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         switch (userdata->pixel_format) {
  85         case bmusb::PixelFormat_10BitYCbCr:
  86                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  87                 break;
  88         case bmusb::PixelFormat_8BitYCbCr:
  89                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
  90                 break;
  91         case bmusb::PixelFormat_8BitRGBA:
  92                 first = userdata->tex_rgba[field] == 0;
  93                 break;
  94         default:
  95                 assert(false);
  96         }
  97
  98         if (first ||
  99             width != userdata->last_width[field] ||
 100             height != userdata->last_height[field]) {
 101                 // We changed resolution since last use of this texture, so we need to create
 102                 // a new object. Note that this each card has its own PBOFrameAllocator,
 103                 // we don't need to worry about these flip-flopping between resolutions.
 104                 switch (userdata->pixel_format) {
 105                 case bmusb::PixelFormat_10BitYCbCr:
 106                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 107                         check_error();
 108                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 109                         check_error();
 110                         break;
 111                 case bmusb::PixelFormat_8BitYCbCr: {
 112                         size_t cbcr_width = width / 2;
 113
 114                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 115                         check_error();
 116                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 117                         check_error();
 118                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 119                         check_error();
 120                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 121                         check_error();
 122                         break;
 123                 }
 124                 case bmusb::PixelFormat_8BitRGBA:
 125                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 126                         check_error();
 127                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
 128                         check_error();
 129                         break;
 130                 }
 131                 userdata->last_width[field] = width;
 132                 userdata->last_height[field] = height;
 133         }
 134         if (global_flags.ten_bit_input &&
 135             (first || v210_width != userdata->last_v210_width[field])) {
 136                 // Same as above; we need to recreate the texture.
 137                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 138                 check_error();
 139                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 140                 check_error();
 141                 userdata->last_v210_width[field] = v210_width;
 142         }
 143 }
 144
 145 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 146 {
 147         if (interlaced_stride) {
 148                 stride *= 2;
 149         }
 150         if (global_flags.flush_pbos) {
 151                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 152                 check_error();
 153         }
 154
 155         glBindTexture(GL_TEXTURE_2D, tex);
 156         check_error();
 157         if (interlaced_stride) {
 158                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 159                 check_error();
 160         } else {
 161                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 162                 check_error();
 163         }
 164
 165         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 166         check_error();
 167         glBindTexture(GL_TEXTURE_2D, 0);
 168         check_error();
 169         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 170         check_error();
 171 }
 172
 173 }  // namespace
 174
 175 void QueueLengthPolicy::update_policy(unsigned queue_length)
 176 {
 177         if (queue_length == 0) {  // Starvation.
 178                 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
 179                         ++safe_queue_length;
 180                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
 181                                 card_index, safe_queue_length);
 182                 }
 183                 frames_with_at_least_one = 0;
 184                 been_at_safe_point_since_last_starvation = false;
 185                 return;
 186         }
 187         if (queue_length >= safe_queue_length) {
 188                 been_at_safe_point_since_last_starvation = true;
 189         }
 190         if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
 191                 --safe_queue_length;
 192                 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
 193                         card_index, safe_queue_length);
 194                 frames_with_at_least_one = 0;
 195         }
 196 }
 197
 198 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 199         : httpd(),
 200           num_cards(num_cards),
 201           mixer_surface(create_surface(format)),
 202           h264_encoder_surface(create_surface(format)),
 203           decklink_output_surface(create_surface(format)),
 204           audio_mixer(num_cards)
 205 {
 206         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 207         check_error();
 208
 209         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 210         // will be halved when sampling them, and we need to compensate here.
 211         movit_texel_subpixel_precision /= 2.0;
 212
 213         resource_pool.reset(new ResourcePool);
 214         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 215                 output_channel[i].parent = this;
 216                 output_channel[i].channel = i;
 217         }
 218
 219         ImageFormat inout_format;
 220         inout_format.color_space = COLORSPACE_sRGB;
 221         inout_format.gamma_curve = GAMMA_sRGB;
 222
 223         // Matches the 4:2:0 format created by the main chain.
 224         YCbCrFormat ycbcr_format;
 225         ycbcr_format.chroma_subsampling_x = 2;
 226         ycbcr_format.chroma_subsampling_y = 2;
 227         if (global_flags.ycbcr_rec709_coefficients) {
 228                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 229         } else {
 230                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 231         }
 232         ycbcr_format.full_range = false;
 233         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 234         ycbcr_format.cb_x_position = 0.0f;
 235         ycbcr_format.cr_x_position = 0.0f;
 236         ycbcr_format.cb_y_position = 0.5f;
 237         ycbcr_format.cr_y_position = 0.5f;
 238
 239         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 240         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 241         check_error();
 242         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 243         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 244         display_chain->add_input(display_input);
 245         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 246         display_chain->set_dither_bits(0);  // Don't bother.
 247         display_chain->finalize();
 248
 249         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));
 250
 251         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 252         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 253
 254         // Start listening for clients only once VideoEncoder has written its header, if any.
 255         httpd.start(9095);
 256
 257         // First try initializing the then PCI devices, then USB, then
 258         // fill up with fake cards until we have the desired number of cards.
