespec start, now;
 652         clock_gettime(CLOCK_MONOTONIC, &start);
 653
 654         int frame = 0;
 655         int stats_dropped_frames = 0;
 656
 657         while (!should_quit) {
 658                 CaptureCard::NewFrame new_frames[MAX_CARDS];
 659                 bool has_new_frame[MAX_CARDS] = { false };
 660                 int num_samples[MAX_CARDS] = { 0 };
 661
 662                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 663                 assert(master_card_index < num_cards);
 664
 665                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 666                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 667                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 668                 send_audio_level_callback();
 669
 670                 handle_hotplugged_cards();
 671
 672                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 673                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 674                                 continue;
 675                         }
 676                         if (new_frames[card_index].frame->len == 0) {
 677                                 ++new_frames[card_index].dropped_frames;
 678                         }
 679                         if (new_frames[card_index].dropped_frames > 0) {
 680                                 printf("Card %u dropped %d frames before this\n",
 681                                         card_index, int(new_frames[card_index].dropped_frames));
 682                         }
 683                 }
 684
 685                 // If the first card is reporting a corrupted or otherwise dropped frame,
 686                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 687                 if (new_frames[master_card_index].frame->len == 0) {
 688                         ++stats_dropped_frames;
 689                         pts_int += new_frames[master_card_index].length;
 690                         continue;
 691                 }
 692
 693                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 694                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 695                                 continue;
 696
 697                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 698                         assert(new_frame->frame != nullptr);
 699                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 700                         check_error();
 701
 702                         // The new texture might need uploading before use.
 703                         if (new_frame->upload_func) {
 704                                 new_frame->upload_func();
 705                                 new_frame->upload_func = nullptr;
 706                         }
 707                 }
 708
 709                 int64_t duration = new_frames[master_card_index].length;
 710                 render_one_frame(duration);
 711                 ++frame;
 712                 pts_int += duration;
 713
 714                 clock_gettime(CLOCK_MONOTONIC, &now);
 715                 double elapsed = now.tv_sec - start.tv_sec +
 716                         1e-9 * (now.tv_nsec - start.tv_nsec);
 717                 if (frame % 100 == 0) {
 718                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 719                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 720                                 1e3 * elapsed / frame);
 721                 //      chain->print_phase_timing();
 722
 723                         // Check our memory usage, to see if we are close to our mlockall()
 724                         // limit (if at all set).
 725                         rusage used;
 726                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 727                                 perror("getrusage(RUSAGE_SELF)");
 728                                 assert(false);
 729                         }
 730
 731                         if (uses_mlock) {
 732                                 rlimit limit;
 733                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 734                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 735                                         assert(false);
 736                                 }
 737
 738                                 printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 739                                         long(used.ru_maxrss / 1024),
 740                                         long(limit.rlim_cur / 1048576),
 741                                         float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 742                         } else {
 743                                 printf(", using %ld MB memory (not locked)",
 744                                         long(used.ru_maxrss / 1024));
 745                         }
 746
 747                         printf("\n");
 748                 }
 749
 750
 751                 if (should_cut.exchange(false)) {  // Test and clear.
 752                         video_encoder->do_cut(frame);
 753                 }
 754
 755 #if 0
 756                 // Reset every 100 frames, so that local variations in frame times
 757                 // (especially for the first few frames, when the shaders are
 758                 // compiled etc.) don't make it hard to measure for the entire
 759                 // remaining duration of the program.
 760                 if (frame == 10000) {
 761                         frame = 0;
 762                         start = now;
 763                 }
 764 #endif
 765                 check_error();
 766         }
 767
 768         resource_pool->clean_context();
 769 }
 770
 771 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])
 772 {
 773 start:
 774         // The first card is the master timer, so wait for it to have a new frame.
 775         // TODO: Add a timeout.
 776         unique_lock<mutex> lock(bmusb_mutex);
 777         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(); });
 778
 779         if (cards[master_card_index].new_frames.empty()) {
 780                 // We were woken up, but not due to a new frame. Deal with it
 781                 // and then restart.
 782                 assert(cards[master_card_index].capture->get_disconnected());
 783                 handle_hotplugged_cards();
 784                 goto start;
 785         }
 786
 787         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 788                 CaptureCard *card = &cards[card_index];
 789                 if (card->new_frames.empty()) {
 790                         assert(card_index != master_card_index);
 791                         card->queue_length_policy.update_policy(-1);
 792                         continue;
 793                 }
 794                 new_frames[card_index] = move(card->new_frames.front());
 795                 has_new_frame[card_index] = true;
 796                 card->new_frames.pop();
 797                 card->new_frames_changed.notify_all();
 798
 799                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 800                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 801                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 802                 assert(num_samples[card_index] >= 0);
 803
 804                 if (card_index == master_card_index) {
 805                         // We don't use the queue length policy for the master card,
 806                         // but we will if it stops being the master. Thus, clear out
 807                         // the policy in case we switch in the future.
