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