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 "ffmpeg_capture.h"
  43 #include "flags.h"
  44 #include "input_mapping.h"
  45 #include "pbo_frame_allocator.h"
  46 #include "ref_counted_gl_sync.h"
  47 #include "resampling_queue.h"
  48 #include "timebase.h"
  49 #include "timecode_renderer.h"
  50 #include "v210_converter.h"
  51 #include "video_encoder.h"
  52
  53 class IDeckLink;
  54 class QOpenGLContext;
  55
  56 using namespace movit;
  57 using namespace std;
  58 using namespace std::chrono;
  59 using namespace std::placeholders;
  60 using namespace bmusb;
  61
  62 Mixer *global_mixer = nullptr;
  63 bool uses_mlock = false;
  64
  65 namespace {
  66
  67 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  68 {
  69         if (interlaced) {
  70                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  71                         input_state->buffered_frames[card_index][frame_num] =
  72                                 input_state->buffered_frames[card_index][frame_num - 1];
  73                 }
  74                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  75         } else {
  76                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  77                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  78                 }
  79         }
  80 }
  81
  82 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned v210_width)
  83 {
  84         bool first;
  85         switch (userdata->pixel_format) {
  86         case PixelFormat_10BitYCbCr:
  87                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  88                 break;
  89         case PixelFormat_8BitYCbCr:
  90                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
  91                 break;
  92         case PixelFormat_8BitBGRA:
  93                 first = userdata->tex_rgba[field] == 0;
  94                 break;
  95         default:
  96                 assert(false);
  97         }
  98
  99         if (first ||
 100             width != userdata->last_width[field] ||
 101             height != userdata->last_height[field]) {
 102                 // We changed resolution since last use of this texture, so we need to create
 103                 // a new object. Note that this each card has its own PBOFrameAllocator,
 104                 // we don't need to worry about these flip-flopping between resolutions.
 105                 switch (userdata->pixel_format) {
 106                 case PixelFormat_10BitYCbCr:
 107                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 108                         check_error();
 109                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 110                         check_error();
 111                         break;
 112                 case PixelFormat_8BitYCbCr: {
 113                         size_t cbcr_width = width / 2;
 114
 115                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 116                         check_error();
 117                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 118                         check_error();
 119                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 120                         check_error();
 121                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 122                         check_error();
 123                         break;
 124                 }
 125                 case PixelFormat_8BitBGRA:
 126                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 127                         check_error();
 128                         if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
 129                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 130                         } else {
 131                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 132                         }
 133                         check_error();
 134                         break;
 135                 }
 136                 userdata->last_width[field] = width;
 137                 userdata->last_height[field] = height;
 138         }
 139         if (global_flags.ten_bit_input &&
 140             (first || v210_width != userdata->last_v210_width[field])) {
 141                 // Same as above; we need to recreate the texture.
 142                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 143                 check_error();
 144                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 145                 check_error();
 146                 userdata->last_v210_width[field] = v210_width;
 147         }
 148 }
 149
 150 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 151 {
 152         if (interlaced_stride) {
 153                 stride *= 2;
 154         }
 155         if (global_flags.flush_pbos) {
 156                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 157                 check_error();
 158         }
 159
 160         glBindTexture(GL_TEXTURE_2D, tex);
 161         check_error();
 162         if (interlaced_stride) {
 163                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 164                 check_error();
 165         } else {
 166                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 167                 check_error();
 168         }
 169
 170         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 171         check_error();
 172         glBindTexture(GL_TEXTURE_2D, 0);
 173         check_error();
 174         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 175         check_error();
 176 }
 177
 178 }  // namespace
 179
 180 void QueueLengthPolicy::update_policy(unsigned queue_length)
 181 {
 182         if (queue_length == 0) {  // Starvation.
 183                 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
 184                         ++safe_queue_length;
 185                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
 186                                 card_index, safe_queue_length);
 187                 }
 188                 frames_with_at_least_one = 0;
 189                 been_at_safe_point_since_last_starvation = false;
 190                 return;
 191         }
 192         if (queue_length >= safe_queue_length) {
 193                 been_at_safe_point_since_last_starvation = true;
 194         }
 195         if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
 196                 --safe_queue_length;
 197                 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
 198                         card_index, safe_queue_length);
 199                 frames_with_at_least_one = 0;
 200         }
 201 }
 202
 203 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 204         : httpd(),
 205           num_cards(num_cards),
 206           mixer_surface(create_surface(format)),
 207           h264_encoder_surface(create_surface(format)),
 208           decklink_output_surface(create_surface(format)),
 209           ycbcr_interpretation(global_flags.ycbcr_interpretation),
 210           audio_mixer(num_cards)
 211 {
 212         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 213         check_error();
 214
 215         // This nearly always should be true.
 216         global_flags.can_disable_srgb_decoder =
 217                 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
 218                 epoxy_has_gl_extension("GL_ARB_sampler_objects");
 219
 220         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 221         // will be halved when sampling them, and we need to compensate here.
 222         movit_texel_subpixel_precision /= 2.0;
 223
 224         resource_pool.reset(new ResourcePool);
 225         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 226                 output_channel[i].parent = this;
 227                 output_channel[i].channel = i;
 228         }
 229
 230         ImageFormat inout_format;
 231         inout_format.color_space = COLORSPACE_sRGB;
 232         inout_format.gamma_curve = GAMMA_sRGB;
 233
 234         // Matches the 4:2:0 format created by the main chain.