 808                         card->queue_length_policy.reset(card_index);
 809                 } else {
 810                         // If we have excess frames compared to the policy for this card,
 811                         // drop frames from the head.
 812                         card->queue_length_policy.update_policy(card->new_frames.size());
 813                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 814                                 card->new_frames.pop();
 815                         }
 816                 }
 817         }
 818 }
 819
 820 void Mixer::handle_hotplugged_cards()
 821 {
 822         // Check for cards that have been disconnected since last frame.
 823         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 824                 CaptureCard *card = &cards[card_index];
 825                 if (card->capture->get_disconnected()) {
 826                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 827                         configure_card(card_index, new FakeCapture(card_index), /*is_fake_capture=*/true);
 828                         card->queue_length_policy.reset(card_index);
 829                         card->capture->start_bm_capture();
 830                 }
 831         }
 832
 833         // Check for cards that have been connected since last frame.
 834         vector<libusb_device *> hotplugged_cards_copy;
 835         {
 836                 lock_guard<mutex> lock(hotplug_mutex);
 837                 swap(hotplugged_cards, hotplugged_cards_copy);
 838         }
 839         for (libusb_device *new_dev : hotplugged_cards_copy) {
 840                 // Look for a fake capture card where we can stick this in.
 841                 int free_card_index = -1;
 842                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 843                         if (cards[card_index].is_fake_capture) {
 844                                 free_card_index = int(card_index);
 845                                 break;
 846                         }
 847                 }
 848
 849                 if (free_card_index == -1) {
 850                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 851                         libusb_unref_device(new_dev);
 852                 } else {
 853                         // BMUSBCapture takes ownership.
 854                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 855                         CaptureCard *card = &cards[free_card_index];
 856                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 857                         configure_card(free_card_index, capture, /*is_fake_capture=*/false);
 858                         card->queue_length_policy.reset(free_card_index);
 859                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 860                         capture->start_bm_capture();
 861                 }
 862         }
 863 }
 864
 865
 866 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 867 {
 868         // Resample the audio as needed, including from previously dropped frames.
 869         assert(num_cards > 0);
 870         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 871                 {
 872                         // Signal to the audio thread to process this frame.
 873                         unique_lock<mutex> lock(audio_mutex);
 874                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
 875                         audio_task_queue_changed.notify_one();
 876                 }
 877                 if (frame_num != dropped_frames) {
 878                         // For dropped frames, increase the pts. Note that if the format changed
 879                         // in the meantime, we have no way of detecting that; we just have to
 880                         // assume the frame length is always the same.
 881                         pts_int += length_per_frame;
 882                 }
 883         }
 884 }
 885
 886 void Mixer::render_one_frame(int64_t duration)
 887 {
 888         // Get the main chain from the theme, and set its state immediately.
 889         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 890         EffectChain *chain = theme_main_chain.chain;
 891         theme_main_chain.setup_chain();
 892         //theme_main_chain.chain->enable_phase_timing(true);
 893
 894         GLuint y_tex, cbcr_tex;
 895         bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
 896         assert(got_frame);
 897
 898         // Render main chain.
 899         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 900         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 901         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 902         check_error();
 903         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 904         resource_pool->release_fbo(fbo);
 905
 906         subsample_chroma(cbcr_full_tex, cbcr_tex);
 907         resource_pool->release_2d_texture(cbcr_full_tex);
 908
 909         // Set the right state for rgba_tex.
 910         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 911         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 912         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 913         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 914         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 915
 916         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 917         RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
 918
 919         // The live frame just shows the RGBA texture we just rendered.
 920         // It owns rgba_tex now.
 921         DisplayFrame live_frame;
 922         live_frame.chain = display_chain.get();
 923         live_frame.setup_chain = [this, rgba_tex]{
 924                 display_input->set_texture_num(rgba_tex);
 925         };
 926         live_frame.ready_fence = fence;
 927         live_frame.input_frames = {};
 928         live_frame.temp_textures = { rgba_tex };
 929         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 930
 931         // Set up preview and any additional channels.
