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 "metrics.h"
  46 #include "pbo_frame_allocator.h"
  47 #include "ref_counted_gl_sync.h"
  48 #include "resampling_queue.h"
  49 #include "timebase.h"
  50 #include "timecode_renderer.h"
  51 #include "v210_converter.h"
  52 #include "video_encoder.h"
  53
  54 class IDeckLink;
  55 class QOpenGLContext;
  56
  57 using namespace movit;
  58 using namespace std;
  59 using namespace std::chrono;
  60 using namespace std::placeholders;
  61 using namespace bmusb;
  62
  63 Mixer *global_mixer = nullptr;
  64 bool uses_mlock = false;
  65
  66 namespace {
  67
  68 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  69 {
  70         if (interlaced) {
  71                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  72                         input_state->buffered_frames[card_index][frame_num] =
  73                                 input_state->buffered_frames[card_index][frame_num - 1];
  74                 }
  75                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  76         } else {
  77                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  78                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  79                 }
  80         }
  81 }
  82
  83 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
  84 {
  85         bool first;
  86         switch (userdata->pixel_format) {
  87         case PixelFormat_10BitYCbCr:
  88                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  89                 break;
  90         case PixelFormat_8BitYCbCr:
  91                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
  92                 break;
  93         case PixelFormat_8BitBGRA:
  94                 first = userdata->tex_rgba[field] == 0;
  95                 break;
  96         case PixelFormat_8BitYCbCrPlanar:
  97                 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
  98                 break;
  99         default:
 100                 assert(false);
 101         }
 102
 103         if (first ||
 104             width != userdata->last_width[field] ||
 105             height != userdata->last_height[field] ||
 106             cbcr_width != userdata->last_cbcr_width[field] ||
 107             cbcr_height != userdata->last_cbcr_height[field]) {
 108                 // We changed resolution since last use of this texture, so we need to create
 109                 // a new object. Note that this each card has its own PBOFrameAllocator,
 110                 // we don't need to worry about these flip-flopping between resolutions.
 111                 switch (userdata->pixel_format) {
 112                 case PixelFormat_10BitYCbCr:
 113                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 114                         check_error();
 115                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 116                         check_error();
 117                         break;
 118                 case PixelFormat_8BitYCbCr: {
 119                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 120                         check_error();
 121                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 122                         check_error();
 123                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 124                         check_error();
 125                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 126                         check_error();
 127                         break;
 128                 }
 129                 case PixelFormat_8BitYCbCrPlanar: {
 130                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 131                         check_error();
 132                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 133                         check_error();
 134                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
 135                         check_error();
 136                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 137                         check_error();
 138                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
 139                         check_error();
 140                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 141                         check_error();
 142                         break;
 143                 }
 144                 case PixelFormat_8BitBGRA:
 145                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 146                         check_error();
 147                         if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
 148                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 149                         } else {
 150                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 151                         }
 152                         check_error();
 153                         break;
 154                 }
 155                 userdata->last_width[field] = width;
 156                 userdata->last_height[field] = height;
 157                 userdata->last_cbcr_width[field] = cbcr_width;
 158                 userdata->last_cbcr_height[field] = cbcr_height;
 159         }
 160         if (global_flags.ten_bit_input &&
 161             (first || v210_width != userdata->last_v210_width[field])) {
 162                 // Same as above; we need to recreate the texture.
 163                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 164                 check_error();
 165                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 166                 check_error();
 167                 userdata->last_v210_width[field] = v210_width;
 168         }
 169 }
 170
 171 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 172 {
 173         if (interlaced_stride) {
 174                 stride *= 2;
 175         }
 176         if (global_flags.flush_pbos) {
 177                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 178                 check_error();
 179         }
 180
 181         glBindTexture(GL_TEXTURE_2D, tex);
 182         check_error();
 183         if (interlaced_stride) {
 184                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 185                 check_error();
 186         } else {
 187                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 188                 check_error();
 189         }
 190
 191         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 192         check_error();
 193         glBindTexture(GL_TEXTURE_2D, 0);
 194         check_error();
 195         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 196         check_error();
 197 }
 198
 199 }  // namespace
 200
 201 void QueueLengthPolicy::update_policy(unsigned queue_length)
 202 {
 203         if (queue_length == 0) {  // Starvation.