 259         unsigned num_pci_devices = 0;
 260         unsigned card_index = 0;
 261
 262         {
 263                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 264                 if (decklink_iterator != nullptr) {
 265                         for ( ; card_index < num_cards; ++card_index) {
 266                                 IDeckLink *decklink;
 267                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 268                                         break;
 269                                 }
 270
 271                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 272                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 273                                 output->set_device(decklink);
 274                                 configure_card(card_index, capture, /*is_fake_capture=*/false, output);
 275                                 ++num_pci_devices;
 276                         }
 277                         decklink_iterator->Release();
 278                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 279                 } else {
 280                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 281                 }
 282         }
 283
 284         unsigned num_usb_devices = BMUSBCapture::num_cards();
 285         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 286                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 287                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 288                 configure_card(card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
 289         }
 290         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 291
 292         unsigned num_fake_cards = 0;
 293         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 294                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 295                 configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
 296         }
 297
 298         if (num_fake_cards > 0) {
 299                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 300         }
 301
 302         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 303         BMUSBCapture::start_bm_thread();
 304
 305         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 306                 cards[card_index].queue_length_policy.reset(card_index);
 307         }
 308
 309         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 310
 311         if (global_flags.ten_bit_input) {
 312                 if (!v210Converter::has_hardware_support()) {
 313                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 314                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 315                         exit(1);
 316                 }
 317                 v210_converter.reset(new v210Converter());
 318
 319                 // These are all the widths listed in the Blackmagic SDK documentation
 320                 // (section 2.7.3, “Display Modes”).
 321                 v210_converter->precompile_shader(720);
 322                 v210_converter->precompile_shader(1280);
 323                 v210_converter->precompile_shader(1920);
 324                 v210_converter->precompile_shader(2048);
 325                 v210_converter->precompile_shader(3840);
 326                 v210_converter->precompile_shader(4096);
 327         }
 328         if (global_flags.ten_bit_output) {
 329                 if (!v210Converter::has_hardware_support()) {
 330                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 331                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 332                         exit(1);
 333                 }
 334         }
 335
 336         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 337         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 338         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 339
 340         if (global_flags.enable_alsa_output) {
 341                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 342         }
 343         if (global_flags.output_card != -1) {
 344                 desired_output_card_index = global_flags.output_card;
 345                 set_output_card_internal(global_flags.output_card);
 346         }
 347 }
 348
 349 Mixer::~Mixer()
 350 {
 351         BMUSBCapture::stop_bm_thread();
 352
 353         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 354                 {
 355                         unique_lock<mutex> lock(card_mutex);
 356                         cards[card_index].should_quit = true;  // Unblock thread.
 357                         cards[card_index].new_frames_changed.notify_all();
 358                 }
 359                 cards[card_index].capture->stop_dequeue_thread();
 360                 if (cards[card_index].output) {
 361                         cards[card_index].output->end_output();
 362                         cards[card_index].output.reset();
 363                 }
 364         }
 365
 366         video_encoder.reset(nullptr);
 367 }
 368
 369 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output)
 370 {
 371         printf("Configuring card %d...\n", card_index);
 372
 373         CaptureCard *card = &cards[card_index];
 374         if (card->capture != nullptr) {
 375                 card->capture->stop_dequeue_thread();
 376         }
 377         card->capture.reset(capture);
 378         card->is_fake_capture = is_fake_capture;
 379         if (card->output.get() != output) {
 380                 card->output.reset(output);
 381         }
 382
 383         bmusb::PixelFormat pixel_format = global_flags.ten_bit_input ? PixelFormat_10BitYCbCr : PixelFormat_8BitYCbCr;
 384
 385         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 386         if (card->frame_allocator == nullptr) {
 387                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 388         }
 389         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 390         if (card->surface == nullptr) {
 391                 card->surface = create_surface_with_same_format(mixer_surface);
 392         }
 393         while (!card->new_frames.empty()) card->new_frames.pop_front();
 394         card->last_timecode = -1;
 395         card->capture->set_pixel_format(pixel_format);
 396         card->capture->configure_card();
 397
 398         // NOTE: start_bm_capture() happens in thread_func().
 399
 400         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 401         audio_mixer.reset_resampler(device);
 402         audio_mixer.set_display_name(device, card->capture->get_description());
 403         audio_mixer.trigger_state_changed_callback();
 404 }
 405
 406 void Mixer::set_output_card_internal(int card_index)
 407 {
 408         // We don't really need to take card_mutex, since we're in the mixer
 409         // thread and don't mess with any queues (which is the only thing that happens
 410         // from other threads), but it's probably the safest in the long run.
 411         unique_lock<mutex> lock(card_mutex);
 412         if (output_card_index != -1) {
 413                 // Switch the old card from output to input.
 414                 CaptureCard *old_card = &cards[output_card_index];
 415                 old_card->output->end_output();
 416
 417                 // Stop the fake card that we put into place.
 418                 // This needs to _not_ happen under the mutex, to avoid deadlock
 419                 // (delivering the last frame needs to take the mutex).
