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         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 197                 output_channel[i].parent = this;
 198                 output_channel[i].channel = i;
 199         }
 200
 201         ImageFormat inout_format;
 202         inout_format.color_space = COLORSPACE_sRGB;
 203         inout_format.gamma_curve = GAMMA_sRGB;
 204
 205         // Matches the 4:2:0 format created by the main chain.
 206         YCbCrFormat ycbcr_format;
 207         ycbcr_format.chroma_subsampling_x = 2;
 208         ycbcr_format.chroma_subsampling_y = 2;
 209         if (global_flags.ycbcr_rec709_coefficients) {
 210                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 211         } else {
 212                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 213         }
 214         ycbcr_format.full_range = false;
 215         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 216         ycbcr_format.cb_x_position = 0.0f;
 217         ycbcr_format.cr_x_position = 0.0f;
 218         ycbcr_format.cb_y_position = 0.5f;
 219         ycbcr_format.cr_y_position = 0.5f;
 220
 221         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 222         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 223         check_error();
 224         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 225         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 226         display_chain->add_input(display_input);
 227         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 228         display_chain->set_dither_bits(0);  // Don't bother.
 229         display_chain->finalize();
 230
 231         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));
 232
 233         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 234         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 235
 236         // Start listening for clients only once VideoEncoder has written its header, if any.
 237         httpd.start(9095);
 238
 239         // First try initializing the then PCI devices, then USB, then
 240         // fill up with fake cards until we have the desired number of cards.
 241         unsigned num_pci_devices = 0;
 242         unsigned card_index = 0;
 243
 244         {
 245                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 246                 if (decklink_iterator != nullptr) {
 247                         for ( ; card_index < num_cards; ++card_index) {
 248                                 IDeckLink *decklink;
 249                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 250                                         break;
 251                                 }
 252
 253                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 254                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 255                                 output->set_device(decklink);
 256                                 configure_card(card_index, capture, /*is_fake_capture=*/false, output);
 257                                 ++num_pci_devices;
 258                         }
 259                         decklink_iterator->Release();
 260                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 261                 } else {
 262                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 263                 }
 264         }
 265
 266         unsigned num_usb_devices = BMUSBCapture::num_cards();
 267         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 268                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 269                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 270                 configure_card(card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
 271         }
 272         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 273
 274         unsigned num_fake_cards = 0;
 275         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 276                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 277                 configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
 278         }
 279
 280         if (num_fake_cards > 0) {
 281                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 282         }
 283
 284         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 285         BMUSBCapture::start_bm_thread();
 286
 287         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 288                 cards[card_index].queue_length_policy.reset(card_index);
 289         }
 290
 291         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 292
 293         if (global_flags.ten_bit_input) {
 294                 if (!v210Converter::has_hardware_support()) {
 295                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 296                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 297                         exit(1);
 298                 }
 299                 v210_converter.reset(new v210Converter());
 300
 301                 // These are all the widths listed in the Blackmagic SDK documentation
 302                 // (section 2.7.3, “Display Modes”).
 303                 v210_converter->precompile_shader(720);
 304                 v210_converter->precompile_shader(1280);
 305                 v210_converter->precompile_shader(1920);
 306                 v210_converter->precompile_shader(2048);
 307                 v210_converter->precompile_shader(3840);
 308                 v210_converter->precompile_shader(4096);
 309         }
 310         if (global_flags.ten_bit_output) {
 311                 if (!v210Converter::has_hardware_support()) {
 312                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 313                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 314                         exit(1);
 315                 }
 316         }
 317
 318         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 319         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 320         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 321
 322         if (global_flags.enable_alsa_output) {
 323                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 324         }
 325         if (global_flags.output_card != -1) {
 326                 desired_output_card_index = global_flags.output_card;
 327                 set_output_card_internal(global_flags.output_card);
 328         }
 329 }
 330
 331 Mixer::~Mixer()
 332 {
 333         BMUSBCapture::stop_bm_thread();
 334
 335         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 336                 {
 337                         unique_lock<mutex> lock(card_mutex);
 338                         cards[card_index].should_quit = true;  // Unblock thread.
