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                         if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
 128                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
 129                         } else {
 130                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
 131                         }
 132                         check_error();
 133                         break;
 134                 }
 135                 userdata->last_width[field] = width;
 136                 userdata->last_height[field] = height;
 137         }
 138         if (global_flags.ten_bit_input &&
 139             (first || v210_width != userdata->last_v210_width[field])) {
 140                 // Same as above; we need to recreate the texture.
 141                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 142                 check_error();
 143                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 144                 check_error();
 145                 userdata->last_v210_width[field] = v210_width;
 146         }
 147 }
 148
 149 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 150 {
 151         if (interlaced_stride) {
 152                 stride *= 2;
 153         }
 154         if (global_flags.flush_pbos) {
 155                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 156                 check_error();
 157         }
 158
 159         glBindTexture(GL_TEXTURE_2D, tex);
 160         check_error();
 161         if (interlaced_stride) {
 162                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 163                 check_error();
 164         } else {
 165                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 166                 check_error();
 167         }
 168
 169         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 170         check_error();
 171         glBindTexture(GL_TEXTURE_2D, 0);
 172         check_error();
 173         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 174         check_error();
 175 }
 176
 177 }  // namespace
 178
 179 void QueueLengthPolicy::update_policy(unsigned queue_length)
 180 {
 181         if (queue_length == 0) {  // Starvation.
 182                 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
 183                         ++safe_queue_length;
 184                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
 185                                 card_index, safe_queue_length);
 186                 }
 187                 frames_with_at_least_one = 0;
 188                 been_at_safe_point_since_last_starvation = false;
 189                 return;
 190         }
 191         if (queue_length >= safe_queue_length) {
 192                 been_at_safe_point_since_last_starvation = true;
 193         }
 194         if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
 195                 --safe_queue_length;
 196                 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
 197                         card_index, safe_queue_length);
 198                 frames_with_at_least_one = 0;
 199         }
 200 }
 201
 202 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 203         : httpd(),
 204           num_cards(num_cards),
 205           mixer_surface(create_surface(format)),
 206           h264_encoder_surface(create_surface(format)),
 207           decklink_output_surface(create_surface(format)),
 208           audio_mixer(num_cards)
 209 {
 210         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 211         check_error();
 212
 213         // This nearly always should be true.
 214         global_flags.can_disable_srgb_decoder = epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode");
 215
 216         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 217         // will be halved when sampling them, and we need to compensate here.
 218         movit_texel_subpixel_precision /= 2.0;
 219
 220         resource_pool.reset(new ResourcePool);
 221         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 222                 output_channel[i].parent = this;
 223                 output_channel[i].channel = i;
 224         }
 225
 226         ImageFormat inout_format;
 227         inout_format.color_space = COLORSPACE_sRGB;
 228         inout_format.gamma_curve = GAMMA_sRGB;
 229
 230         // Matches the 4:2:0 format created by the main chain.
 231         YCbCrFormat ycbcr_format;
 232         ycbcr_format.chroma_subsampling_x = 2;
 233         ycbcr_format.chroma_subsampling_y = 2;
 234         if (global_flags.ycbcr_rec709_coefficients) {
 235                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 236         } else {
 237                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 238         }
 239         ycbcr_format.full_range = false;
 240         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 241         ycbcr_format.cb_x_position = 0.0f;
 242         ycbcr_format.cr_x_position = 0.0f;
 243         ycbcr_format.cb_y_position = 0.5f;
 244         ycbcr_format.cr_y_position = 0.5f;
 245
 246         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 247         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 248         check_error();
 249         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 250         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 251         display_chain->add_input(display_input);
 252         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 253         display_chain->set_dither_bits(0);  // Don't bother.
 254         display_chain->finalize();
 255
 256         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));
 257
 258         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 259         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 260
 261         // Start listening for clients only once VideoEncoder has written its header, if any.
 262         httpd.start(9095);
 263
 264         // First try initializing the then PCI devices, then USB, then
 265         // fill up with fake cards until we have the desired number of cards.
 266         unsigned num_pci_devices = 0;
 267         unsigned card_index = 0;
 268
 269         {
 270                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 271                 if (decklink_iterator != nullptr) {
 272                         for ( ; card_index < num_cards; ++card_index) {
 273                                 IDeckLink *decklink;
 274                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 275                                         break;
 276                                 }
 277
 278                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 279                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 280                                 output->set_device(decklink);
 281                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 282                                 ++num_pci_devices;
 283                         }
 284                         decklink_iterator->Release();
 285                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 286                 } else {
 287                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 288                 }
 289         }
 290
 291         unsigned num_usb_devices = BMUSBCapture::num_cards();
 292         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 293                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 294                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 295                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 296         }
 297         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 298
 299         unsigned num_fake_cards = 0;
 300         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 301                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 302                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 303         }
 304
 305         if (num_fake_cards > 0) {
 306                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 307         }
 308
 309         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 310         BMUSBCapture::start_bm_thread();
 311
 312         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 313                 cards[card_index].queue_length_policy.reset(card_index);
 314         }
 315
 316         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 317
 318         if (global_flags.ten_bit_input) {
 319                 if (!v210Converter::has_hardware_support()) {
 320                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 321                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 322                         exit(1);
 323                 }
 324                 v210_converter.reset(new v210Converter());
 325
 326                 // These are all the widths listed in the Blackmagic SDK documentation
 327                 // (section 2.7.3, “Display Modes”).