 420                 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
 421                 lock.unlock();
 422                 fake_capture->stop_dequeue_thread();
 423                 lock.lock();
 424                 old_card->capture = move(old_card->parked_capture);
 425                 old_card->is_fake_capture = false;
 426                 old_card->capture->start_bm_capture();
 427         }
 428         if (card_index != -1) {
 429                 CaptureCard *card = &cards[card_index];
 430                 bmusb::CaptureInterface *capture = card->capture.get();
 431                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 432                 // several seconds to complete (blocking on DisableVideoInput);
 433                 // see if we can maybe do it asynchronously.
 434                 lock.unlock();
 435                 capture->stop_dequeue_thread();
 436                 lock.lock();
 437                 card->parked_capture = move(card->capture);
 438                 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 439                 configure_card(card_index, fake_capture, /*is_fake_capture=*/true, card->output.release());
 440                 card->queue_length_policy.reset(card_index);
 441                 card->capture->start_bm_capture();
 442                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 443                 card->output->start_output(desired_output_video_mode, pts_int);
 444         }
 445         output_card_index = card_index;
 446 }
 447
 448 namespace {
 449
 450 int unwrap_timecode(uint16_t current_wrapped, int last)
 451 {
 452         uint16_t last_wrapped = last & 0xffff;
 453         if (current_wrapped > last_wrapped) {
 454                 return (last & ~0xffff) | current_wrapped;
 455         } else {
 456                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 457         }
 458 }
 459
 460 }  // namespace
 461
 462 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 463                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 464                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 465 {
 466         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 467         CaptureCard *card = &cards[card_index];
 468
 469         if (is_mode_scanning[card_index]) {
 470                 if (video_format.has_signal) {
 471                         // Found a stable signal, so stop scanning.
 472                         is_mode_scanning[card_index] = false;
 473                 } else {
 474                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 475                         steady_clock::time_point now = steady_clock::now();
 476                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 477                         if (sec_since_last_switch > switch_time_s) {
 478                                 // It isn't this mode; try the next one.
 479                                 mode_scanlist_index[card_index]++;
 480                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 481                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 482                                 last_mode_scan_change[card_index] = now;
 483                         }
 484                 }
 485         }
 486
 487         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 488         assert(frame_length > 0);
 489
 490         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;
 491         if (num_samples > OUTPUT_FREQUENCY / 10) {
 492                 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",
 493                         card_index, int(audio_frame.len), int(audio_offset),
 494                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 495                 if (video_frame.owner) {
 496                         video_frame.owner->release_frame(video_frame);
 497                 }
 498                 if (audio_frame.owner) {
 499                         audio_frame.owner->release_frame(audio_frame);
 500                 }
 501                 return;
 502         }
 503
 504         int dropped_frames = 0;
 505         if (card->last_timecode != -1) {
 506                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 507         }
 508
 509         // Number of samples per frame if we need to insert silence.
 510         // (Could be nonintegral, but resampling will save us then.)
 511         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 512
 513         if (dropped_frames > MAX_FPS * 2) {
 514                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 515                         card_index, card->last_timecode, timecode);
 516                 audio_mixer.reset_resampler(device);
 517                 dropped_frames = 0;
 518         } else if (dropped_frames > 0) {
 519                 // Insert silence as needed.
 520                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 521                         card_index, dropped_frames, timecode);
 522
 523                 bool success;
 524                 do {
 525                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 526                 } while (!success);
 527         }
 528
 529         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 530
 531         // Done with the audio, so release it.
 532         if (audio_frame.owner) {
 533                 audio_frame.owner->release_frame(audio_frame);
 534         }
 535
 536         card->last_timecode = timecode;
 537
 538         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 539         if (video_frame.len - video_offset == 0 ||
 540             video_frame.len - video_offset != expected_length) {
 541                 if (video_frame.len != 0) {
 542                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 543                                 card_index, video_frame.len - video_offset, expected_length);
 544                 }
 545                 if (video_frame.owner) {
 546                         video_frame.owner->release_frame(video_frame);
 547                 }
 548
 549                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 550                 // so that pts can go up accordingly.
 551                 {
 552                         unique_lock<mutex> lock(card_mutex);
 553                         CaptureCard::NewFrame new_frame;
 554                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 555                         new_frame.length = frame_length;
 556                         new_frame.interlaced = false;
 557                         new_frame.dropped_frames = dropped_frames;
 558                         new_frame.received_timestamp = video_frame.received_timestamp;
 559                         card->new_frames.push_back(move(new_frame));
 560                         card->new_frames_changed.notify_all();
 561                 }
 562                 return;
 563         }
 564
 565         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 566
 567         unsigned num_fields = video_format.interlaced ? 2 : 1;
 568         steady_clock::time_point frame_upload_start;
 569         bool interlaced_stride = false;
 570         if (video_format.interlaced) {
 571                 // Send the two fields along as separate frames; the other side will need to add
 572                 // a deinterlacer to actually get this right.
 573                 assert(video_format.height % 2 == 0);
 574                 video_format.height /= 2;
 575                 assert(frame_length % 2 == 0);
 576                 frame_length /= 2;
 577                 num_fields = 2;
 578                 if (video_format.second_field_start == 1) {
 579                         interlaced_stride = true;
 580                 }
 581                 frame_upload_start = steady_clock::now();
 582         }
 583         userdata->last_interlaced = video_format.interlaced;
 584         userdata->last_has_signal = video_format.has_signal;
 585         userdata->last_is_connected = video_format.is_connected;
 586         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 587         userdata->last_frame_rate_den = video_format.frame_rate_den;
 588         RefCountedFrame frame(video_frame);
 589
 590         // Upload the textures.
