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
  31 #include "bmusb/bmusb.h"
  32 #include "context.h"
  33 #include "decklink_capture.h"
  34 #include "defs.h"
  35 #include "flags.h"
  36 #include "h264encode.h"
  37 #include "pbo_frame_allocator.h"
  38 #include "ref_counted_gl_sync.h"
  39 #include "timebase.h"
  40
  41 class QOpenGLContext;
  42
  43 using namespace movit;
  44 using namespace std;
  45 using namespace std::placeholders;
  46
  47 Mixer *global_mixer = nullptr;
  48
  49 namespace {
  50
  51 void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  52 {
  53         assert(in_channels >= out_channels);
  54         for (size_t i = 0; i < num_samples; ++i) {
  55                 for (size_t j = 0; j < out_channels; ++j) {
  56                         uint32_t s1 = *src++;
  57                         uint32_t s2 = *src++;
  58                         uint32_t s3 = *src++;
  59                         uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24);
  60                         dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f);
  61                 }
  62                 src += 3 * (in_channels - out_channels);
  63         }
  64 }
  65
  66 void convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples)
  67 {
  68         assert(in_channels >= out_channels);
  69         for (size_t i = 0; i < num_samples; ++i) {
  70                 for (size_t j = 0; j < out_channels; ++j) {
  71                         // Note: Assumes little-endian.
  72                         int32_t s = *(int32_t *)src;
  73                         dst[i * out_channels + j] = s * (1.0f / 4294967296.0f);
  74                         src += 4;
  75                 }
  76                 src += 4 * (in_channels - out_channels);
  77         }
  78 }
  79
  80 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  81 {
  82         if (interlaced) {
  83                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  84                         input_state->buffered_frames[card_index][frame_num] =
  85                                 input_state->buffered_frames[card_index][frame_num - 1];
  86                 }
  87                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  88         } else {
  89                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  90                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  91                 }
  92         }
  93 }
  94
  95 string generate_local_dump_filename(int frame)
  96 {
  97         time_t now = time(NULL);
  98         tm now_tm;
  99         localtime_r(&now, &now_tm);
 100
 101         char timestamp[256];
 102         strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm);
 103
 104         // Use the frame number to disambiguate between two cuts starting
 105         // on the same second.
 106         char filename[256];
 107         snprintf(filename, sizeof(filename), "%s%s-f%02d%s",
 108                 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
 109         return filename;
 110 }
 111
 112 }  // namespace
 113
 114 void QueueLengthPolicy::update_policy(int queue_length)
 115 {
 116         if (queue_length < 0) {  // Starvation.
 117                 if (safe_queue_length < 5) {
 118                         ++safe_queue_length;
 119                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
 120                                 card_index, safe_queue_length);
 121                 }
 122                 frames_with_at_least_one = 0;
 123                 return;
 124         }
 125         if (queue_length > 0) {
 126                 if (++frames_with_at_least_one >= 50 && safe_queue_length > 0) {
 127                         --safe_queue_length;
 128                         fprintf(stderr, "Card %u: Spare frames for more than 50 frames, reducing safe limit to %u frames\n",
 129                                 card_index, safe_queue_length);
 130                         frames_with_at_least_one = 0;
 131                 }
 132         } else {
 133                 frames_with_at_least_one = 0;
 134         }
 135 }
 136
 137 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 138         : httpd(WIDTH, HEIGHT),
 139           num_cards(num_cards),
 140           mixer_surface(create_surface(format)),
 141           h264_encoder_surface(create_surface(format)),
 142           correlation(OUTPUT_FREQUENCY),
 143           level_compressor(OUTPUT_FREQUENCY),
 144           limiter(OUTPUT_FREQUENCY),
 145           compressor(OUTPUT_FREQUENCY)
 146 {
 147         httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str());
 148         httpd.start(9095);
 149
 150         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 151         check_error();
 152
 153         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 154         // will be halved when sampling them, and we need to compensate here.
 155         movit_texel_subpixel_precision /= 2.0;
 156
 157         resource_pool.reset(new ResourcePool);
 158         theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
 159         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 160                 output_channel[i].parent = this;
 161         }
 162
 163         ImageFormat inout_format;
 164         inout_format.color_space = COLORSPACE_sRGB;
 165         inout_format.gamma_curve = GAMMA_sRGB;
 166
 167         // Display chain; shows the live output produced by the main chain (its RGBA version).
 168         display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get()));
 169         check_error();
 170         display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
 171         display_chain->add_input(display_input);
 172         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 173         display_chain->set_dither_bits(0);  // Don't bother.
 174         display_chain->finalize();
 175
 176         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 177
 178         // First try initializing the PCI devices, then USB, until we have the desired number of cards.