 235         YCbCrFormat ycbcr_format;
 236         ycbcr_format.chroma_subsampling_x = 2;
 237         ycbcr_format.chroma_subsampling_y = 2;
 238         if (global_flags.ycbcr_rec709_coefficients) {
 239                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 240         } else {
 241                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 242         }
 243         ycbcr_format.full_range = false;
 244         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 245         ycbcr_format.cb_x_position = 0.0f;
 246         ycbcr_format.cr_x_position = 0.0f;
 247         ycbcr_format.cb_y_position = 0.5f;
 248         ycbcr_format.cr_y_position = 0.5f;
 249
 250         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 251         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 252         check_error();
 253         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 254         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 255         display_chain->add_input(display_input);
 256         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 257         display_chain->set_dither_bits(0);  // Don't bother.
 258         display_chain->finalize();
 259
 260         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));
 261
 262         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 263         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 264
 265         // Start listening for clients only once VideoEncoder has written its header, if any.
 266         httpd.start(9095);
 267
 268         // First try initializing the then PCI devices, then USB, then
 269         // fill up with fake cards until we have the desired number of cards.
 270         unsigned num_pci_devices = 0;
 271         unsigned card_index = 0;
 272
 273         {
 274                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 275                 if (decklink_iterator != nullptr) {
 276                         for ( ; card_index < num_cards; ++card_index) {
 277                                 IDeckLink *decklink;
 278                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 279                                         break;
 280                                 }
 281
 282                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 283                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 284                                 output->set_device(decklink);
 285                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 286                                 ++num_pci_devices;
 287                         }
 288                         decklink_iterator->Release();
 289                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 290                 } else {
 291                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 292                 }
 293         }
 294
 295         unsigned num_usb_devices = BMUSBCapture::num_cards();
 296         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 297                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 298                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 299                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 300         }
 301         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 302
 303         unsigned num_fake_cards = 0;
 304         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 305                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 306                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 307         }
 308
 309         if (num_fake_cards > 0) {
 310                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 311         }
 312
 313         // Initialize all video inputs the theme asked for. Note that these are
 314         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 315         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 316         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 317                 if (card_index >= MAX_VIDEO_CARDS) {
 318                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 319                         exit(1);
 320                 }
 321                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 322                 video_inputs[video_card_index]->set_card_index(card_index);
 323         }
 324         num_video_inputs = video_inputs.size();
 325
 326         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 327         BMUSBCapture::start_bm_thread();
 328
 329         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 330                 cards[card_index].queue_length_policy.reset(card_index);
 331         }
 332
 333         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 334
 335         if (global_flags.ten_bit_input) {
 336                 if (!v210Converter::has_hardware_support()) {
 337                         fprintf(stderr, "ERROR: --ten-bit-input 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                 v210_converter.reset(new v210Converter());
 342
 343                 // These are all the widths listed in the Blackmagic SDK documentation
 344                 // (section 2.7.3, “Display Modes”).
 345                 v210_converter->precompile_shader(720);
 346                 v210_converter->precompile_shader(1280);
 347                 v210_converter->precompile_shader(1920);
 348                 v210_converter->precompile_shader(2048);
 349                 v210_converter->precompile_shader(3840);
 350                 v210_converter->precompile_shader(4096);
 351         }
 352         if (global_flags.ten_bit_output) {
 353                 if (!v210Converter::has_hardware_support()) {
 354                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 355                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 356                         exit(1);
 357                 }
 358         }
 359
 360         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 361         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 362         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 363
 364         if (global_flags.enable_alsa_output) {
 365                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 366         }
 367         if (global_flags.output_card != -1) {
 368                 desired_output_card_index = global_flags.output_card;
 369                 set_output_card_internal(global_flags.output_card);
 370         }
 371 }
 372
 373 Mixer::~Mixer()
 374 {
 375         BMUSBCapture::stop_bm_thread();
 376
 377         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 378                 {
 379                         unique_lock<mutex> lock(card_mutex);
 380                         cards[card_index].should_quit = true;  // Unblock thread.
 381                         cards[card_index].new_frames_changed.notify_all();
 382                 }
 383                 cards[card_index].capture->stop_dequeue_thread();
 384                 if (cards[card_index].output) {
 385                         cards[card_index].output->end_output();
 386                         cards[card_index].output.reset();
 387                 }
 388         }
 389
 390         video_encoder.reset(nullptr);
 391 }
 392
 393 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 394 {
 395         printf("Configuring card %d...\n", card_index);
 396
 397         CaptureCard *card = &cards[card_index];
 398         if (card->capture != nullptr) {
 399                 card->capture->stop_dequeue_thread();
 400         }
 401         card->capture.reset(capture);
 402         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 403         if (card->output.get() != output) {
 404                 card->output.reset(output);
 405         }
 406
 407         PixelFormat pixel_format;
 408         if (card_type == CardType::FFMPEG_INPUT) {
 409                 pixel_format = PixelFormat_8BitBGRA;
 410         } else if (global_flags.ten_bit_input) {
 411                 pixel_format = PixelFormat_10BitYCbCr;
 412         } else {
 413                 pixel_format = PixelFormat_8BitYCbCr;
 414         }
 415
 416         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 417         if (card->frame_allocator == nullptr) {
 418                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 419         }
 420         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 421         if (card->surface == nullptr) {
 422                 card->surface = create_surface_with_same_format(mixer_surface);
 423         }
 424         while (!card->new_frames.empty()) card->new_frames.pop_front();
 425         card->last_timecode = -1;
 426         card->capture->set_pixel_format(pixel_format);
 427         card->capture->configure_card();
 428
 429         // NOTE: start_bm_capture() happens in thread_func().