 204                 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
 205                         ++safe_queue_length;
 206                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
 207                                 card_index, safe_queue_length);
 208                 }
 209                 frames_with_at_least_one = 0;
 210                 been_at_safe_point_since_last_starvation = false;
 211                 return;
 212         }
 213         if (queue_length >= safe_queue_length) {
 214                 been_at_safe_point_since_last_starvation = true;
 215         }
 216         if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
 217                 --safe_queue_length;
 218                 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
 219                         card_index, safe_queue_length);
 220                 frames_with_at_least_one = 0;
 221         }
 222 }
 223
 224 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 225         : httpd(),
 226           num_cards(num_cards),
 227           mixer_surface(create_surface(format)),
 228           h264_encoder_surface(create_surface(format)),
 229           decklink_output_surface(create_surface(format)),
 230           ycbcr_interpretation(global_flags.ycbcr_interpretation),
 231           audio_mixer(num_cards)
 232 {
 233         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 234         check_error();
 235
 236         // This nearly always should be true.
 237         global_flags.can_disable_srgb_decoder =
 238                 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
 239                 epoxy_has_gl_extension("GL_ARB_sampler_objects");
 240
 241         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 242         // will be halved when sampling them, and we need to compensate here.
 243         movit_texel_subpixel_precision /= 2.0;
 244
 245         resource_pool.reset(new ResourcePool);
 246         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 247                 output_channel[i].parent = this;
 248                 output_channel[i].channel = i;
 249         }
 250
 251         ImageFormat inout_format;
 252         inout_format.color_space = COLORSPACE_sRGB;
 253         inout_format.gamma_curve = GAMMA_sRGB;
 254
 255         // Matches the 4:2:0 format created by the main chain.
 256         YCbCrFormat ycbcr_format;
 257         ycbcr_format.chroma_subsampling_x = 2;
 258         ycbcr_format.chroma_subsampling_y = 2;
 259         if (global_flags.ycbcr_rec709_coefficients) {
 260                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 261         } else {
 262                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 263         }
 264         ycbcr_format.full_range = false;
 265         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 266         ycbcr_format.cb_x_position = 0.0f;
 267         ycbcr_format.cr_x_position = 0.0f;
 268         ycbcr_format.cb_y_position = 0.5f;
 269         ycbcr_format.cr_y_position = 0.5f;
 270
 271         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 272         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 273         check_error();
 274         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 275         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 276         display_chain->add_input(display_input);
 277         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 278         display_chain->set_dither_bits(0);  // Don't bother.
 279         display_chain->finalize();
 280
 281         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));
 282
 283         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 284         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 285
 286         // Start listening for clients only once VideoEncoder has written its header, if any.
 287         httpd.start(9095);
 288
 289         // First try initializing the then PCI devices, then USB, then
 290         // fill up with fake cards until we have the desired number of cards.
 291         unsigned num_pci_devices = 0;
 292         unsigned card_index = 0;
 293
 294         {
 295                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 296                 if (decklink_iterator != nullptr) {
 297                         for ( ; card_index < num_cards; ++card_index) {
 298                                 IDeckLink *decklink;
 299                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 300                                         break;
 301                                 }
 302
 303                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 304                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 305                                 output->set_device(decklink);
 306                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 307                                 ++num_pci_devices;
 308                         }
 309                         decklink_iterator->Release();
 310                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 311                 } else {
 312                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 313                 }
 314         }
 315
 316         unsigned num_usb_devices = BMUSBCapture::num_cards();
 317         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 318                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 319                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 320                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 321         }
 322         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 323
 324         unsigned num_fake_cards = 0;
 325         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 326                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 327                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 328         }
 329
 330         if (num_fake_cards > 0) {
 331                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 332         }
 333
 334         // Initialize all video inputs the theme asked for. Note that these are
 335         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 336         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 337         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 338                 if (card_index >= MAX_VIDEO_CARDS) {
 339                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 340                         exit(1);
 341                 }
 342                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 343                 video_inputs[video_card_index]->set_card_index(card_index);
 344         }
 345         num_video_inputs = video_inputs.size();
 346
 347         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 348         BMUSBCapture::start_bm_thread();
 349
 350         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 351                 cards[card_index].queue_length_policy.reset(card_index);
 352         }
 353
 354         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 355
 356         if (global_flags.ten_bit_input) {
 357                 if (!v210Converter::has_hardware_support()) {
 358                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 359                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 360                         exit(1);
 361                 }
 362                 v210_converter.reset(new v210Converter());
 363
 364                 // These are all the widths listed in the Blackmagic SDK documentation
 365                 // (section 2.7.3, “Display Modes”).