 179         unsigned num_pci_devices = 0, num_usb_devices = 0;
 180         unsigned card_index = 0;
 181
 182         IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 183         if (decklink_iterator != nullptr) {
 184                 for ( ; card_index < num_cards; ++card_index) {
 185                         IDeckLink *decklink;
 186                         if (decklink_iterator->Next(&decklink) != S_OK) {
 187                                 break;
 188                         }
 189
 190                         configure_card(card_index, format, new DeckLinkCapture(decklink, card_index));
 191                         ++num_pci_devices;
 192                 }
 193                 decklink_iterator->Release();
 194                 fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices);
 195         } else {
 196                 fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 197         }
 198         for ( ; card_index < num_cards; ++card_index) {
 199                 configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices));
 200                 ++num_usb_devices;
 201         }
 202
 203         if (num_usb_devices > 0) {
 204                 BMUSBCapture::start_bm_thread();
 205         }
 206
 207         for (card_index = 0; card_index < num_cards; ++card_index) {
 208                 cards[card_index].queue_length_policy.reset(card_index);
 209                 cards[card_index].capture->start_bm_capture();
 210         }
 211
 212         // Set up stuff for NV12 conversion.
 213
 214         // Cb/Cr shader.
 215         string cbcr_vert_shader =
 216                 "#version 130 \n"
 217                 " \n"
 218                 "in vec2 position; \n"
 219                 "in vec2 texcoord; \n"
 220                 "out vec2 tc0; \n"
 221                 "uniform vec2 foo_chroma_offset_0; \n"
 222                 " \n"
 223                 "void main() \n"
 224                 "{ \n"
 225                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 226                 "    // \n"
 227                 "    //   2.000  0.000  0.000 -1.000 \n"
 228                 "    //   0.000  2.000  0.000 -1.000 \n"
 229                 "    //   0.000  0.000 -2.000 -1.000 \n"
 230                 "    //   0.000  0.000  0.000  1.000 \n"
 231                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 232                 "    vec2 flipped_tc = texcoord; \n"
 233                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
 234                 "} \n";
 235         string cbcr_frag_shader =
 236                 "#version 130 \n"
 237                 "in vec2 tc0; \n"
 238                 "uniform sampler2D cbcr_tex; \n"
 239                 "out vec4 FragColor; \n"
 240                 "void main() { \n"
 241                 "    FragColor = texture(cbcr_tex, tc0); \n"
 242                 "} \n";
 243         vector<string> frag_shader_outputs;
 244         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 245
 246         float vertices[] = {
 247                 0.0f, 2.0f,
 248                 0.0f, 0.0f,
 249                 2.0f, 0.0f
 250         };
 251         cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 252         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 253         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 254
 255         r128.init(2, OUTPUT_FREQUENCY);
 256         r128.integr_start();
 257
 258         locut.init(FILTER_HPF, 2);
 259
 260         // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
 261         // and there's a limit to how important the peak meter is.
 262         peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0);
 263
 264         alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 265 }
 266
 267 Mixer::~Mixer()
 268 {
 269         resource_pool->release_glsl_program(cbcr_program_num);
 270         glDeleteBuffers(1, &cbcr_vbo);
 271         BMUSBCapture::stop_bm_thread();
 272
 273         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 274                 {
 275                         unique_lock<mutex> lock(bmusb_mutex);
 276                         cards[card_index].should_quit = true;  // Unblock thread.
 277                         cards[card_index].new_frames_changed.notify_all();
 278                 }
 279                 cards[card_index].capture->stop_dequeue_thread();
 280         }
 281
 282         h264_encoder.reset(nullptr);
 283 }
 284
 285 void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture)
 286 {
 287         printf("Configuring card %d...\n", card_index);
 288
 289         CaptureCard *card = &cards[card_index];
 290         card->capture = capture;
 291         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 292         card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT));  // 8 MB.
 293         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 294         card->surface = create_surface(format);
 295         card->capture->set_dequeue_thread_callbacks(
 296                 [card]{
 297                         eglBindAPI(EGL_OPENGL_API);
 298                         card->context = create_context(card->surface);
 299                         if (!make_current(card->context, card->surface)) {
 300                                 printf("failed to create bmusb context\n");
 301                                 exit(1);
 302                         }
 303                 },
 304                 [this]{
 305                         resource_pool->clean_context();
 306                 });
 307         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 308         card->capture->configure_card();
 309 }
 310
 311
 312 namespace {
 313
 314 int unwrap_timecode(uint16_t current_wrapped, int last)
 315 {
 316         uint16_t last_wrapped = last & 0xffff;
 317         if (current_wrapped > last_wrapped) {
 318                 return (last & ~0xffff) | current_wrapped;
 319         } else {
 320                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 321         }
 322 }
 323
 324 float find_peak(const float *samples, size_t num_samples)
 325 {
 326         float m = fabs(samples[0]);
 327         for (size_t i = 1; i < num_samples; ++i) {
 328                 m = max(m, fabs(samples[i]));
 329         }
 330         return m;
 331 }
 332
 333 void deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *out_r)
 334 {
 335         size_t num_samples = in.size() / 2;
 336         out_l->resize(num_samples);
 337         out_r->resize(num_samples);
 338
 339         const float *inptr = in.data();
 340         float *lptr = &(*out_l)[0];
 341         float *rptr = &(*out_r)[0];
 342         for (size_t i = 0; i < num_samples; ++i) {
 343                 *lptr++ = *inptr++;
 344                 *rptr++ = *inptr++;
 345         }
 346 }
 347
 348 }  // namespace
 349
 350 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 351                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 352                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 353 {
 354         CaptureCard *card = &cards[card_index];
 355
 356         if (is_mode_scanning[card_index]) {
 357                 if (video_format.has_signal) {
 358                         // Found a stable signal, so stop scanning.
