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 <pthread.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <sys/resource.h>
  18 #include <algorithm>
  19 #include <chrono>
  20 #include <condition_variable>
  21 #include <cstddef>
  22 #include <cstdint>
  23 #include <memory>
  24 #include <mutex>
  25 #include <ratio>
  26 #include <string>
  27 #include <thread>
  28 #include <utility>
  29 #include <vector>
  30
  31 #include "DeckLinkAPI.h"
  32 #include "LinuxCOM.h"
  33 #include "alsa_output.h"
  34 #include "bmusb/bmusb.h"
  35 #include "bmusb/fake_capture.h"
  36 #include "chroma_subsampler.h"
  37 #include "context.h"
  38 #include "decklink_capture.h"
  39 #include "decklink_output.h"
  40 #include "defs.h"
  41 #include "disk_space_estimator.h"
  42 #include "flags.h"
  43 #include "input_mapping.h"
  44 #include "pbo_frame_allocator.h"
  45 #include "ref_counted_gl_sync.h"
  46 #include "resampling_queue.h"
  47 #include "timebase.h"
  48 #include "timecode_renderer.h"
  49 #include "v210_converter.h"
  50 #include "video_encoder.h"
  51
  52 class IDeckLink;
  53 class QOpenGLContext;
  54
  55 using namespace movit;
  56 using namespace std;
  57 using namespace std::chrono;
  58 using namespace std::placeholders;
  59 using namespace bmusb;
  60
  61 Mixer *global_mixer = nullptr;
  62 bool uses_mlock = false;
  63
  64 namespace {
  65
  66 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  67 {
  68         if (interlaced) {
  69                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  70                         input_state->buffered_frames[card_index][frame_num] =
  71                                 input_state->buffered_frames[card_index][frame_num - 1];
  72                 }
  73                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  74         } else {
  75                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  76                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  77                 }
  78         }
  79 }
  80
  81 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned v210_width)
  82 {
  83         bool first;
  84         if (global_flags.ten_bit_input) {
  85                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  86         } else {
  87                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
  88         }
  89
  90         if (first ||
  91             width != userdata->last_width[field] ||
  92             height != userdata->last_height[field]) {
  93                 // We changed resolution since last use of this texture, so we need to create
  94                 // a new object. Note that this each card has its own PBOFrameAllocator,
  95                 // we don't need to worry about these flip-flopping between resolutions.
  96                 if (global_flags.ten_bit_input) {
  97                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
  98                         check_error();
  99                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 100                         check_error();
 101                 } else {
 102                         size_t cbcr_width = width / 2;
 103
 104                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 105                         check_error();
 106                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 107                         check_error();
 108                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 109                         check_error();
 110                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 111                         check_error();
 112                 }
 113                 userdata->last_width[field] = width;
 114                 userdata->last_height[field] = height;
 115         }
 116         if (global_flags.ten_bit_input &&
 117             (first || v210_width != userdata->last_v210_width[field])) {
 118                 // Same as above; we need to recreate the texture.
 119                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 120                 check_error();
 121                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 122                 check_error();
 123                 userdata->last_v210_width[field] = v210_width;
 124         }
 125 }
 126
 127 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 128 {
 129         if (interlaced_stride) {
 130                 stride *= 2;
 131         }
 132         if (global_flags.flush_pbos) {
 133                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 134                 check_error();
 135         }
 136
 137         glBindTexture(GL_TEXTURE_2D, tex);
 138         check_error();
 139         if (interlaced_stride) {
 140                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 141                 check_error();
 142         } else {
 143                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 144                 check_error();
 145         }
 146
 147         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 148         check_error();
 149         glBindTexture(GL_TEXTURE_2D, 0);
 150         check_error();
 151         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 152         check_error();
 153 }
 154
 155 }  // namespace
 156
 157 void QueueLengthPolicy::update_policy(unsigned queue_length)
 158 {
 159         if (queue_length == 0) {  // Starvation.
 160                 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
 161                         ++safe_queue_length;
 162                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
 163                                 card_index, safe_queue_length);
 164                 }
 165                 frames_with_at_least_one = 0;
 166                 been_at_safe_point_since_last_starvation = false;
 167                 return;
 168         }
 169         if (queue_length >= safe_queue_length) {
 170                 been_at_safe_point_since_last_starvation = true;
 171         }
 172         if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
 173                 --safe_queue_length;
 174                 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
 175                         card_index, safe_queue_length);
 176                 frames_with_at_least_one = 0;
 177         }
 178 }
 179
 180 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 181         : httpd(),
 182           num_cards(num_cards),
 183           mixer_surface(create_surface(format)),
 184           h264_encoder_surface(create_surface(format)),
 185           decklink_output_surface(create_surface(format)),
 186           audio_mixer(num_cards)
 187 {
 188         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 189         check_error();
 190
 191         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 192         // will be halved when sampling them, and we need to compensate here.
 193         movit_texel_subpixel_precision /= 2.0;
 194
 195         resource_pool.reset(new ResourcePool);
 196         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 197         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 198                 output_channel[i].parent = this;
 199                 output_channel[i].channel = i;
 200         }
 201
 202         ImageFormat inout_format;
 203         inout_format.color_space = COLORSPACE_sRGB;
 204         inout_format.gamma_curve = GAMMA_sRGB;
 205
 206         // Matches the 4:2:0 format created by the main chain.
