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 (userdata->pixel_format == bmusb::PixelFormat_10BitYCbCr) {
  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 (userdata->pixel_format == bmusb::PixelFormat_10BitYCbCr) {
  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
 365         bmusb::PixelFormat pixel_format = global_flags.ten_bit_input ? PixelFormat_10BitYCbCr : PixelFormat_8BitYCbCr;
 366
 367         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 368         if (card->frame_allocator == nullptr) {
 369                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 370         }
 371         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 372         if (card->surface == nullptr) {
 373                 card->surface = create_surface_with_same_format(mixer_surface);
 374         }
 375         while (!card->new_frames.empty()) card->new_frames.pop_front();
 376         card->last_timecode = -1;
 377         card->capture->set_pixel_format(pixel_format);
 378         card->capture->configure_card();
 379
 380         // NOTE: start_bm_capture() happens in thread_func().
 381
 382         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 383         audio_mixer.reset_resampler(device);
 384         audio_mixer.set_display_name(device, card->capture->get_description());
 385         audio_mixer.trigger_state_changed_callback();
 386 }
 387
 388 void Mixer::set_output_card_internal(int card_index)
 389 {
 390         // We don't really need to take card_mutex, since we're in the mixer
 391         // thread and don't mess with any queues (which is the only thing that happens
 392         // from other threads), but it's probably the safest in the long run.
 393         unique_lock<mutex> lock(card_mutex);
 394         if (output_card_index != -1) {
 395                 // Switch the old card from output to input.
 396                 CaptureCard *old_card = &cards[output_card_index];
 397                 old_card->output->end_output();
 398
 399                 // Stop the fake card that we put into place.
 400                 // This needs to _not_ happen under the mutex, to avoid deadlock
 401                 // (delivering the last frame needs to take the mutex).
 402                 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
 403                 lock.unlock();
 404                 fake_capture->stop_dequeue_thread();
 405                 lock.lock();
 406                 old_card->capture = move(old_card->parked_capture);
 407                 old_card->is_fake_capture = false;
 408                 old_card->capture->start_bm_capture();
 409         }
 410         if (card_index != -1) {
 411                 CaptureCard *card = &cards[card_index];
 412                 bmusb::CaptureInterface *capture = card->capture.get();
 413                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 414                 // several seconds to complete (blocking on DisableVideoInput);
 415                 // see if we can maybe do it asynchronously.
 416                 lock.unlock();
 417                 capture->stop_dequeue_thread();
 418                 lock.lock();
 419                 card->parked_capture = move(card->capture);
 420                 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 421                 configure_card(card_index, fake_capture, /*is_fake_capture=*/true, card->output.release());
 422                 card->queue_length_policy.reset(card_index);
 423                 card->capture->start_bm_capture();
 424                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 425                 card->output->start_output(desired_output_video_mode, pts_int);
 426         }
 427         output_card_index = card_index;
 428 }
 429
 430 namespace {
 431
 432 int unwrap_timecode(uint16_t current_wrapped, int last)
 433 {
 434         uint16_t last_wrapped = last & 0xffff;
 435         if (current_wrapped > last_wrapped) {
 436                 return (last & ~0xffff) | current_wrapped;
 437         } else {
 438                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 439         }
 440 }
 441
 442 }  // namespace
 443
 444 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 445                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 446                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 447 {
 448         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 449         CaptureCard *card = &cards[card_index];
 450
 451         if (is_mode_scanning[card_index]) {
 452                 if (video_format.has_signal) {
 453                         // Found a stable signal, so stop scanning.
 454                         is_mode_scanning[card_index] = false;
 455                 } else {
 456                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 457                         steady_clock::time_point now = steady_clock::now();
 458                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 459                         if (sec_since_last_switch > switch_time_s) {
 460                                 // It isn't this mode; try the next one.