 339                         cards[card_index].new_frames_changed.notify_all();
 340                 }
 341                 cards[card_index].capture->stop_dequeue_thread();
 342                 if (cards[card_index].output) {
 343                         cards[card_index].output->end_output();
 344                         cards[card_index].output.reset();
 345                 }
 346         }
 347
 348         video_encoder.reset(nullptr);
 349 }
 350
 351 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output)
 352 {
 353         printf("Configuring card %d...\n", card_index);
 354
 355         CaptureCard *card = &cards[card_index];
 356         if (card->capture != nullptr) {
 357                 card->capture->stop_dequeue_thread();
 358         }
 359         card->capture.reset(capture);
 360         card->is_fake_capture = is_fake_capture;
 361         if (card->output.get() != output) {
 362                 card->output.reset(output);
 363         }
 364         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 365         if (card->frame_allocator == nullptr) {
 366                 card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 367         }
 368         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 369         if (card->surface == nullptr) {
 370                 card->surface = create_surface_with_same_format(mixer_surface);
 371         }
 372         while (!card->new_frames.empty()) card->new_frames.pop_front();
 373         card->last_timecode = -1;
 374         card->capture->set_pixel_format(global_flags.ten_bit_input ? PixelFormat_10BitYCbCr : PixelFormat_8BitYCbCr);
 375         card->capture->configure_card();
 376
 377         // NOTE: start_bm_capture() happens in thread_func().
 378
 379         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 380         audio_mixer.reset_resampler(device);
 381         audio_mixer.set_display_name(device, card->capture->get_description());
 382         audio_mixer.trigger_state_changed_callback();
 383 }
 384
 385 void Mixer::set_output_card_internal(int card_index)
 386 {
 387         // We don't really need to take card_mutex, since we're in the mixer
 388         // thread and don't mess with any queues (which is the only thing that happens
 389         // from other threads), but it's probably the safest in the long run.
 390         unique_lock<mutex> lock(card_mutex);
 391         if (output_card_index != -1) {
 392                 // Switch the old card from output to input.
 393                 CaptureCard *old_card = &cards[output_card_index];
 394                 old_card->output->end_output();
 395
 396                 // Stop the fake card that we put into place.
 397                 // This needs to _not_ happen under the mutex, to avoid deadlock
 398                 // (delivering the last frame needs to take the mutex).
 399                 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
 400                 lock.unlock();
 401                 fake_capture->stop_dequeue_thread();
 402                 lock.lock();
 403                 old_card->capture = move(old_card->parked_capture);
 404                 old_card->is_fake_capture = false;
 405                 old_card->capture->start_bm_capture();
 406         }
 407         if (card_index != -1) {
 408                 CaptureCard *card = &cards[card_index];
 409                 bmusb::CaptureInterface *capture = card->capture.get();
 410                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 411                 // several seconds to complete (blocking on DisableVideoInput);
 412                 // see if we can maybe do it asynchronously.
 413                 lock.unlock();
 414                 capture->stop_dequeue_thread();
 415                 lock.lock();
 416                 card->parked_capture = move(card->capture);
 417                 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 418                 configure_card(card_index, fake_capture, /*is_fake_capture=*/true, card->output.release());
 419                 card->queue_length_policy.reset(card_index);
 420                 card->capture->start_bm_capture();
 421                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 422                 card->output->start_output(desired_output_video_mode, pts_int);
 423         }
 424         output_card_index = card_index;
 425 }
 426
 427 namespace {
 428
 429 int unwrap_timecode(uint16_t current_wrapped, int last)
 430 {
 431         uint16_t last_wrapped = last & 0xffff;
 432         if (current_wrapped > last_wrapped) {
 433                 return (last & ~0xffff) | current_wrapped;
 434         } else {
 435                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 436         }
 437 }
 438
 439 }  // namespace
 440
 441 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 442                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 443                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 444 {
 445         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 446         CaptureCard *card = &cards[card_index];
 447
 448         if (is_mode_scanning[card_index]) {
 449                 if (video_format.has_signal) {
 450                         // Found a stable signal, so stop scanning.
 451                         is_mode_scanning[card_index] = false;
 452                 } else {
 453                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 454                         steady_clock::time_point now = steady_clock::now();
 455                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 456                         if (sec_since_last_switch > switch_time_s) {
 457                                 // It isn't this mode; try the next one.
 458                                 mode_scanlist_index[card_index]++;
 459                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 460                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 461                                 last_mode_scan_change[card_index] = now;
 462                         }
 463                 }
 464         }
 465
 466         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 467         assert(frame_length > 0);
 468
 469         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;
 470         if (num_samples > OUTPUT_FREQUENCY / 10) {
 471                 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",
 472                         card_index, int(audio_frame.len), int(audio_offset),
 473                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 474                 if (video_frame.owner) {
 475                         video_frame.owner->release_frame(video_frame);
 476                 }
 477                 if (audio_frame.owner) {
 478                         audio_frame.owner->release_frame(audio_frame);
 479                 }
 480                 return;
 481         }
 482
 483         int dropped_frames = 0;
 484         if (card->last_timecode != -1) {
 485                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 486         }
 487
 488         // Number of samples per frame if we need to insert silence.