 207         YCbCrFormat ycbcr_format;
 208         ycbcr_format.chroma_subsampling_x = 2;
 209         ycbcr_format.chroma_subsampling_y = 2;
 210         if (global_flags.ycbcr_rec709_coefficients) {
 211                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 212         } else {
 213                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 214         }
 215         ycbcr_format.full_range = false;
 216         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 217         ycbcr_format.cb_x_position = 0.0f;
 218         ycbcr_format.cr_x_position = 0.0f;
 219         ycbcr_format.cb_y_position = 0.5f;
 220         ycbcr_format.cr_y_position = 0.5f;
 221
 222         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 223         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 224         check_error();
 225         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 226         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 227         display_chain->add_input(display_input);
 228         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 229         display_chain->set_dither_bits(0);  // Don't bother.
 230         display_chain->finalize();
 231
 232         video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
 233
 234         // Start listening for clients only once VideoEncoder has written its header, if any.
 235         httpd.start(9095);
 236
 237         // First try initializing the then PCI devices, then USB, then
 238         // fill up with fake cards until we have the desired number of cards.
 239         unsigned num_pci_devices = 0;
 240         unsigned card_index = 0;
 241
 242         {
 243                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 244                 if (decklink_iterator != nullptr) {
 245                         for ( ; card_index < num_cards; ++card_index) {
 246                                 IDeckLink *decklink;
 247                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 248                                         break;
 249                                 }
 250
 251                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 252                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 253                                 output->set_device(decklink);
 254                                 configure_card(card_index, capture, /*is_fake_capture=*/false, output);
 255                                 ++num_pci_devices;
 256                         }
 257                         decklink_iterator->Release();
 258                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 259                 } else {
 260                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 261                 }
 262         }
 263
 264         unsigned num_usb_devices = BMUSBCapture::num_cards();
 265         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 266                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 267                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 268                 configure_card(card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
 269         }
 270         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 271
 272         unsigned num_fake_cards = 0;
 273         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 274                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 275                 configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
 276         }
 277
 278         if (num_fake_cards > 0) {
 279                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 280         }
 281
 282         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 283         BMUSBCapture::start_bm_thread();
 284
 285         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 286                 cards[card_index].queue_length_policy.reset(card_index);
 287         }
 288
 289         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 290
 291         if (global_flags.ten_bit_input) {
 292                 if (!v210Converter::has_hardware_support()) {
 293                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 294                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 295                         exit(1);
 296                 }
 297                 v210_converter.reset(new v210Converter());
 298
 299                 // These are all the widths listed in the Blackmagic SDK documentation
 300                 // (section 2.7.3, “Display Modes”).
 301                 v210_converter->precompile_shader(720);
 302                 v210_converter->precompile_shader(1280);
 303                 v210_converter->precompile_shader(1920);
 304                 v210_converter->precompile_shader(2048);
 305                 v210_converter->precompile_shader(3840);
 306                 v210_converter->precompile_shader(4096);
 307         }
 308
 309         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 310         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 311         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 312
 313         if (global_flags.enable_alsa_output) {
 314                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 315         }
 316         if (global_flags.output_card != -1) {
 317                 desired_output_card_index = global_flags.output_card;
 318                 set_output_card_internal(global_flags.output_card);
 319         }
 320 }
 321
 322 Mixer::~Mixer()
 323 {
 324         BMUSBCapture::stop_bm_thread();
 325
 326         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 327                 {
 328                         unique_lock<mutex> lock(card_mutex);
 329                         cards[card_index].should_quit = true;  // Unblock thread.
 330                         cards[card_index].new_frames_changed.notify_all();
 331                 }
 332                 cards[card_index].capture->stop_dequeue_thread();
 333                 if (cards[card_index].output) {
 334                         cards[card_index].output->end_output();
 335                         cards[card_index].output.reset();
 336                 }
 337         }
 338
 339         video_encoder.reset(nullptr);
 340 }
 341
 342 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output)
 343 {
 344         printf("Configuring card %d...\n", card_index);
 345
 346         CaptureCard *card = &cards[card_index];
 347         if (card->capture != nullptr) {
 348                 card->capture->stop_dequeue_thread();
 349         }
 350         card->capture.reset(capture);
 351         card->is_fake_capture = is_fake_capture;
 352         if (card->output.get() != output) {
 353                 card->output.reset(output);
 354         }
 355         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 356         if (card->frame_allocator == nullptr) {
 357                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 358         }
 359         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 360         if (card->surface == nullptr) {
 361                 card->surface = create_surface_with_same_format(mixer_surface);
 362         }
 363         while (!card->new_frames.empty()) card->new_frames.pop_front();
 364         card->last_timecode = -1;
 365         card->capture->set_pixel_format(global_flags.ten_bit_input ? PixelFormat_10BitYCbCr : PixelFormat_8BitYCbCr);
 366         card->capture->configure_card();
 367
 368         // NOTE: start_bm_capture() happens in thread_func().
 369
 370         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 371         audio_mixer.reset_resampler(device);
 372         audio_mixer.set_display_name(device, card->capture->get_description());
 373         audio_mixer.trigger_state_changed_callback();
 374 }
 375
 376 void Mixer::set_output_card_internal(int card_index)
 377 {
 378         // We don't really need to take card_mutex, since we're in the mixer
 379         // thread and don't mess with any queues (which is the only thing that happens
 380         // from other threads), but it's probably the safest in the long run.