 461                                 mode_scanlist_index[card_index]++;
 462                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 463                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 464                                 last_mode_scan_change[card_index] = now;
 465                         }
 466                 }
 467         }
 468
 469         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 470         assert(frame_length > 0);
 471
 472         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;
 473         if (num_samples > OUTPUT_FREQUENCY / 10) {
 474                 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",
 475                         card_index, int(audio_frame.len), int(audio_offset),
 476                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 477                 if (video_frame.owner) {
 478                         video_frame.owner->release_frame(video_frame);
 479                 }
 480                 if (audio_frame.owner) {
 481                         audio_frame.owner->release_frame(audio_frame);
 482                 }
 483                 return;
 484         }
 485
 486         int dropped_frames = 0;
 487         if (card->last_timecode != -1) {
 488                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 489         }
 490
 491         // Number of samples per frame if we need to insert silence.
 492         // (Could be nonintegral, but resampling will save us then.)
 493         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 494
 495         if (dropped_frames > MAX_FPS * 2) {
 496                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 497                         card_index, card->last_timecode, timecode);
 498                 audio_mixer.reset_resampler(device);
 499                 dropped_frames = 0;
 500         } else if (dropped_frames > 0) {
 501                 // Insert silence as needed.
 502                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 503                         card_index, dropped_frames, timecode);
 504
 505                 bool success;
 506                 do {
 507                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 508                 } while (!success);
 509         }
 510
 511         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 512
 513         // Done with the audio, so release it.
 514         if (audio_frame.owner) {
 515                 audio_frame.owner->release_frame(audio_frame);
 516         }
 517
 518         card->last_timecode = timecode;
 519
 520         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 521         if (video_frame.len - video_offset == 0 ||
 522             video_frame.len - video_offset != expected_length) {
 523                 if (video_frame.len != 0) {
 524                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 525                                 card_index, video_frame.len - video_offset, expected_length);
 526                 }
 527                 if (video_frame.owner) {
 528                         video_frame.owner->release_frame(video_frame);
 529                 }
 530
 531                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 532                 // so that pts can go up accordingly.
 533                 {
 534                         unique_lock<mutex> lock(card_mutex);
 535                         CaptureCard::NewFrame new_frame;
 536                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 537                         new_frame.length = frame_length;
 538                         new_frame.interlaced = false;
 539                         new_frame.dropped_frames = dropped_frames;
 540                         new_frame.received_timestamp = video_frame.received_timestamp;
 541                         card->new_frames.push_back(move(new_frame));
 542                         card->new_frames_changed.notify_all();
 543                 }
 544                 return;
 545         }
 546
 547         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 548
 549         unsigned num_fields = video_format.interlaced ? 2 : 1;
 550         steady_clock::time_point frame_upload_start;
 551         bool interlaced_stride = false;
 552         if (video_format.interlaced) {
 553                 // Send the two fields along as separate frames; the other side will need to add
 554                 // a deinterlacer to actually get this right.
 555                 assert(video_format.height % 2 == 0);
 556                 video_format.height /= 2;
 557                 assert(frame_length % 2 == 0);
 558                 frame_length /= 2;
 559                 num_fields = 2;
 560                 if (video_format.second_field_start == 1) {
 561                         interlaced_stride = true;
 562                 }
 563                 frame_upload_start = steady_clock::now();
 564         }
 565         userdata->last_interlaced = video_format.interlaced;
 566         userdata->last_has_signal = video_format.has_signal;
 567         userdata->last_is_connected = video_format.is_connected;
 568         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 569         userdata->last_frame_rate_den = video_format.frame_rate_den;
 570         RefCountedFrame frame(video_frame);
 571
 572         // Upload the textures.
 573         const size_t cbcr_width = video_format.width / 2;
 574         const size_t cbcr_offset = video_offset / 2;
 575         const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 576
 577         for (unsigned field = 0; field < num_fields; ++field) {
 578                 // Put the actual texture upload in a lambda that is executed in the main thread.
 579                 // It is entirely possible to do this in the same thread (and it might even be
 580                 // faster, depending on the GPU and driver), but it appears to be trickling
 581                 // driver bugs very easily.
 582                 //
 583                 // Note that this means we must hold on to the actual frame data in <userdata>
 584                 // until the upload command is run, but we hold on to <frame> much longer than that
 585                 // (in fact, all the way until we no longer use the texture in rendering).