 328                 v210_converter->precompile_shader(720);
 329                 v210_converter->precompile_shader(1280);
 330                 v210_converter->precompile_shader(1920);
 331                 v210_converter->precompile_shader(2048);
 332                 v210_converter->precompile_shader(3840);
 333                 v210_converter->precompile_shader(4096);
 334         }
 335         if (global_flags.ten_bit_output) {
 336                 if (!v210Converter::has_hardware_support()) {
 337                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 338                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 339                         exit(1);
 340                 }
 341         }
 342
 343         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 344         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 345         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 346
 347         if (global_flags.enable_alsa_output) {
 348                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 349         }
 350         if (global_flags.output_card != -1) {
 351                 desired_output_card_index = global_flags.output_card;
 352                 set_output_card_internal(global_flags.output_card);
 353         }
 354 }
 355
 356 Mixer::~Mixer()
 357 {
 358         BMUSBCapture::stop_bm_thread();
 359
 360         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 361                 {
 362                         unique_lock<mutex> lock(card_mutex);
 363                         cards[card_index].should_quit = true;  // Unblock thread.
 364                         cards[card_index].new_frames_changed.notify_all();
 365                 }
 366                 cards[card_index].capture->stop_dequeue_thread();
 367                 if (cards[card_index].output) {
 368                         cards[card_index].output->end_output();
 369                         cards[card_index].output.reset();
 370                 }
 371         }
 372
 373         video_encoder.reset(nullptr);
 374 }
 375
 376 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 377 {
 378         printf("Configuring card %d...\n", card_index);
 379
 380         CaptureCard *card = &cards[card_index];
 381         if (card->capture != nullptr) {
 382                 card->capture->stop_dequeue_thread();
 383         }
 384         card->capture.reset(capture);
 385         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 386         if (card->output.get() != output) {
 387                 card->output.reset(output);
 388         }
 389
 390         bmusb::PixelFormat pixel_format = global_flags.ten_bit_input ? PixelFormat_10BitYCbCr : PixelFormat_8BitYCbCr;
 391
 392         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 393         if (card->frame_allocator == nullptr) {
 394                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 395         }
 396         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 397         if (card->surface == nullptr) {
 398                 card->surface = create_surface_with_same_format(mixer_surface);
 399         }
 400         while (!card->new_frames.empty()) card->new_frames.pop_front();
 401         card->last_timecode = -1;
 402         card->capture->set_pixel_format(pixel_format);
 403         card->capture->configure_card();
 404
 405         // NOTE: start_bm_capture() happens in thread_func().
 406
 407         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 408         audio_mixer.reset_resampler(device);
 409         audio_mixer.set_display_name(device, card->capture->get_description());
 410         audio_mixer.trigger_state_changed_callback();
 411 }
 412
 413 void Mixer::set_output_card_internal(int card_index)
 414 {
 415         // We don't really need to take card_mutex, since we're in the mixer
 416         // thread and don't mess with any queues (which is the only thing that happens
 417         // from other threads), but it's probably the safest in the long run.
 418         unique_lock<mutex> lock(card_mutex);
 419         if (output_card_index != -1) {
 420                 // Switch the old card from output to input.
 421                 CaptureCard *old_card = &cards[output_card_index];
 422                 old_card->output->end_output();
 423
 424                 // Stop the fake card that we put into place.
 425                 // This needs to _not_ happen under the mutex, to avoid deadlock
 426                 // (delivering the last frame needs to take the mutex).
 427                 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
 428                 lock.unlock();
 429                 fake_capture->stop_dequeue_thread();
 430                 lock.lock();
 431                 old_card->capture = move(old_card->parked_capture);
 432                 old_card->is_fake_capture = false;
 433                 old_card->capture->start_bm_capture();
 434         }
 435         if (card_index != -1) {
 436                 CaptureCard *card = &cards[card_index];
 437                 bmusb::CaptureInterface *capture = card->capture.get();
 438                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 439                 // several seconds to complete (blocking on DisableVideoInput);
 440                 // see if we can maybe do it asynchronously.
 441                 lock.unlock();
 442                 capture->stop_dequeue_thread();
 443                 lock.lock();
 444                 card->parked_capture = move(card->capture);
 445                 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 446                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 447                 card->queue_length_policy.reset(card_index);
 448                 card->capture->start_bm_capture();
 449                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 450                 card->output->start_output(desired_output_video_mode, pts_int);
 451         }
 452         output_card_index = card_index;
 453 }
 454
 455 namespace {
 456
 457 int unwrap_timecode(uint16_t current_wrapped, int last)
 458 {
 459         uint16_t last_wrapped = last & 0xffff;
 460         if (current_wrapped > last_wrapped) {
 461                 return (last & ~0xffff) | current_wrapped;
 462         } else {
 463                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 464         }
 465 }
 466
 467 }  // namespace
 468
 469 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 470                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 471                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 472 {
 473         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 474         CaptureCard *card = &cards[card_index];
 475
 476         if (is_mode_scanning[card_index]) {
 477                 if (video_format.has_signal) {
 478                         // Found a stable signal, so stop scanning.