 381         unique_lock<mutex> lock(card_mutex);
 382         if (output_card_index != -1) {
 383                 // Switch the old card from output to input.
 384                 CaptureCard *old_card = &cards[output_card_index];
 385                 old_card->output->end_output();
 386
 387                 // Stop the fake card that we put into place.
 388                 // This needs to _not_ happen under the mutex, to avoid deadlock
 389                 // (delivering the last frame needs to take the mutex).
 390                 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
 391                 lock.unlock();
 392                 fake_capture->stop_dequeue_thread();
 393                 lock.lock();
 394                 old_card->capture = move(old_card->parked_capture);
 395                 old_card->is_fake_capture = false;
 396                 old_card->capture->start_bm_capture();
 397         }
 398         if (card_index != -1) {
 399                 CaptureCard *card = &cards[card_index];
 400                 bmusb::CaptureInterface *capture = card->capture.get();
 401                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 402                 // several seconds to complete (blocking on DisableVideoInput);
 403                 // see if we can maybe do it asynchronously.
 404                 lock.unlock();
 405                 capture->stop_dequeue_thread();
 406                 lock.lock();
 407                 card->parked_capture = move(card->capture);
 408                 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 409                 configure_card(card_index, fake_capture, /*is_fake_capture=*/true, card->output.release());
 410                 card->queue_length_policy.reset(card_index);
 411                 card->capture->start_bm_capture();
 412                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 413                 card->output->start_output(desired_output_video_mode, pts_int);
 414         }
 415         output_card_index = card_index;
 416 }
 417
 418 namespace {
 419
 420 int unwrap_timecode(uint16_t current_wrapped, int last)
 421 {
 422         uint16_t last_wrapped = last & 0xffff;
 423         if (current_wrapped > last_wrapped) {
 424                 return (last & ~0xffff) | current_wrapped;
 425         } else {
 426                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 427         }
 428 }
 429
 430 }  // namespace
 431
 432 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 433                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 434                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 435 {
 436         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 437         CaptureCard *card = &cards[card_index];
 438
 439         if (is_mode_scanning[card_index]) {
 440                 if (video_format.has_signal) {
 441                         // Found a stable signal, so stop scanning.
 442                         is_mode_scanning[card_index] = false;
 443                 } else {
 444                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 445                         steady_clock::time_point now = steady_clock::now();
 446                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 447                         if (sec_since_last_switch > switch_time_s) {
 448                                 // It isn't this mode; try the next one.
 449                                 mode_scanlist_index[card_index]++;
 450                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 451                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 452                                 last_mode_scan_change[card_index] = now;
 453                         }
 454                 }
 455         }
 456
 457         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 458         assert(frame_length > 0);
 459
 460         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;
 461         if (num_samples > OUTPUT_FREQUENCY / 10) {
 462                 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",
 463                         card_index, int(audio_frame.len), int(audio_offset),
 464                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 465                 if (video_frame.owner) {
 466                         video_frame.owner->release_frame(video_frame);
 467                 }
 468                 if (audio_frame.owner) {
 469                         audio_frame.owner->release_frame(audio_frame);
 470                 }
 471                 return;
 472         }
 473
 474         int dropped_frames = 0;
 475         if (card->last_timecode != -1) {
 476                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 477         }
 478
 479         // Number of samples per frame if we need to insert silence.
 480         // (Could be nonintegral, but resampling will save us then.)
 481         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 482
 483         if (dropped_frames > MAX_FPS * 2) {
 484                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 485                         card_index, card->last_timecode, timecode);
 486                 audio_mixer.reset_resampler(device);
 487                 dropped_frames = 0;
 488         } else if (dropped_frames > 0) {
 489                 // Insert silence as needed.
 490                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 491                         card_index, dropped_frames, timecode);
 492
 493                 bool success;
 494                 do {
 495                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 496                 } while (!success);
 497         }
 498
 499         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 500
 501         // Done with the audio, so release it.
 502         if (audio_frame.owner) {
 503                 audio_frame.owner->release_frame(audio_frame);
 504         }
 505
 506         card->last_timecode = timecode;
 507
 508         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 509         if (video_frame.len - video_offset == 0 ||
 510             video_frame.len - video_offset != expected_length) {
 511                 if (video_frame.len != 0) {
 512                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 513                                 card_index, video_frame.len - video_offset, expected_length);
 514                 }
 515                 if (video_frame.owner) {
 516                         video_frame.owner->release_frame(video_frame);
 517                 }
 518
 519                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 520                 // so that pts can go up accordingly.
 521                 {
 522                         unique_lock<mutex> lock(card_mutex);
 523                         CaptureCard::NewFrame new_frame;
 524                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 525                         new_frame.length = frame_length;
 526                         new_frame.interlaced = false;
 527                         new_frame.dropped_frames = dropped_frames;
 528                         new_frame.received_timestamp = video_frame.received_timestamp;
 529                         card->new_frames.push_back(move(new_frame));
 530                         card->new_frames_changed.notify_all();
 531                 }
 532                 return;
 533         }
 534
 535         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 536
 537         unsigned num_fields = video_format.interlaced ? 2 : 1;
 538         steady_clock::time_point frame_upload_start;
 539         bool interlaced_stride = false;
 540         if (video_format.interlaced) {
 541                 // Send the two fields along as separate frames; the other side will need to add
 542                 // a deinterlacer to actually get this right.
