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         if (global_flags.ten_bit_output) {
 309                 if (!v210Converter::has_hardware_support()) {
 310                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 311                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 312                         exit(1);
 313                 }
 314         }
 315
 316         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 317         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 318         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 319
 320         if (global_flags.enable_alsa_output) {
 321                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 322         }
 323         if (global_flags.output_card != -1) {
 324                 desired_output_card_index = global_flags.output_card;
 325                 set_output_card_internal(global_flags.output_card);
 326         }
 327 }
 328
 329 Mixer::~Mixer()
 330 {
 331         BMUSBCapture::stop_bm_thread();
 332
 333         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 334                 {
 335                         unique_lock<mutex> lock(card_mutex);
 336                         cards[card_index].should_quit = true;  // Unblock thread.
 337                         cards[card_index].new_frames_changed.notify_all();
 338                 }
 339                 cards[card_index].capture->stop_dequeue_thread();
 340                 if (cards[card_index].output) {
 341                         cards[card_index].output->end_output();
 342                         cards[card_index].output.reset();
 343                 }
 344         }
 345
 346         video_encoder.reset(nullptr);
 347 }
 348
 349 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output)
 350 {
 351         printf("Configuring card %d...\n", card_index);
 352
 353         CaptureCard *card = &cards[card_index];
 354         if (card->capture != nullptr) {
 355                 card->capture->stop_dequeue_thread();
 356         }
 357         card->capture.reset(capture);
 358         card->is_fake_capture = is_fake_capture;
 359         if (card->output.get() != output) {
 360                 card->output.reset(output);
 361         }
 362         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 363         if (card->frame_allocator == nullptr) {
 364                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 365         }
 366         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 367         if (card->surface == nullptr) {
 368                 card->surface = create_surface_with_same_format(mixer_surface);
 369         }
 370         while (!card->new_frames.empty()) card->new_frames.pop_front();
 371         card->last_timecode = -1;
 372         card->capture->set_pixel_format(global_flags.ten_bit_input ? PixelFormat_10BitYCbCr : PixelFormat_8BitYCbCr);
 373         card->capture->configure_card();
 374
 375         // NOTE: start_bm_capture() happens in thread_func().
 376
 377         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 378         audio_mixer.reset_resampler(device);
 379         audio_mixer.set_display_name(device, card->capture->get_description());
 380         audio_mixer.trigger_state_changed_callback();
 381 }
 382
 383 void Mixer::set_output_card_internal(int card_index)
 384 {
 385         // We don't really need to take card_mutex, since we're in the mixer
 386         // thread and don't mess with any queues (which is the only thing that happens
 387         // from other threads), but it's probably the safest in the long run.
 388         unique_lock<mutex> lock(card_mutex);
 389         if (output_card_index != -1) {
 390                 // Switch the old card from output to input.
 391                 CaptureCard *old_card = &cards[output_card_index];
 392                 old_card->output->end_output();
 393
 394                 // Stop the fake card that we put into place.
 395                 // This needs to _not_ happen under the mutex, to avoid deadlock
 396                 // (delivering the last frame needs to take the mutex).
 397                 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
 398                 lock.unlock();
 399                 fake_capture->stop_dequeue_thread();
 400                 lock.lock();
 401                 old_card->capture = move(old_card->parked_capture);
 402                 old_card->is_fake_capture = false;
 403                 old_card->capture->start_bm_capture();
 404         }
 405         if (card_index != -1) {
 406                 CaptureCard *card = &cards[card_index];
 407                 bmusb::CaptureInterface *capture = card->capture.get();
 408                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 409                 // several seconds to complete (blocking on DisableVideoInput);
 410                 // see if we can maybe do it asynchronously.
 411                 lock.unlock();
 412                 capture->stop_dequeue_thread();
 413                 lock.lock();
 414                 card->parked_capture = move(card->capture);
 415                 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 416                 configure_card(card_index, fake_capture, /*is_fake_capture=*/true, card->output.release());
 417                 card->queue_length_policy.reset(card_index);
 418                 card->capture->start_bm_capture();
 419                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 420                 card->output->start_output(desired_output_video_mode, pts_int);
 421         }
 422         output_card_index = card_index;
 423 }
 424
 425 namespace {
 426
 427 int unwrap_timecode(uint16_t current_wrapped, int last)
 428 {
 429         uint16_t last_wrapped = last & 0xffff;
 430         if (current_wrapped > last_wrapped) {
 431                 return (last & ~0xffff) | current_wrapped;
 432         } else {
 433                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 434         }
 435 }
 436
 437 }  // namespace
 438
 439 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 440                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 441                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 442 {
 443         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 444         CaptureCard *card = &cards[card_index];
 445
 446         if (is_mode_scanning[card_index]) {
 447                 if (video_format.has_signal) {
 448                         // Found a stable signal, so stop scanning.