 479                         is_mode_scanning[card_index] = false;
 480                 } else {
 481                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 482                         steady_clock::time_point now = steady_clock::now();
 483                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 484                         if (sec_since_last_switch > switch_time_s) {
 485                                 // It isn't this mode; try the next one.
 486                                 mode_scanlist_index[card_index]++;
 487                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 488                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 489                                 last_mode_scan_change[card_index] = now;
 490                         }
 491                 }
 492         }
 493
 494         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 495         assert(frame_length > 0);
 496
 497         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;
 498         if (num_samples > OUTPUT_FREQUENCY / 10) {
 499                 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",
 500                         card_index, int(audio_frame.len), int(audio_offset),
 501                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 502                 if (video_frame.owner) {
 503                         video_frame.owner->release_frame(video_frame);
 504                 }
 505                 if (audio_frame.owner) {
 506                         audio_frame.owner->release_frame(audio_frame);
 507                 }
 508                 return;
 509         }
 510
 511         int dropped_frames = 0;
 512         if (card->last_timecode != -1) {
 513                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 514         }
 515
 516         // Number of samples per frame if we need to insert silence.
 517         // (Could be nonintegral, but resampling will save us then.)
 518         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 519
 520         if (dropped_frames > MAX_FPS * 2) {
 521                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 522                         card_index, card->last_timecode, timecode);
 523                 audio_mixer.reset_resampler(device);
 524                 dropped_frames = 0;
 525         } else if (dropped_frames > 0) {
 526                 // Insert silence as needed.
 527                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 528                         card_index, dropped_frames, timecode);
 529
 530                 bool success;
 531                 do {
 532                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 533                 } while (!success);
 534         }
 535
 536         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 537
 538         // Done with the audio, so release it.
 539         if (audio_frame.owner) {
 540                 audio_frame.owner->release_frame(audio_frame);
 541         }
 542
 543         card->last_timecode = timecode;
 544
 545         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 546         if (video_frame.len - video_offset == 0 ||
 547             video_frame.len - video_offset != expected_length) {
 548                 if (video_frame.len != 0) {
 549                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 550                                 card_index, video_frame.len - video_offset, expected_length);
 551                 }
 552                 if (video_frame.owner) {
 553                         video_frame.owner->release_frame(video_frame);
 554                 }
 555
 556                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 557                 // so that pts can go up accordingly.
 558                 {
 559                         unique_lock<mutex> lock(card_mutex);
 560                         CaptureCard::NewFrame new_frame;
 561                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 562                         new_frame.length = frame_length;
 563                         new_frame.interlaced = false;
 564                         new_frame.dropped_frames = dropped_frames;
 565                         new_frame.received_timestamp = video_frame.received_timestamp;
 566                         card->new_frames.push_back(move(new_frame));
 567                         card->new_frames_changed.notify_all();
 568                 }
 569                 return;
 570         }
 571
 572         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 573
 574         unsigned num_fields = video_format.interlaced ? 2 : 1;
 575         steady_clock::time_point frame_upload_start;
 576         bool interlaced_stride = false;
 577         if (video_format.interlaced) {
 578                 // Send the two fields along as separate frames; the other side will need to add
 579                 // a deinterlacer to actually get this right.
 580                 assert(video_format.height % 2 == 0);
 581                 video_format.height /= 2;
 582                 assert(frame_length % 2 == 0);
 583                 frame_length /= 2;
 584                 num_fields = 2;
 585                 if (video_format.second_field_start == 1) {
 586                         interlaced_stride = true;
 587                 }
 588                 frame_upload_start = steady_clock::now();
 589         }
 590         userdata->last_interlaced = video_format.interlaced;
 591         userdata->last_has_signal = video_format.has_signal;
 592         userdata->last_is_connected = video_format.is_connected;
 593         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 594         userdata->last_frame_rate_den = video_format.frame_rate_den;
 595         RefCountedFrame frame(video_frame);
 596
 597         // Upload the textures.
 598         const size_t cbcr_width = video_format.width / 2;
 599         const size_t cbcr_offset = video_offset / 2;
 600         const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 601
 602         for (unsigned field = 0; field < num_fields; ++field) {
 603                 // Put the actual texture upload in a lambda that is executed in the main thread.
 604                 // It is entirely possible to do this in the same thread (and it might even be
 605                 // faster, depending on the GPU and driver), but it appears to be trickling
 606                 // driver bugs very easily.