 430
 431         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 432         audio_mixer.reset_resampler(device);
 433         audio_mixer.set_display_name(device, card->capture->get_description());
 434         audio_mixer.trigger_state_changed_callback();
 435 }
 436
 437 void Mixer::set_output_card_internal(int card_index)
 438 {
 439         // We don't really need to take card_mutex, since we're in the mixer
 440         // thread and don't mess with any queues (which is the only thing that happens
 441         // from other threads), but it's probably the safest in the long run.
 442         unique_lock<mutex> lock(card_mutex);
 443         if (output_card_index != -1) {
 444                 // Switch the old card from output to input.
 445                 CaptureCard *old_card = &cards[output_card_index];
 446                 old_card->output->end_output();
 447
 448                 // Stop the fake card that we put into place.
 449                 // This needs to _not_ happen under the mutex, to avoid deadlock
 450                 // (delivering the last frame needs to take the mutex).
 451                 CaptureInterface *fake_capture = old_card->capture.get();
 452                 lock.unlock();
 453                 fake_capture->stop_dequeue_thread();
 454                 lock.lock();
 455                 old_card->capture = move(old_card->parked_capture);
 456                 old_card->is_fake_capture = false;
 457                 old_card->capture->start_bm_capture();
 458         }
 459         if (card_index != -1) {
 460                 CaptureCard *card = &cards[card_index];
 461                 CaptureInterface *capture = card->capture.get();
 462                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 463                 // several seconds to complete (blocking on DisableVideoInput);
 464                 // see if we can maybe do it asynchronously.
 465                 lock.unlock();
 466                 capture->stop_dequeue_thread();
 467                 lock.lock();
 468                 card->parked_capture = move(card->capture);
 469                 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 470                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 471                 card->queue_length_policy.reset(card_index);
 472                 card->capture->start_bm_capture();
 473                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 474                 card->output->start_output(desired_output_video_mode, pts_int);
 475         }
 476         output_card_index = card_index;
 477 }
 478
 479 namespace {
 480
 481 int unwrap_timecode(uint16_t current_wrapped, int last)
 482 {
 483         uint16_t last_wrapped = last & 0xffff;
 484         if (current_wrapped > last_wrapped) {
 485                 return (last & ~0xffff) | current_wrapped;
 486         } else {
 487                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 488         }
 489 }
 490
 491 }  // namespace
 492
 493 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 494                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 495                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 496 {
 497         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 498         CaptureCard *card = &cards[card_index];
 499
 500         if (is_mode_scanning[card_index]) {
 501                 if (video_format.has_signal) {
 502                         // Found a stable signal, so stop scanning.
 503                         is_mode_scanning[card_index] = false;
 504                 } else {
 505                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 506                         steady_clock::time_point now = steady_clock::now();
 507                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 508                         if (sec_since_last_switch > switch_time_s) {
 509                                 // It isn't this mode; try the next one.
 510                                 mode_scanlist_index[card_index]++;
 511                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 512                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 513                                 last_mode_scan_change[card_index] = now;
 514                         }
 515                 }
 516         }
 517
 518         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 519         assert(frame_length > 0);
 520
 521         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;
 522         if (num_samples > OUTPUT_FREQUENCY / 10) {
 523                 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",
 524                         card_index, int(audio_frame.len), int(audio_offset),
 525                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 526                 if (video_frame.owner) {
 527                         video_frame.owner->release_frame(video_frame);
 528                 }
 529                 if (audio_frame.owner) {
 530                         audio_frame.owner->release_frame(audio_frame);
 531                 }
 532                 return;
 533         }
 534
 535         int dropped_frames = 0;
 536         if (card->last_timecode != -1) {
 537                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 538         }
 539
 540         // Number of samples per frame if we need to insert silence.
 541         // (Could be nonintegral, but resampling will save us then.)
 542         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 543
 544         if (dropped_frames > MAX_FPS * 2) {
 545                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 546                         card_index, card->last_timecode, timecode);
 547                 audio_mixer.reset_resampler(device);
 548                 dropped_frames = 0;
 549         } else if (dropped_frames > 0) {
 550                 // Insert silence as needed.
 551                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 552                         card_index, dropped_frames, timecode);
 553
 554                 bool success;
 555                 do {
 556                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 557                 } while (!success);
 558         }
 559
 560         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 561
 562         // Done with the audio, so release it.
 563         if (audio_frame.owner) {
 564                 audio_frame.owner->release_frame(audio_frame);
 565         }
 566
 567         card->last_timecode = timecode;
 568
 569         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 570         if (video_frame.len - video_offset == 0 ||
 571             video_frame.len - video_offset != expected_length) {
 572                 if (video_frame.len != 0) {
 573                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 574                                 card_index, video_frame.len - video_offset, expected_length);
 575                 }
 576                 if (video_frame.owner) {
 577                         video_frame.owner->release_frame(video_frame);
 578                 }
 579
 580                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 581                 // so that pts can go up accordingly.
 582                 {
 583                         unique_lock<mutex> lock(card_mutex);
 584                         CaptureCard::NewFrame new_frame;
 585                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 586                         new_frame.length = frame_length;
 587                         new_frame.interlaced = false;
 588                         new_frame.dropped_frames = dropped_frames;
 589                         new_frame.received_timestamp = video_frame.received_timestamp;
 590                         card->new_frames.push_back(move(new_frame));
 591                         card->new_frames_changed.notify_all();
 592                 }
 593                 return;
 594         }
 595
 596         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 597
 598         unsigned num_fields = video_format.interlaced ? 2 : 1;
 599         steady_clock::time_point frame_upload_start;
 600         bool interlaced_stride = false;
 601         if (video_format.interlaced) {
 602                 // Send the two fields along as separate frames; the other side will need to add
 603                 // a deinterlacer to actually get this right.