 586                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 587                         unsigned field_start_line;
 588                         if (field == 1) {
 589                                 field_start_line = video_format.second_field_start;
 590                         } else {
 591                                 field_start_line = video_format.extra_lines_top;
 592                         }
 593
 594                         // For 8-bit input, v210_width will be nonsensical but not used.
 595                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 596                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 597
 598                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 599                         check_error();
 600
 601                         if (userdata->pixel_format == bmusb::PixelFormat_10BitYCbCr) {
 602                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 603                                 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);
 604                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 605                         } else {
 606                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 607                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 608
 609                                 // Make up our own strides, since we are interleaving.
 610                                 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);
 611                                 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);
 612                         }
 613
 614                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 615                         check_error();
 616                 };
 617
 618                 if (field == 1) {
 619                         // Don't upload the second field as fast as we can; wait until
 620                         // the field time has approximately passed. (Otherwise, we could
 621                         // get timing jitter against the other sources, and possibly also
 622                         // against the video display, although the latter is not as critical.)
 623                         // This requires our system clock to be reasonably close to the
 624                         // video clock, but that's not an unreasonable assumption.
 625                         steady_clock::time_point second_field_start = frame_upload_start +
 626                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 627                         this_thread::sleep_until(second_field_start);
 628                 }
 629
 630                 {
 631                         unique_lock<mutex> lock(card_mutex);
 632                         CaptureCard::NewFrame new_frame;
 633                         new_frame.frame = frame;
 634                         new_frame.length = frame_length;
 635                         new_frame.field = field;
 636                         new_frame.interlaced = video_format.interlaced;
 637                         new_frame.upload_func = upload_func;
 638                         new_frame.dropped_frames = dropped_frames;
 639                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 640                         card->new_frames.push_back(move(new_frame));
 641                         card->new_frames_changed.notify_all();
 642                 }
 643         }
 644 }
 645
 646 void Mixer::bm_hotplug_add(libusb_device *dev)
 647 {
 648         lock_guard<mutex> lock(hotplug_mutex);
 649         hotplugged_cards.push_back(dev);
 650 }
 651
 652 void Mixer::bm_hotplug_remove(unsigned card_index)
 653 {
 654         cards[card_index].new_frames_changed.notify_all();
 655 }
 656
 657 void Mixer::thread_func()
 658 {
 659         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 660
 661         eglBindAPI(EGL_OPENGL_API);
 662         QOpenGLContext *context = create_context(mixer_surface);
 663         if (!make_current(context, mixer_surface)) {
 664                 printf("oops\n");
 665                 exit(1);
 666         }
 667
 668         // Start the actual capture. (We don't want to do it before we're actually ready
 669         // to process output frames.)
 670         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 671                 if (int(card_index) != output_card_index) {
 672                         cards[card_index].capture->start_bm_capture();
 673                 }
 674         }
 675
 676         steady_clock::time_point start, now;
 677         start = steady_clock::now();
 678
 679         int stats_dropped_frames = 0;
 680
 681         while (!should_quit) {
 682                 if (desired_output_card_index != output_card_index) {
 683                         set_output_card_internal(desired_output_card_index);
 684                 }
 685                 if (output_card_index != -1 &&
 686                     desired_output_video_mode != output_video_mode) {
 687                         DeckLinkOutput *output = cards[output_card_index].output.get();
 688                         output->end_output();
 689                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 690                         output->start_output(desired_output_video_mode, pts_int);
 691                 }
 692
 693                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 694                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 695
 696                 bool master_card_is_output;
 697                 unsigned master_card_index;
 698                 if (output_card_index != -1) {
 699                         master_card_is_output = true;
 700                         master_card_index = output_card_index;
 701                 } else {
 702                         master_card_is_output = false;
 703                         master_card_index = theme->map_signal(master_clock_channel);
 704                         assert(master_card_index < num_cards);
 705                 }
 706
 707                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 708                 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);
 709                 stats_dropped_frames += output_frame_info.dropped_frames;
 710
 711                 handle_hotplugged_cards();
 712
 713                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 714                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 715                                 continue;
 716                         }
 717                         if (new_frames[card_index].frame->len == 0) {
 718                                 ++new_frames[card_index].dropped_frames;
 719                         }
 720                         if (new_frames[card_index].dropped_frames > 0) {
 721                                 printf("Card %u dropped %d frames before this\n",
 722                                         card_index, int(new_frames[card_index].dropped_frames));
 723                         }
 724                 }
 725
 726                 // If the first card is reporting a corrupted or otherwise dropped frame,
 727                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 728                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 729                         ++stats_dropped_frames;
 730                         pts_int += new_frames[master_card_index].length;
 731                         continue;
 732                 }
 733
 734                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 735                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 736                                 continue;
 737
 738                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 739                         assert(new_frame->frame != nullptr);
 740                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 741                         check_error();
 742
 743                         // The new texture might need uploading before use.