 359                         is_mode_scanning[card_index] = false;
 360                 } else {
 361                         static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
 362                         timespec now;
 363                         clock_gettime(CLOCK_MONOTONIC, &now);
 364                         double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) +
 365                                 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec);
 366                         if (sec_since_last_switch > switch_time_s) {
 367                                 // It isn't this mode; try the next one.
 368                                 mode_scanlist_index[card_index]++;
 369                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 370                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 371                                 last_mode_scan_change[card_index] = now;
 372                         }
 373                 }
 374         }
 375
 376         int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom;
 377
 378         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;
 379         if (num_samples > OUTPUT_FREQUENCY / 10) {
 380                 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",
 381                         card_index, int(audio_frame.len), int(audio_offset),
 382                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 383                 if (video_frame.owner) {
 384                         video_frame.owner->release_frame(video_frame);
 385                 }
 386                 if (audio_frame.owner) {
 387                         audio_frame.owner->release_frame(audio_frame);
 388                 }
 389                 return;
 390         }
 391
 392         int64_t local_pts = card->next_local_pts;
 393         int dropped_frames = 0;
 394         if (card->last_timecode != -1) {
 395                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 396         }
 397
 398         // Convert the audio to stereo fp32 and add it.
 399         vector<float> audio;
 400         audio.resize(num_samples * 2);
 401         switch (audio_format.bits_per_sample) {
 402         case 0:
 403                 assert(num_samples == 0);
 404                 break;
 405         case 24:
 406                 convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 407                 break;
 408         case 32:
 409                 convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples);
 410                 break;
 411         default:
 412                 fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample);
 413                 assert(false);
 414         }
 415
 416         // Add the audio.
 417         {
 418                 unique_lock<mutex> lock(card->audio_mutex);
 419
 420                 // Number of samples per frame if we need to insert silence.
 421                 // (Could be nonintegral, but resampling will save us then.)
 422                 int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 423
 424                 if (dropped_frames > MAX_FPS * 2) {
 425                         fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 426                                 card_index, card->last_timecode, timecode);
 427                         card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
 428                         dropped_frames = 0;
 429                 } else if (dropped_frames > 0) {
 430                         // Insert silence as needed.
 431                         fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 432                                 card_index, dropped_frames, timecode);
 433                         vector<float> silence(silence_samples * 2, 0.0f);
 434                         for (int i = 0; i < dropped_frames; ++i) {
 435                                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
 436                                 // Note that if the format changed in the meantime, we have
 437                                 // no way of detecting that; we just have to assume the frame length
 438                                 // is always the same.
 439                                 local_pts += frame_length;
 440                         }
 441                 }
 442                 if (num_samples == 0) {
 443                         audio.resize(silence_samples * 2);
 444                         num_samples = silence_samples;
 445                 }
 446                 card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
 447                 card->next_local_pts = local_pts + frame_length;
 448         }
 449
 450         card->last_timecode = timecode;
 451
 452         // Done with the audio, so release it.
 453         if (audio_frame.owner) {
 454                 audio_frame.owner->release_frame(audio_frame);
 455         }
 456
 457         size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
 458         if (video_frame.len - video_offset == 0 ||
 459             video_frame.len - video_offset != expected_length) {
 460                 if (video_frame.len != 0) {
 461                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 462                                 card_index, video_frame.len - video_offset, expected_length);
 463                 }
 464                 if (video_frame.owner) {
 465                         video_frame.owner->release_frame(video_frame);
 466                 }
 467
 468                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 469                 // so that pts can go up accordingly.
 470                 {
 471                         unique_lock<mutex> lock(bmusb_mutex);
 472                         CaptureCard::NewFrame new_frame;
 473                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 474                         new_frame.length = frame_length;
 475                         new_frame.interlaced = false;
 476                         new_frame.ready_fence = nullptr;
 477                         new_frame.dropped_frames = dropped_frames;
 478                         card->new_frames.push(move(new_frame));
 479                         card->new_frames_changed.notify_all();
 480                 }
 481                 return;
 482         }
 483
 484         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 485
 486         unsigned num_fields = video_format.interlaced ? 2 : 1;
 487         timespec frame_upload_start;
 488         if (video_format.interlaced) {
 489                 // Send the two fields along as separate frames; the other side will need to add
 490                 // a deinterlacer to actually get this right.