 604                 assert(video_format.height % 2 == 0);
 605                 video_format.height /= 2;
 606                 assert(frame_length % 2 == 0);
 607                 frame_length /= 2;
 608                 num_fields = 2;
 609                 if (video_format.second_field_start == 1) {
 610                         interlaced_stride = true;
 611                 }
 612                 frame_upload_start = steady_clock::now();
 613         }
 614         userdata->last_interlaced = video_format.interlaced;
 615         userdata->last_has_signal = video_format.has_signal;
 616         userdata->last_is_connected = video_format.is_connected;
 617         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 618         userdata->last_frame_rate_den = video_format.frame_rate_den;
 619         RefCountedFrame frame(video_frame);
 620
 621         // Upload the textures.
 622         const size_t cbcr_width = video_format.width / 2;
 623         const size_t cbcr_offset = video_offset / 2;
 624         const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 625
 626         for (unsigned field = 0; field < num_fields; ++field) {
 627                 // Put the actual texture upload in a lambda that is executed in the main thread.
 628                 // It is entirely possible to do this in the same thread (and it might even be
 629                 // faster, depending on the GPU and driver), but it appears to be trickling
 630                 // driver bugs very easily.
 631                 //
 632                 // Note that this means we must hold on to the actual frame data in <userdata>
 633                 // until the upload command is run, but we hold on to <frame> much longer than that
 634                 // (in fact, all the way until we no longer use the texture in rendering).
 635                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 636                         unsigned field_start_line;
 637                         if (field == 1) {
 638                                 field_start_line = video_format.second_field_start;
 639                         } else {
 640                                 field_start_line = video_format.extra_lines_top;
 641                         }
 642
 643                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 644                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 645                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 646
 647                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 648                         check_error();
 649
 650                         switch (userdata->pixel_format) {
 651                         case PixelFormat_10BitYCbCr: {
 652                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 653                                 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);
 654                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 655                                 break;
 656                         }
 657                         case PixelFormat_8BitYCbCr: {
 658                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 659                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 660
 661                                 // Make up our own strides, since we are interleaving.
 662                                 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);
 663                                 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);
 664                                 break;
 665                         }
 666                         case PixelFormat_8BitBGRA: {
 667                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 668                                 upload_texture(userdata->tex_rgba[field], video_format.width, video_format.height, video_format.stride, interlaced_stride, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, field_start);
 669                                 break;
 670                         }
 671                         default:
 672                                 assert(false);
 673                         }
 674
 675                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 676                         check_error();
 677                 };
 678
 679                 if (field == 1) {
 680                         // Don't upload the second field as fast as we can; wait until
 681                         // the field time has approximately passed. (Otherwise, we could
 682                         // get timing jitter against the other sources, and possibly also
 683                         // against the video display, although the latter is not as critical.)
 684                         // This requires our system clock to be reasonably close to the
 685                         // video clock, but that's not an unreasonable assumption.
 686                         steady_clock::time_point second_field_start = frame_upload_start +
 687                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 688                         this_thread::sleep_until(second_field_start);
 689                 }
 690
 691                 {
 692                         unique_lock<mutex> lock(card_mutex);
 693                         CaptureCard::NewFrame new_frame;
 694                         new_frame.frame = frame;
 695                         new_frame.length = frame_length;
 696                         new_frame.field = field;
 697                         new_frame.interlaced = video_format.interlaced;
 698                         new_frame.upload_func = upload_func;
 699                         new_frame.dropped_frames = dropped_frames;
 700                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 701                         card->new_frames.push_back(move(new_frame));
 702                         card->new_frames_changed.notify_all();
 703                 }
 704         }
 705 }
 706
 707 void Mixer::bm_hotplug_add(libusb_device *dev)
 708 {
 709         lock_guard<mutex> lock(hotplug_mutex);
 710         hotplugged_cards.push_back(dev);
 711 }
 712
 713 void Mixer::bm_hotplug_remove(unsigned card_index)
 714 {
 715         cards[card_index].new_frames_changed.notify_all();
 716 }
 717
 718 void Mixer::thread_func()
 719 {
 720         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 721
 722         eglBindAPI(EGL_OPENGL_API);
 723         QOpenGLContext *context = create_context(mixer_surface);
 724         if (!make_current(context, mixer_surface)) {
 725                 printf("oops\n");
 726                 exit(1);
 727         }
 728
 729         // Start the actual capture. (We don't want to do it before we're actually ready
 730         // to process output frames.)