 744                         if (new_frame->upload_func) {
 745                                 new_frame->upload_func();
 746                                 new_frame->upload_func = nullptr;
 747                         }
 748                 }
 749
 750                 int64_t frame_duration = output_frame_info.frame_duration;
 751                 render_one_frame(frame_duration);
 752                 ++frame_num;
 753                 pts_int += frame_duration;
 754
 755                 now = steady_clock::now();
 756                 double elapsed = duration<double>(now - start).count();
 757                 if (frame_num % 100 == 0) {
 758                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 759                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 760                                 1e3 * elapsed / frame_num);
 761                 //      chain->print_phase_timing();
 762
 763                         // Check our memory usage, to see if we are close to our mlockall()
 764                         // limit (if at all set).
 765                         rusage used;
 766                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 767                                 perror("getrusage(RUSAGE_SELF)");
 768                                 assert(false);
 769                         }
 770
 771                         if (uses_mlock) {
 772                                 rlimit limit;
 773                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 774                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 775                                         assert(false);
 776                                 }
 777
 778                                 if (limit.rlim_cur == 0) {
 779                                         printf(", using %ld MB memory (locked)",
 780                                                 long(used.ru_maxrss / 1024));
 781                                 } else {
 782                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 783                                                 long(used.ru_maxrss / 1024),
 784                                                 long(limit.rlim_cur / 1048576),
 785                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 786                                 }
 787                         } else {
 788                                 printf(", using %ld MB memory (not locked)",
 789                                         long(used.ru_maxrss / 1024));
 790                         }
 791
 792                         printf("\n");
 793                 }
 794
 795
 796                 if (should_cut.exchange(false)) {  // Test and clear.
 797                         video_encoder->do_cut(frame_num);
 798                 }
 799
 800 #if 0
 801                 // Reset every 100 frames, so that local variations in frame times
 802                 // (especially for the first few frames, when the shaders are
 803                 // compiled etc.) don't make it hard to measure for the entire
 804                 // remaining duration of the program.
 805                 if (frame == 10000) {
 806                         frame = 0;
 807                         start = now;
 808                 }
 809 #endif
 810                 check_error();
 811         }
 812
 813         resource_pool->clean_context();
 814 }
 815
 816 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 817 {
 818         if (output_card_index != -1) {
 819                 // The output card (ie., cards[output_card_index].output) is the master clock,
 820                 // so no input card (ie., cards[card_index].capture) is.
 821                 return false;
 822         }
 823         return (card_index == master_card_index);
 824 }
 825
 826 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 827 {
 828         // Count the number of frames in the queue, including any frames
 829         // we dropped. It's hard to know exactly how we should deal with
 830         // dropped (corrupted) input frames; they don't help our goal of
 831         // avoiding starvation, but they still add to the problem of latency.
 832         // Since dropped frames is going to mean a bump in the signal anyway,
 833         // we err on the side of having more stable latency instead.
 834         unsigned queue_length = 0;
 835         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 836                 queue_length += frame.dropped_frames + 1;
 837         }
 838         card->queue_length_policy.update_policy(queue_length);
 839
 840         // If needed, drop frames until the queue is below the safe limit.
 841         // We prefer to drop from the head, because all else being equal,
 842         // we'd like more recent frames (less latency).