 449                         is_mode_scanning[card_index] = false;
 450                 } else {
 451                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 452                         steady_clock::time_point now = steady_clock::now();
 453                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 454                         if (sec_since_last_switch > switch_time_s) {
 455                                 // It isn't this mode; try the next one.
 456                                 mode_scanlist_index[card_index]++;
 457                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 458                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 459                                 last_mode_scan_change[card_index] = now;
 460                         }
 461                 }
 462         }
 463
 464         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 465         assert(frame_length > 0);
 466
 467         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;
 468         if (num_samples > OUTPUT_FREQUENCY / 10) {
 469                 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",
 470                         card_index, int(audio_frame.len), int(audio_offset),
 471                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 472                 if (video_frame.owner) {
 473                         video_frame.owner->release_frame(video_frame);
 474                 }
 475                 if (audio_frame.owner) {
 476                         audio_frame.owner->release_frame(audio_frame);
 477                 }
 478                 return;
 479         }
 480
 481         int dropped_frames = 0;
 482         if (card->last_timecode != -1) {
 483                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 484         }
 485
 486         // Number of samples per frame if we need to insert silence.
 487         // (Could be nonintegral, but resampling will save us then.)
 488         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 489
 490         if (dropped_frames > MAX_FPS * 2) {
 491                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 492                         card_index, card->last_timecode, timecode);
 493                 audio_mixer.reset_resampler(device);
 494                 dropped_frames = 0;
 495         } else if (dropped_frames > 0) {
 496                 // Insert silence as needed.
 497                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 498                         card_index, dropped_frames, timecode);
 499
 500                 bool success;
 501                 do {
 502                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 503                 } while (!success);
 504         }
 505
 506         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 507
 508         // Done with the audio, so release it.
 509         if (audio_frame.owner) {
 510                 audio_frame.owner->release_frame(audio_frame);
 511         }
 512
 513         card->last_timecode = timecode;
 514
 515         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 516         if (video_frame.len - video_offset == 0 ||
 517             video_frame.len - video_offset != expected_length) {
 518                 if (video_frame.len != 0) {
 519                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 520                                 card_index, video_frame.len - video_offset, expected_length);
 521                 }
 522                 if (video_frame.owner) {
 523                         video_frame.owner->release_frame(video_frame);
 524                 }
 525
 526                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 527                 // so that pts can go up accordingly.
 528                 {
 529                         unique_lock<mutex> lock(card_mutex);
 530                         CaptureCard::NewFrame new_frame;
 531                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 532                         new_frame.length = frame_length;
 533                         new_frame.interlaced = false;
 534                         new_frame.dropped_frames = dropped_frames;
 535                         new_frame.received_timestamp = video_frame.received_timestamp;
 536                         card->new_frames.push_back(move(new_frame));
 537                         card->new_frames_changed.notify_all();
 538                 }
 539                 return;
 540         }
 541
 542         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 543
 544         unsigned num_fields = video_format.interlaced ? 2 : 1;
 545         steady_clock::time_point frame_upload_start;
 546         bool interlaced_stride = false;
 547         if (video_format.interlaced) {
 548                 // Send the two fields along as separate frames; the other side will need to add
 549                 // a deinterlacer to actually get this right.
 550                 assert(video_format.height % 2 == 0);
 551                 video_format.height /= 2;
 552                 assert(frame_length % 2 == 0);
 553                 frame_length /= 2;
 554                 num_fields = 2;
 555                 if (video_format.second_field_start == 1) {
 556                         interlaced_stride = true;
 557                 }
 558                 frame_upload_start = steady_clock::now();
 559         }
 560         userdata->last_interlaced = video_format.interlaced;
 561         userdata->last_has_signal = video_format.has_signal;
 562         userdata->last_is_connected = video_format.is_connected;
 563         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 564         userdata->last_frame_rate_den = video_format.frame_rate_den;
 565         RefCountedFrame frame(video_frame);
 566
 567         // Upload the textures.