 731         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 732                 if (int(card_index) != output_card_index) {
 733                         cards[card_index].capture->start_bm_capture();
 734                 }
 735         }
 736
 737         steady_clock::time_point start, now;
 738         start = steady_clock::now();
 739
 740         int stats_dropped_frames = 0;
 741
 742         while (!should_quit) {
 743                 if (desired_output_card_index != output_card_index) {
 744                         set_output_card_internal(desired_output_card_index);
 745                 }
 746                 if (output_card_index != -1 &&
 747                     desired_output_video_mode != output_video_mode) {
 748                         DeckLinkOutput *output = cards[output_card_index].output.get();
 749                         output->end_output();
 750                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 751                         output->start_output(desired_output_video_mode, pts_int);
 752                 }
 753
 754                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 755                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 756
 757                 bool master_card_is_output;
 758                 unsigned master_card_index;
 759                 if (output_card_index != -1) {
 760                         master_card_is_output = true;
 761                         master_card_index = output_card_index;
 762                 } else {
 763                         master_card_is_output = false;
 764                         master_card_index = theme->map_signal(master_clock_channel);
 765                         assert(master_card_index < num_cards);
 766                 }
 767
 768                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 769                 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);
 770                 stats_dropped_frames += output_frame_info.dropped_frames;
 771
 772                 handle_hotplugged_cards();
 773
 774                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 775                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 776                                 continue;
 777                         }
 778                         if (new_frames[card_index].frame->len == 0) {
 779                                 ++new_frames[card_index].dropped_frames;
 780                         }
 781                         if (new_frames[card_index].dropped_frames > 0) {
 782                                 printf("Card %u dropped %d frames before this\n",
 783                                         card_index, int(new_frames[card_index].dropped_frames));
 784                         }
 785                 }
 786
 787                 // If the first card is reporting a corrupted or otherwise dropped frame,
 788                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 789                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 790                         ++stats_dropped_frames;
 791                         pts_int += new_frames[master_card_index].length;
 792                         continue;
 793                 }
 794
 795                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 796                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 797                                 continue;
 798
 799                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 800                         assert(new_frame->frame != nullptr);
 801                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 802                         check_error();
 803
 804                         // The new texture might need uploading before use.
 805                         if (new_frame->upload_func) {
 806                                 new_frame->upload_func();
 807                                 new_frame->upload_func = nullptr;
 808                         }
 809                 }
 810
 811                 int64_t frame_duration = output_frame_info.frame_duration;
 812                 render_one_frame(frame_duration);
 813                 ++frame_num;
 814                 pts_int += frame_duration;
 815
 816                 now = steady_clock::now();
 817                 double elapsed = duration<double>(now - start).count();
 818                 if (frame_num % 100 == 0) {
 819                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 820                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 821                                 1e3 * elapsed / frame_num);
 822                 //      chain->print_phase_timing();
 823
 824                         // Check our memory usage, to see if we are close to our mlockall()
 825                         // limit (if at all set).
 826                         rusage used;
 827                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 828                                 perror("getrusage(RUSAGE_SELF)");
 829                                 assert(false);
 830                         }
 831
 832                         if (uses_mlock) {
 833                                 rlimit limit;
 834                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 835                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 836                                         assert(false);
 837                                 }
 838
 839                                 if (limit.rlim_cur == 0) {
 840                                         printf(", using %ld MB memory (locked)",
 841                                                 long(used.ru_maxrss / 1024));
 842                                 } else {
 843                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 844                                                 long(used.ru_maxrss / 1024),
 845                                                 long(limit.rlim_cur / 1048576),
 846                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 847                                 }
 848                         } else {
 849                                 printf(", using %ld MB memory (not locked)",
 850                                         long(used.ru_maxrss / 1024));
 851                         }
 852
 853                         printf("\n");
 854                 }
 855
 856
 857                 if (should_cut.exchange(false)) {  // Test and clear.
 858                         video_encoder->do_cut(frame_num);
 859                 }
 860
 861 #if 0
 862                 // Reset every 100 frames, so that local variations in frame times
 863                 // (especially for the first few frames, when the shaders are
 864                 // compiled etc.) don't make it hard to measure for the entire
 865                 // remaining duration of the program.
 866                 if (frame == 10000) {
 867                         frame = 0;
 868                         start = now;
 869                 }
 870 #endif
 871                 check_error();
 872         }
 873
 874         resource_pool->clean_context();
 875 }
 876
 877 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 878 {
 879         if (output_card_index != -1) {
 880                 // The output card (ie., cards[output_card_index].output) is the master clock,
 881                 // so no input card (ie., cards[card_index].capture) is.
 882                 return false;
 883         }
 884         return (card_index == master_card_index);
 885 }
 886
 887 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 888 {
 889         // Count the number of frames in the queue, including any frames
 890         // we dropped. It's hard to know exactly how we should deal with
 891         // dropped (corrupted) input frames; they don't help our goal of
 892         // avoiding starvation, but they still add to the problem of latency.
 893         // Since dropped frames is going to mean a bump in the signal anyway,
 894         // we err on the side of having more stable latency instead.
 895         unsigned queue_length = 0;
 896         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 897                 queue_length += frame.dropped_frames + 1;
 898         }
 899         card->queue_length_policy.update_policy(queue_length);
 900
 901         // If needed, drop frames until the queue is below the safe limit.
 902         // We prefer to drop from the head, because all else being equal,
 903         // we'd like more recent frames (less latency).
 904         unsigned dropped_frames = 0;
 905         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 906                 assert(!card->new_frames.empty());
 907                 assert(queue_length > card->new_frames.front().dropped_frames);
 908                 queue_length -= card->new_frames.front().dropped_frames;
 909
 910                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 911                         // No need to drop anything.
 912                         break;
 913                 }
 914
 915                 card->new_frames.pop_front();
 916                 card->new_frames_changed.notify_all();
 917                 --queue_length;
 918                 ++dropped_frames;
 919         }
 920
 921 #if 0
 922         if (dropped_frames > 0) {
 923                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 924                         card_index, dropped_frames);
 925         }
 926 #endif
 927 }
 928
 929
 930 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])
 931 {
 932         OutputFrameInfo output_frame_info;
 933 start:
 934         unique_lock<mutex> lock(card_mutex, defer_lock);
 935         if (master_card_is_output) {
 936                 // Clocked to the output, so wait for it to be ready for the next frame.