 932         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 933                 DisplayFrame display_frame;
 934                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 935                 display_frame.chain = chain.chain;
 936                 display_frame.setup_chain = chain.setup_chain;
 937                 display_frame.ready_fence = fence;
 938                 display_frame.input_frames = chain.input_frames;
 939                 display_frame.temp_textures = {};
 940                 output_channel[i].output_frame(display_frame);
 941         }
 942 }
 943
 944 void Mixer::send_audio_level_callback()
 945 {
 946         if (audio_level_callback == nullptr) {
 947                 return;
 948         }
 949
 950         unique_lock<mutex> lock(compressor_mutex);
 951         double loudness_s = r128.loudness_S();
 952         double loudness_i = r128.integrated();
 953         double loudness_range_low = r128.range_min();
 954         double loudness_range_high = r128.range_max();
 955
 956         audio_level_callback(loudness_s, 20.0 * log10(peak),
 957                 loudness_i, loudness_range_low, loudness_range_high,
 958                 gain_staging_db, 20.0 * log10(final_makeup_gain),
 959                 correlation.get_correlation());
 960 }
 961
 962 void Mixer::audio_thread_func()
 963 {
 964         while (!should_quit) {
 965                 AudioTask task;
 966
 967                 {
 968                         unique_lock<mutex> lock(audio_mutex);
 969                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
 970                         if (should_quit) {
 971                                 return;
 972                         }
 973                         task = audio_task_queue.front();
 974                         audio_task_queue.pop();
 975                 }
 976
 977                 process_audio_one_frame(task.pts_int, task.num_samples);
 978         }
 979 }
 980
 981 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 982 {
 983         vector<float> samples_card;
 984         vector<float> samples_out;
 985
 986         // TODO: Allow mixing audio from several sources.
 987         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 988         assert(selected_audio_card < num_cards);
 989
 990         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 991                 samples_card.resize(num_samples * 2);
 992                 {
 993                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 994                         cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples);
 995                 }
 996                 if (card_index == selected_audio_card) {
 997                         samples_out = move(samples_card);
 998                 }
 999         }
1000
1001         // Cut away everything under 120 Hz (or whatever the cutoff is);
1002         // we don't need it for voice, and it will reduce headroom
1003         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
1004         // should be dampened.)
1005         if (locut_enabled) {
1006                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
1007         }
1008
1009         // Apply a level compressor to get the general level right.
1010         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
1011         // (or more precisely, near it, since we don't use infinite ratio),
1012         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
1013         // entirely arbitrary, but from practical tests with speech, it seems to
1014         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
1015         {
1016                 unique_lock<mutex> lock(compressor_mutex);
1017                 if (level_compressor_enabled) {
1018                         float threshold = 0.01f;   // -40 dBFS.
1019                         float ratio = 20.0f;
1020                         float attack_time = 0.5f;
1021                         float release_time = 20.0f;
1022                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
1023                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1024                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
1025                 } else {
1026                         // Just apply the gain we already had.
1027                         float g = pow(10.0f, gain_staging_db / 20.0f);
1028                         for (size_t i = 0; i < samples_out.size(); ++i) {
1029                                 samples_out[i] *= g;
1030                         }
1031                 }
1032         }
1033
1034 #if 0
1035         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
1036                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
1037                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
1038                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
1039 #endif
1040
1041 //      float limiter_att, compressor_att;
1042
1043         // The real compressor.
1044         if (compressor_enabled) {
1045                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
1046                 float ratio = 20.0f;
1047                 float attack_time = 0.005f;
1048                 float release_time = 0.040f;
1049                 float makeup_gain = 2.0f;  // +6 dB.
1050                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1051 //              compressor_att = compressor.get_attenuation();
1052         }
1053
1054         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
1055         // Note that since ratio is not infinite, we could go slightly higher than this.
1056         if (limiter_enabled) {
1057                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
1058                 float ratio = 30.0f;
1059                 float attack_time = 0.0f;  // Instant.
1060                 float release_time = 0.020f;
1061                 float makeup_gain = 1.0f;  // 0 dB.
1062                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
1063 //              limiter_att = limiter.get_attenuation();
1064         }
1065
1066 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
1067
1068         // Upsample 4x to find interpolated peak.
1069         peak_resampler.inp_data = samples_out.data();
1070         peak_resampler.inp_count = samples_out.size() / 2;
1071
1072         vector<float> interpolated_samples_out;
1073         interpolated_samples_out.resize(samples_out.size());
1074         while (peak_resampler.inp_count > 0) {  // About four iterations.