 491                 assert(video_format.height % 2 == 0);
 492                 video_format.height /= 2;
 493                 assert(frame_length % 2 == 0);
 494                 frame_length /= 2;
 495                 num_fields = 2;
 496                 clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
 497         }
 498         userdata->last_interlaced = video_format.interlaced;
 499         userdata->last_has_signal = video_format.has_signal;
 500         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 501         userdata->last_frame_rate_den = video_format.frame_rate_den;
 502         RefCountedFrame frame(video_frame);
 503
 504         // Upload the textures.
 505         size_t cbcr_width = video_format.width / 2;
 506         size_t cbcr_offset = video_offset / 2;
 507         size_t y_offset = video_frame.size / 2 + video_offset / 2;
 508
 509         for (unsigned field = 0; field < num_fields; ++field) {
 510                 unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);
 511
 512                 if (userdata->tex_y[field] == 0 ||
 513                     userdata->tex_cbcr[field] == 0 ||
 514                     video_format.width != userdata->last_width[field] ||
 515                     video_format.height != userdata->last_height[field]) {
 516                         // We changed resolution since last use of this texture, so we need to create
 517                         // a new object. Note that this each card has its own PBOFrameAllocator,
 518                         // we don't need to worry about these flip-flopping between resolutions.
 519                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 520                         check_error();
 521                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 522                         check_error();
 523                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 524                         check_error();
 525                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 526                         check_error();
 527                         userdata->last_width[field] = video_format.width;
 528                         userdata->last_height[field] = video_format.height;
 529                 }
 530
 531                 GLuint pbo = userdata->pbo;
 532                 check_error();
 533                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
 534                 check_error();
 535                 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 536                 check_error();
 537
 538                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 539                 check_error();
 540                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
 541                 check_error();
 542                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 543                 check_error();
 544                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + video_format.width * field_start_line));
 545                 check_error();
 546                 glBindTexture(GL_TEXTURE_2D, 0);
 547                 check_error();
 548                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 549                 check_error();
 550                 GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 551                 check_error();
 552                 assert(fence != nullptr);
 553
 554                 if (field == 1) {
 555                         // Don't upload the second field as fast as we can; wait until
 556                         // the field time has approximately passed. (Otherwise, we could
 557                         // get timing jitter against the other sources, and possibly also
 558                         // against the video display, although the latter is not as critical.)
 559                         // This requires our system clock to be reasonably close to the
 560                         // video clock, but that's not an unreasonable assumption.
 561                         timespec second_field_start;
 562                         second_field_start.tv_nsec = frame_upload_start.tv_nsec +
 563                                 frame_length * 1000000000 / TIMEBASE;
 564                         second_field_start.tv_sec = frame_upload_start.tv_sec +
 565                                 second_field_start.tv_nsec / 1000000000;
 566                         second_field_start.tv_nsec %= 1000000000;
 567
 568                         while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
 569                                                &second_field_start, nullptr) == -1 &&
 570                                errno == EINTR) ;
 571                 }
 572
 573                 {
 574                         unique_lock<mutex> lock(bmusb_mutex);
 575                         CaptureCard::NewFrame new_frame;
 576                         new_frame.frame = frame;
 577                         new_frame.length = frame_length;
 578                         new_frame.field = field;
 579                         new_frame.interlaced = video_format.interlaced;
 580                         new_frame.ready_fence = fence;
 581                         new_frame.dropped_frames = dropped_frames;
 582                         card->new_frames.push(move(new_frame));
 583                         card->new_frames_changed.notify_all();
 584                 }
 585         }
 586 }
 587
 588 void Mixer::thread_func()
 589 {
 590         eglBindAPI(EGL_OPENGL_API);
 591         QOpenGLContext *context = create_context(mixer_surface);
 592         if (!make_current(context, mixer_surface)) {
 593                 printf("oops\n");
 594                 exit(1);
 595         }
 596
 597         struct timespec start, now;
 598         clock_gettime(CLOCK_MONOTONIC, &start);
 599
 600         int frame = 0;
 601         int stats_dropped_frames = 0;
 602
 603         while (!should_quit) {
 604                 CaptureCard::NewFrame new_frames[MAX_CARDS];
 605                 bool has_new_frame[MAX_CARDS] = { false };
 606                 int num_samples[MAX_CARDS] = { 0 };
 607
 608                 // TODO: Add a timeout.