 543                 assert(video_format.height % 2 == 0);
 544                 video_format.height /= 2;
 545                 assert(frame_length % 2 == 0);
 546                 frame_length /= 2;
 547                 num_fields = 2;
 548                 if (video_format.second_field_start == 1) {
 549                         interlaced_stride = true;
 550                 }
 551                 frame_upload_start = steady_clock::now();
 552         }
 553         userdata->last_interlaced = video_format.interlaced;
 554         userdata->last_has_signal = video_format.has_signal;
 555         userdata->last_is_connected = video_format.is_connected;
 556         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 557         userdata->last_frame_rate_den = video_format.frame_rate_den;
 558         RefCountedFrame frame(video_frame);
 559
 560         // Upload the textures.
 561         const size_t cbcr_width = video_format.width / 2;
 562         const size_t cbcr_offset = video_offset / 2;
 563         const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 564
 565         for (unsigned field = 0; field < num_fields; ++field) {
 566                 // Put the actual texture upload in a lambda that is executed in the main thread.
 567                 // It is entirely possible to do this in the same thread (and it might even be
 568                 // faster, depending on the GPU and driver), but it appears to be trickling
 569                 // driver bugs very easily.
 570                 //
 571                 // Note that this means we must hold on to the actual frame data in <userdata>
 572                 // until the upload command is run, but we hold on to <frame> much longer than that
 573                 // (in fact, all the way until we no longer use the texture in rendering).
 574                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 575                         unsigned field_start_line;
 576                         if (field == 1) {
 577                                 field_start_line = video_format.second_field_start;
 578                         } else {
 579                                 field_start_line = video_format.extra_lines_top;
 580                         }
 581
 582                         // For 8-bit input, v210_width will be nonsensical but not used.
 583                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 584                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 585
 586                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 587                         check_error();
 588
 589                         if (global_flags.ten_bit_input) {
 590                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 591                                 upload_texture(userdata->tex_v210[field], v210_width, video_format.height, video_format.stride, interlaced_stride, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, field_start);
 592                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 593                         } else {
 594                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 595                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 596
 597                                 // Make up our own strides, since we are interleaving.
 598                                 upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start);
 599                                 upload_texture(userdata->tex_cbcr[field], cbcr_width, video_format.height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
 600                         }
 601
 602                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 603                         check_error();
 604                 };
 605
 606                 if (field == 1) {
 607                         // Don't upload the second field as fast as we can; wait until
 608                         // the field time has approximately passed. (Otherwise, we could
 609                         // get timing jitter against the other sources, and possibly also
 610                         // against the video display, although the latter is not as critical.)
 611                         // This requires our system clock to be reasonably close to the
 612                         // video clock, but that's not an unreasonable assumption.
 613                         steady_clock::time_point second_field_start = frame_upload_start +
 614                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 615                         this_thread::sleep_until(second_field_start);
 616                 }
 617
 618                 {
 619                         unique_lock<mutex> lock(card_mutex);
 620                         CaptureCard::NewFrame new_frame;
 621                         new_frame.frame = frame;
 622                         new_frame.length = frame_length;
 623                         new_frame.field = field;
 624                         new_frame.interlaced = video_format.interlaced;
 625                         new_frame.upload_func = upload_func;
 626                         new_frame.dropped_frames = dropped_frames;
 627                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 628                         card->new_frames.push_back(move(new_frame));
 629                         card->new_frames_changed.notify_all();
 630                 }
 631         }
 632 }
 633
 634 void Mixer::bm_hotplug_add(libusb_device *dev)
 635 {
 636         lock_guard<mutex> lock(hotplug_mutex);
 637         hotplugged_cards.push_back(dev);
 638 }
 639
 640 void Mixer::bm_hotplug_remove(unsigned card_index)
 641 {
 642         cards[card_index].new_frames_changed.notify_all();
 643 }
 644
 645 void Mixer::thread_func()
 646 {
 647         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 648
 649         eglBindAPI(EGL_OPENGL_API);
 650         QOpenGLContext *context = create_context(mixer_surface);
 651         if (!make_current(context, mixer_surface)) {
 652                 printf("oops\n");
 653                 exit(1);
 654         }
 655
 656         // Start the actual capture. (We don't want to do it before we're actually ready
 657         // to process output frames.)