 366                 v210_converter->precompile_shader(720);
 367                 v210_converter->precompile_shader(1280);
 368                 v210_converter->precompile_shader(1920);
 369                 v210_converter->precompile_shader(2048);
 370                 v210_converter->precompile_shader(3840);
 371                 v210_converter->precompile_shader(4096);
 372         }
 373         if (global_flags.ten_bit_output) {
 374                 if (!v210Converter::has_hardware_support()) {
 375                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 376                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 377                         exit(1);
 378                 }
 379         }
 380
 381         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 382         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 383         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 384
 385         if (global_flags.enable_alsa_output) {
 386                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 387         }
 388         if (global_flags.output_card != -1) {
 389                 desired_output_card_index = global_flags.output_card;
 390                 set_output_card_internal(global_flags.output_card);
 391         }
 392
 393         global_metrics.register_int_metric("num_frames", &metrics_num_frames);
 394         global_metrics.register_int_metric("dropped_frames", &metrics_dropped_frames);
 395         global_metrics.register_double_metric("uptime", &metrics_uptime);
 396 }
 397
 398 Mixer::~Mixer()
 399 {
 400         BMUSBCapture::stop_bm_thread();
 401
 402         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 403                 {
 404                         unique_lock<mutex> lock(card_mutex);
 405                         cards[card_index].should_quit = true;  // Unblock thread.
 406                         cards[card_index].new_frames_changed.notify_all();
 407                 }
 408                 cards[card_index].capture->stop_dequeue_thread();
 409                 if (cards[card_index].output) {
 410                         cards[card_index].output->end_output();
 411                         cards[card_index].output.reset();
 412                 }
 413         }
 414
 415         video_encoder.reset(nullptr);
 416 }
 417
 418 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 419 {
 420         printf("Configuring card %d...\n", card_index);
 421
 422         CaptureCard *card = &cards[card_index];
 423         if (card->capture != nullptr) {
 424                 card->capture->stop_dequeue_thread();
 425         }
 426         card->capture.reset(capture);
 427         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 428         card->type = card_type;
 429         if (card->output.get() != output) {
 430                 card->output.reset(output);
 431         }
 432
 433         PixelFormat pixel_format;
 434         if (card_type == CardType::FFMPEG_INPUT) {
 435                 pixel_format = capture->get_current_pixel_format();
 436         } else if (global_flags.ten_bit_input) {
 437                 pixel_format = PixelFormat_10BitYCbCr;
 438         } else {
 439                 pixel_format = PixelFormat_8BitYCbCr;
 440         }
 441
 442         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 443         if (card->frame_allocator == nullptr) {
 444                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 445         }
 446         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 447         if (card->surface == nullptr) {
 448                 card->surface = create_surface_with_same_format(mixer_surface);
 449         }
 450         while (!card->new_frames.empty()) card->new_frames.pop_front();
 451         card->last_timecode = -1;
 452         card->capture->set_pixel_format(pixel_format);
 453         card->capture->configure_card();
 454
 455         // NOTE: start_bm_capture() happens in thread_func().
 456
 457         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 458         audio_mixer.reset_resampler(device);
 459         audio_mixer.set_display_name(device, card->capture->get_description());
 460         audio_mixer.trigger_state_changed_callback();
 461 }
 462
 463 void Mixer::set_output_card_internal(int card_index)
 464 {
 465         // We don't really need to take card_mutex, since we're in the mixer
 466         // thread and don't mess with any queues (which is the only thing that happens
 467         // from other threads), but it's probably the safest in the long run.
 468         unique_lock<mutex> lock(card_mutex);
 469         if (output_card_index != -1) {
 470                 // Switch the old card from output to input.
 471                 CaptureCard *old_card = &cards[output_card_index];
 472                 old_card->output->end_output();
 473
 474                 // Stop the fake card that we put into place.
 475                 // This needs to _not_ happen under the mutex, to avoid deadlock
 476                 // (delivering the last frame needs to take the mutex).
 477                 CaptureInterface *fake_capture = old_card->capture.get();
 478                 lock.unlock();
 479                 fake_capture->stop_dequeue_thread();
 480                 lock.lock();
 481                 old_card->capture = move(old_card->parked_capture);
 482                 old_card->is_fake_capture = false;
 483                 old_card->capture->start_bm_capture();
 484         }
 485         if (card_index != -1) {
 486                 CaptureCard *card = &cards[card_index];
 487                 CaptureInterface *capture = card->capture.get();
 488                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 489                 // several seconds to complete (blocking on DisableVideoInput);
 490                 // see if we can maybe do it asynchronously.