 609                 unsigned master_card_index = theme->map_signal(master_clock_channel);
 610                 assert(master_card_index < num_cards);
 611
 612                 get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
 613                 schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
 614                 stats_dropped_frames += new_frames[master_card_index].dropped_frames;
 615                 send_audio_level_callback();
 616
 617                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 618                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 619                                 continue;
 620                         }
 621                         if (new_frames[card_index].frame->len == 0) {
 622                                 ++new_frames[card_index].dropped_frames;
 623                         }
 624                         if (new_frames[card_index].dropped_frames > 0) {
 625                                 printf("Card %u dropped %d frames before this\n",
 626                                         card_index, int(new_frames[card_index].dropped_frames));
 627                         }
 628                 }
 629
 630                 // If the first card is reporting a corrupted or otherwise dropped frame,
 631                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 632                 if (new_frames[master_card_index].frame->len == 0) {
 633                         ++stats_dropped_frames;
 634                         pts_int += new_frames[master_card_index].length;
 635                         continue;
 636                 }
 637
 638                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 639                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 640                                 continue;
 641
 642                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 643                         assert(new_frame->frame != nullptr);
 644                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 645                         check_error();
 646
 647                         // The new texture might still be uploaded,
 648                         // tell the GPU to wait until it's there.
 649                         if (new_frame->ready_fence) {
 650                                 glWaitSync(new_frame->ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED);
 651                                 check_error();
 652                                 glDeleteSync(new_frame->ready_fence);
 653                                 check_error();
 654                         }
 655                 }
 656
 657                 render_one_frame();
 658                 ++frame;
 659                 pts_int += new_frames[master_card_index].length;
 660
 661                 clock_gettime(CLOCK_MONOTONIC, &now);
 662                 double elapsed = now.tv_sec - start.tv_sec +
 663                         1e-9 * (now.tv_nsec - start.tv_nsec);
 664                 if (frame % 100 == 0) {
 665                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
 666                                 frame, stats_dropped_frames, elapsed, frame / elapsed,
 667                                 1e3 * elapsed / frame);
 668                 //      chain->print_phase_timing();
 669                 }
 670
 671                 if (should_cut.exchange(false)) {  // Test and clear.
 672                         string filename = generate_local_dump_filename(frame);
 673                         printf("Starting new recording: %s\n", filename.c_str());
 674                         h264_encoder->shutdown();
 675                         httpd.close_output_file();
 676                         httpd.open_output_file(filename.c_str());
 677                         h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd));
 678                 }
 679
 680 #if 0
 681                 // Reset every 100 frames, so that local variations in frame times
 682                 // (especially for the first few frames, when the shaders are
 683                 // compiled etc.) don't make it hard to measure for the entire
 684                 // remaining duration of the program.
 685                 if (frame == 10000) {
 686                         frame = 0;
 687                         start = now;
 688                 }
 689 #endif
 690                 check_error();
 691         }
 692
 693         resource_pool->clean_context();
 694 }
 695
 696 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])
 697 {
 698         // The first card is the master timer, so wait for it to have a new frame.
 699         unique_lock<mutex> lock(bmusb_mutex);
 700         cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty(); });
 701
 702         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 703                 CaptureCard *card = &cards[card_index];
 704                 if (card->new_frames.empty()) {
 705                         assert(card_index != master_card_index);
 706                         card->queue_length_policy.update_policy(-1);
 707                         continue;
 708                 }
 709                 new_frames[card_index] = move(card->new_frames.front());
 710                 has_new_frame[card_index] = true;
 711                 card->new_frames.pop();
 712                 card->new_frames_changed.notify_all();
 713
 714                 int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].length + card->fractional_samples;
 715                 num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
 716                 card->fractional_samples = num_samples_times_timebase % TIMEBASE;
 717                 assert(num_samples[card_index] >= 0);
 718
 719                 if (card_index == master_card_index) {
 720                         // We don't use the queue length policy for the master card,
 721                         // but we will if it stops being the master. Thus, clear out
 722                         // the policy in case we switch in the future.
 723                         card->queue_length_policy.reset(card_index);
 724                 } else {
 725                         // If we have excess frames compared to the policy for this card,
 726                         // drop frames from the head.
 727                         card->queue_length_policy.update_policy(card->new_frames.size());
 728                         while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
 729                                 card->new_frames.pop();
 730                         }
 731                 }
 732         }
 733 }
 734
 735 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
 736 {
 737         // Resample the audio as needed, including from previously dropped frames.
 738         assert(num_cards > 0);
 739         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 740                 {
 741                         // Signal to the audio thread to process this frame.
 742                         unique_lock<mutex> lock(audio_mutex);
 743                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame});
 744                         audio_task_queue_changed.notify_one();
 745                 }
 746                 if (frame_num != dropped_frames) {
 747                         // For dropped frames, increase the pts. Note that if the format changed
 748                         // in the meantime, we have no way of detecting that; we just have to
 749                         // assume the frame length is always the same.