 591         const size_t cbcr_width = video_format.width / 2;
 592         const size_t cbcr_offset = video_offset / 2;
 593         const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 594
 595         for (unsigned field = 0; field < num_fields; ++field) {
 596                 // Put the actual texture upload in a lambda that is executed in the main thread.
 597                 // It is entirely possible to do this in the same thread (and it might even be
 598                 // faster, depending on the GPU and driver), but it appears to be trickling
 599                 // driver bugs very easily.
 600                 //
 601                 // Note that this means we must hold on to the actual frame data in <userdata>
 602                 // until the upload command is run, but we hold on to <frame> much longer than that
 603                 // (in fact, all the way until we no longer use the texture in rendering).
 604                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 605                         unsigned field_start_line;
 606                         if (field == 1) {
 607                                 field_start_line = video_format.second_field_start;
 608                         } else {
 609                                 field_start_line = video_format.extra_lines_top;
 610                         }
 611
 612                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 613                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 614                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 615
 616                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 617                         check_error();
 618
 619                         switch (userdata->pixel_format) {
 620                         case bmusb::PixelFormat_10BitYCbCr: {
 621                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 622                                 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);
 623                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 624                                 break;
 625                         }
 626                         case bmusb::PixelFormat_8BitYCbCr: {
 627                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 628                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 629
 630                                 // Make up our own strides, since we are interleaving.
 631                                 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);
 632                                 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);
 633                                 break;
 634                         }
 635                         case bmusb::PixelFormat_8BitRGBA: {
 636                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 637                                 upload_texture(userdata->tex_rgba[field], video_format.width, video_format.height, video_format.stride, interlaced_stride, GL_RGBA, GL_UNSIGNED_BYTE, field_start);
 638                                 break;
 639                         }
 640                         default:
 641                                 assert(false);
 642                         }
 643
 644                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 645                         check_error();
 646                 };
 647
 648                 if (field == 1) {
 649                         // Don't upload the second field as fast as we can; wait until
 650                         // the field time has approximately passed. (Otherwise, we could
 651                         // get timing jitter against the other sources, and possibly also
 652                         // against the video display, although the latter is not as critical.)
 653                         // This requires our system clock to be reasonably close to the
 654                         // video clock, but that's not an unreasonable assumption.
 655                         steady_clock::time_point second_field_start = frame_upload_start +
 656                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 657                         this_thread::sleep_until(second_field_start);
 658                 }
 659
 660                 {
 661                         unique_lock<mutex> lock(card_mutex);
 662                         CaptureCard::NewFrame new_frame;
 663                         new_frame.frame = frame;
 664                         new_frame.length = frame_length;
 665                         new_frame.field = field;
 666                         new_frame.interlaced = video_format.interlaced;
 667                         new_frame.upload_func = upload_func;
 668                         new_frame.dropped_frames = dropped_frames;
 669                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 670                         card->new_frames.push_back(move(new_frame));
 671                         card->new_frames_changed.notify_all();
 672                 }
 673         }
 674 }
 675
 676 void Mixer::bm_hotplug_add(libusb_device *dev)
 677 {
 678         lock_guard<mutex> lock(hotplug_mutex);
 679         hotplugged_cards.push_back(dev);
 680 }
 681
 682 void Mixer::bm_hotplug_remove(unsigned card_index)
 683 {
 684         cards[card_index].new_frames_changed.notify_all();
 685 }
 686
 687 void Mixer::thread_func()
 688 {
 689         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 690
 691         eglBindAPI(EGL_OPENGL_API);
 692         QOpenGLContext *context = create_context(mixer_surface);
 693         if (!make_current(context, mixer_surface)) {
 694                 printf("oops\n");
 695                 exit(1);
 696         }
 697
 698         // Start the actual capture. (We don't want to do it before we're actually ready
 699         // to process output frames.)