 843         unsigned dropped_frames = 0;
 844         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 845                 assert(!card->new_frames.empty());
 846                 assert(queue_length > card->new_frames.front().dropped_frames);
 847                 queue_length -= card->new_frames.front().dropped_frames;
 848
 849                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 850                         // No need to drop anything.
 851                         break;
 852                 }
 853
 854                 card->new_frames.pop_front();
 855                 card->new_frames_changed.notify_all();
 856                 --queue_length;
 857                 ++dropped_frames;
 858         }
 859
 860 #if 0
 861         if (dropped_frames > 0) {
 862                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 863                         card_index, dropped_frames);
 864         }
 865 #endif
 866 }
 867
 868
 869 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])
 870 {
 871         OutputFrameInfo output_frame_info;
 872 start:
 873         unique_lock<mutex> lock(card_mutex, defer_lock);
 874         if (master_card_is_output) {
 875                 // Clocked to the output, so wait for it to be ready for the next frame.
 876                 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);
 877                 lock.lock();
 878         } else {
 879                 // Wait for the master card to have a new frame.
 880                 // TODO: Add a timeout.
 881                 output_frame_info.is_preroll = false;
 882                 lock.lock();
 883                 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(); });
 884         }
 885
 886         if (master_card_is_output) {
 887                 handle_hotplugged_cards();
 888         } else if (cards[master_card_index].new_frames.empty()) {
 889                 // We were woken up, but not due to a new frame. Deal with it
 890                 // and then restart.
 891                 assert(cards[master_card_index].capture->get_disconnected());
 892                 handle_hotplugged_cards();
 893                 goto start;
 894         }
 895
 896         if (!master_card_is_output) {
 897                 output_frame_info.frame_timestamp =
 898                         cards[master_card_index].new_frames.front().received_timestamp;
 899         }
 900
 901         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 902                 CaptureCard *card = &cards[card_index];
 903                 if (input_card_is_master_clock(card_index, master_card_index)) {
 904                         // We don't use the queue length policy for the master card,
 905                         // but we will if it stops being the master. Thus, clear out
 906                         // the policy in case we switch in the future.
 907                         card->queue_length_policy.reset(card_index);
 908                         assert(!card->new_frames.empty());
 909                 } else {
 910                         trim_queue(card, card_index);
 911                 }
 912                 if (!card->new_frames.empty()) {
 913                         new_frames[card_index] = move(card->new_frames.front());
 914                         has_new_frame[card_index] = true;
 915                         card->new_frames.pop_front();
 916                         card->new_frames_changed.notify_all();
 917                 }
 918         }
 919
 920         if (!master_card_is_output) {
 921                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 922                 output_frame_info.frame_duration = new_frames[master_card_index].length;
 923         }
 924
 925         // This might get off by a fractional sample when changing master card
 926         // between ones with different frame rates, but that's fine.
 927         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 928         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 929         fractional_samples = num_samples_times_timebase % TIMEBASE;
 930         assert(output_frame_info.num_samples >= 0);
 931
 932         return output_frame_info;
 933 }
 934
 935 void Mixer::handle_hotplugged_cards()
 936 {
 937         // Check for cards that have been disconnected since last frame.
 938         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 939                 CaptureCard *card = &cards[card_index];
 940                 if (card->capture->get_disconnected()) {
 941                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 942                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 943                         configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
 944                         card->queue_length_policy.reset(card_index);
 945                         card->capture->start_bm_capture();
 946                 }
 947         }
 948
 949         // Check for cards that have been connected since last frame.
 950         vector<libusb_device *> hotplugged_cards_copy;
 951         {
 952                 lock_guard<mutex> lock(hotplug_mutex);
 953                 swap(hotplugged_cards, hotplugged_cards_copy);
 954         }
 955         for (libusb_device *new_dev : hotplugged_cards_copy) {
 956                 // Look for a fake capture card where we can stick this in.
 957                 int free_card_index = -1;
 958                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 959                         if (cards[card_index].is_fake_capture) {
 960                                 free_card_index = card_index;
 961                                 break;
 962                         }
 963                 }
 964
 965                 if (free_card_index == -1) {
 966                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 967                         libusb_unref_device(new_dev);
 968                 } else {
 969                         // BMUSBCapture takes ownership.