 937                 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);
 938                 lock.lock();
 939         } else {
 940                 // Wait for the master card to have a new frame.
 941                 // TODO: Add a timeout.
 942                 output_frame_info.is_preroll = false;
 943                 lock.lock();
 944                 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(); });
 945         }
 946
 947         if (master_card_is_output) {
 948                 handle_hotplugged_cards();
 949         } else if (cards[master_card_index].new_frames.empty()) {
 950                 // We were woken up, but not due to a new frame. Deal with it
 951                 // and then restart.
 952                 assert(cards[master_card_index].capture->get_disconnected());
 953                 handle_hotplugged_cards();
 954                 goto start;
 955         }
 956
 957         if (!master_card_is_output) {
 958                 output_frame_info.frame_timestamp =
 959                         cards[master_card_index].new_frames.front().received_timestamp;
 960         }
 961
 962         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 963                 CaptureCard *card = &cards[card_index];
 964                 if (input_card_is_master_clock(card_index, master_card_index)) {
 965                         // We don't use the queue length policy for the master card,
 966                         // but we will if it stops being the master. Thus, clear out
 967                         // the policy in case we switch in the future.
 968                         card->queue_length_policy.reset(card_index);
 969                         assert(!card->new_frames.empty());
 970                 } else {
 971                         trim_queue(card, card_index);
 972                 }
 973                 if (!card->new_frames.empty()) {
 974                         new_frames[card_index] = move(card->new_frames.front());
 975                         has_new_frame[card_index] = true;
 976                         card->new_frames.pop_front();
 977                         card->new_frames_changed.notify_all();
 978                 }
 979         }
 980
 981         if (!master_card_is_output) {
 982                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 983                 output_frame_info.frame_duration = new_frames[master_card_index].length;
 984         }
 985
 986         // This might get off by a fractional sample when changing master card
 987         // between ones with different frame rates, but that's fine.
 988         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 989         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 990         fractional_samples = num_samples_times_timebase % TIMEBASE;
 991         assert(output_frame_info.num_samples >= 0);
 992
 993         return output_frame_info;
 994 }
 995
 996 void Mixer::handle_hotplugged_cards()
 997 {
 998         // Check for cards that have been disconnected since last frame.
 999         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1000                 CaptureCard *card = &cards[card_index];
1001                 if (card->capture->get_disconnected()) {
1002                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1003                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1004                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1005                         card->queue_length_policy.reset(card_index);
1006                         card->capture->start_bm_capture();
1007                 }
1008         }
1009
1010         // Check for cards that have been connected since last frame.
1011         vector<libusb_device *> hotplugged_cards_copy;
1012         {
1013                 lock_guard<mutex> lock(hotplug_mutex);
1014                 swap(hotplugged_cards, hotplugged_cards_copy);
1015         }
1016         for (libusb_device *new_dev : hotplugged_cards_copy) {
1017                 // Look for a fake capture card where we can stick this in.
1018                 int free_card_index = -1;
1019                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1020                         if (cards[card_index].is_fake_capture) {
1021                                 free_card_index = card_index;
1022                                 break;
1023                         }
1024                 }
1025
1026                 if (free_card_index == -1) {
1027                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1028                         libusb_unref_device(new_dev);
1029                 } else {
1030                         // BMUSBCapture takes ownership.
1031                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1032                         CaptureCard *card = &cards[free_card_index];
1033                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1034                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1035                         card->queue_length_policy.reset(free_card_index);
1036                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1037                         capture->start_bm_capture();
1038                 }
1039         }
1040 }
1041
1042
1043 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)
1044 {
1045         // Resample the audio as needed, including from previously dropped frames.
1046         assert(num_cards > 0);
1047         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1048                 const bool dropped_frame = (frame_num != dropped_frames);
1049                 {
1050                         // Signal to the audio thread to process this frame.
1051                         // Note that if the frame is a dropped frame, we signal that
1052                         // we don't want to use this frame as base for adjusting
1053                         // the resampler rate. The reason for this is that the timing
1054                         // of these frames is often way too late; they typically don't
1055                         // “arrive” before we synthesize them. Thus, we could end up
1056                         // in a situation where we have inserted e.g. five audio frames
1057                         // into the queue before we then start pulling five of them
1058                         // back out. This makes ResamplingQueue overestimate the delay,
1059                         // causing undue resampler changes. (We _do_ use the last,
1060                         // non-dropped frame; perhaps we should just discard that as well,
1061                         // since dropped frames are expected to be rare, and it might be
1062                         // better to just wait until we have a slightly more normal situation).
1063                         unique_lock<mutex> lock(audio_mutex);
1064                         bool adjust_rate = !dropped_frame && !is_preroll;
1065                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1066                         audio_task_queue_changed.notify_one();
1067                 }
1068                 if (dropped_frame) {
1069                         // For dropped frames, increase the pts. Note that if the format changed
1070                         // in the meantime, we have no way of detecting that; we just have to
1071                         // assume the frame length is always the same.
1072                         pts_int += length_per_frame;
1073                 }
1074         }
1075 }
1076
1077 void Mixer::render_one_frame(int64_t duration)
1078 {
1079         // Determine the time code for this frame before we start rendering.
1080         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1081         if (display_timecode_on_stdout) {
1082                 printf("Timecode: '%s'\n", timecode_text.c_str());
1083         }
1084
1085         // Update Y'CbCr settings for all cards.