 658         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 659                 if (int(card_index) != output_card_index) {
 660                         cards[card_index].capture->start_bm_capture();
 661                 }
 662         }
 663
 664         steady_clock::time_point start, now;
 665         start = steady_clock::now();
 666
 667         int stats_dropped_frames = 0;
 668
 669         while (!should_quit) {
 670                 if (desired_output_card_index != output_card_index) {
 671                         set_output_card_internal(desired_output_card_index);
 672                 }
 673                 if (output_card_index != -1 &&
 674                     desired_output_video_mode != output_video_mode) {
 675                         DeckLinkOutput *output = cards[output_card_index].output.get();
 676                         output->end_output();
 677                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 678                         output->start_output(desired_output_video_mode, pts_int);
 679                 }
 680
 681                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 682                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 683
 684                 bool master_card_is_output;
 685                 unsigned master_card_index;
 686                 if (output_card_index != -1) {
 687                         master_card_is_output = true;
 688                         master_card_index = output_card_index;
 689                 } else {
 690                         master_card_is_output = false;
 691                         master_card_index = theme->map_signal(master_clock_channel);
 692                         assert(master_card_index < num_cards);
 693                 }
 694
 695                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 696                 schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration, output_frame_info.is_preroll, output_frame_info.frame_timestamp);
 697                 stats_dropped_frames += output_frame_info.dropped_frames;
 698
 699                 handle_hotplugged_cards();
 700
 701                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 702                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 703                                 continue;
 704                         }
 705                         if (new_frames[card_index].frame->len == 0) {
 706                                 ++new_frames[card_index].dropped_frames;
 707                         }
 708                         if (new_frames[card_index].dropped_frames > 0) {
 709                                 printf("Card %u dropped %d frames before this\n",
 710                                         card_index, int(new_frames[card_index].dropped_frames));
 711                         }
 712                 }
 713
 714                 // If the first card is reporting a corrupted or otherwise dropped frame,
 715                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 716                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 717                         ++stats_dropped_frames;
 718                         pts_int += new_frames[master_card_index].length;
 719                         continue;
 720                 }
 721
 722                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 723                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 724                                 continue;
 725
 726                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 727                         assert(new_frame->frame != nullptr);
 728                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 729                         check_error();
 730
 731                         // The new texture might need uploading before use.
 732                         if (new_frame->upload_func) {
 733                                 new_frame->upload_func();
 734                                 new_frame->upload_func = nullptr;
 735                         }
 736                 }
 737
 738                 int64_t frame_duration = output_frame_info.frame_duration;
 739                 render_one_frame(frame_duration);
 740                 ++frame_num;
 741                 pts_int += frame_duration;
 742
 743                 now = steady_clock::now();
 744                 double elapsed = duration<double>(now - start).count();
 745                 if (frame_num % 100 == 0) {
 746                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 747                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 748                                 1e3 * elapsed / frame_num);
 749                 //      chain->print_phase_timing();
 750
 751                         // Check our memory usage, to see if we are close to our mlockall()
 752                         // limit (if at all set).
 753                         rusage used;
 754                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 755                                 perror("getrusage(RUSAGE_SELF)");
 756                                 assert(false);
 757                         }
 758
 759                         if (uses_mlock) {
 760                                 rlimit limit;
 761                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 762                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 763                                         assert(false);
 764                                 }
 765
 766                                 if (limit.rlim_cur == 0) {
 767                                         printf(", using %ld MB memory (locked)",
 768                                                 long(used.ru_maxrss / 1024));
 769                                 } else {
 770                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 771                                                 long(used.ru_maxrss / 1024),
 772                                                 long(limit.rlim_cur / 1048576),
 773                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 774                                 }
 775                         } else {
 776                                 printf(", using %ld MB memory (not locked)",
 777                                         long(used.ru_maxrss / 1024));
 778                         }
 779
 780                         printf("\n");
 781                 }
 782
 783
 784                 if (should_cut.exchange(false)) {  // Test and clear.
 785                         video_encoder->do_cut(frame_num);
 786                 }
 787
 788 #if 0
 789                 // Reset every 100 frames, so that local variations in frame times
 790                 // (especially for the first few frames, when the shaders are
 791                 // compiled etc.) don't make it hard to measure for the entire
 792                 // remaining duration of the program.
 793                 if (frame == 10000) {
 794                         frame = 0;
 795                         start = now;
 796                 }
 797 #endif
 798                 check_error();
 799         }
 800
 801         resource_pool->clean_context();
 802 }
 803
 804 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 805 {
 806         if (output_card_index != -1) {
 807                 // The output card (ie., cards[output_card_index].output) is the master clock,
 808                 // so no input card (ie., cards[card_index].capture) is.
 809                 return false;
 810         }
 811         return (card_index == master_card_index);
 812 }
 813
 814 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 815 {
 816         // Count the number of frames in the queue, including any frames
 817         // we dropped. It's hard to know exactly how we should deal with
 818         // dropped (corrupted) input frames; they don't help our goal of
 819         // avoiding starvation, but they still add to the problem of latency.
 820         // Since dropped frames is going to mean a bump in the signal anyway,
 821         // we err on the side of having more stable latency instead.
 822         unsigned queue_length = 0;
 823         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 824                 queue_length += frame.dropped_frames + 1;
 825         }
 826         card->queue_length_policy.update_policy(queue_length);
 827
 828         // If needed, drop frames until the queue is below the safe limit.
 829         // We prefer to drop from the head, because all else being equal,
 830         // we'd like more recent frames (less latency).
 831         unsigned dropped_frames = 0;
 832         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 833                 assert(!card->new_frames.empty());
 834                 assert(queue_length > card->new_frames.front().dropped_frames);
 835                 queue_length -= card->new_frames.front().dropped_frames;
 836
 837                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 838                         // No need to drop anything.
 839                         break;
 840                 }
 841
 842                 card->new_frames.pop_front();
 843                 card->new_frames_changed.notify_all();
 844                 --queue_length;
 845                 ++dropped_frames;
 846         }
 847
 848         if (dropped_frames > 0) {
 849                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 850                         card_index, dropped_frames);
 851         }
 852 }
 853
 854
 855 Mixer::OutputFrameInfo Mixer::get_one_frame_from_each_card(unsigned master_card_index, bool master_card_is_output, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS])
 856 {
 857         OutputFrameInfo output_frame_info;
 858 start:
 859         unique_lock<mutex> lock(card_mutex, defer_lock);
 860         if (master_card_is_output) {
 861                 // Clocked to the output, so wait for it to be ready for the next frame.