1075                 peak_resampler.out_data = &interpolated_samples_out[0];
1076                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
1077                 peak_resampler.process();
1078                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
1079                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
1080                 peak_resampler.out_data = nullptr;
1081         }
1082
1083         // At this point, we are most likely close to +0 LU, but all of our
1084         // measurements have been on raw sample values, not R128 values.
1085         // So we have a final makeup gain to get us to +0 LU; the gain
1086         // adjustments required should be relatively small, and also, the
1087         // offset shouldn't change much (only if the type of audio changes
1088         // significantly). Thus, we shoot for updating this value basically
1089         // “whenever we process buffers”, since the R128 calculation isn't exactly
1090         // something we get out per-sample.
1091         //
1092         // Note that there's a feedback loop here, so we choose a very slow filter
1093         // (half-time of 100 seconds).
1094         double target_loudness_factor, alpha;
1095         {
1096                 unique_lock<mutex> lock(compressor_mutex);
1097                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
1098                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
1099                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
1100
1101                 // If we're outside +/- 5 LU uncorrected, we don't count it as
1102                 // a normal signal (probably silence) and don't change the
1103                 // correction factor; just apply what we already have.
1104                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
1105                         alpha = 0.0;
1106                 } else {
1107                         // Formula adapted from
1108                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
1109                         const double half_time_s = 100.0;
1110                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
1111                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
1112                 }
1113
1114                 double m = final_makeup_gain;
1115                 for (size_t i = 0; i < samples_out.size(); i += 2) {
1116                         samples_out[i + 0] *= m;
1117                         samples_out[i + 1] *= m;
1118                         m += (target_loudness_factor - m) * alpha;
1119                 }
1120                 final_makeup_gain = m;
1121         }
1122
1123         // Find R128 levels and L/R correlation.
1124         vector<float> left, right;
1125         deinterleave_samples(samples_out, &left, &right);
1126         float *ptrs[] = { left.data(), right.data() };
1127         {
1128                 unique_lock<mutex> lock(compressor_mutex);
1129                 r128.process(left.size(), ptrs);
1130                 correlation.process_samples(samples_out);
1131         }
1132
1133         // Send the samples to the sound card.
1134         if (alsa) {
1135                 alsa->write(samples_out);
1136         }
1137
1138         // And finally add them to the output.
1139         video_encoder->add_audio(frame_pts_int, move(samples_out));
1140 }
1141
1142 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1143 {
1144         GLuint vao;
1145         glGenVertexArrays(1, &vao);
1146         check_error();
1147
1148         glBindVertexArray(vao);
1149         check_error();
1150
1151         // Extract Cb/Cr.
1152         GLuint fbo = resource_pool->create_fbo(dst_tex);
1153         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1154         glViewport(0, 0, WIDTH/2, HEIGHT/2);
1155         check_error();
1156
1157         glUseProgram(cbcr_program_num);
1158         check_error();
1159
1160         glActiveTexture(GL_TEXTURE0);
1161         check_error();
1162         glBindTexture(GL_TEXTURE_2D, src_tex);
1163         check_error();
1164         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1165         check_error();
1166         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1167         check_error();
1168         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1169         check_error();
1170
1171         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1172         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1173
1174         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1175         check_error();
1176
1177         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1178                 glEnableVertexAttribArray(attr_index);
1179                 check_error();
1180                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1181                 check_error();
1182         }
1183
1184         glDrawArrays(GL_TRIANGLES, 0, 3);
1185         check_error();
1186
1187         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1188                 glDisableVertexAttribArray(attr_index);
1189                 check_error();
1190         }
1191
1192         glUseProgram(0);
1193         check_error();
1194         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1195         check_error();
1196
1197         resource_pool->release_fbo(fbo);
1198         glDeleteVertexArrays(1, &vao);
1199 }
1200
1201 void Mixer::release_display_frame(DisplayFrame *frame)
1202 {
1203         for (GLuint texnum : frame->temp_textures) {
1204                 resource_pool->release_2d_texture(texnum);
1205         }
1206         frame->temp_textures.clear();
1207         frame->ready_fence.reset();
1208         frame->input_frames.clear();
1209 }
1210
1211 void Mixer::start()
1212 {
1213         mixer_thread = thread(&Mixer::thread_func, this);