 489         // (Could be nonintegral, but resampling will save us then.)
 490         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 491
 492         if (dropped_frames > MAX_FPS * 2) {
 493                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 494                         card_index, card->last_timecode, timecode);
 495                 audio_mixer.reset_resampler(device);
 496                 dropped_frames = 0;
 497         } else if (dropped_frames > 0) {
 498                 // Insert silence as needed.
 499                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 500                         card_index, dropped_frames, timecode);
 501
 502                 bool success;
 503                 do {
 504                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 505                 } while (!success);
 506         }
 507
 508         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 509
 510         // Done with the audio, so release it.
 511         if (audio_frame.owner) {
 512                 audio_frame.owner->release_frame(audio_frame);
 513         }
 514
 515         card->last_timecode = timecode;
 516
 517         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 518         if (video_frame.len - video_offset == 0 ||
 519             video_frame.len - video_offset != expected_length) {
 520                 if (video_frame.len != 0) {
 521                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 522                                 card_index, video_frame.len - video_offset, expected_length);
 523                 }
 524                 if (video_frame.owner) {
 525                         video_frame.owner->release_frame(video_frame);
 526                 }
 527
 528                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 529                 // so that pts can go up accordingly.
 530                 {
 531                         unique_lock<mutex> lock(card_mutex);
 532                         CaptureCard::NewFrame new_frame;
 533                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 534                         new_frame.length = frame_length;
 535                         new_frame.interlaced = false;
 536                         new_frame.dropped_frames = dropped_frames;
 537                         new_frame.received_timestamp = video_frame.received_timestamp;
 538                         card->new_frames.push_back(move(new_frame));
 539                         card->new_frames_changed.notify_all();
 540                 }
 541                 return;
 542         }
 543
 544         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 545
 546         unsigned num_fields = video_format.interlaced ? 2 : 1;
 547         steady_clock::time_point frame_upload_start;
 548         bool interlaced_stride = false;
 549         if (video_format.interlaced) {
 550                 // Send the two fields along as separate frames; the other side will need to add
 551                 // a deinterlacer to actually get this right.
 552                 assert(video_format.height % 2 == 0);
 553                 video_format.height /= 2;
 554                 assert(frame_length % 2 == 0);
 555                 frame_length /= 2;
 556                 num_fields = 2;
 557                 if (video_format.second_field_start == 1) {
 558                         interlaced_stride = true;
 559                 }
 560                 frame_upload_start = steady_clock::now();
 561         }
 562         userdata->last_interlaced = video_format.interlaced;
 563         userdata->last_has_signal = video_format.has_signal;
 564         userdata->last_is_connected = video_format.is_connected;
 565         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 566         userdata->last_frame_rate_den = video_format.frame_rate_den;
 567         RefCountedFrame frame(video_frame);
 568
 569         // Upload the textures.
 570         const size_t cbcr_width = video_format.width / 2;
 571         const size_t cbcr_offset = video_offset / 2;
 572         const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 573
 574         for (unsigned field = 0; field < num_fields; ++field) {
 575                 // Put the actual texture upload in a lambda that is executed in the main thread.
 576                 // It is entirely possible to do this in the same thread (and it might even be
 577                 // faster, depending on the GPU and driver), but it appears to be trickling
 578                 // driver bugs very easily.
 579                 //
 580                 // Note that this means we must hold on to the actual frame data in <userdata>
 581                 // until the upload command is run, but we hold on to <frame> much longer than that
 582                 // (in fact, all the way until we no longer use the texture in rendering).
 583                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 584                         unsigned field_start_line;
 585                         if (field == 1) {
 586                                 field_start_line = video_format.second_field_start;
 587                         } else {
 588                                 field_start_line = video_format.extra_lines_top;
 589                         }
 590
 591                         // For 8-bit input, v210_width will be nonsensical but not used.
 592                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 593                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 594
 595                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 596                         check_error();
 597
 598                         if (global_flags.ten_bit_input) {
 599                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 600                                 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);
 601                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 602                         } else {
 603                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 604                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 605
 606                                 // Make up our own strides, since we are interleaving.
 607                                 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);
 608                                 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);
 609                         }
 610
 611                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 612                         check_error();
 613                 };
 614
 615                 if (field == 1) {
 616                         // Don't upload the second field as fast as we can; wait until
 617                         // the field time has approximately passed. (Otherwise, we could
 618                         // get timing jitter against the other sources, and possibly also
 619                         // against the video display, although the latter is not as critical.)
 620                         // This requires our system clock to be reasonably close to the
 621                         // video clock, but that's not an unreasonable assumption.