 862                 cards[master_card_index].output->wait_for_frame(pts_int, &output_frame_info.dropped_frames, &output_frame_info.frame_duration, &output_frame_info.is_preroll, &output_frame_info.frame_timestamp);
 863                 lock.lock();
 864         } else {
 865                 // Wait for the master card to have a new frame.
 866                 // TODO: Add a timeout.
 867                 output_frame_info.is_preroll = false;
 868                 lock.lock();
 869                 cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
 870         }
 871
 872         if (master_card_is_output) {
 873                 handle_hotplugged_cards();
 874         } else if (cards[master_card_index].new_frames.empty()) {
 875                 // We were woken up, but not due to a new frame. Deal with it
 876                 // and then restart.
 877                 assert(cards[master_card_index].capture->get_disconnected());
 878                 handle_hotplugged_cards();
 879                 goto start;
 880         }
 881
 882         if (!master_card_is_output) {
 883                 output_frame_info.frame_timestamp =
 884                         cards[master_card_index].new_frames.front().received_timestamp;
 885         }
 886
 887         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 888                 CaptureCard *card = &cards[card_index];
 889                 if (input_card_is_master_clock(card_index, master_card_index)) {
 890                         // We don't use the queue length policy for the master card,
 891                         // but we will if it stops being the master. Thus, clear out
 892                         // the policy in case we switch in the future.
 893                         card->queue_length_policy.reset(card_index);
 894                         assert(!card->new_frames.empty());
 895                 } else {
 896                         trim_queue(card, card_index);
 897                 }
 898                 if (!card->new_frames.empty()) {
 899                         new_frames[card_index] = move(card->new_frames.front());
 900                         has_new_frame[card_index] = true;
 901                         card->new_frames.pop_front();
 902                         card->new_frames_changed.notify_all();
 903                 }
 904         }
 905
 906         if (!master_card_is_output) {
 907                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 908                 output_frame_info.frame_duration = new_frames[master_card_index].length;
 909         }
 910
 911         // This might get off by a fractional sample when changing master card
 912         // between ones with different frame rates, but that's fine.
 913         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 914         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 915         fractional_samples = num_samples_times_timebase % TIMEBASE;
 916         assert(output_frame_info.num_samples >= 0);
 917
 918         return output_frame_info;
 919 }
 920
 921 void Mixer::handle_hotplugged_cards()
 922 {
 923         // Check for cards that have been disconnected since last frame.
 924         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 925                 CaptureCard *card = &cards[card_index];
 926                 if (card->capture->get_disconnected()) {
 927                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 928                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 929                         configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
 930                         card->queue_length_policy.reset(card_index);
 931                         card->capture->start_bm_capture();
 932                 }
 933         }
 934
 935         // Check for cards that have been connected since last frame.
 936         vector<libusb_device *> hotplugged_cards_copy;
 937         {
 938                 lock_guard<mutex> lock(hotplug_mutex);
 939                 swap(hotplugged_cards, hotplugged_cards_copy);
 940         }
 941         for (libusb_device *new_dev : hotplugged_cards_copy) {
 942                 // Look for a fake capture card where we can stick this in.
 943                 int free_card_index = -1;
 944                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 945                         if (cards[card_index].is_fake_capture) {
 946                                 free_card_index = card_index;
 947                                 break;
 948                         }
 949                 }
 950
 951                 if (free_card_index == -1) {
 952                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 953                         libusb_unref_device(new_dev);
 954                 } else {
 955                         // BMUSBCapture takes ownership.
 956                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 957                         CaptureCard *card = &cards[free_card_index];
 958                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 959                         configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
 960                         card->queue_length_policy.reset(free_card_index);
 961                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 962                         capture->start_bm_capture();
 963                 }
 964         }
 965 }
 966
 967
 968 void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame, bool is_preroll, steady_clock::time_point frame_timestamp)
 969 {
 970         // Resample the audio as needed, including from previously dropped frames.
 971         assert(num_cards > 0);
 972         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 973                 const bool dropped_frame = (frame_num != dropped_frames);
 974                 {
 975                         // Signal to the audio thread to process this frame.
 976                         // Note that if the frame is a dropped frame, we signal that
 977                         // we don't want to use this frame as base for adjusting
 978                         // the resampler rate. The reason for this is that the timing
 979                         // of these frames is often way too late; they typically don't
 980                         // “arrive” before we synthesize them. Thus, we could end up
 981                         // in a situation where we have inserted e.g. five audio frames
 982                         // into the queue before we then start pulling five of them
 983                         // back out. This makes ResamplingQueue overestimate the delay,
 984                         // causing undue resampler changes. (We _do_ use the last,
 985                         // non-dropped frame; perhaps we should just discard that as well,
 986                         // since dropped frames are expected to be rare, and it might be
 987                         // better to just wait until we have a slightly more normal situation).
 988                         unique_lock<mutex> lock(audio_mutex);
 989                         bool adjust_rate = !dropped_frame && !is_preroll;
 990                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
 991                         audio_task_queue_changed.notify_one();
 992                 }
 993                 if (dropped_frame) {
 994                         // For dropped frames, increase the pts. Note that if the format changed
 995                         // in the meantime, we have no way of detecting that; we just have to
 996                         // assume the frame length is always the same.