 568         const size_t cbcr_width = video_format.width / 2;
 569         const size_t cbcr_offset = video_offset / 2;
 570         const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 571
 572         for (unsigned field = 0; field < num_fields; ++field) {
 573                 // Put the actual texture upload in a lambda that is executed in the main thread.
 574                 // It is entirely possible to do this in the same thread (and it might even be
 575                 // faster, depending on the GPU and driver), but it appears to be trickling
 576                 // driver bugs very easily.
 577                 //
 578                 // Note that this means we must hold on to the actual frame data in <userdata>
 579                 // until the upload command is run, but we hold on to <frame> much longer than that
 580                 // (in fact, all the way until we no longer use the texture in rendering).
 581                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 582                         unsigned field_start_line;
 583                         if (field == 1) {
 584                                 field_start_line = video_format.second_field_start;
 585                         } else {
 586                                 field_start_line = video_format.extra_lines_top;
 587                         }
 588
 589                         // For 8-bit input, v210_width will be nonsensical but not used.
 590                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 591                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 592
 593                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 594                         check_error();
 595
 596                         if (global_flags.ten_bit_input) {
 597                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 598                                 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);
 599                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 600                         } else {
 601                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 602                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 603
 604                                 // Make up our own strides, since we are interleaving.
 605                                 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);
 606                                 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);
 607                         }
 608
 609                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 610                         check_error();
 611                 };
 612
 613                 if (field == 1) {
 614                         // Don't upload the second field as fast as we can; wait until
 615                         // the field time has approximately passed. (Otherwise, we could
 616                         // get timing jitter against the other sources, and possibly also
 617                         // against the video display, although the latter is not as critical.)
 618                         // This requires our system clock to be reasonably close to the
 619                         // video clock, but that's not an unreasonable assumption.
 620                         steady_clock::time_point second_field_start = frame_upload_start +
 621                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 622                         this_thread::sleep_until(second_field_start);
 623                 }
 624
 625                 {
 626                         unique_lock<mutex> lock(card_mutex);
 627                         CaptureCard::NewFrame new_frame;
 628                         new_frame.frame = frame;
 629                         new_frame.length = frame_length;
 630                         new_frame.field = field;
 631                         new_frame.interlaced = video_format.interlaced;
 632                         new_frame.upload_func = upload_func;
 633                         new_frame.dropped_frames = dropped_frames;
 634                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 635                         card->new_frames.push_back(move(new_frame));
 636                         card->new_frames_changed.notify_all();
 637                 }
 638         }
 639 }
 640
 641 void Mixer::bm_hotplug_add(libusb_device *dev)
 642 {
 643         lock_guard<mutex> lock(hotplug_mutex);
 644         hotplugged_cards.push_back(dev);
 645 }
 646
 647 void Mixer::bm_hotplug_remove(unsigned card_index)
 648 {
 649         cards[card_index].new_frames_changed.notify_all();
 650 }
 651
 652 void Mixer::thread_func()
 653 {
 654         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 655
 656         eglBindAPI(EGL_OPENGL_API);
 657         QOpenGLContext *context = create_context(mixer_surface);
 658         if (!make_current(context, mixer_surface)) {
 659                 printf("oops\n");
 660                 exit(1);
 661         }
 662
 663         // Start the actual capture. (We don't want to do it before we're actually ready
 664         // to process output frames.)