 622                         steady_clock::time_point second_field_start = frame_upload_start +
 623                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 624                         this_thread::sleep_until(second_field_start);
 625                 }
 626
 627                 {
 628                         unique_lock<mutex> lock(card_mutex);
 629                         CaptureCard::NewFrame new_frame;
 630                         new_frame.frame = frame;
 631                         new_frame.length = frame_length;
 632                         new_frame.field = field;
 633                         new_frame.interlaced = video_format.interlaced;
 634                         new_frame.upload_func = upload_func;
 635                         new_frame.dropped_frames = dropped_frames;
 636                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 637                         card->new_frames.push_back(move(new_frame));
 638                         card->new_frames_changed.notify_all();
 639                 }
 640         }
 641 }
 642
 643 void Mixer::bm_hotplug_add(libusb_device *dev)
 644 {
 645         lock_guard<mutex> lock(hotplug_mutex);
 646         hotplugged_cards.push_back(dev);
 647 }
 648
 649 void Mixer::bm_hotplug_remove(unsigned card_index)
 650 {
 651         cards[card_index].new_frames_changed.notify_all();
 652 }
 653
 654 void Mixer::thread_func()
 655 {
 656         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 657
 658         eglBindAPI(EGL_OPENGL_API);
 659         QOpenGLContext *context = create_context(mixer_surface);
 660         if (!make_current(context, mixer_surface)) {
 661                 printf("oops\n");
 662                 exit(1);
 663         }
 664
 665         // Start the actual capture. (We don't want to do it before we're actually ready
 666         // to process output frames.)
 667         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 668                 if (int(card_index) != output_card_index) {
 669                         cards[card_index].capture->start_bm_capture();
 670                 }
 671         }
 672
 673         steady_clock::time_point start, now;
 674         start = steady_clock::now();
 675
 676         int stats_dropped_frames = 0;
 677
 678         while (!should_quit) {
 679                 if (desired_output_card_index != output_card_index) {
 680                         set_output_card_internal(desired_output_card_index);
 681                 }
 682                 if (output_card_index != -1 &&
 683                     desired_output_video_mode != output_video_mode) {
 684                         DeckLinkOutput *output = cards[output_card_index].output.get();
 685                         output->end_output();
 686                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 687                         output->start_output(desired_output_video_mode, pts_int);
 688                 }
 689
 690                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 691                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 692
 693                 bool master_card_is_output;
 694                 unsigned master_card_index;
 695                 if (output_card_index != -1) {
 696                         master_card_is_output = true;
 697                         master_card_index = output_card_index;
 698                 } else {
 699                         master_card_is_output = false;
 700                         master_card_index = theme->map_signal(master_clock_channel);
 701                         assert(master_card_index < num_cards);
 702                 }
 703
 704                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 705                 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);
 706                 stats_dropped_frames += output_frame_info.dropped_frames;
 707
 708                 handle_hotplugged_cards();
 709
 710                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 711                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 712                                 continue;
 713                         }
 714                         if (new_frames[card_index].frame->len == 0) {
 715                                 ++new_frames[card_index].dropped_frames;
 716                         }
 717                         if (new_frames[card_index].dropped_frames > 0) {
 718                                 printf("Card %u dropped %d frames before this\n",
 719                                         card_index, int(new_frames[card_index].dropped_frames));
 720                         }
 721                 }
 722
 723                 // If the first card is reporting a corrupted or otherwise dropped frame,
 724                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 725                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 726                         ++stats_dropped_frames;
 727                         pts_int += new_frames[master_card_index].length;
 728                         continue;
 729                 }
 730
 731                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 732                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 733                                 continue;
 734
 735                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 736                         assert(new_frame->frame != nullptr);
 737                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 738                         check_error();
 739
 740                         // The new texture might need uploading before use.
 741                         if (new_frame->upload_func) {
 742                                 new_frame->upload_func();
 743                                 new_frame->upload_func = nullptr;
 744                         }
 745                 }
 746
 747                 int64_t frame_duration = output_frame_info.frame_duration;
 748                 render_one_frame(frame_duration);
 749                 ++frame_num;
 750                 pts_int += frame_duration;
 751
 752                 now = steady_clock::now();
 753                 double elapsed = duration<double>(now - start).count();
 754                 if (frame_num % 100 == 0) {
 755                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 756                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 757                                 1e3 * elapsed / frame_num);
 758                 //      chain->print_phase_timing();
 759
 760                         // Check our memory usage, to see if we are close to our mlockall()
 761                         // limit (if at all set).