 700         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 701                 if (int(card_index) != output_card_index) {
 702                         cards[card_index].capture->start_bm_capture();
 703                 }
 704         }
 705
 706         steady_clock::time_point start, now;
 707         start = steady_clock::now();
 708
 709         int stats_dropped_frames = 0;
 710
 711         while (!should_quit) {
 712                 if (desired_output_card_index != output_card_index) {
 713                         set_output_card_internal(desired_output_card_index);
 714                 }
 715                 if (output_card_index != -1 &&
 716                     desired_output_video_mode != output_video_mode) {
 717                         DeckLinkOutput *output = cards[output_card_index].output.get();
 718                         output->end_output();
 719                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 720                         output->start_output(desired_output_video_mode, pts_int);
 721                 }
 722
 723                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 724                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 725
 726                 bool master_card_is_output;
 727                 unsigned master_card_index;
 728                 if (output_card_index != -1) {
 729                         master_card_is_output = true;
 730                         master_card_index = output_card_index;
 731                 } else {
 732                         master_card_is_output = false;
 733                         master_card_index = theme->map_signal(master_clock_channel);
 734                         assert(master_card_index < num_cards);
 735                 }
 736
 737                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 738                 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);
 739                 stats_dropped_frames += output_frame_info.dropped_frames;
 740
 741                 handle_hotplugged_cards();
 742
 743                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 744                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 745                                 continue;
 746                         }
 747                         if (new_frames[card_index].frame->len == 0) {
 748                                 ++new_frames[card_index].dropped_frames;
 749                         }
 750                         if (new_frames[card_index].dropped_frames > 0) {
 751                                 printf("Card %u dropped %d frames before this\n",
 752                                         card_index, int(new_frames[card_index].dropped_frames));
 753                         }
 754                 }
 755
 756                 // If the first card is reporting a corrupted or otherwise dropped frame,
 757                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 758                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 759                         ++stats_dropped_frames;
 760                         pts_int += new_frames[master_card_index].length;
 761                         continue;
 762                 }
 763
 764                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 765                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 766                                 continue;
 767
 768                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 769                         assert(new_frame->frame != nullptr);
 770                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 771                         check_error();
 772
 773                         // The new texture might need uploading before use.
 774                         if (new_frame->upload_func) {
 775                                 new_frame->upload_func();
 776                                 new_frame->upload_func = nullptr;
 777                         }
 778                 }
 779
 780                 int64_t frame_duration = output_frame_info.frame_duration;
 781                 render_one_frame(frame_duration);
 782                 ++frame_num;
 783                 pts_int += frame_duration;
 784
 785                 now = steady_clock::now();
 786                 double elapsed = duration<double>(now - start).count();
 787                 if (frame_num % 100 == 0) {
 788                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 789                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 790                                 1e3 * elapsed / frame_num);
 791                 //      chain->print_phase_timing();
 792
 793                         // Check our memory usage, to see if we are close to our mlockall()
 794                         // limit (if at all set).
 795                         rusage used;
 796                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 797                                 perror("getrusage(RUSAGE_SELF)");
 798                                 assert(false);
 799                         }
 800
 801                         if (uses_mlock) {
 802                                 rlimit limit;
 803                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 804                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 805                                         assert(false);
 806                                 }
 807
 808                                 if (limit.rlim_cur == 0) {
 809                                         printf(", using %ld MB memory (locked)",
 810                                                 long(used.ru_maxrss / 1024));
 811                                 } else {
 812                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 813                                                 long(used.ru_maxrss / 1024),
 814                                                 long(limit.rlim_cur / 1048576),
 815                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 816                                 }
 817                         } else {
 818                                 printf(", using %ld MB memory (not locked)",
 819                                         long(used.ru_maxrss / 1024));
 820                         }
 821
 822                         printf("\n");
 823                 }
 824
 825
 826                 if (should_cut.exchange(false)) {  // Test and clear.
 827                         video_encoder->do_cut(frame_num);
 828                 }
 829
 830 #if 0
 831                 // Reset every 100 frames, so that local variations in frame times
 832                 // (especially for the first few frames, when the shaders are
 833                 // compiled etc.) don't make it hard to measure for the entire
 834                 // remaining duration of the program.
 835                 if (frame == 10000) {
 836                         frame = 0;
 837                         start = now;
 838                 }
 839 #endif
 840                 check_error();
 841         }
 842
 843         resource_pool->clean_context();
 844 }
 845
 846 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 847 {
 848         if (output_card_index != -1) {
 849                 // The output card (ie., cards[output_card_index].output) is the master clock,
 850                 // so no input card (ie., cards[card_index].capture) is.
 851                 return false;
 852         }
 853         return (card_index == master_card_index);
 854 }
 855
 856 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 857 {
 858         // Count the number of frames in the queue, including any frames
 859         // we dropped. It's hard to know exactly how we should deal with
 860         // dropped (corrupted) input frames; they don't help our goal of
 861         // avoiding starvation, but they still add to the problem of latency.
 862         // Since dropped frames is going to mean a bump in the signal anyway,
 863         // we err on the side of having more stable latency instead.
 864         unsigned queue_length = 0;
 865         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 866                 queue_length += frame.dropped_frames + 1;
 867         }
 868         card->queue_length_policy.update_policy(queue_length);
 869
 870         // If needed, drop frames until the queue is below the safe limit.
 871         // We prefer to drop from the head, because all else being equal,
 872         // we'd like more recent frames (less latency).
 873         unsigned dropped_frames = 0;
 874         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 875                 assert(!card->new_frames.empty());
 876                 assert(queue_length > card->new_frames.front().dropped_frames);
 877                 queue_length -= card->new_frames.front().dropped_frames;
 878
 879                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 880                         // No need to drop anything.
 881                         break;
 882                 }
 883
 884                 card->new_frames.pop_front();
 885                 card->new_frames_changed.notify_all();
 886                 --queue_length;
 887                 ++dropped_frames;
 888         }
 889
 890 #if 0
 891         if (dropped_frames > 0) {
 892                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 893                         card_index, dropped_frames);
 894         }
 895 #endif
 896 }
 897
 898
 899 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])
 900 {
 901         OutputFrameInfo output_frame_info;
 902 start:
 903         unique_lock<mutex> lock(card_mutex, defer_lock);
 904         if (master_card_is_output) {
 905                 // Clocked to the output, so wait for it to be ready for the next frame.