 607                 //
 608                 // Note that this means we must hold on to the actual frame data in <userdata>
 609                 // until the upload command is run, but we hold on to <frame> much longer than that
 610                 // (in fact, all the way until we no longer use the texture in rendering).
 611                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 612                         unsigned field_start_line;
 613                         if (field == 1) {
 614                                 field_start_line = video_format.second_field_start;
 615                         } else {
 616                                 field_start_line = video_format.extra_lines_top;
 617                         }
 618
 619                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 620                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 621                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 622
 623                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 624                         check_error();
 625
 626                         switch (userdata->pixel_format) {
 627                         case bmusb::PixelFormat_10BitYCbCr: {
 628                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 629                                 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);
 630                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 631                                 break;
 632                         }
 633                         case bmusb::PixelFormat_8BitYCbCr: {
 634                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 635                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 636
 637                                 // Make up our own strides, since we are interleaving.
 638                                 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);
 639                                 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);
 640                                 break;
 641                         }
 642                         case bmusb::PixelFormat_8BitRGBA: {
 643                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 644                                 upload_texture(userdata->tex_rgba[field], video_format.width, video_format.height, video_format.stride, interlaced_stride, GL_RGBA, GL_UNSIGNED_BYTE, field_start);
 645                                 break;
 646                         }
 647                         default:
 648                                 assert(false);
 649                         }
 650
 651                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 652                         check_error();
 653                 };
 654
 655                 if (field == 1) {
 656                         // Don't upload the second field as fast as we can; wait until
 657                         // the field time has approximately passed. (Otherwise, we could
 658                         // get timing jitter against the other sources, and possibly also
 659                         // against the video display, although the latter is not as critical.)
 660                         // This requires our system clock to be reasonably close to the
 661                         // video clock, but that's not an unreasonable assumption.
 662                         steady_clock::time_point second_field_start = frame_upload_start +
 663                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 664                         this_thread::sleep_until(second_field_start);
 665                 }
 666
 667                 {
 668                         unique_lock<mutex> lock(card_mutex);
 669                         CaptureCard::NewFrame new_frame;
 670                         new_frame.frame = frame;
 671                         new_frame.length = frame_length;
 672                         new_frame.field = field;
 673                         new_frame.interlaced = video_format.interlaced;
 674                         new_frame.upload_func = upload_func;
 675                         new_frame.dropped_frames = dropped_frames;
 676                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 677                         card->new_frames.push_back(move(new_frame));
 678                         card->new_frames_changed.notify_all();
 679                 }
 680         }
 681 }
 682
 683 void Mixer::bm_hotplug_add(libusb_device *dev)
 684 {
 685         lock_guard<mutex> lock(hotplug_mutex);
 686         hotplugged_cards.push_back(dev);
 687 }
 688
 689 void Mixer::bm_hotplug_remove(unsigned card_index)
 690 {
 691         cards[card_index].new_frames_changed.notify_all();
 692 }
 693
 694 void Mixer::thread_func()
 695 {
 696         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 697
 698         eglBindAPI(EGL_OPENGL_API);
 699         QOpenGLContext *context = create_context(mixer_surface);
 700         if (!make_current(context, mixer_surface)) {
 701                 printf("oops\n");
 702                 exit(1);
 703         }
 704
 705         // Start the actual capture. (We don't want to do it before we're actually ready
 706         // to process output frames.)
 707         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 708                 if (int(card_index) != output_card_index) {
 709                         cards[card_index].capture->start_bm_capture();
 710                 }
 711         }
 712
 713         steady_clock::time_point start, now;
 714         start = steady_clock::now();
 715
 716         int stats_dropped_frames = 0;
 717
 718         while (!should_quit) {
 719                 if (desired_output_card_index != output_card_index) {
 720                         set_output_card_internal(desired_output_card_index);
 721                 }
 722                 if (output_card_index != -1 &&
 723                     desired_output_video_mode != output_video_mode) {
 724                         DeckLinkOutput *output = cards[output_card_index].output.get();
 725                         output->end_output();
 726                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 727                         output->start_output(desired_output_video_mode, pts_int);
 728                 }
 729
 730                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 731                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 732
 733                 bool master_card_is_output;
 734                 unsigned master_card_index;
 735                 if (output_card_index != -1) {
 736                         master_card_is_output = true;
 737                         master_card_index = output_card_index;
 738                 } else {
 739                         master_card_is_output = false;
 740                         master_card_index = theme->map_signal(master_clock_channel);
 741                         assert(master_card_index < num_cards);
 742                 }
 743
 744                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 745                 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);
 746                 stats_dropped_frames += output_frame_info.dropped_frames;
 747
 748                 handle_hotplugged_cards();
 749
 750                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 751                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 752                                 continue;
 753                         }
 754                         if (new_frames[card_index].frame->len == 0) {
 755                                 ++new_frames[card_index].dropped_frames;
 756                         }
 757                         if (new_frames[card_index].dropped_frames > 0) {
 758                                 printf("Card %u dropped %d frames before this\n",
 759                                         card_index, int(new_frames[card_index].dropped_frames));
 760                         }
 761                 }
 762
 763                 // If the first card is reporting a corrupted or otherwise dropped frame,
 764                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 765                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 766                         ++stats_dropped_frames;
 767                         pts_int += new_frames[master_card_index].length;
 768                         continue;
 769                 }
 770
 771                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 772                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 773                                 continue;
 774
 775                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 776                         assert(new_frame->frame != nullptr);
 777                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 778                         check_error();
 779
 780                         // The new texture might need uploading before use.