 762                         rusage used;
 763                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 764                                 perror("getrusage(RUSAGE_SELF)");
 765                                 assert(false);
 766                         }
 767
 768                         if (uses_mlock) {
 769                                 rlimit limit;
 770                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 771                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 772                                         assert(false);
 773                                 }
 774
 775                                 if (limit.rlim_cur == 0) {
 776                                         printf(", using %ld MB memory (locked)",
 777                                                 long(used.ru_maxrss / 1024));
 778                                 } else {
 779                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 780                                                 long(used.ru_maxrss / 1024),
 781                                                 long(limit.rlim_cur / 1048576),
 782                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 783                                 }
 784                         } else {
 785                                 printf(", using %ld MB memory (not locked)",
 786                                         long(used.ru_maxrss / 1024));
 787                         }
 788
 789                         printf("\n");
 790                 }
 791
 792
 793                 if (should_cut.exchange(false)) {  // Test and clear.
 794                         video_encoder->do_cut(frame_num);
 795                 }
 796
 797 #if 0
 798                 // Reset every 100 frames, so that local variations in frame times
 799                 // (especially for the first few frames, when the shaders are
 800                 // compiled etc.) don't make it hard to measure for the entire
 801                 // remaining duration of the program.
 802                 if (frame == 10000) {
 803                         frame = 0;
 804                         start = now;
 805                 }
 806 #endif
 807                 check_error();
 808         }
 809
 810         resource_pool->clean_context();
 811 }
 812
 813 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 814 {
 815         if (output_card_index != -1) {
 816                 // The output card (ie., cards[output_card_index].output) is the master clock,
 817                 // so no input card (ie., cards[card_index].capture) is.
 818                 return false;
 819         }
 820         return (card_index == master_card_index);
 821 }
 822
 823 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 824 {
 825         // Count the number of frames in the queue, including any frames
 826         // we dropped. It's hard to know exactly how we should deal with
 827         // dropped (corrupted) input frames; they don't help our goal of
 828         // avoiding starvation, but they still add to the problem of latency.
 829         // Since dropped frames is going to mean a bump in the signal anyway,
 830         // we err on the side of having more stable latency instead.
 831         unsigned queue_length = 0;
 832         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 833                 queue_length += frame.dropped_frames + 1;
 834         }
 835         card->queue_length_policy.update_policy(queue_length);
 836
 837         // If needed, drop frames until the queue is below the safe limit.
 838         // We prefer to drop from the head, because all else being equal,
 839         // we'd like more recent frames (less latency).
 840         unsigned dropped_frames = 0;
 841         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 842                 assert(!card->new_frames.empty());
 843                 assert(queue_length > card->new_frames.front().dropped_frames);
 844                 queue_length -= card->new_frames.front().dropped_frames;
 845
 846                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 847                         // No need to drop anything.
 848                         break;
 849                 }
 850
 851                 card->new_frames.pop_front();
 852                 card->new_frames_changed.notify_all();
 853                 --queue_length;
 854                 ++dropped_frames;
 855         }
 856
 857 #if 0
 858         if (dropped_frames > 0) {
 859                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 860                         card_index, dropped_frames);
 861         }
 862 #endif
 863 }
 864
 865
 866 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])
 867 {
 868         OutputFrameInfo output_frame_info;
 869 start:
 870         unique_lock<mutex> lock(card_mutex, defer_lock);
 871         if (master_card_is_output) {
 872                 // Clocked to the output, so wait for it to be ready for the next frame.
 873                 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);
 874                 lock.lock();
 875         } else {
 876                 // Wait for the master card to have a new frame.
 877                 // TODO: Add a timeout.
 878                 output_frame_info.is_preroll = false;
 879                 lock.lock();
 880                 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(); });
 881         }
 882
 883         if (master_card_is_output) {
 884                 handle_hotplugged_cards();
 885         } else if (cards[master_card_index].new_frames.empty()) {
 886                 // We were woken up, but not due to a new frame. Deal with it
 887                 // and then restart.
 888                 assert(cards[master_card_index].capture->get_disconnected());
 889                 handle_hotplugged_cards();
 890                 goto start;
 891         }
 892
 893         if (!master_card_is_output) {
 894                 output_frame_info.frame_timestamp =
 895                         cards[master_card_index].new_frames.front().received_timestamp;
 896         }
 897
 898         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 899                 CaptureCard *card = &cards[card_index];
 900                 if (input_card_is_master_clock(card_index, master_card_index)) {
 901                         // We don't use the queue length policy for the master card,
 902                         // but we will if it stops being the master. Thus, clear out
 903                         // the policy in case we switch in the future.