 906                 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);
 907                 lock.lock();
 908         } else {
 909                 // Wait for the master card to have a new frame.
 910                 // TODO: Add a timeout.
 911                 output_frame_info.is_preroll = false;
 912                 lock.lock();
 913                 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(); });
 914         }
 915
 916         if (master_card_is_output) {
 917                 handle_hotplugged_cards();
 918         } else if (cards[master_card_index].new_frames.empty()) {
 919                 // We were woken up, but not due to a new frame. Deal with it
 920                 // and then restart.
 921                 assert(cards[master_card_index].capture->get_disconnected());
 922                 handle_hotplugged_cards();
 923                 goto start;
 924         }
 925
 926         if (!master_card_is_output) {
 927                 output_frame_info.frame_timestamp =
 928                         cards[master_card_index].new_frames.front().received_timestamp;
 929         }
 930
 931         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 932                 CaptureCard *card = &cards[card_index];
 933                 if (input_card_is_master_clock(card_index, master_card_index)) {
 934                         // We don't use the queue length policy for the master card,
 935                         // but we will if it stops being the master. Thus, clear out
 936                         // the policy in case we switch in the future.
 937                         card->queue_length_policy.reset(card_index);
 938                         assert(!card->new_frames.empty());
 939                 } else {
 940                         trim_queue(card, card_index);
 941                 }
 942                 if (!card->new_frames.empty()) {
 943                         new_frames[card_index] = move(card->new_frames.front());
 944                         has_new_frame[card_index] = true;
 945                         card->new_frames.pop_front();
 946                         card->new_frames_changed.notify_all();
 947                 }
 948         }
 949
 950         if (!master_card_is_output) {
 951                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 952                 output_frame_info.frame_duration = new_frames[master_card_index].length;
 953         }
 954
 955         // This might get off by a fractional sample when changing master card
 956         // between ones with different frame rates, but that's fine.
 957         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 958         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 959         fractional_samples = num_samples_times_timebase % TIMEBASE;
 960         assert(output_frame_info.num_samples >= 0);
 961
 962         return output_frame_info;
 963 }
 964
 965 void Mixer::handle_hotplugged_cards()
 966 {
 967         // Check for cards that have been disconnected since last frame.
 968         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 969                 CaptureCard *card = &cards[card_index];
 970                 if (card->capture->get_disconnected()) {
 971                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 972                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 973                         configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
 974                         card->queue_length_policy.reset(card_index);
 975                         card->capture->start_bm_capture();
 976                 }
 977         }
 978
 979         // Check for cards that have been connected since last frame.
 980         vector<libusb_device *> hotplugged_cards_copy;
 981         {
 982                 lock_guard<mutex> lock(hotplug_mutex);
 983                 swap(hotplugged_cards, hotplugged_cards_copy);
 984         }
 985         for (libusb_device *new_dev : hotplugged_cards_copy) {
 986                 // Look for a fake capture card where we can stick this in.
 987                 int free_card_index = -1;
 988                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 989                         if (cards[card_index].is_fake_capture) {
 990                                 free_card_index = card_index;
 991                                 break;
 992                         }
 993                 }
 994
 995                 if (free_card_index == -1) {
 996                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
 997                         libusb_unref_device(new_dev);
 998                 } else {
 999                         // BMUSBCapture takes ownership.
1000                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1001                         CaptureCard *card = &cards[free_card_index];
1002                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1003                         configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
1004                         card->queue_length_policy.reset(free_card_index);
1005                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1006                         capture->start_bm_capture();
1007                 }
1008         }
1009 }
1010
1011
1012 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)
1013 {
1014         // Resample the audio as needed, including from previously dropped frames.
1015         assert(num_cards > 0);
1016         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1017                 const bool dropped_frame = (frame_num != dropped_frames);
1018                 {
1019                         // Signal to the audio thread to process this frame.
1020                         // Note that if the frame is a dropped frame, we signal that
1021                         // we don't want to use this frame as base for adjusting
1022                         // the resampler rate. The reason for this is that the timing
1023                         // of these frames is often way too late; they typically don't
1024                         // “arrive” before we synthesize them. Thus, we could end up
1025                         // in a situation where we have inserted e.g. five audio frames
1026                         // into the queue before we then start pulling five of them
1027                         // back out. This makes ResamplingQueue overestimate the delay,
1028                         // causing undue resampler changes. (We _do_ use the last,
1029                         // non-dropped frame; perhaps we should just discard that as well,
1030                         // since dropped frames are expected to be rare, and it might be
1031                         // better to just wait until we have a slightly more normal situation).
1032                         unique_lock<mutex> lock(audio_mutex);
1033                         bool adjust_rate = !dropped_frame && !is_preroll;
1034                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1035                         audio_task_queue_changed.notify_one();
1036                 }
1037                 if (dropped_frame) {
1038                         // For dropped frames, increase the pts. Note that if the format changed
1039                         // in the meantime, we have no way of detecting that; we just have to
1040                         // assume the frame length is always the same.