 970                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
 971                         CaptureCard *card = &cards[free_card_index];
 972                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
 973                         configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
 974                         card->queue_length_policy.reset(free_card_index);
 975                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
 976                         capture->start_bm_capture();
 977                 }
 978         }
 979 }
 980
 981
 982 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)
 983 {
 984         // Resample the audio as needed, including from previously dropped frames.
 985         assert(num_cards > 0);
 986         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
 987                 const bool dropped_frame = (frame_num != dropped_frames);
 988                 {
 989                         // Signal to the audio thread to process this frame.
 990                         // Note that if the frame is a dropped frame, we signal that
 991                         // we don't want to use this frame as base for adjusting
 992                         // the resampler rate. The reason for this is that the timing
 993                         // of these frames is often way too late; they typically don't
 994                         // “arrive” before we synthesize them. Thus, we could end up
 995                         // in a situation where we have inserted e.g. five audio frames
 996                         // into the queue before we then start pulling five of them
 997                         // back out. This makes ResamplingQueue overestimate the delay,
 998                         // causing undue resampler changes. (We _do_ use the last,
 999                         // non-dropped frame; perhaps we should just discard that as well,
1000                         // since dropped frames are expected to be rare, and it might be
1001                         // better to just wait until we have a slightly more normal situation).
1002                         unique_lock<mutex> lock(audio_mutex);
1003                         bool adjust_rate = !dropped_frame && !is_preroll;
1004                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1005                         audio_task_queue_changed.notify_one();
1006                 }
1007                 if (dropped_frame) {
1008                         // For dropped frames, increase the pts. Note that if the format changed
1009                         // in the meantime, we have no way of detecting that; we just have to
1010                         // assume the frame length is always the same.
1011                         pts_int += length_per_frame;
1012                 }
1013         }
1014 }
1015
1016 void Mixer::render_one_frame(int64_t duration)
1017 {
1018         // Determine the time code for this frame before we start rendering.
1019         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1020         if (display_timecode_on_stdout) {
1021                 printf("Timecode: '%s'\n", timecode_text.c_str());
1022         }
1023
1024         // Get the main chain from the theme, and set its state immediately.
1025         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1026         EffectChain *chain = theme_main_chain.chain;
1027         theme_main_chain.setup_chain();
1028         //theme_main_chain.chain->enable_phase_timing(true);
1029
1030         // If HDMI/SDI output is active and the user has requested auto mode,
1031         // its mode overrides the existing Y'CbCr setting for the chain.
1032         YCbCrLumaCoefficients ycbcr_output_coefficients;
1033         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1034                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1035         } else {
1036                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1037         }
1038
1039         // TODO: Reduce the duplication against theme.cpp.
1040         YCbCrFormat output_ycbcr_format;
1041         output_ycbcr_format.chroma_subsampling_x = 1;
1042         output_ycbcr_format.chroma_subsampling_y = 1;
1043         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1044         output_ycbcr_format.full_range = false;
1045         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1046         chain->change_ycbcr_output_format(output_ycbcr_format);
1047
1048         // Render main chain. If we're using zerocopy Quick Sync encoding
1049         // (the default case), we take an extra copy of the created outputs,
1050         // so that we can display it back to the screen later (it's less memory
1051         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1052         // Ideally, we'd like to avoid taking copies and just use the main textures
1053         // for display as well, but they're just views into VA-API memory and must be
1054         // unmapped during encoding, so we can't use them for display, unfortunately.
1055         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1056         GLuint y_copy_tex, cbcr_copy_tex = 0;
1057         GLuint y_display_tex, cbcr_display_tex;
1058         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1059         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1060         const bool is_zerocopy = video_encoder->is_zerocopy();
1061         if (is_zerocopy) {
1062                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1063                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1064                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1065
1066                 y_display_tex = y_copy_tex;
1067                 cbcr_display_tex = cbcr_copy_tex;
1068
1069                 // y_tex and cbcr_tex will be given by VideoEncoder.