 997                         pts_int += length_per_frame;
 998                 }
 999         }
1000 }
1001
1002 void Mixer::render_one_frame(int64_t duration)
1003 {
1004         // Determine the time code for this frame before we start rendering.
1005         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1006         if (display_timecode_on_stdout) {
1007                 printf("Timecode: '%s'\n", timecode_text.c_str());
1008         }
1009
1010         // Get the main chain from the theme, and set its state immediately.
1011         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1012         EffectChain *chain = theme_main_chain.chain;
1013         theme_main_chain.setup_chain();
1014         //theme_main_chain.chain->enable_phase_timing(true);
1015
1016         // If HDMI/SDI output is active and the user has requested auto mode,
1017         // its mode overrides the existing Y'CbCr setting for the chain.
1018         YCbCrLumaCoefficients ycbcr_output_coefficients;
1019         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1020                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1021         } else {
1022                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1023         }
1024
1025         // TODO: Reduce the duplication against theme.cpp.
1026         YCbCrFormat output_ycbcr_format;
1027         output_ycbcr_format.chroma_subsampling_x = 1;
1028         output_ycbcr_format.chroma_subsampling_y = 1;
1029         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1030         output_ycbcr_format.full_range = false;
1031         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1032         chain->change_ycbcr_output_format(output_ycbcr_format);
1033
1034         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1035         GLuint y_tex, cbcr_tex;
1036         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1037         assert(got_frame);
1038
1039         // Render main chain. We take an extra copy of the created outputs,
1040         // so that we can display it back to the screen later (it's less memory
1041         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1042         // Ideally, we'd like to avoid taking copies and just use the main textures
1043         // for display as well, but if they're used for zero-copy Quick Sync encoding
1044         // (the default case), they're just views into VA-API memory and must be
1045         // unmapped during encoding, so we can't use them for display, unfortunately.
1046         GLuint cbcr_full_tex, cbcr_copy_tex, y_copy_tex;
1047         if (global_flags.x264_bit_depth > 8) {
1048                 cbcr_full_tex = resource_pool->create_2d_texture(GL_RG16, global_flags.width, global_flags.height);
1049                 y_copy_tex = resource_pool->create_2d_texture(GL_R16, global_flags.width, global_flags.height);
1050                 cbcr_copy_tex = resource_pool->create_2d_texture(GL_RG16, global_flags.width / 2, global_flags.height / 2);
1051         } else {
1052                 cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
1053                 y_copy_tex = resource_pool->create_2d_texture(GL_R8, global_flags.width, global_flags.height);
1054                 cbcr_copy_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width / 2, global_flags.height / 2);
1055         }
1056         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1057         check_error();
1058         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1059
1060         if (display_timecode_in_stream) {
1061                 // Render the timecode on top.
1062                 timecode_renderer->render_timecode(fbo, timecode_text);
1063         }
1064
1065         resource_pool->release_fbo(fbo);
1066
1067         chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1068         if (output_card_index != -1) {
1069                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1070         }
1071         resource_pool->release_2d_texture(cbcr_full_tex);
1072
1073         // Set the right state for the Y' and CbCr copies.
1074         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1075         glBindTexture(GL_TEXTURE_2D, y_copy_tex);
1076         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1077         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1078         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1079
1080         glBindTexture(GL_TEXTURE_2D, cbcr_copy_tex);
1081         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1082         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1083         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1084
1085         RefCountedGLsync fence = video_encoder->end_frame();
1086
1087         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1088         // It owns y_copy_tex and cbcr_copy_tex now.
1089         DisplayFrame live_frame;
1090         live_frame.chain = display_chain.get();
1091         live_frame.setup_chain = [this, y_copy_tex, cbcr_copy_tex]{
1092                 display_input->set_texture_num(0, y_copy_tex);
1093                 display_input->set_texture_num(1, cbcr_copy_tex);
1094         };
1095         live_frame.ready_fence = fence;
1096         live_frame.input_frames = {};
1097         live_frame.temp_textures = { y_copy_tex, cbcr_copy_tex };
1098         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1099
1100         // Set up preview and any additional channels.
1101         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1102                 DisplayFrame display_frame;
1103                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1104                 display_frame.chain = chain.chain;
1105                 display_frame.setup_chain = chain.setup_chain;
1106                 display_frame.ready_fence = fence;
1107                 display_frame.input_frames = chain.input_frames;
1108                 display_frame.temp_textures = {};
1109                 output_channel[i].output_frame(display_frame);
1110         }
1111 }
1112
1113 void Mixer::audio_thread_func()
1114 {
1115         pthread_setname_np(pthread_self(), "Mixer_Audio");
1116
1117         while (!should_quit) {
1118                 AudioTask task;
1119
1120                 {
1121                         unique_lock<mutex> lock(audio_mutex);
1122                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1123                         if (should_quit) {
1124                                 return;
1125                         }
1126                         task = audio_task_queue.front();
1127                         audio_task_queue.pop();
1128                 }
1129
1130                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1131                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1132                 vector<float> samples_out = audio_mixer.get_output(
1133                         task.frame_timestamp,
1134                         task.num_samples,
1135                         rate_adjustment_policy);
1136
1137                 // Send the samples to the sound card, then add them to the output.