 750                         pts_int += length_per_frame;
 751                 }
 752         }
 753 }
 754
 755 void Mixer::render_one_frame()
 756 {
 757         // Get the main chain from the theme, and set its state immediately.
 758         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
 759         EffectChain *chain = theme_main_chain.chain;
 760         theme_main_chain.setup_chain();
 761         //theme_main_chain.chain->enable_phase_timing(true);
 762
 763         GLuint y_tex, cbcr_tex;
 764         bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
 765         assert(got_frame);
 766
 767         // Render main chain.
 768         GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT);
 769         GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT);  // Saves texture bandwidth, although dithering gets messed up.
 770         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
 771         check_error();
 772         chain->render_to_fbo(fbo, WIDTH, HEIGHT);
 773         resource_pool->release_fbo(fbo);
 774
 775         subsample_chroma(cbcr_full_tex, cbcr_tex);
 776         resource_pool->release_2d_texture(cbcr_full_tex);
 777
 778         // Set the right state for rgba_tex.
 779         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 780         glBindTexture(GL_TEXTURE_2D, rgba_tex);
 781         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 782         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 783         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 784
 785         RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 786         check_error();
 787
 788         const int64_t av_delay = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
 789         h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
 790
 791         // The live frame just shows the RGBA texture we just rendered.
 792         // It owns rgba_tex now.
 793         DisplayFrame live_frame;
 794         live_frame.chain = display_chain.get();
 795         live_frame.setup_chain = [this, rgba_tex]{
 796                 display_input->set_texture_num(rgba_tex);
 797         };
 798         live_frame.ready_fence = fence;
 799         live_frame.input_frames = {};
 800         live_frame.temp_textures = { rgba_tex };
 801         output_channel[OUTPUT_LIVE].output_frame(live_frame);
 802
 803         // Set up preview and any additional channels.
 804         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
 805                 DisplayFrame display_frame;
 806                 Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state);  // FIXME: dimensions
 807                 display_frame.chain = chain.chain;
 808                 display_frame.setup_chain = chain.setup_chain;
 809                 display_frame.ready_fence = fence;
 810                 display_frame.input_frames = chain.input_frames;
 811                 display_frame.temp_textures = {};
 812                 output_channel[i].output_frame(display_frame);
 813         }
 814 }
 815
 816 void Mixer::send_audio_level_callback()
 817 {
 818         if (audio_level_callback == nullptr) {
 819                 return;
 820         }
 821
 822         unique_lock<mutex> lock(compressor_mutex);
 823         double loudness_s = r128.loudness_S();
 824         double loudness_i = r128.integrated();
 825         double loudness_range_low = r128.range_min();
 826         double loudness_range_high = r128.range_max();
 827
 828         audio_level_callback(loudness_s, 20.0 * log10(peak),
 829                 loudness_i, loudness_range_low, loudness_range_high,
 830                 gain_staging_db, 20.0 * log10(final_makeup_gain),
 831                 correlation.get_correlation());
 832 }
 833
 834 void Mixer::audio_thread_func()
 835 {
 836         while (!should_quit) {
 837                 AudioTask task;
 838
 839                 {
 840                         unique_lock<mutex> lock(audio_mutex);
 841                         audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
 842                         task = audio_task_queue.front();
 843                         audio_task_queue.pop();
 844                 }
 845
 846                 process_audio_one_frame(task.pts_int, task.num_samples);
 847         }
 848 }
 849
 850 void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
 851 {
 852         vector<float> samples_card;
 853         vector<float> samples_out;
 854
 855         // TODO: Allow mixing audio from several sources.
 856         unsigned selected_audio_card = theme->map_signal(audio_source_channel);
 857         assert(selected_audio_card < num_cards);
 858
 859         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 860                 samples_card.resize(num_samples * 2);
 861                 {
 862                         unique_lock<mutex> lock(cards[card_index].audio_mutex);
 863                         if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
 864                                 printf("Card %d reported previous underrun.\n", card_index);
 865                         }
 866                 }
 867                 if (card_index == selected_audio_card) {
 868                         samples_out = move(samples_card);
 869                 }
 870         }
 871
 872         // Cut away everything under 120 Hz (or whatever the cutoff is);
 873         // we don't need it for voice, and it will reduce headroom
 874         // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
 875         // should be dampened.)
 876         if (locut_enabled) {
 877                 locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
 878         }
 879
 880         // Apply a level compressor to get the general level right.
 881         // Basically, if it's over about -40 dBFS, we squeeze it down to that level
 882         // (or more precisely, near it, since we don't use infinite ratio),
 883         // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
 884         // entirely arbitrary, but from practical tests with speech, it seems to
 885         // put ut around -23 LUFS, so it's a reasonable starting point for later use.