1070         } else {
1071                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1072                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1073                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1074
1075                 y_display_tex = y_tex;
1076                 cbcr_display_tex = cbcr_tex;
1077         }
1078
1079         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1080         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1081         assert(got_frame);
1082
1083         GLuint fbo;
1084         if (is_zerocopy) {
1085                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1086         } else {
1087                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1088         }
1089         check_error();
1090         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1091
1092         if (display_timecode_in_stream) {
1093                 // Render the timecode on top.
1094                 timecode_renderer->render_timecode(fbo, timecode_text);
1095         }
1096
1097         resource_pool->release_fbo(fbo);
1098
1099         if (is_zerocopy) {
1100                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1101         } else {
1102                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1103         }
1104         if (output_card_index != -1) {
1105                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1106         }
1107         resource_pool->release_2d_texture(cbcr_full_tex);
1108
1109         // Set the right state for the Y' and CbCr textures we use for display.
1110         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1111         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1112         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1113         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1114         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1115
1116         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1117         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1118         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1119         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1120
1121         RefCountedGLsync fence = video_encoder->end_frame();
1122
1123         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1124         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1125         DisplayFrame live_frame;
1126         live_frame.chain = display_chain.get();
1127         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1128                 display_input->set_texture_num(0, y_display_tex);
1129                 display_input->set_texture_num(1, cbcr_display_tex);
1130         };
1131         live_frame.ready_fence = fence;
1132         live_frame.input_frames = {};
1133         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1134         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1135
1136         // Set up preview and any additional channels.
1137         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1138                 DisplayFrame display_frame;
1139                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1140                 display_frame.chain = chain.chain;
1141                 display_frame.setup_chain = chain.setup_chain;
1142                 display_frame.ready_fence = fence;
1143                 display_frame.input_frames = chain.input_frames;
1144                 display_frame.temp_textures = {};
1145                 output_channel[i].output_frame(display_frame);
1146         }
1147 }
1148
1149 void Mixer::audio_thread_func()
1150 {
1151         pthread_setname_np(pthread_self(), "Mixer_Audio");
1152
1153         while (!should_quit) {
1154                 AudioTask task;
1155
1156                 {
1157                         unique_lock<mutex> lock(audio_mutex);
1158                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1159                         if (should_quit) {
1160                                 return;
1161                         }
1162                         task = audio_task_queue.front();
1163                         audio_task_queue.pop();
1164                 }
1165
1166                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1167                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1168                 vector<float> samples_out = audio_mixer.get_output(
1169                         task.frame_timestamp,
1170                         task.num_samples,
1171                         rate_adjustment_policy);
1172
1173                 // Send the samples to the sound card, then add them to the output.
1174                 if (alsa) {
1175                         alsa->write(samples_out);
1176                 }
1177                 if (output_card_index != -1) {
1178                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1179                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1180                 }
1181                 video_encoder->add_audio(task.pts_int, move(samples_out));
1182         }
1183 }
1184
1185 void Mixer::release_display_frame(DisplayFrame *frame)
1186 {
1187         for (GLuint texnum : frame->temp_textures) {
1188                 resource_pool->release_2d_texture(texnum);
1189         }
1190         frame->temp_textures.clear();
1191         frame->ready_fence.reset();
1192         frame->input_frames.clear();
1193 }
1194
1195 void Mixer::start()
1196 {
1197         mixer_thread = thread(&Mixer::thread_func, this);
1198         audio_thread = thread(&Mixer::audio_thread_func, this);
1199 }
1200
1201 void Mixer::quit()
1202 {
1203         should_quit = true;
1204         audio_task_queue_changed.notify_one();
1205         mixer_thread.join();
1206         audio_thread.join();
1207 }
1208
1209 void Mixer::transition_clicked(int transition_num)
1210 {
1211         theme->transition_clicked(transition_num, pts());
1212 }
1213
1214 void Mixer::channel_clicked(int preview_num)
1215 {
1216         theme->channel_clicked(preview_num);
1217 }
1218
1219 void Mixer::start_mode_scanning(unsigned card_index)
1220 {
1221         assert(card_index < num_cards);
1222         if (is_mode_scanning[card_index]) {
1223                 return;
1224         }
1225         is_mode_scanning[card_index] = true;
1226         mode_scanlist[card_index].clear();
1227         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1228                 mode_scanlist[card_index].push_back(mode.first);
1229         }
1230         assert(!mode_scanlist[card_index].empty());
1231         mode_scanlist_index[card_index] = 0;
1232         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1233         last_mode_scan_change[card_index] = steady_clock::now();
1234 }
1235
1236 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1237 {
1238         assert(desired_output_card_index != -1);
1239         unique_lock<mutex> lock(card_mutex);
1240         return cards[desired_output_card_index].output->get_available_video_modes();
1241 }
1242
1243 Mixer::OutputChannel::~OutputChannel()
1244 {
1245         if (has_current_frame) {
1246                 parent->release_display_frame(&current_frame);
1247         }
1248         if (has_ready_frame) {
1249                 parent->release_display_frame(&ready_frame);
1250         }
1251 }
1252
1253 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1254 {
1255         // Store this frame for display. Remove the ready frame if any
1256         // (it was seemingly never used).