 904                         card->queue_length_policy.reset(card_index);
 905                         assert(!card->new_frames.empty());
 906                 } else {
 907                         trim_queue(card, card_index);
 908                 }
 909                 if (!card->new_frames.empty()) {
 910                         new_frames[card_index] = move(card->new_frames.front());
 911                         has_new_frame[card_index] = true;
 912                         card->new_frames.pop_front();
 913                         card->new_frames_changed.notify_all();
 914                 }
 915         }
 916
 917         if (!master_card_is_output) {
 918                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 919                 output_frame_info.frame_duration = new_frames[master_card_index].length;
 920         }
 921
 922         // This might get off by a fractional sample when changing master card
 923         // between ones with different frame rates, but that's fine.
 924         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 925         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 926         fractional_samples = num_samples_times_timebase % TIMEBASE;
 927         assert(output_frame_info.num_samples >= 0);
 928
 929         return output_frame_info;
 930 }
 931
 932 void Mixer::handle_hotplugged_cards()
 933 {
 934         // Check for cards that have been disconnected since last frame.
 935         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 936                 CaptureCard *card = &cards[card_index];
 937                 if (card->capture->get_disconnected()) {
 938                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 939                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 940                         configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
 941                         card->queue_length_policy.reset(card_index);
 942                         card->capture->start_bm_capture();
 943                 }
 944         }
 945
 946         // Check for cards that have been connected since last frame.
 947         vector<libusb_device *> hotplugged_cards_copy;
 948         {
 949                 lock_guard<mutex> lock(hotplug_mutex);
 950                 swap(hotplugged_cards, hotplugged_cards_copy);
 951         }
 952         for (libusb_device *new_dev : hotplugged_cards_copy) {
 953                 // Look for a fake capture card where we can stick this in.
 954                 int free_card_index = -1;
 955                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 956                         if (cards[card_index].is_fake_capture) {
 957                                 free_card_index = card_index;
 958                                 break;
 959                         }
 960                 }
 961
 962                 if (free_card_index == -1) {
 963                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 964                         libusb_unref_device(new_dev);
 965                 } else {
 966                         // BMUSBCapture takes ownership.
 967                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 968                         CaptureCard *card = &cards[free_card_index];
 969                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 970                         configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
 971                         card->queue_length_policy.reset(free_card_index);
 972                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 973                         capture->start_bm_capture();
 974                 }
 975         }
 976 }
 977
 978
 979 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)
 980 {
 981         // Resample the audio as needed, including from previously dropped frames.
 982         assert(num_cards > 0);
 983         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 984                 const bool dropped_frame = (frame_num != dropped_frames);
 985                 {
 986                         // Signal to the audio thread to process this frame.
 987                         // Note that if the frame is a dropped frame, we signal that
 988                         // we don't want to use this frame as base for adjusting
 989                         // the resampler rate. The reason for this is that the timing
 990                         // of these frames is often way too late; they typically don't
 991                         // “arrive” before we synthesize them. Thus, we could end up
 992                         // in a situation where we have inserted e.g. five audio frames
 993                         // into the queue before we then start pulling five of them
 994                         // back out. This makes ResamplingQueue overestimate the delay,
 995                         // causing undue resampler changes. (We _do_ use the last,
 996                         // non-dropped frame; perhaps we should just discard that as well,
 997                         // since dropped frames are expected to be rare, and it might be
 998                         // better to just wait until we have a slightly more normal situation).
 999                         unique_lock<mutex> lock(audio_mutex);
1000                         bool adjust_rate = !dropped_frame && !is_preroll;
1001                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1002                         audio_task_queue_changed.notify_one();
1003                 }
1004                 if (dropped_frame) {
1005                         // For dropped frames, increase the pts. Note that if the format changed
1006                         // in the meantime, we have no way of detecting that; we just have to
1007                         // assume the frame length is always the same.
1008                         pts_int += length_per_frame;
1009                 }
1010         }
1011 }
1012
1013 void Mixer::render_one_frame(int64_t duration)
1014 {
1015         // Determine the time code for this frame before we start rendering.
1016         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1017         if (display_timecode_on_stdout) {
1018                 printf("Timecode: '%s'\n", timecode_text.c_str());
1019         }
1020
1021         // Get the main chain from the theme, and set its state immediately.
1022         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1023         EffectChain *chain = theme_main_chain.chain;
1024         theme_main_chain.setup_chain();
1025         //theme_main_chain.chain->enable_phase_timing(true);
1026
1027         // If HDMI/SDI output is active and the user has requested auto mode,
1028         // its mode overrides the existing Y'CbCr setting for the chain.
1029         YCbCrLumaCoefficients ycbcr_output_coefficients;
1030         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1031                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1032         } else {
1033                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1034         }
1035
1036         // TODO: Reduce the duplication against theme.cpp.