1138                 if (alsa) {
1139                         alsa->write(samples_out);
1140                 }
1141                 if (output_card_index != -1) {
1142                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1143                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1144                 }
1145                 video_encoder->add_audio(task.pts_int, move(samples_out));
1146         }
1147 }
1148
1149 void Mixer::release_display_frame(DisplayFrame *frame)
1150 {
1151         for (GLuint texnum : frame->temp_textures) {
1152                 resource_pool->release_2d_texture(texnum);
1153         }
1154         frame->temp_textures.clear();
1155         frame->ready_fence.reset();
1156         frame->input_frames.clear();
1157 }
1158
1159 void Mixer::start()
1160 {
1161         mixer_thread = thread(&Mixer::thread_func, this);
1162         audio_thread = thread(&Mixer::audio_thread_func, this);
1163 }
1164
1165 void Mixer::quit()
1166 {
1167         should_quit = true;
1168         audio_task_queue_changed.notify_one();
1169         mixer_thread.join();
1170         audio_thread.join();
1171 }
1172
1173 void Mixer::transition_clicked(int transition_num)
1174 {
1175         theme->transition_clicked(transition_num, pts());
1176 }
1177
1178 void Mixer::channel_clicked(int preview_num)
1179 {
1180         theme->channel_clicked(preview_num);
1181 }
1182
1183 void Mixer::start_mode_scanning(unsigned card_index)
1184 {
1185         assert(card_index < num_cards);
1186         if (is_mode_scanning[card_index]) {
1187                 return;
1188         }
1189         is_mode_scanning[card_index] = true;
1190         mode_scanlist[card_index].clear();
1191         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1192                 mode_scanlist[card_index].push_back(mode.first);
1193         }
1194         assert(!mode_scanlist[card_index].empty());
1195         mode_scanlist_index[card_index] = 0;
1196         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1197         last_mode_scan_change[card_index] = steady_clock::now();
1198 }
1199
1200 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1201 {
1202         assert(desired_output_card_index != -1);
1203         unique_lock<mutex> lock(card_mutex);
1204         return cards[desired_output_card_index].output->get_available_video_modes();
1205 }
1206
1207 Mixer::OutputChannel::~OutputChannel()
1208 {
1209         if (has_current_frame) {
1210                 parent->release_display_frame(&current_frame);
1211         }
1212         if (has_ready_frame) {
1213                 parent->release_display_frame(&ready_frame);
1214         }
1215 }
1216
1217 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1218 {
1219         // Store this frame for display. Remove the ready frame if any
1220         // (it was seemingly never used).
1221         {
1222                 unique_lock<mutex> lock(frame_mutex);
1223                 if (has_ready_frame) {
1224                         parent->release_display_frame(&ready_frame);
1225                 }
1226                 ready_frame = frame;
1227                 has_ready_frame = true;
1228         }
1229
1230         if (new_frame_ready_callback) {
1231                 new_frame_ready_callback();
1232         }
1233
1234         // Reduce the number of callbacks by filtering duplicates. The reason
1235         // why we bother doing this is that Qt seemingly can get into a state
1236         // where its builds up an essentially unbounded queue of signals,
1237         // consuming more and more memory, and there's no good way of collapsing
1238         // user-defined signals or limiting the length of the queue.
1239         if (transition_names_updated_callback) {
1240                 vector<string> transition_names = global_mixer->get_transition_names();
1241                 bool changed = false;
1242                 if (transition_names.size() != last_transition_names.size()) {
1243                         changed = true;
1244                 } else {
1245                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1246                                 if (transition_names[i] != last_transition_names[i]) {
1247                                         changed = true;
1248                                         break;
1249                                 }
1250                         }
1251                 }
1252                 if (changed) {
1253                         transition_names_updated_callback(transition_names);
1254                         last_transition_names = transition_names;
1255                 }
1256         }
1257         if (name_updated_callback) {
1258                 string name = global_mixer->get_channel_name(channel);
1259                 if (name != last_name) {
1260                         name_updated_callback(name);
1261                         last_name = name;
1262                 }
1263         }
1264         if (color_updated_callback) {
1265                 string color = global_mixer->get_channel_color(channel);
1266                 if (color != last_color) {
1267                         color_updated_callback(color);
1268                         last_color = color;
1269                 }
1270         }
1271 }
1272
1273 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1274 {
1275         unique_lock<mutex> lock(frame_mutex);
1276         if (!has_current_frame && !has_ready_frame) {
1277                 return false;
1278         }
1279
1280         if (has_current_frame && has_ready_frame) {
1281                 // We have a new ready frame. Toss the current one.
1282                 parent->release_display_frame(&current_frame);
1283                 has_current_frame = false;
1284         }
1285         if (has_ready_frame) {
1286                 assert(!has_current_frame);
1287                 current_frame = ready_frame;
1288                 ready_frame.ready_fence.reset();  // Drop the refcount.
1289                 ready_frame.input_frames.clear();  // Drop the refcounts.
1290                 has_current_frame = true;
1291                 has_ready_frame = false;
1292         }
1293
1294         *frame = current_frame;
1295         return true;
1296 }
1297
1298 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1299 {
1300         new_frame_ready_callback = callback;
1301 }
1302
1303 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1304 {
1305         transition_names_updated_callback = callback;
1306 }
1307
1308 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1309 {
1310         name_updated_callback = callback;
1311 }
1312
1313 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1314 {
1315         color_updated_callback = callback;
1316 }
1317
1318 mutex RefCountedGLsync::fence_lock;