1257         {
1258                 unique_lock<mutex> lock(frame_mutex);
1259                 if (has_ready_frame) {
1260                         parent->release_display_frame(&ready_frame);
1261                 }
1262                 ready_frame = frame;
1263                 has_ready_frame = true;
1264         }
1265
1266         if (new_frame_ready_callback) {
1267                 new_frame_ready_callback();
1268         }
1269
1270         // Reduce the number of callbacks by filtering duplicates. The reason
1271         // why we bother doing this is that Qt seemingly can get into a state
1272         // where its builds up an essentially unbounded queue of signals,
1273         // consuming more and more memory, and there's no good way of collapsing
1274         // user-defined signals or limiting the length of the queue.
1275         if (transition_names_updated_callback) {
1276                 vector<string> transition_names = global_mixer->get_transition_names();
1277                 bool changed = false;
1278                 if (transition_names.size() != last_transition_names.size()) {
1279                         changed = true;
1280                 } else {
1281                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1282                                 if (transition_names[i] != last_transition_names[i]) {
1283                                         changed = true;
1284                                         break;
1285                                 }
1286                         }
1287                 }
1288                 if (changed) {
1289                         transition_names_updated_callback(transition_names);
1290                         last_transition_names = transition_names;
1291                 }
1292         }
1293         if (name_updated_callback) {
1294                 string name = global_mixer->get_channel_name(channel);
1295                 if (name != last_name) {
1296                         name_updated_callback(name);
1297                         last_name = name;
1298                 }
1299         }
1300         if (color_updated_callback) {
1301                 string color = global_mixer->get_channel_color(channel);
1302                 if (color != last_color) {
1303                         color_updated_callback(color);
1304                         last_color = color;
1305                 }
1306         }
1307 }
1308
1309 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1310 {
1311         unique_lock<mutex> lock(frame_mutex);
1312         if (!has_current_frame && !has_ready_frame) {
1313                 return false;
1314         }
1315
1316         if (has_current_frame && has_ready_frame) {
1317                 // We have a new ready frame. Toss the current one.
1318                 parent->release_display_frame(&current_frame);
1319                 has_current_frame = false;
1320         }
1321         if (has_ready_frame) {
1322                 assert(!has_current_frame);
1323                 current_frame = ready_frame;
1324                 ready_frame.ready_fence.reset();  // Drop the refcount.
1325                 ready_frame.input_frames.clear();  // Drop the refcounts.
1326                 has_current_frame = true;
1327                 has_ready_frame = false;
1328         }
1329
1330         *frame = current_frame;
1331         return true;
1332 }
1333
1334 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1335 {
1336         new_frame_ready_callback = callback;
1337 }
1338
1339 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1340 {
1341         transition_names_updated_callback = callback;
1342 }
1343
1344 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1345 {
1346         name_updated_callback = callback;
1347 }
1348
1349 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1350 {
1351         color_updated_callback = callback;
1352 }
1353
1354 mutex RefCountedGLsync::fence_lock;