 665         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 666                 if (int(card_index) != output_card_index) {
 667                         cards[card_index].capture->start_bm_capture();
 668                 }
 669         }
 670
 671         steady_clock::time_point start, now;
 672         start = steady_clock::now();
 673
 674         int stats_dropped_frames = 0;
 675
 676         while (!should_quit) {
 677                 if (desired_output_card_index != output_card_index) {
 678                         set_output_card_internal(desired_output_card_index);
 679                 }
 680                 if (output_card_index != -1 &&
 681                     desired_output_video_mode != output_video_mode) {
 682                         DeckLinkOutput *output = cards[output_card_index].output.get();
 683                         output->end_output();
 684                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 685                         output->start_output(desired_output_video_mode, pts_int);
 686                 }
 687
 688                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 689                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 690
 691                 bool master_card_is_output;
 692                 unsigned master_card_index;
 693                 if (output_card_index != -1) {
 694                         master_card_is_output = true;
 695                         master_card_index = output_card_index;
 696                 } else {
 697                         master_card_is_output = false;
 698                         master_card_index = theme->map_signal(master_clock_channel);
 699                         assert(master_card_index < num_cards);
 700                 }
 701
 702                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 703                 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);
 704                 stats_dropped_frames += output_frame_info.dropped_frames;
 705
 706                 handle_hotplugged_cards();
 707
 708                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 709                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 710                                 continue;
 711                         }
 712                         if (new_frames[card_index].frame->len == 0) {
 713                                 ++new_frames[card_index].dropped_frames;
 714                         }
 715                         if (new_frames[card_index].dropped_frames > 0) {
 716                                 printf("Card %u dropped %d frames before this\n",
 717                                         card_index, int(new_frames[card_index].dropped_frames));
 718                         }
 719                 }
 720
 721                 // If the first card is reporting a corrupted or otherwise dropped frame,
 722                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 723                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 724                         ++stats_dropped_frames;
 725                         pts_int += new_frames[master_card_index].length;
 726                         continue;
 727                 }
 728
 729                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 730                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 731                                 continue;
 732
 733                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 734                         assert(new_frame->frame != nullptr);
 735                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 736                         check_error();
 737
 738                         // The new texture might need uploading before use.
 739                         if (new_frame->upload_func) {
 740                                 new_frame->upload_func();
 741                                 new_frame->upload_func = nullptr;
 742                         }
 743                 }
 744
 745                 int64_t frame_duration = output_frame_info.frame_duration;
 746                 render_one_frame(frame_duration);
 747                 ++frame_num;
 748                 pts_int += frame_duration;
 749
 750                 now = steady_clock::now();
 751                 double elapsed = duration<double>(now - start).count();
 752                 if (frame_num % 100 == 0) {
 753                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 754                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 755                                 1e3 * elapsed / frame_num);
 756                 //      chain->print_phase_timing();
 757
 758                         // Check our memory usage, to see if we are close to our mlockall()
 759                         // limit (if at all set).
 760                         rusage used;
 761                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 762                                 perror("getrusage(RUSAGE_SELF)");
 763                                 assert(false);
 764                         }
 765
 766                         if (uses_mlock) {
 767                                 rlimit limit;
 768                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 769                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 770                                         assert(false);
 771                                 }
 772
 773                                 if (limit.rlim_cur == 0) {
 774                                         printf(", using %ld MB memory (locked)",
 775                                                 long(used.ru_maxrss / 1024));
 776                                 } else {
 777                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 778                                                 long(used.ru_maxrss / 1024),
 779                                                 long(limit.rlim_cur / 1048576),
 780                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 781                                 }
 782                         } else {
 783                                 printf(", using %ld MB memory (not locked)",
 784                                         long(used.ru_maxrss / 1024));
 785                         }
 786
 787                         printf("\n");
 788                 }
 789
 790
 791                 if (should_cut.exchange(false)) {  // Test and clear.
 792                         video_encoder->do_cut(frame_num);
 793                 }
 794
 795 #if 0
 796                 // Reset every 100 frames, so that local variations in frame times
 797                 // (especially for the first few frames, when the shaders are
 798                 // compiled etc.) don't make it hard to measure for the entire
 799                 // remaining duration of the program.
 800                 if (frame == 10000) {
 801                         frame = 0;
 802                         start = now;
 803                 }
 804 #endif
 805                 check_error();
 806         }
 807
 808         resource_pool->clean_context();
 809 }
 810
 811 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 812 {
 813         if (output_card_index != -1) {
 814                 // The output card (ie., cards[output_card_index].output) is the master clock,
 815                 // so no input card (ie., cards[card_index].capture) is.
 816                 return false;
 817         }
 818         return (card_index == master_card_index);
 819 }
 820
 821 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 822 {
 823         // Count the number of frames in the queue, including any frames
 824         // we dropped. It's hard to know exactly how we should deal with
 825         // dropped (corrupted) input frames; they don't help our goal of
 826         // avoiding starvation, but they still add to the problem of latency.
 827         // Since dropped frames is going to mean a bump in the signal anyway,
 828         // we err on the side of having more stable latency instead.