1041                         pts_int += length_per_frame;
1042                 }
1043         }
1044 }
1045
1046 void Mixer::render_one_frame(int64_t duration)
1047 {
1048         // Determine the time code for this frame before we start rendering.
1049         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1050         if (display_timecode_on_stdout) {
1051                 printf("Timecode: '%s'\n", timecode_text.c_str());
1052         }
1053
1054         // Get the main chain from the theme, and set its state immediately.
1055         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1056         EffectChain *chain = theme_main_chain.chain;
1057         theme_main_chain.setup_chain();
1058         //theme_main_chain.chain->enable_phase_timing(true);
1059
1060         // If HDMI/SDI output is active and the user has requested auto mode,
1061         // its mode overrides the existing Y'CbCr setting for the chain.
1062         YCbCrLumaCoefficients ycbcr_output_coefficients;
1063         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1064                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1065         } else {
1066                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1067         }
1068
1069         // TODO: Reduce the duplication against theme.cpp.
1070         YCbCrFormat output_ycbcr_format;
1071         output_ycbcr_format.chroma_subsampling_x = 1;
1072         output_ycbcr_format.chroma_subsampling_y = 1;
1073         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1074         output_ycbcr_format.full_range = false;
1075         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1076         chain->change_ycbcr_output_format(output_ycbcr_format);
1077
1078         // Render main chain. If we're using zerocopy Quick Sync encoding
1079         // (the default case), we take an extra copy of the created outputs,
1080         // so that we can display it back to the screen later (it's less memory
1081         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1082         // Ideally, we'd like to avoid taking copies and just use the main textures
1083         // for display as well, but they're just views into VA-API memory and must be
1084         // unmapped during encoding, so we can't use them for display, unfortunately.
1085         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1086         GLuint y_copy_tex, cbcr_copy_tex = 0;
1087         GLuint y_display_tex, cbcr_display_tex;
1088         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1089         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1090         const bool is_zerocopy = video_encoder->is_zerocopy();
1091         if (is_zerocopy) {
1092                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1093                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1094                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1095
1096                 y_display_tex = y_copy_tex;
1097                 cbcr_display_tex = cbcr_copy_tex;
1098
1099                 // y_tex and cbcr_tex will be given by VideoEncoder.
1100         } else {
1101                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1102                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1103                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1104
1105                 y_display_tex = y_tex;
1106                 cbcr_display_tex = cbcr_tex;
1107         }
1108
1109         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1110         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1111         assert(got_frame);
1112
1113         GLuint fbo;
1114         if (is_zerocopy) {
1115                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1116         } else {
1117                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1118         }
1119         check_error();
1120         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1121
1122         if (display_timecode_in_stream) {
1123                 // Render the timecode on top.
1124                 timecode_renderer->render_timecode(fbo, timecode_text);
1125         }
1126
1127         resource_pool->release_fbo(fbo);
1128
1129         if (is_zerocopy) {
1130                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1131         } else {
1132                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1133         }
1134         if (output_card_index != -1) {
1135                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1136         }
1137         resource_pool->release_2d_texture(cbcr_full_tex);
1138
1139         // Set the right state for the Y' and CbCr textures we use for display.
1140         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1141         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1142         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1143         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1144         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1145
1146         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1147         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1148         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1149         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1150
1151         RefCountedGLsync fence = video_encoder->end_frame();
1152
1153         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1154         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1155         DisplayFrame live_frame;
1156         live_frame.chain = display_chain.get();
1157         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1158                 display_input->set_texture_num(0, y_display_tex);
1159                 display_input->set_texture_num(1, cbcr_display_tex);
1160         };
1161         live_frame.ready_fence = fence;
1162         live_frame.input_frames = {};
1163         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1164         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1165
1166         // Set up preview and any additional channels.
1167         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1168                 DisplayFrame display_frame;
1169                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1170                 display_frame.chain = chain.chain;
1171                 display_frame.setup_chain = chain.setup_chain;
1172                 display_frame.ready_fence = fence;
1173                 display_frame.input_frames = chain.input_frames;
1174                 display_frame.temp_textures = {};
1175                 output_channel[i].output_frame(display_frame);
1176         }
1177 }
1178
1179 void Mixer::audio_thread_func()
1180 {
1181         pthread_setname_np(pthread_self(), "Mixer_Audio");
1182
1183         while (!should_quit) {
1184                 AudioTask task;
1185
1186                 {
1187                         unique_lock<mutex> lock(audio_mutex);
1188                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1189                         if (should_quit) {
1190                                 return;
1191                         }
1192                         task = audio_task_queue.front();
1193                         audio_task_queue.pop();
1194                 }
1195
1196                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1197                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1198                 vector<float> samples_out = audio_mixer.get_output(
1199                         task.frame_timestamp,
1200                         task.num_samples,
1201                         rate_adjustment_policy);
1202
1203                 // Send the samples to the sound card, then add them to the output.
