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