 829         unsigned queue_length = 0;
 830         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 831                 queue_length += frame.dropped_frames + 1;
 832         }
 833         card->queue_length_policy.update_policy(queue_length);
 834
 835         // If needed, drop frames until the queue is below the safe limit.
 836         // We prefer to drop from the head, because all else being equal,
 837         // we'd like more recent frames (less latency).
 838         unsigned dropped_frames = 0;
 839         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 840                 assert(!card->new_frames.empty());
 841                 assert(queue_length > card->new_frames.front().dropped_frames);
 842                 queue_length -= card->new_frames.front().dropped_frames;
 843
 844                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 845                         // No need to drop anything.
 846                         break;
 847                 }
 848
 849                 card->new_frames.pop_front();
 850                 card->new_frames_changed.notify_all();
 851                 --queue_length;
 852                 ++dropped_frames;
 853         }
 854
 855         if (dropped_frames > 0) {
 856                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 857                         card_index, dropped_frames);
 858         }
 859 }
 860
 861
 862 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])
 863 {
 864         OutputFrameInfo output_frame_info;
 865 start:
 866         unique_lock<mutex> lock(card_mutex, defer_lock);
 867         if (master_card_is_output) {
 868                 // Clocked to the output, so wait for it to be ready for the next frame.
 869                 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);
 870                 lock.lock();
 871         } else {
 872                 // Wait for the master card to have a new frame.
 873                 // TODO: Add a timeout.
 874                 output_frame_info.is_preroll = false;
 875                 lock.lock();
 876                 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(); });
 877         }
 878
 879         if (master_card_is_output) {
 880                 handle_hotplugged_cards();
 881         } else if (cards[master_card_index].new_frames.empty()) {
 882                 // We were woken up, but not due to a new frame. Deal with it
 883                 // and then restart.
 884                 assert(cards[master_card_index].capture->get_disconnected());
 885                 handle_hotplugged_cards();
 886                 goto start;
 887         }
 888
 889         if (!master_card_is_output) {
 890                 output_frame_info.frame_timestamp =
 891                         cards[master_card_index].new_frames.front().received_timestamp;
 892         }
 893
 894         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 895                 CaptureCard *card = &cards[card_index];
 896                 if (input_card_is_master_clock(card_index, master_card_index)) {
 897                         // We don't use the queue length policy for the master card,
 898                         // but we will if it stops being the master. Thus, clear out
 899                         // the policy in case we switch in the future.
 900                         card->queue_length_policy.reset(card_index);
 901                         assert(!card->new_frames.empty());
 902                 } else {
 903                         trim_queue(card, card_index);
 904                 }
 905                 if (!card->new_frames.empty()) {
 906                         new_frames[card_index] = move(card->new_frames.front());
 907                         has_new_frame[card_index] = true;
 908                         card->new_frames.pop_front();
 909                         card->new_frames_changed.notify_all();
 910                 }
 911         }
 912
 913         if (!master_card_is_output) {
 914                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 915                 output_frame_info.frame_duration = new_frames[master_card_index].length;
 916         }
 917
 918         // This might get off by a fractional sample when changing master card
 919         // between ones with different frame rates, but that's fine.
 920         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 921         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 922         fractional_samples = num_samples_times_timebase % TIMEBASE;
 923         assert(output_frame_info.num_samples >= 0);
 924
 925         return output_frame_info;
 926 }
 927
 928 void Mixer::handle_hotplugged_cards()
 929 {
 930         // Check for cards that have been disconnected since last frame.
 931         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 932                 CaptureCard *card = &cards[card_index];
 933                 if (card->capture->get_disconnected()) {
 934                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 935                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 936                         configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
 937                         card->queue_length_policy.reset(card_index);
 938                         card->capture->start_bm_capture();
 939                 }
 940         }
 941
 942         // Check for cards that have been connected since last frame.
 943         vector<libusb_device *> hotplugged_cards_copy;
 944         {
 945                 lock_guard<mutex> lock(hotplug_mutex);
 946                 swap(hotplugged_cards, hotplugged_cards_copy);
 947         }
 948         for (libusb_device *new_dev : hotplugged_cards_copy) {
 949                 // Look for a fake capture card where we can stick this in.