1086         {
1087                 unique_lock<mutex> lock(card_mutex);
1088                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1089                         YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1090                         input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1091                         input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1092                         input_state.full_range[card_index] = interpretation->full_range;
1093                 }
1094         }
1095
1096         // Get the main chain from the theme, and set its state immediately.
1097         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1098         EffectChain *chain = theme_main_chain.chain;
1099         theme_main_chain.setup_chain();
1100         //theme_main_chain.chain->enable_phase_timing(true);
1101
1102         // The theme can't (or at least shouldn't!) call connect_signal() on
1103         // each FFmpeg input, so we'll do it here.
1104         for (const pair<LiveInputWrapper *, FFmpegCapture *> &conn : theme->get_signal_connections()) {
1105                 conn.first->connect_signal_raw(conn.second->get_card_index());
1106         }
1107
1108         // If HDMI/SDI output is active and the user has requested auto mode,
1109         // its mode overrides the existing Y'CbCr setting for the chain.
1110         YCbCrLumaCoefficients ycbcr_output_coefficients;
1111         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1112                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1113         } else {
1114                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1115         }
1116
1117         // TODO: Reduce the duplication against theme.cpp.
1118         YCbCrFormat output_ycbcr_format;
1119         output_ycbcr_format.chroma_subsampling_x = 1;
1120         output_ycbcr_format.chroma_subsampling_y = 1;
1121         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1122         output_ycbcr_format.full_range = false;
1123         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1124         chain->change_ycbcr_output_format(output_ycbcr_format);
1125
1126         // Render main chain. If we're using zerocopy Quick Sync encoding
1127         // (the default case), we take an extra copy of the created outputs,
1128         // so that we can display it back to the screen later (it's less memory
1129         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1130         // Ideally, we'd like to avoid taking copies and just use the main textures
1131         // for display as well, but they're just views into VA-API memory and must be
1132         // unmapped during encoding, so we can't use them for display, unfortunately.
1133         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1134         GLuint y_copy_tex, cbcr_copy_tex = 0;
1135         GLuint y_display_tex, cbcr_display_tex;
1136         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1137         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1138         const bool is_zerocopy = video_encoder->is_zerocopy();
1139         if (is_zerocopy) {
1140                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1141                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1142                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1143
1144                 y_display_tex = y_copy_tex;
1145                 cbcr_display_tex = cbcr_copy_tex;
1146
1147                 // y_tex and cbcr_tex will be given by VideoEncoder.
1148         } else {
1149                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1150                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1151                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1152
1153                 y_display_tex = y_tex;
1154                 cbcr_display_tex = cbcr_tex;
1155         }
1156
1157         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1158         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1159         assert(got_frame);
1160
1161         GLuint fbo;
1162         if (is_zerocopy) {
1163                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1164         } else {
1165                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1166         }
1167         check_error();
1168         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1169
1170         if (display_timecode_in_stream) {
1171                 // Render the timecode on top.
1172                 timecode_renderer->render_timecode(fbo, timecode_text);
1173         }
1174
1175         resource_pool->release_fbo(fbo);
1176
1177         if (is_zerocopy) {
1178                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1179         } else {
1180                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1181         }
1182         if (output_card_index != -1) {
1183                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1184         }
1185         resource_pool->release_2d_texture(cbcr_full_tex);
1186
1187         // Set the right state for the Y' and CbCr textures we use for display.
1188         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1189         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1190         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1191         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1192         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1193
1194         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1195         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1196         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1197         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1198
1199         RefCountedGLsync fence = video_encoder->end_frame();
1200
1201         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1202         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1203         DisplayFrame live_frame;
1204         live_frame.chain = display_chain.get();
1205         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1206                 display_input->set_texture_num(0, y_display_tex);
1207                 display_input->set_texture_num(1, cbcr_display_tex);
1208         };
1209         live_frame.ready_fence = fence;
1210         live_frame.input_frames = {};
1211         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1212         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1213
1214         // Set up preview and any additional channels.
1215         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1216                 DisplayFrame display_frame;
1217                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1218                 display_frame.chain = chain.chain;
1219                 display_frame.setup_chain = chain.setup_chain;
1220                 display_frame.ready_fence = fence;
1221                 display_frame.input_frames = chain.input_frames;
1222                 display_frame.temp_textures = {};
1223                 output_channel[i].output_frame(display_frame);
1224         }
1225 }
1226
1227 void Mixer::audio_thread_func()
1228 {
1229         pthread_setname_np(pthread_self(), "Mixer_Audio");
1230
1231         while (!should_quit) {
1232                 AudioTask task;
1233
1234                 {
1235                         unique_lock<mutex> lock(audio_mutex);
1236                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1237                         if (should_quit) {
1238                                 return;
1239                         }
1240                         task = audio_task_queue.front();
1241                         audio_task_queue.pop();
1242                 }
1243
1244                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1245                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1246                 vector<float> samples_out = audio_mixer.get_output(
1247                         task.frame_timestamp,
1248                         task.num_samples,
1249                         rate_adjustment_policy);
1250
1251                 // Send the samples to the sound card, then add them to the output.