1037         YCbCrFormat output_ycbcr_format;
1038         output_ycbcr_format.chroma_subsampling_x = 1;
1039         output_ycbcr_format.chroma_subsampling_y = 1;
1040         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1041         output_ycbcr_format.full_range = false;
1042         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1043         chain->change_ycbcr_output_format(output_ycbcr_format);
1044
1045         // Render main chain. If we're using zerocopy Quick Sync encoding
1046         // (the default case), 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 they're just views into VA-API memory and must be
1051         // unmapped during encoding, so we can't use them for display, unfortunately.
1052         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1053         GLuint y_copy_tex, cbcr_copy_tex = 0;
1054         GLuint y_display_tex, cbcr_display_tex;
1055         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1056         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1057         const bool is_zerocopy = video_encoder->is_zerocopy();
1058         if (is_zerocopy) {
1059                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1060                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1061                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1062
1063                 y_display_tex = y_copy_tex;
1064                 cbcr_display_tex = cbcr_copy_tex;
1065
1066                 // y_tex and cbcr_tex will be given by VideoEncoder.
1067         } else {
1068                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1069                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1070                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1071
1072                 y_display_tex = y_tex;
1073                 cbcr_display_tex = cbcr_tex;
1074         }
1075
1076         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1077         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1078         assert(got_frame);
1079
1080         GLuint fbo;
1081         if (is_zerocopy) {
1082                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1083         } else {
1084                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1085         }
1086         check_error();
1087         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1088
1089         if (display_timecode_in_stream) {
1090                 // Render the timecode on top.
1091                 timecode_renderer->render_timecode(fbo, timecode_text);
1092         }
1093
1094         resource_pool->release_fbo(fbo);
1095
1096         if (is_zerocopy) {
1097                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1098         } else {
1099                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1100         }
1101         if (output_card_index != -1) {
1102                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1103         }
1104         resource_pool->release_2d_texture(cbcr_full_tex);
1105
1106         // Set the right state for the Y' and CbCr textures we use for display.
1107         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1108         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1109         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1110         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1111         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1112
1113         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1114         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1115         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1116         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1117
1118         RefCountedGLsync fence = video_encoder->end_frame();
1119
1120         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1121         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1122         DisplayFrame live_frame;
1123         live_frame.chain = display_chain.get();
1124         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1125                 display_input->set_texture_num(0, y_display_tex);
1126                 display_input->set_texture_num(1, cbcr_display_tex);
1127         };
1128         live_frame.ready_fence = fence;
1129         live_frame.input_frames = {};
1130         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1131         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1132
1133         // Set up preview and any additional channels.
1134         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1135                 DisplayFrame display_frame;
1136                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1137                 display_frame.chain = chain.chain;
1138                 display_frame.setup_chain = chain.setup_chain;
1139                 display_frame.ready_fence = fence;
1140                 display_frame.input_frames = chain.input_frames;
1141                 display_frame.temp_textures = {};
1142                 output_channel[i].output_frame(display_frame);
1143         }
1144 }
1145
1146 void Mixer::audio_thread_func()
1147 {
1148         pthread_setname_np(pthread_self(), "Mixer_Audio");
1149
1150         while (!should_quit) {
1151                 AudioTask task;
1152
1153                 {
1154                         unique_lock<mutex> lock(audio_mutex);
1155                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1156                         if (should_quit) {
1157                                 return;
1158                         }
1159                         task = audio_task_queue.front();
1160                         audio_task_queue.pop();
1161                 }
1162
1163                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1164                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1165                 vector<float> samples_out = audio_mixer.get_output(
1166                         task.frame_timestamp,
1167                         task.num_samples,
1168                         rate_adjustment_policy);
1169
1170                 // Send the samples to the sound card, then add them to the output.