 781                         if (new_frame->upload_func) {
 782                                 new_frame->upload_func();
 783                                 new_frame->upload_func = nullptr;
 784                         }
 785                 }
 786
 787                 int64_t frame_duration = output_frame_info.frame_duration;
 788                 render_one_frame(frame_duration);
 789                 ++frame_num;
 790                 pts_int += frame_duration;
 791
 792                 now = steady_clock::now();
 793                 double elapsed = duration<double>(now - start).count();
 794                 if (frame_num % 100 == 0) {
 795                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 796                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 797                                 1e3 * elapsed / frame_num);
 798                 //      chain->print_phase_timing();
 799
 800                         // Check our memory usage, to see if we are close to our mlockall()
 801                         // limit (if at all set).
 802                         rusage used;
 803                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 804                                 perror("getrusage(RUSAGE_SELF)");
 805                                 assert(false);
 806                         }
 807
 808                         if (uses_mlock) {
 809                                 rlimit limit;
 810                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 811                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 812                                         assert(false);
 813                                 }
 814
 815                                 if (limit.rlim_cur == 0) {
 816                                         printf(", using %ld MB memory (locked)",
 817                                                 long(used.ru_maxrss / 1024));
 818                                 } else {
 819                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 820                                                 long(used.ru_maxrss / 1024),
 821                                                 long(limit.rlim_cur / 1048576),
 822                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 823                                 }
 824                         } else {
 825                                 printf(", using %ld MB memory (not locked)",
 826                                         long(used.ru_maxrss / 1024));
 827                         }
 828
 829                         printf("\n");
 830                 }
 831
 832
 833                 if (should_cut.exchange(false)) {  // Test and clear.
 834                         video_encoder->do_cut(frame_num);
 835                 }
 836
 837 #if 0
 838                 // Reset every 100 frames, so that local variations in frame times
 839                 // (especially for the first few frames, when the shaders are
 840                 // compiled etc.) don't make it hard to measure for the entire
 841                 // remaining duration of the program.
 842                 if (frame == 10000) {
 843                         frame = 0;
 844                         start = now;
 845                 }
 846 #endif
 847                 check_error();
 848         }
 849
 850         resource_pool->clean_context();
 851 }
 852
 853 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 854 {
 855         if (output_card_index != -1) {
 856                 // The output card (ie., cards[output_card_index].output) is the master clock,
 857                 // so no input card (ie., cards[card_index].capture) is.
 858                 return false;
 859         }
 860         return (card_index == master_card_index);
 861 }
 862
 863 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 864 {
 865         // Count the number of frames in the queue, including any frames
 866         // we dropped. It's hard to know exactly how we should deal with
 867         // dropped (corrupted) input frames; they don't help our goal of
 868         // avoiding starvation, but they still add to the problem of latency.
 869         // Since dropped frames is going to mean a bump in the signal anyway,
 870         // we err on the side of having more stable latency instead.
 871         unsigned queue_length = 0;
 872         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 873                 queue_length += frame.dropped_frames + 1;
 874         }
 875         card->queue_length_policy.update_policy(queue_length);
 876
 877         // If needed, drop frames until the queue is below the safe limit.
 878         // We prefer to drop from the head, because all else being equal,
 879         // we'd like more recent frames (less latency).
 880         unsigned dropped_frames = 0;
 881         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 882                 assert(!card->new_frames.empty());
 883                 assert(queue_length > card->new_frames.front().dropped_frames);
 884                 queue_length -= card->new_frames.front().dropped_frames;
 885
 886                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 887                         // No need to drop anything.
 888                         break;
 889                 }
 890
 891                 card->new_frames.pop_front();
 892                 card->new_frames_changed.notify_all();
 893                 --queue_length;
 894                 ++dropped_frames;
 895         }
 896
 897 #if 0
 898         if (dropped_frames > 0) {
 899                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 900                         card_index, dropped_frames);
 901         }
 902 #endif
 903 }
 904
 905
 906 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])
 907 {
 908         OutputFrameInfo output_frame_info;
 909 start:
 910         unique_lock<mutex> lock(card_mutex, defer_lock);
 911         if (master_card_is_output) {
 912                 // Clocked to the output, so wait for it to be ready for the next frame.
 913                 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);
 914                 lock.lock();
 915         } else {
 916                 // Wait for the master card to have a new frame.
 917                 // TODO: Add a timeout.
 918                 output_frame_info.is_preroll = false;
 919                 lock.lock();
 920                 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(); });
 921         }
 922
 923         if (master_card_is_output) {
 924                 handle_hotplugged_cards();
 925         } else if (cards[master_card_index].new_frames.empty()) {
 926                 // We were woken up, but not due to a new frame. Deal with it
 927                 // and then restart.