1204                 if (alsa) {
1205                         alsa->write(samples_out);
1206                 }
1207                 if (output_card_index != -1) {
1208                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1209                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1210                 }
1211                 video_encoder->add_audio(task.pts_int, move(samples_out));
1212         }
1213 }
1214
1215 void Mixer::release_display_frame(DisplayFrame *frame)
1216 {
1217         for (GLuint texnum : frame->temp_textures) {
1218                 resource_pool->release_2d_texture(texnum);
1219         }
1220         frame->temp_textures.clear();
1221         frame->ready_fence.reset();
1222         frame->input_frames.clear();
1223 }
1224
1225 void Mixer::start()
1226 {
1227         mixer_thread = thread(&Mixer::thread_func, this);
1228         audio_thread = thread(&Mixer::audio_thread_func, this);
1229 }
1230
1231 void Mixer::quit()
1232 {
1233         should_quit = true;
1234         audio_task_queue_changed.notify_one();
1235         mixer_thread.join();
1236         audio_thread.join();
1237 }
1238
1239 void Mixer::transition_clicked(int transition_num)
1240 {
1241         theme->transition_clicked(transition_num, pts());
1242 }
1243
1244 void Mixer::channel_clicked(int preview_num)
1245 {
1246         theme->channel_clicked(preview_num);
1247 }
1248
1249 void Mixer::start_mode_scanning(unsigned card_index)
1250 {
1251         assert(card_index < num_cards);
1252         if (is_mode_scanning[card_index]) {
1253                 return;
1254         }
1255         is_mode_scanning[card_index] = true;
1256         mode_scanlist[card_index].clear();
1257         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1258                 mode_scanlist[card_index].push_back(mode.first);
1259         }
1260         assert(!mode_scanlist[card_index].empty());
1261         mode_scanlist_index[card_index] = 0;
1262         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1263         last_mode_scan_change[card_index] = steady_clock::now();
1264 }
1265
1266 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1267 {
1268         assert(desired_output_card_index != -1);
1269         unique_lock<mutex> lock(card_mutex);
1270         return cards[desired_output_card_index].output->get_available_video_modes();
1271 }
1272
1273 Mixer::OutputChannel::~OutputChannel()
1274 {
1275         if (has_current_frame) {
1276                 parent->release_display_frame(&current_frame);
1277         }
1278         if (has_ready_frame) {
1279                 parent->release_display_frame(&ready_frame);
1280         }
1281 }
1282
1283 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1284 {
1285         // Store this frame for display. Remove the ready frame if any
1286         // (it was seemingly never used).
1287         {
1288                 unique_lock<mutex> lock(frame_mutex);
1289                 if (has_ready_frame) {
1290                         parent->release_display_frame(&ready_frame);
1291                 }
1292                 ready_frame = frame;
1293                 has_ready_frame = true;
1294         }
1295
1296         if (new_frame_ready_callback) {
1297                 new_frame_ready_callback();
1298         }
1299
1300         // Reduce the number of callbacks by filtering duplicates. The reason
1301         // why we bother doing this is that Qt seemingly can get into a state
1302         // where its builds up an essentially unbounded queue of signals,
1303         // consuming more and more memory, and there's no good way of collapsing
1304         // user-defined signals or limiting the length of the queue.
1305         if (transition_names_updated_callback) {
1306                 vector<string> transition_names = global_mixer->get_transition_names();
1307                 bool changed = false;
1308                 if (transition_names.size() != last_transition_names.size()) {
1309                         changed = true;
1310                 } else {
1311                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1312                                 if (transition_names[i] != last_transition_names[i]) {
1313                                         changed = true;
1314                                         break;
1315                                 }
1316                         }
1317                 }
1318                 if (changed) {
1319                         transition_names_updated_callback(transition_names);
1320                         last_transition_names = transition_names;
1321                 }
1322         }
1323         if (name_updated_callback) {
1324                 string name = global_mixer->get_channel_name(channel);
1325                 if (name != last_name) {
1326                         name_updated_callback(name);
1327                         last_name = name;
1328                 }
1329         }
1330         if (color_updated_callback) {
1331                 string color = global_mixer->get_channel_color(channel);
1332                 if (color != last_color) {
1333                         color_updated_callback(color);
1334                         last_color = color;
1335                 }
1336         }
1337 }
1338
1339 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1340 {
1341         unique_lock<mutex> lock(frame_mutex);
1342         if (!has_current_frame && !has_ready_frame) {
1343                 return false;
1344         }
1345
1346         if (has_current_frame && has_ready_frame) {
1347                 // We have a new ready frame. Toss the current one.
1348                 parent->release_display_frame(&current_frame);
1349                 has_current_frame = false;
1350         }
1351         if (has_ready_frame) {
1352                 assert(!has_current_frame);
1353                 current_frame = ready_frame;
1354                 ready_frame.ready_fence.reset();  // Drop the refcount.
1355                 ready_frame.input_frames.clear();  // Drop the refcounts.
1356                 has_current_frame = true;
1357                 has_ready_frame = false;
1358         }
1359
1360         *frame = current_frame;
1361         return true;
1362 }
1363
1364 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1365 {
1366         new_frame_ready_callback = callback;
1367 }
1368
1369 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1370 {
1371         transition_names_updated_callback = callback;
1372 }
1373
1374 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1375 {
1376         name_updated_callback = callback;
1377 }
1378
1379 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1380 {
1381         color_updated_callback = callback;
1382 }
1383
1384 mutex RefCountedGLsync::fence_lock;