 886         {
 887                 unique_lock<mutex> lock(compressor_mutex);
 888                 if (level_compressor_enabled) {
 889                         float threshold = 0.01f;   // -40 dBFS.
 890                         float ratio = 20.0f;
 891                         float attack_time = 0.5f;
 892                         float release_time = 20.0f;
 893                         float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f);  // +26 dB.
 894                         level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 895                         gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
 896                 } else {
 897                         // Just apply the gain we already had.
 898                         float g = pow(10.0f, gain_staging_db / 20.0f);
 899                         for (size_t i = 0; i < samples_out.size(); ++i) {
 900                                 samples_out[i] *= g;
 901                         }
 902                 }
 903         }
 904
 905 #if 0
 906         printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
 907                 level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
 908                 level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
 909                 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
 910 #endif
 911
 912 //      float limiter_att, compressor_att;
 913
 914         // The real compressor.
 915         if (compressor_enabled) {
 916                 float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
 917                 float ratio = 20.0f;
 918                 float attack_time = 0.005f;
 919                 float release_time = 0.040f;
 920                 float makeup_gain = 2.0f;  // +6 dB.
 921                 compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 922 //              compressor_att = compressor.get_attenuation();
 923         }
 924
 925         // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
 926         // Note that since ratio is not infinite, we could go slightly higher than this.
 927         if (limiter_enabled) {
 928                 float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
 929                 float ratio = 30.0f;
 930                 float attack_time = 0.0f;  // Instant.
 931                 float release_time = 0.020f;
 932                 float makeup_gain = 1.0f;  // 0 dB.
 933                 limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
 934 //              limiter_att = limiter.get_attenuation();
 935         }
 936
 937 //      printf("limiter=%+5.1f  compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
 938
 939         // Upsample 4x to find interpolated peak.
 940         peak_resampler.inp_data = samples_out.data();
 941         peak_resampler.inp_count = samples_out.size() / 2;
 942
 943         vector<float> interpolated_samples_out;
 944         interpolated_samples_out.resize(samples_out.size());
 945         while (peak_resampler.inp_count > 0) {  // About four iterations.
 946                 peak_resampler.out_data = &interpolated_samples_out[0];
 947                 peak_resampler.out_count = interpolated_samples_out.size() / 2;
 948                 peak_resampler.process();
 949                 size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
 950                 peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
 951                 peak_resampler.out_data = nullptr;
 952         }
 953
 954         // At this point, we are most likely close to +0 LU, but all of our
 955         // measurements have been on raw sample values, not R128 values.
 956         // So we have a final makeup gain to get us to +0 LU; the gain
 957         // adjustments required should be relatively small, and also, the
 958         // offset shouldn't change much (only if the type of audio changes
 959         // significantly). Thus, we shoot for updating this value basically
 960         // “whenever we process buffers”, since the R128 calculation isn't exactly
 961         // something we get out per-sample.
 962         //
 963         // Note that there's a feedback loop here, so we choose a very slow filter
 964         // (half-time of 100 seconds).
 965         double target_loudness_factor, alpha;
 966         {
 967                 unique_lock<mutex> lock(compressor_mutex);
 968                 double loudness_lu = r128.loudness_M() - ref_level_lufs;
 969                 double current_makeup_lu = 20.0f * log10(final_makeup_gain);
 970                 target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
 971
 972                 // If we're outside +/- 5 LU uncorrected, we don't count it as
 973                 // a normal signal (probably silence) and don't change the
 974                 // correction factor; just apply what we already have.
 975                 if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
 976                         alpha = 0.0;
 977                 } else {
 978                         // Formula adapted from
 979                         // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
 980                         const double half_time_s = 100.0;
 981                         const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
 982                         alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
 983                 }
 984
 985                 double m = final_makeup_gain;
 986                 for (size_t i = 0; i < samples_out.size(); i += 2) {
 987                         samples_out[i + 0] *= m;
 988                         samples_out[i + 1] *= m;
 989                         m += (target_loudness_factor - m) * alpha;
 990                 }
 991                 final_makeup_gain = m;
 992         }
 993
 994         // Find R128 levels and L/R correlation.
 995         vector<float> left, right;
 996         deinterleave_samples(samples_out, &left, &right);
 997         float *ptrs[] = { left.data(), right.data() };
 998         {
 999                 unique_lock<mutex> lock(compressor_mutex);
1000                 r128.process(left.size(), ptrs);
1001                 correlation.process_samples(samples_out);
1002         }
1003
1004         // Send the samples to the sound card.
1005         if (alsa) {
1006                 alsa->write(samples_out);
1007         }
1008
1009         // And finally add them to the output.
1010         h264_encoder->add_audio(frame_pts_int, move(samples_out));
1011 }
1012
1013 void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
1014 {
1015         GLuint vao;
1016         glGenVertexArrays(1, &vao);
1017         check_error();
1018
1019         glBindVertexArray(vao);
1020         check_error();
1021
1022         // Extract Cb/Cr.