 928                 assert(cards[master_card_index].capture->get_disconnected());
 929                 handle_hotplugged_cards();
 930                 goto start;
 931         }
 932
 933         if (!master_card_is_output) {
 934                 output_frame_info.frame_timestamp =
 935                         cards[master_card_index].new_frames.front().received_timestamp;
 936         }
 937
 938         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 939                 CaptureCard *card = &cards[card_index];
 940                 if (input_card_is_master_clock(card_index, master_card_index)) {
 941                         // We don't use the queue length policy for the master card,
 942                         // but we will if it stops being the master. Thus, clear out
 943                         // the policy in case we switch in the future.
 944                         card->queue_length_policy.reset(card_index);
 945                         assert(!card->new_frames.empty());
 946                 } else {
 947                         trim_queue(card, card_index);
 948                 }
 949                 if (!card->new_frames.empty()) {
 950                         new_frames[card_index] = move(card->new_frames.front());
 951                         has_new_frame[card_index] = true;
 952                         card->new_frames.pop_front();
 953                         card->new_frames_changed.notify_all();
 954                 }
 955         }
 956
 957         if (!master_card_is_output) {
 958                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 959                 output_frame_info.frame_duration = new_frames[master_card_index].length;
 960         }
 961
 962         // This might get off by a fractional sample when changing master card
 963         // between ones with different frame rates, but that's fine.
 964         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 965         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 966         fractional_samples = num_samples_times_timebase % TIMEBASE;
 967         assert(output_frame_info.num_samples >= 0);
 968
 969         return output_frame_info;
 970 }
 971
 972 void Mixer::handle_hotplugged_cards()
 973 {
 974         // Check for cards that have been disconnected since last frame.
 975         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 976                 CaptureCard *card = &cards[card_index];
 977                 if (card->capture->get_disconnected()) {
 978                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
 979                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 980                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 981                         card->queue_length_policy.reset(card_index);
 982                         card->capture->start_bm_capture();
 983                 }
 984         }
 985
 986         // Check for cards that have been connected since last frame.
 987         vector<libusb_device *> hotplugged_cards_copy;
 988         {
 989                 lock_guard<mutex> lock(hotplug_mutex);
 990                 swap(hotplugged_cards, hotplugged_cards_copy);
 991         }
 992         for (libusb_device *new_dev : hotplugged_cards_copy) {
 993                 // Look for a fake capture card where we can stick this in.
 994                 int free_card_index = -1;
 995                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 996                         if (cards[card_index].is_fake_capture) {
 997                                 free_card_index = card_index;
 998                                 break;
 999                         }
1000                 }
1001
1002                 if (free_card_index == -1) {
1003                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1004                         libusb_unref_device(new_dev);
1005                 } else {
1006                         // BMUSBCapture takes ownership.
1007                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1008                         CaptureCard *card = &cards[free_card_index];
1009                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1010                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1011                         card->queue_length_policy.reset(free_card_index);
1012                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1013                         capture->start_bm_capture();
1014                 }
1015         }
1016 }
1017
1018
1019 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)
1020 {
1021         // Resample the audio as needed, including from previously dropped frames.
1022         assert(num_cards > 0);
1023         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1024                 const bool dropped_frame = (frame_num != dropped_frames);
1025                 {
1026                         // Signal to the audio thread to process this frame.
1027                         // Note that if the frame is a dropped frame, we signal that
1028                         // we don't want to use this frame as base for adjusting
1029                         // the resampler rate. The reason for this is that the timing
1030                         // of these frames is often way too late; they typically don't
1031                         // “arrive” before we synthesize them. Thus, we could end up
1032                         // in a situation where we have inserted e.g. five audio frames
1033                         // into the queue before we then start pulling five of them
1034                         // back out. This makes ResamplingQueue overestimate the delay,
1035                         // causing undue resampler changes. (We _do_ use the last,
1036                         // non-dropped frame; perhaps we should just discard that as well,
1037                         // since dropped frames are expected to be rare, and it might be
1038                         // better to just wait until we have a slightly more normal situation).
1039                         unique_lock<mutex> lock(audio_mutex);
1040                         bool adjust_rate = !dropped_frame && !is_preroll;
1041                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1042                         audio_task_queue_changed.notify_one();
1043                 }
1044                 if (dropped_frame) {
1045                         // For dropped frames, increase the pts. Note that if the format changed
1046                         // in the meantime, we have no way of detecting that; we just have to
1047                         // assume the frame length is always the same.
1048                         pts_int += length_per_frame;
1049                 }
1050         }
1051 }
1052
1053 void Mixer::render_one_frame(int64_t duration)
1054 {
1055         // Determine the time code for this frame before we start rendering.
1056         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1057         if (display_timecode_on_stdout) {
1058                 printf("Timecode: '%s'\n", timecode_text.c_str());
1059         }
1060
1061         // Get the main chain from the theme, and set its state immediately.