1171                 if (alsa) {
1172                         alsa->write(samples_out);
1173                 }
1174                 if (output_card_index != -1) {
1175                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1176                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1177                 }
1178                 video_encoder->add_audio(task.pts_int, move(samples_out));
1179         }
1180 }
1181
1182 void Mixer::release_display_frame(DisplayFrame *frame)
1183 {
1184         for (GLuint texnum : frame->temp_textures) {
1185                 resource_pool->release_2d_texture(texnum);
1186         }
1187         frame->temp_textures.clear();
1188         frame->ready_fence.reset();
1189         frame->input_frames.clear();
1190 }
1191
1192 void Mixer::start()
1193 {
1194         mixer_thread = thread(&Mixer::thread_func, this);
1195         audio_thread = thread(&Mixer::audio_thread_func, this);
1196 }
1197
1198 void Mixer::quit()
1199 {
1200         should_quit = true;
1201         audio_task_queue_changed.notify_one();
1202         mixer_thread.join();
1203         audio_thread.join();
1204 }
1205
1206 void Mixer::transition_clicked(int transition_num)
1207 {
1208         theme->transition_clicked(transition_num, pts());
1209 }
1210
1211 void Mixer::channel_clicked(int preview_num)
1212 {
1213         theme->channel_clicked(preview_num);
1214 }
1215
1216 void Mixer::start_mode_scanning(unsigned card_index)
1217 {
1218         assert(card_index < num_cards);
1219         if (is_mode_scanning[card_index]) {
1220                 return;
1221         }
1222         is_mode_scanning[card_index] = true;
1223         mode_scanlist[card_index].clear();
1224         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1225                 mode_scanlist[card_index].push_back(mode.first);
1226         }
1227         assert(!mode_scanlist[card_index].empty());
1228         mode_scanlist_index[card_index] = 0;
1229         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1230         last_mode_scan_change[card_index] = steady_clock::now();
1231 }
1232
1233 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1234 {
1235         assert(desired_output_card_index != -1);
1236         unique_lock<mutex> lock(card_mutex);
1237         return cards[desired_output_card_index].output->get_available_video_modes();
1238 }
1239
1240 Mixer::OutputChannel::~OutputChannel()
1241 {
1242         if (has_current_frame) {
1243                 parent->release_display_frame(&current_frame);
1244         }
1245         if (has_ready_frame) {
1246                 parent->release_display_frame(&ready_frame);
1247         }
1248 }
1249
1250 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1251 {
1252         // Store this frame for display. Remove the ready frame if any
1253         // (it was seemingly never used).
1254         {
1255                 unique_lock<mutex> lock(frame_mutex);
1256                 if (has_ready_frame) {
1257                         parent->release_display_frame(&ready_frame);
1258                 }
1259                 ready_frame = frame;
1260                 has_ready_frame = true;
1261         }
1262
1263         if (new_frame_ready_callback) {
1264                 new_frame_ready_callback();
1265         }
1266
1267         // Reduce the number of callbacks by filtering duplicates. The reason
1268         // why we bother doing this is that Qt seemingly can get into a state
1269         // where its builds up an essentially unbounded queue of signals,
1270         // consuming more and more memory, and there's no good way of collapsing
1271         // user-defined signals or limiting the length of the queue.
1272         if (transition_names_updated_callback) {
1273                 vector<string> transition_names = global_mixer->get_transition_names();
1274                 bool changed = false;
1275                 if (transition_names.size() != last_transition_names.size()) {
1276                         changed = true;
1277                 } else {
1278                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1279                                 if (transition_names[i] != last_transition_names[i]) {
1280                                         changed = true;
1281                                         break;
1282                                 }
1283                         }
1284                 }
1285                 if (changed) {
1286                         transition_names_updated_callback(transition_names);
1287                         last_transition_names = transition_names;
1288                 }
1289         }
1290         if (name_updated_callback) {
1291                 string name = global_mixer->get_channel_name(channel);
1292                 if (name != last_name) {
1293                         name_updated_callback(name);
1294                         last_name = name;
1295                 }
1296         }
1297         if (color_updated_callback) {
1298                 string color = global_mixer->get_channel_color(channel);
1299                 if (color != last_color) {
1300                         color_updated_callback(color);
1301                         last_color = color;
1302                 }
1303         }
1304 }
1305
1306 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1307 {
1308         unique_lock<mutex> lock(frame_mutex);
1309         if (!has_current_frame && !has_ready_frame) {
1310                 return false;
1311         }
1312
1313         if (has_current_frame && has_ready_frame) {
1314                 // We have a new ready frame. Toss the current one.
1315                 parent->release_display_frame(&current_frame);
1316                 has_current_frame = false;
1317         }
1318         if (has_ready_frame) {
1319                 assert(!has_current_frame);
1320                 current_frame = ready_frame;
1321                 ready_frame.ready_fence.reset();  // Drop the refcount.
1322                 ready_frame.input_frames.clear();  // Drop the refcounts.
1323                 has_current_frame = true;
1324                 has_ready_frame = false;
1325         }
1326
1327         *frame = current_frame;
1328         return true;
1329 }
1330
1331 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1332 {
1333         new_frame_ready_callback = callback;
1334 }
1335
1336 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1337 {
1338         transition_names_updated_callback = callback;
1339 }
1340
1341 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1342 {
1343         name_updated_callback = callback;
1344 }
1345
1346 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1347 {
1348         color_updated_callback = callback;
1349 }
1350
1351 mutex RefCountedGLsync::fence_lock;