1023         GLuint fbo = resource_pool->create_fbo(dst_tex);
1024         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
1025         glViewport(0, 0, WIDTH/2, HEIGHT/2);
1026         check_error();
1027
1028         glUseProgram(cbcr_program_num);
1029         check_error();
1030
1031         glActiveTexture(GL_TEXTURE0);
1032         check_error();
1033         glBindTexture(GL_TEXTURE_2D, src_tex);
1034         check_error();
1035         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1036         check_error();
1037         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1038         check_error();
1039         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1040         check_error();
1041
1042         float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
1043         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
1044
1045         glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
1046         check_error();
1047
1048         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1049                 glEnableVertexAttribArray(attr_index);
1050                 check_error();
1051                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
1052                 check_error();
1053         }
1054
1055         glDrawArrays(GL_TRIANGLES, 0, 3);
1056         check_error();
1057
1058         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
1059                 glDisableVertexAttribArray(attr_index);
1060                 check_error();
1061         }
1062
1063         glUseProgram(0);
1064         check_error();
1065         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1066         check_error();
1067
1068         resource_pool->release_fbo(fbo);
1069         glDeleteVertexArrays(1, &vao);
1070 }
1071
1072 void Mixer::release_display_frame(DisplayFrame *frame)
1073 {
1074         for (GLuint texnum : frame->temp_textures) {
1075                 resource_pool->release_2d_texture(texnum);
1076         }
1077         frame->temp_textures.clear();
1078         frame->ready_fence.reset();
1079         frame->input_frames.clear();
1080 }
1081
1082 void Mixer::start()
1083 {
1084         mixer_thread = thread(&Mixer::thread_func, this);
1085         audio_thread = thread(&Mixer::audio_thread_func, this);
1086 }
1087
1088 void Mixer::quit()
1089 {
1090         should_quit = true;
1091         mixer_thread.join();
1092         audio_thread.join();
1093 }
1094
1095 void Mixer::transition_clicked(int transition_num)
1096 {
1097         theme->transition_clicked(transition_num, pts());
1098 }
1099
1100 void Mixer::channel_clicked(int preview_num)
1101 {
1102         theme->channel_clicked(preview_num);
1103 }
1104
1105 void Mixer::reset_meters()
1106 {
1107         peak_resampler.reset();
1108         peak = 0.0f;
1109         r128.reset();
1110         r128.integr_start();
1111         correlation.reset();
1112 }
1113
1114 void Mixer::start_mode_scanning(unsigned card_index)
1115 {
1116         assert(card_index < num_cards);
1117         if (is_mode_scanning[card_index]) {
1118                 return;
1119         }
1120         is_mode_scanning[card_index] = true;
1121         mode_scanlist[card_index].clear();
1122         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1123                 mode_scanlist[card_index].push_back(mode.first);
1124         }
1125         assert(!mode_scanlist[card_index].empty());
1126         mode_scanlist_index[card_index] = 0;
1127         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1128         clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]);
1129 }
1130
1131 Mixer::OutputChannel::~OutputChannel()
1132 {
1133         if (has_current_frame) {
1134                 parent->release_display_frame(&current_frame);
1135         }
1136         if (has_ready_frame) {
1137                 parent->release_display_frame(&ready_frame);
1138         }
1139 }
1140
1141 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1142 {
1143         // Store this frame for display. Remove the ready frame if any
1144         // (it was seemingly never used).
1145         {
1146                 unique_lock<mutex> lock(frame_mutex);
1147                 if (has_ready_frame) {
1148                         parent->release_display_frame(&ready_frame);
1149                 }
1150                 ready_frame = frame;
1151                 has_ready_frame = true;
1152         }
1153
1154         if (has_new_frame_ready_callback) {
1155                 new_frame_ready_callback();
1156         }
1157 }
1158
1159 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1160 {
1161         unique_lock<mutex> lock(frame_mutex);
1162         if (!has_current_frame && !has_ready_frame) {
1163                 return false;
1164         }
1165
1166         if (has_current_frame && has_ready_frame) {
1167                 // We have a new ready frame. Toss the current one.
1168                 parent->release_display_frame(&current_frame);
1169                 has_current_frame = false;
1170         }
1171         if (has_ready_frame) {
1172                 assert(!has_current_frame);
1173                 current_frame = ready_frame;
1174                 ready_frame.ready_fence.reset();  // Drop the refcount.
1175                 ready_frame.input_frames.clear();  // Drop the refcounts.
1176                 has_current_frame = true;
1177                 has_ready_frame = false;
1178         }
1179
1180         *frame = current_frame;
1181         return true;
1182 }
1183
1184 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1185 {
1186         new_frame_ready_callback = callback;
1187         has_new_frame_ready_callback = true;
1188 }