 950                 int free_card_index = -1;
 951                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 952                         if (cards[card_index].is_fake_capture) {
 953                                 free_card_index = card_index;
 954                                 break;
 955                         }
 956                 }
 957
 958                 if (free_card_index == -1) {
 959                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 960                         libusb_unref_device(new_dev);
 961                 } else {
 962                         // BMUSBCapture takes ownership.
 963                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 964                         CaptureCard *card = &cards[free_card_index];
 965                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 966                         configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
 967                         card->queue_length_policy.reset(free_card_index);
 968                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 969                         capture->start_bm_capture();
 970                 }
 971         }
 972 }
 973
 974
 975 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)
 976 {
 977         // Resample the audio as needed, including from previously dropped frames.
 978         assert(num_cards > 0);
 979         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 980                 const bool dropped_frame = (frame_num != dropped_frames);
 981                 {
 982                         // Signal to the audio thread to process this frame.
 983                         // Note that if the frame is a dropped frame, we signal that
 984                         // we don't want to use this frame as base for adjusting
 985                         // the resampler rate. The reason for this is that the timing
 986                         // of these frames is often way too late; they typically don't
 987                         // “arrive” before we synthesize them. Thus, we could end up
 988                         // in a situation where we have inserted e.g. five audio frames
 989                         // into the queue before we then start pulling five of them
 990                         // back out. This makes ResamplingQueue overestimate the delay,
 991                         // causing undue resampler changes. (We _do_ use the last,
 992                         // non-dropped frame; perhaps we should just discard that as well,
 993                         // since dropped frames are expected to be rare, and it might be
 994                         // better to just wait until we have a slightly more normal situation).
 995                         unique_lock<mutex> lock(audio_mutex);
 996                         bool adjust_rate = !dropped_frame && !is_preroll;
 997                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
 998                         audio_task_queue_changed.notify_one();
 999                 }
1000                 if (dropped_frame) {
1001                         // For dropped frames, increase the pts. Note that if the format changed
1002                         // in the meantime, we have no way of detecting that; we just have to
1003                         // assume the frame length is always the same.
1004                         pts_int += length_per_frame;
1005                 }
1006         }
1007 }
1008
1009 void Mixer::render_one_frame(int64_t duration)
1010 {
1011         // Determine the time code for this frame before we start rendering.
1012         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1013         if (display_timecode_on_stdout) {
1014                 printf("Timecode: '%s'\n", timecode_text.c_str());
1015         }
1016
1017         // Get the main chain from the theme, and set its state immediately.
1018         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1019         EffectChain *chain = theme_main_chain.chain;
1020         theme_main_chain.setup_chain();
1021         //theme_main_chain.chain->enable_phase_timing(true);
1022
1023         // If HDMI/SDI output is active and the user has requested auto mode,
1024         // its mode overrides the existing Y'CbCr setting for the chain.
1025         YCbCrLumaCoefficients ycbcr_output_coefficients;
1026         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1027                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1028         } else {
1029                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1030         }
1031
1032         // TODO: Reduce the duplication against theme.cpp.
1033         YCbCrFormat output_ycbcr_format;
1034         output_ycbcr_format.chroma_subsampling_x = 1;
1035         output_ycbcr_format.chroma_subsampling_y = 1;
1036         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1037         output_ycbcr_format.full_range = false;
1038         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1039         chain->change_ycbcr_output_format(output_ycbcr_format);
1040
1041         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1042         GLuint y_tex, cbcr_tex;
1043         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1044         assert(got_frame);
1045
1046         // Render main chain. We take an extra copy of the created outputs,
1047         // so that we can display it back to the screen later (it's less memory
1048         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1049         // Ideally, we'd like to avoid taking copies and just use the main textures
1050         // for display as well, but if they're used for zero-copy Quick Sync encoding
1051         // (the default case), they're just views into VA-API memory and must be
1052         // unmapped during encoding, so we can't use them for display, unfortunately.