1214         audio_thread = thread(&Mixer::audio_thread_func, this);
1215 }
1216
1217 void Mixer::quit()
1218 {
1219         should_quit = true;
1220         audio_task_queue_changed.notify_one();
1221         mixer_thread.join();
1222         audio_thread.join();
1223 }
1224
1225 void Mixer::transition_clicked(int transition_num)
1226 {
1227         theme->transition_clicked(transition_num, pts());
1228 }
1229
1230 void Mixer::channel_clicked(int preview_num)
1231 {
1232         theme->channel_clicked(preview_num);
1233 }
1234
1235 void Mixer::reset_meters()
1236 {
1237         peak_resampler.reset();
1238         peak = 0.0f;
1239         r128.reset();
1240         r128.integr_start();
1241         correlation.reset();
1242 }
1243
1244 void Mixer::start_mode_scanning(unsigned card_index)
1245 {
1246         assert(card_index < num_cards);
1247         if (is_mode_scanning[card_index]) {
1248                 return;
1249         }
1250         is_mode_scanning[card_index] = true;
1251         mode_scanlist[card_index].clear();
1252         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1253                 mode_scanlist[card_index].push_back(mode.first);
1254         }
1255         assert(!mode_scanlist[card_index].empty());
1256         mode_scanlist_index[card_index] = 0;
1257         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1258         clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1259 }
1260
1261 Mixer::OutputChannel::~OutputChannel()
1262 {
1263         if (has_current_frame) {
1264                 parent->release_display_frame(&current_frame);
1265         }
1266         if (has_ready_frame) {
1267                 parent->release_display_frame(&ready_frame);
1268         }
1269 }
1270
1271 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1272 {
1273         // Store this frame for display. Remove the ready frame if any
1274         // (it was seemingly never used).
1275         {
1276                 unique_lock<mutex> lock(frame_mutex);
1277                 if (has_ready_frame) {
1278                         parent->release_display_frame(&ready_frame);
1279                 }
1280                 ready_frame = frame;
1281                 has_ready_frame = true;
1282         }
1283
1284         if (new_frame_ready_callback) {
1285                 new_frame_ready_callback();
1286         }
1287
1288         // Reduce the number of callbacks by filtering duplicates. The reason
1289         // why we bother doing this is that Qt seemingly can get into a state
1290         // where its builds up an essentially unbounded queue of signals,
1291         // consuming more and more memory, and there's no good way of collapsing
1292         // user-defined signals or limiting the length of the queue.
1293         if (transition_names_updated_callback) {
1294                 vector<string> transition_names = global_mixer->get_transition_names();
1295                 bool changed = false;
1296                 if (transition_names.size() != last_transition_names.size()) {
1297                         changed = true;
1298                 } else {
1299                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1300                                 if (transition_names[i] != last_transition_names[i]) {
1301                                         changed = true;
1302                                         break;
1303                                 }
1304                         }
1305                 }
1306                 if (changed) {
1307                         transition_names_updated_callback(transition_names);
1308                         last_transition_names = transition_names;
1309                 }
1310         }
1311         if (name_updated_callback) {
1312                 string name = global_mixer->get_channel_name(channel);
1313                 if (name != last_name) {
1314                         name_updated_callback(name);
1315                         last_name = name;
1316                 }
1317         }
1318         if (color_updated_callback) {
1319                 string color = global_mixer->get_channel_color(channel);
1320                 if (color != last_color) {
1321                         color_updated_callback(color);
1322                         last_color = color;
1323                 }
1324         }
1325 }
1326
1327 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1328 {
1329         unique_lock<mutex> lock(frame_mutex);
1330         if (!has_current_frame && !has_ready_frame) {
1331                 return false;
1332         }
1333
1334         if (has_current_frame && has_ready_frame) {
1335                 // We have a new ready frame. Toss the current one.
1336                 parent->release_display_frame(&current_frame);
1337                 has_current_frame = false;
1338         }
1339         if (has_ready_frame) {
1340                 assert(!has_current_frame);
1341                 current_frame = ready_frame;
1342                 ready_frame.ready_fence.reset();  // Drop the refcount.
1343                 ready_frame.input_frames.clear();  // Drop the refcounts.
1344                 has_current_frame = true;
1345                 has_ready_frame = false;
1346         }
1347
1348         *frame = current_frame;
1349         return true;
1350 }
1351
1352 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1353 {
1354         new_frame_ready_callback = callback;
1355 }
1356
1357 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1358 {
1359         transition_names_updated_callback = callback;
1360 }
1361
1362 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1363 {
1364         name_updated_callback = callback;
1365 }
1366
1367 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1368 {
1369         color_updated_callback = callback;
1370 }
1371
1372 mutex RefCountedGLsync::fence_lock;