1062         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1063         EffectChain *chain = theme_main_chain.chain;
1064         theme_main_chain.setup_chain();
1065         //theme_main_chain.chain->enable_phase_timing(true);
1066
1067         // If HDMI/SDI output is active and the user has requested auto mode,
1068         // its mode overrides the existing Y'CbCr setting for the chain.
1069         YCbCrLumaCoefficients ycbcr_output_coefficients;
1070         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1071                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1072         } else {
1073                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1074         }
1075
1076         // TODO: Reduce the duplication against theme.cpp.
1077         YCbCrFormat output_ycbcr_format;
1078         output_ycbcr_format.chroma_subsampling_x = 1;
1079         output_ycbcr_format.chroma_subsampling_y = 1;
1080         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1081         output_ycbcr_format.full_range = false;
1082         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1083         chain->change_ycbcr_output_format(output_ycbcr_format);
1084
1085         // Render main chain. If we're using zerocopy Quick Sync encoding
1086         // (the default case), we take an extra copy of the created outputs,
1087         // so that we can display it back to the screen later (it's less memory
1088         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1089         // Ideally, we'd like to avoid taking copies and just use the main textures
1090         // for display as well, but they're just views into VA-API memory and must be
1091         // unmapped during encoding, so we can't use them for display, unfortunately.
1092         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1093         GLuint y_copy_tex, cbcr_copy_tex = 0;
1094         GLuint y_display_tex, cbcr_display_tex;
1095         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1096         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1097         const bool is_zerocopy = video_encoder->is_zerocopy();
1098         if (is_zerocopy) {
1099                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1100                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1101                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1102
1103                 y_display_tex = y_copy_tex;
1104                 cbcr_display_tex = cbcr_copy_tex;
1105
1106                 // y_tex and cbcr_tex will be given by VideoEncoder.
1107         } else {
1108                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1109                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1110                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1111
1112                 y_display_tex = y_tex;
1113                 cbcr_display_tex = cbcr_tex;
1114         }
1115
1116         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1117         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1118         assert(got_frame);
1119
1120         GLuint fbo;
1121         if (is_zerocopy) {
1122                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1123         } else {
1124                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1125         }
1126         check_error();
1127         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1128
1129         if (display_timecode_in_stream) {
1130                 // Render the timecode on top.
1131                 timecode_renderer->render_timecode(fbo, timecode_text);
1132         }
1133
1134         resource_pool->release_fbo(fbo);
1135
1136         if (is_zerocopy) {
1137                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1138         } else {
1139                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1140         }
1141         if (output_card_index != -1) {
1142                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1143         }
1144         resource_pool->release_2d_texture(cbcr_full_tex);
1145
1146         // Set the right state for the Y' and CbCr textures we use for display.
1147         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1148         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1149         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1150         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1151         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1152
1153         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1154         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1155         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1156         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1157
1158         RefCountedGLsync fence = video_encoder->end_frame();
1159
1160         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1161         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1162         DisplayFrame live_frame;
1163         live_frame.chain = display_chain.get();
1164         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1165                 display_input->set_texture_num(0, y_display_tex);
1166                 display_input->set_texture_num(1, cbcr_display_tex);
1167         };
1168         live_frame.ready_fence = fence;
1169         live_frame.input_frames = {};
1170         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1171         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1172
1173         // Set up preview and any additional channels.
1174         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1175                 DisplayFrame display_frame;
1176                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1177                 display_frame.chain = chain.chain;
1178                 display_frame.setup_chain = chain.setup_chain;
1179                 display_frame.ready_fence = fence;
1180                 display_frame.input_frames = chain.input_frames;
1181                 display_frame.temp_textures = {};
1182                 output_channel[i].output_frame(display_frame);
1183         }
1184 }
1185
1186 void Mixer::audio_thread_func()
1187 {
1188         pthread_setname_np(pthread_self(), "Mixer_Audio");
1189
1190         while (!should_quit) {
1191                 AudioTask task;
1192
1193                 {
1194                         unique_lock<mutex> lock(audio_mutex);
1195                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1196                         if (should_quit) {
1197                                 return;
1198                         }
1199                         task = audio_task_queue.front();
1200                         audio_task_queue.pop();
1201                 }
1202
1203                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1204                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1205                 vector<float> samples_out = audio_mixer.get_output(
1206                         task.frame_timestamp,
1207                         task.num_samples,
1208                         rate_adjustment_policy);
1209
1210                 // Send the samples to the sound card, then add them to the output.