1252                 if (alsa) {
1253                         alsa->write(samples_out);
1254                 }
1255                 if (output_card_index != -1) {
1256                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1257                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1258                 }
1259                 video_encoder->add_audio(task.pts_int, move(samples_out));
1260         }
1261 }
1262
1263 void Mixer::release_display_frame(DisplayFrame *frame)
1264 {
1265         for (GLuint texnum : frame->temp_textures) {
1266                 resource_pool->release_2d_texture(texnum);
1267         }
1268         frame->temp_textures.clear();
1269         frame->ready_fence.reset();
1270         frame->input_frames.clear();
1271 }
1272
1273 void Mixer::start()
1274 {
1275         mixer_thread = thread(&Mixer::thread_func, this);
1276         audio_thread = thread(&Mixer::audio_thread_func, this);
1277 }
1278
1279 void Mixer::quit()
1280 {
1281         should_quit = true;
1282         audio_task_queue_changed.notify_one();
1283         mixer_thread.join();
1284         audio_thread.join();
1285 }
1286
1287 void Mixer::transition_clicked(int transition_num)
1288 {
1289         theme->transition_clicked(transition_num, pts());
1290 }
1291
1292 void Mixer::channel_clicked(int preview_num)
1293 {
1294         theme->channel_clicked(preview_num);
1295 }
1296
1297 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1298 {
1299         unique_lock<mutex> lock(card_mutex);
1300         return ycbcr_interpretation[card_index];
1301 }
1302
1303 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1304 {
1305         unique_lock<mutex> lock(card_mutex);
1306         ycbcr_interpretation[card_index] = interpretation;
1307 }
1308
1309 void Mixer::start_mode_scanning(unsigned card_index)
1310 {
1311         assert(card_index < num_cards);
1312         if (is_mode_scanning[card_index]) {
1313                 return;
1314         }
1315         is_mode_scanning[card_index] = true;
1316         mode_scanlist[card_index].clear();
1317         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1318                 mode_scanlist[card_index].push_back(mode.first);
1319         }
1320         assert(!mode_scanlist[card_index].empty());
1321         mode_scanlist_index[card_index] = 0;
1322         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1323         last_mode_scan_change[card_index] = steady_clock::now();
1324 }
1325
1326 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1327 {
1328         assert(desired_output_card_index != -1);
1329         unique_lock<mutex> lock(card_mutex);
1330         return cards[desired_output_card_index].output->get_available_video_modes();
1331 }
1332
1333 Mixer::OutputChannel::~OutputChannel()
1334 {
1335         if (has_current_frame) {
1336                 parent->release_display_frame(&current_frame);
1337         }
1338         if (has_ready_frame) {
1339                 parent->release_display_frame(&ready_frame);
1340         }
1341 }
1342
1343 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1344 {
1345         // Store this frame for display. Remove the ready frame if any
1346         // (it was seemingly never used).
1347         {
1348                 unique_lock<mutex> lock(frame_mutex);
1349                 if (has_ready_frame) {
1350                         parent->release_display_frame(&ready_frame);
1351                 }
1352                 ready_frame = frame;
1353                 has_ready_frame = true;
1354
1355                 // Call the callbacks under the mutex (they should be short),
1356                 // so that we don't race against a callback removal.
1357                 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1358                         key_and_callback.second();
1359                 }
1360         }
1361
1362         // Reduce the number of callbacks by filtering duplicates. The reason
1363         // why we bother doing this is that Qt seemingly can get into a state
1364         // where its builds up an essentially unbounded queue of signals,
1365         // consuming more and more memory, and there's no good way of collapsing
1366         // user-defined signals or limiting the length of the queue.
1367         if (transition_names_updated_callback) {
1368                 vector<string> transition_names = global_mixer->get_transition_names();
1369                 bool changed = false;
1370                 if (transition_names.size() != last_transition_names.size()) {
1371                         changed = true;
1372                 } else {
1373                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1374                                 if (transition_names[i] != last_transition_names[i]) {
1375                                         changed = true;
1376                                         break;
1377                                 }
1378                         }
1379                 }
1380                 if (changed) {
1381                         transition_names_updated_callback(transition_names);
1382                         last_transition_names = transition_names;
1383                 }
1384         }
1385         if (name_updated_callback) {
1386                 string name = global_mixer->get_channel_name(channel);
1387                 if (name != last_name) {
1388                         name_updated_callback(name);
1389                         last_name = name;
1390                 }
1391         }
1392         if (color_updated_callback) {
1393                 string color = global_mixer->get_channel_color(channel);
1394                 if (color != last_color) {
1395                         color_updated_callback(color);
1396                         last_color = color;
1397                 }
1398         }
1399 }
1400
1401 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1402 {
1403         unique_lock<mutex> lock(frame_mutex);
1404         if (!has_current_frame && !has_ready_frame) {
1405                 return false;
1406         }
1407
1408         if (has_current_frame && has_ready_frame) {
1409                 // We have a new ready frame. Toss the current one.
1410                 parent->release_display_frame(&current_frame);
1411                 has_current_frame = false;
1412         }
1413         if (has_ready_frame) {
1414                 assert(!has_current_frame);
1415                 current_frame = ready_frame;
1416                 ready_frame.ready_fence.reset();  // Drop the refcount.
1417                 ready_frame.input_frames.clear();  // Drop the refcounts.
1418                 has_current_frame = true;
1419                 has_ready_frame = false;
1420         }
1421
1422         *frame = current_frame;
1423         return true;
1424 }
1425
1426 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1427 {
1428         unique_lock<mutex> lock(frame_mutex);
1429         new_frame_ready_callbacks[key] = callback;
1430 }
1431
1432 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1433 {
1434         unique_lock<mutex> lock(frame_mutex);
1435         new_frame_ready_callbacks.erase(key);
1436 }
1437
1438 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1439 {
1440         transition_names_updated_callback = callback;
1441 }
1442
1443 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1444 {
1445         name_updated_callback = callback;
1446 }
1447
1448 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1449 {
1450         color_updated_callback = callback;
1451 }
1452
1453 mutex RefCountedGLsync::fence_lock;