1053         GLuint cbcr_full_tex, cbcr_copy_tex, y_copy_tex;
1054         if (global_flags.x264_bit_depth > 8) {
1055                 cbcr_full_tex = resource_pool->create_2d_texture(GL_RG16, global_flags.width, global_flags.height);
1056                 y_copy_tex = resource_pool->create_2d_texture(GL_R16, global_flags.width, global_flags.height);
1057                 cbcr_copy_tex = resource_pool->create_2d_texture(GL_RG16, global_flags.width / 2, global_flags.height / 2);
1058         } else {
1059                 cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width, global_flags.height);
1060                 y_copy_tex = resource_pool->create_2d_texture(GL_R8, global_flags.width, global_flags.height);
1061                 cbcr_copy_tex = resource_pool->create_2d_texture(GL_RG8, global_flags.width / 2, global_flags.height / 2);
1062         }
1063         GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1064         check_error();
1065         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1066
1067         if (display_timecode_in_stream) {
1068                 // Render the timecode on top.
1069                 timecode_renderer->render_timecode(fbo, timecode_text);
1070         }
1071
1072         resource_pool->release_fbo(fbo);
1073
1074         chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1075         if (output_card_index != -1) {
1076                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1077         }
1078         resource_pool->release_2d_texture(cbcr_full_tex);
1079
1080         // Set the right state for the Y' and CbCr copies.
1081         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1082         glBindTexture(GL_TEXTURE_2D, y_copy_tex);
1083         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1084         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1085         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1086
1087         glBindTexture(GL_TEXTURE_2D, cbcr_copy_tex);
1088         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1089         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1090         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1091
1092         RefCountedGLsync fence = video_encoder->end_frame();
1093
1094         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1095         // It owns y_copy_tex and cbcr_copy_tex now.
1096         DisplayFrame live_frame;
1097         live_frame.chain = display_chain.get();
1098         live_frame.setup_chain = [this, y_copy_tex, cbcr_copy_tex]{
1099                 display_input->set_texture_num(0, y_copy_tex);
1100                 display_input->set_texture_num(1, cbcr_copy_tex);
1101         };
1102         live_frame.ready_fence = fence;
1103         live_frame.input_frames = {};
1104         live_frame.temp_textures = { y_copy_tex, cbcr_copy_tex };
1105         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1106
1107         // Set up preview and any additional channels.
1108         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1109                 DisplayFrame display_frame;
1110                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1111                 display_frame.chain = chain.chain;
1112                 display_frame.setup_chain = chain.setup_chain;
1113                 display_frame.ready_fence = fence;
1114                 display_frame.input_frames = chain.input_frames;
1115                 display_frame.temp_textures = {};
1116                 output_channel[i].output_frame(display_frame);
1117         }
1118 }
1119
1120 void Mixer::audio_thread_func()
1121 {
1122         pthread_setname_np(pthread_self(), "Mixer_Audio");
1123
1124         while (!should_quit) {
1125                 AudioTask task;
1126
1127                 {
1128                         unique_lock<mutex> lock(audio_mutex);
1129                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1130                         if (should_quit) {
1131                                 return;
1132                         }
1133                         task = audio_task_queue.front();
1134                         audio_task_queue.pop();
1135                 }
1136
1137                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1138                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1139                 vector<float> samples_out = audio_mixer.get_output(
1140                         task.frame_timestamp,
1141                         task.num_samples,
1142                         rate_adjustment_policy);
1143
1144                 // Send the samples to the sound card, then add them to the output.