1211                 if (alsa) {
1212                         alsa->write(samples_out);
1213                 }
1214                 if (output_card_index != -1) {
1215                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1216                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1217                 }
1218                 video_encoder->add_audio(task.pts_int, move(samples_out));
1219         }
1220 }
1221
1222 void Mixer::release_display_frame(DisplayFrame *frame)
1223 {
1224         for (GLuint texnum : frame->temp_textures) {
1225                 resource_pool->release_2d_texture(texnum);
1226         }
1227         frame->temp_textures.clear();
1228         frame->ready_fence.reset();
1229         frame->input_frames.clear();
1230 }
1231
1232 void Mixer::start()
1233 {
1234         mixer_thread = thread(&Mixer::thread_func, this);
1235         audio_thread = thread(&Mixer::audio_thread_func, this);
1236 }
1237
1238 void Mixer::quit()
1239 {
1240         should_quit = true;
1241         audio_task_queue_changed.notify_one();
1242         mixer_thread.join();
1243         audio_thread.join();
1244 }
1245
1246 void Mixer::transition_clicked(int transition_num)
1247 {
1248         theme->transition_clicked(transition_num, pts());
1249 }
1250
1251 void Mixer::channel_clicked(int preview_num)
1252 {
1253         theme->channel_clicked(preview_num);
1254 }
1255
1256 void Mixer::start_mode_scanning(unsigned card_index)
1257 {
1258         assert(card_index < num_cards);
1259         if (is_mode_scanning[card_index]) {
1260                 return;
1261         }
1262         is_mode_scanning[card_index] = true;
1263         mode_scanlist[card_index].clear();
1264         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1265                 mode_scanlist[card_index].push_back(mode.first);
1266         }
1267         assert(!mode_scanlist[card_index].empty());
1268         mode_scanlist_index[card_index] = 0;
1269         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1270         last_mode_scan_change[card_index] = steady_clock::now();
1271 }
1272
1273 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1274 {
1275         assert(desired_output_card_index != -1);
1276         unique_lock<mutex> lock(card_mutex);
1277         return cards[desired_output_card_index].output->get_available_video_modes();
1278 }
1279
1280 Mixer::OutputChannel::~OutputChannel()
1281 {
1282         if (has_current_frame) {
1283                 parent->release_display_frame(&current_frame);
1284         }
1285         if (has_ready_frame) {
1286                 parent->release_display_frame(&ready_frame);
1287         }
1288 }
1289
1290 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1291 {
1292         // Store this frame for display. Remove the ready frame if any
1293         // (it was seemingly never used).
1294         {
1295                 unique_lock<mutex> lock(frame_mutex);
1296                 if (has_ready_frame) {
1297                         parent->release_display_frame(&ready_frame);
1298                 }
1299                 ready_frame = frame;
1300                 has_ready_frame = true;
1301         }
1302
1303         if (new_frame_ready_callback) {
1304                 new_frame_ready_callback();
1305         }
1306
1307         // Reduce the number of callbacks by filtering duplicates. The reason
1308         // why we bother doing this is that Qt seemingly can get into a state
1309         // where its builds up an essentially unbounded queue of signals,
1310         // consuming more and more memory, and there's no good way of collapsing
1311         // user-defined signals or limiting the length of the queue.
1312         if (transition_names_updated_callback) {
1313                 vector<string> transition_names = global_mixer->get_transition_names();
1314                 bool changed = false;
1315                 if (transition_names.size() != last_transition_names.size()) {
1316                         changed = true;
1317                 } else {
1318                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1319                                 if (transition_names[i] != last_transition_names[i]) {
1320                                         changed = true;
1321                                         break;
1322                                 }
1323                         }
1324                 }
1325                 if (changed) {
1326                         transition_names_updated_callback(transition_names);
1327                         last_transition_names = transition_names;
1328                 }
1329         }
1330         if (name_updated_callback) {
1331                 string name = global_mixer->get_channel_name(channel);
1332                 if (name != last_name) {
1333                         name_updated_callback(name);
1334                         last_name = name;
1335                 }
1336         }
1337         if (color_updated_callback) {
1338                 string color = global_mixer->get_channel_color(channel);
1339                 if (color != last_color) {
1340                         color_updated_callback(color);
1341                         last_color = color;
1342                 }
1343         }
1344 }
1345
1346 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1347 {
1348         unique_lock<mutex> lock(frame_mutex);
1349         if (!has_current_frame && !has_ready_frame) {
1350                 return false;
1351         }
1352
1353         if (has_current_frame && has_ready_frame) {
1354                 // We have a new ready frame. Toss the current one.
1355                 parent->release_display_frame(&current_frame);
1356                 has_current_frame = false;
1357         }
1358         if (has_ready_frame) {
1359                 assert(!has_current_frame);
1360                 current_frame = ready_frame;
1361                 ready_frame.ready_fence.reset();  // Drop the refcount.
1362                 ready_frame.input_frames.clear();  // Drop the refcounts.
1363                 has_current_frame = true;
1364                 has_ready_frame = false;
1365         }
1366
1367         *frame = current_frame;
1368         return true;
1369 }
1370
1371 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1372 {
1373         new_frame_ready_callback = callback;
1374 }
1375
1376 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1377 {
1378         transition_names_updated_callback = callback;
1379 }
1380
1381 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1382 {
1383         name_updated_callback = callback;
1384 }
1385
1386 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1387 {
1388         color_updated_callback = callback;
1389 }
1390
1391 mutex RefCountedGLsync::fence_lock;