1145                 if (alsa) {
1146                         alsa->write(samples_out);
1147                 }
1148                 if (output_card_index != -1) {
1149                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1150                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1151                 }
1152                 video_encoder->add_audio(task.pts_int, move(samples_out));
1153         }
1154 }
1155
1156 void Mixer::release_display_frame(DisplayFrame *frame)
1157 {
1158         for (GLuint texnum : frame->temp_textures) {
1159                 resource_pool->release_2d_texture(texnum);
1160         }
1161         frame->temp_textures.clear();
1162         frame->ready_fence.reset();
1163         frame->input_frames.clear();
1164 }
1165
1166 void Mixer::start()
1167 {
1168         mixer_thread = thread(&Mixer::thread_func, this);
1169         audio_thread = thread(&Mixer::audio_thread_func, this);
1170 }
1171
1172 void Mixer::quit()
1173 {
1174         should_quit = true;
1175         audio_task_queue_changed.notify_one();
1176         mixer_thread.join();
1177         audio_thread.join();
1178 }
1179
1180 void Mixer::transition_clicked(int transition_num)
1181 {
1182         theme->transition_clicked(transition_num, pts());
1183 }
1184
1185 void Mixer::channel_clicked(int preview_num)
1186 {
1187         theme->channel_clicked(preview_num);
1188 }
1189
1190 void Mixer::start_mode_scanning(unsigned card_index)
1191 {
1192         assert(card_index < num_cards);
1193         if (is_mode_scanning[card_index]) {
1194                 return;
1195         }
1196         is_mode_scanning[card_index] = true;
1197         mode_scanlist[card_index].clear();
1198         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1199                 mode_scanlist[card_index].push_back(mode.first);
1200         }
1201         assert(!mode_scanlist[card_index].empty());
1202         mode_scanlist_index[card_index] = 0;
1203         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1204         last_mode_scan_change[card_index] = steady_clock::now();
1205 }
1206
1207 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1208 {
1209         assert(desired_output_card_index != -1);
1210         unique_lock<mutex> lock(card_mutex);
1211         return cards[desired_output_card_index].output->get_available_video_modes();
1212 }
1213
1214 Mixer::OutputChannel::~OutputChannel()
1215 {
1216         if (has_current_frame) {
1217                 parent->release_display_frame(&current_frame);
1218         }
1219         if (has_ready_frame) {
1220                 parent->release_display_frame(&ready_frame);
1221         }
1222 }
1223
1224 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1225 {
1226         // Store this frame for display. Remove the ready frame if any
1227         // (it was seemingly never used).
1228         {
1229                 unique_lock<mutex> lock(frame_mutex);
1230                 if (has_ready_frame) {
1231                         parent->release_display_frame(&ready_frame);
1232                 }
1233                 ready_frame = frame;
1234                 has_ready_frame = true;
1235         }
1236
1237         if (new_frame_ready_callback) {
1238                 new_frame_ready_callback();
1239         }
1240
1241         // Reduce the number of callbacks by filtering duplicates. The reason
1242         // why we bother doing this is that Qt seemingly can get into a state
1243         // where its builds up an essentially unbounded queue of signals,
1244         // consuming more and more memory, and there's no good way of collapsing
1245         // user-defined signals or limiting the length of the queue.
1246         if (transition_names_updated_callback) {
1247                 vector<string> transition_names = global_mixer->get_transition_names();
1248                 bool changed = false;
1249                 if (transition_names.size() != last_transition_names.size()) {
1250                         changed = true;
1251                 } else {
1252                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1253                                 if (transition_names[i] != last_transition_names[i]) {
1254                                         changed = true;
1255                                         break;
1256                                 }
1257                         }
1258                 }
1259                 if (changed) {
1260                         transition_names_updated_callback(transition_names);
1261                         last_transition_names = transition_names;
1262                 }
1263         }
1264         if (name_updated_callback) {
1265                 string name = global_mixer->get_channel_name(channel);
1266                 if (name != last_name) {
1267                         name_updated_callback(name);
1268                         last_name = name;
1269                 }
1270         }
1271         if (color_updated_callback) {
1272                 string color = global_mixer->get_channel_color(channel);
1273                 if (color != last_color) {
1274                         color_updated_callback(color);
1275                         last_color = color;
1276                 }
1277         }
1278 }
1279
1280 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1281 {
1282         unique_lock<mutex> lock(frame_mutex);
1283         if (!has_current_frame && !has_ready_frame) {
1284                 return false;
1285         }
1286
1287         if (has_current_frame && has_ready_frame) {
1288                 // We have a new ready frame. Toss the current one.
1289                 parent->release_display_frame(&current_frame);
1290                 has_current_frame = false;
1291         }
1292         if (has_ready_frame) {
1293                 assert(!has_current_frame);
1294                 current_frame = ready_frame;
1295                 ready_frame.ready_fence.reset();  // Drop the refcount.
1296                 ready_frame.input_frames.clear();  // Drop the refcounts.
1297                 has_current_frame = true;
1298                 has_ready_frame = false;
1299         }
1300
1301         *frame = current_frame;
1302         return true;
1303 }
1304
1305 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1306 {
1307         new_frame_ready_callback = callback;
1308 }
1309
1310 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1311 {
1312         transition_names_updated_callback = callback;
1313 }
1314
1315 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1316 {
1317         name_updated_callback = callback;
1318 }
1319
1320 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1321 {
1322         color_updated_callback = callback;
1323 }
1324
1325 mutex RefCountedGLsync::fence_lock;