 491                 lock.unlock();
 492                 capture->stop_dequeue_thread();
 493                 lock.lock();
 494                 card->parked_capture = move(card->capture);
 495                 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 496                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 497                 card->queue_length_policy.reset(card_index);
 498                 card->capture->start_bm_capture();
 499                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 500                 card->output->start_output(desired_output_video_mode, pts_int);
 501         }
 502         output_card_index = card_index;
 503 }
 504
 505 namespace {
 506
 507 int unwrap_timecode(uint16_t current_wrapped, int last)
 508 {
 509         uint16_t last_wrapped = last & 0xffff;
 510         if (current_wrapped > last_wrapped) {
 511                 return (last & ~0xffff) | current_wrapped;
 512         } else {
 513                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 514         }
 515 }
 516
 517 }  // namespace
 518
 519 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 520                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 521                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 522 {
 523         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 524         CaptureCard *card = &cards[card_index];
 525
 526         if (is_mode_scanning[card_index]) {
 527                 if (video_format.has_signal) {
 528                         // Found a stable signal, so stop scanning.
 529                         is_mode_scanning[card_index] = false;
 530                 } else {
 531                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 532                         steady_clock::time_point now = steady_clock::now();
 533                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 534                         if (sec_since_last_switch > switch_time_s) {
 535                                 // It isn't this mode; try the next one.
 536                                 mode_scanlist_index[card_index]++;
 537                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 538                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 539                                 last_mode_scan_change[card_index] = now;
 540                         }
 541                 }
 542         }
 543
 544         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 545         assert(frame_length > 0);
 546
 547         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;
 548         if (num_samples > OUTPUT_FREQUENCY / 10) {
 549                 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",
 550                         card_index, int(audio_frame.len), int(audio_offset),
 551                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 552                 if (video_frame.owner) {
 553                         video_frame.owner->release_frame(video_frame);
 554                 }
 555                 if (audio_frame.owner) {
 556                         audio_frame.owner->release_frame(audio_frame);
 557                 }
 558                 return;
 559         }
 560
 561         int dropped_frames = 0;
 562         if (card->last_timecode != -1) {
 563                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 564         }
 565
 566         // Number of samples per frame if we need to insert silence.
 567         // (Could be nonintegral, but resampling will save us then.)
 568         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 569
 570         if (dropped_frames > MAX_FPS * 2) {
 571                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 572                         card_index, card->last_timecode, timecode);
 573                 audio_mixer.reset_resampler(device);
 574                 dropped_frames = 0;
 575         } else if (dropped_frames > 0) {
 576                 // Insert silence as needed.
 577                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 578                         card_index, dropped_frames, timecode);
 579
 580                 bool success;
 581                 do {
 582                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 583                 } while (!success);
 584         }
 585
 586         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 587
 588         // Done with the audio, so release it.
 589         if (audio_frame.owner) {
 590                 audio_frame.owner->release_frame(audio_frame);
 591         }
 592
 593         card->last_timecode = timecode;
 594
 595         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 596
 597         size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
 598         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 599         if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
 600                 // The calculation above is wrong for planar Y'CbCr, so just override it.
 601                 assert(card->type == CardType::FFMPEG_INPUT);
 602                 assert(video_offset == 0);
 603                 expected_length = video_frame.len;
 604
 605                 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
 606                 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
 607                 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
 608                 cbcr_offset = video_format.width * video_format.height;
 609                 y_offset = 0;
 610         } else {
 611                 // All the other Y'CbCr formats are 4:2:2.
 612                 cbcr_width = video_format.width / 2;
 613                 cbcr_height = video_format.height;
 614                 cbcr_offset = video_offset / 2;
 615                 y_offset = video_frame.size / 2 + video_offset / 2;
 616         }
 617         if (video_frame.len - video_offset == 0 ||
 618             video_frame.len - video_offset != expected_length) {
 619                 if (video_frame.len != 0) {
 620                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 621                                 card_index, video_frame.len - video_offset, expected_length);
 622                 }
 623                 if (video_frame.owner) {
 624                         video_frame.owner->release_frame(video_frame);
 625                 }
 626
 627                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 628                 // so that pts can go up accordingly.
 629                 {
 630                         unique_lock<mutex> lock(card_mutex);
 631                         CaptureCard::NewFrame new_frame;
 632                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 633                         new_frame.length = frame_length;
 634                         new_frame.interlaced = false;
 635                         new_frame.dropped_frames = dropped_frames;
 636                         new_frame.received_timestamp = video_frame.received_timestamp;
 637                         card->new_frames.push_back(move(new_frame));
 638                         card->new_frames_changed.notify_all();
 639                 }
 640                 return;
 641         }
 642
 643         unsigned num_fields = video_format.interlaced ? 2 : 1;
 644         steady_clock::time_point frame_upload_start;
 645         bool interlaced_stride = false;
 646         if (video_format.interlaced) {
 647                 // Send the two fields along as separate frames; the other side will need to add
 648                 // a deinterlacer to actually get this right.
 649                 assert(video_format.height % 2 == 0);
 650                 video_format.height /= 2;
 651                 cbcr_height /= 2;
 652                 assert(frame_length % 2 == 0);
 653                 frame_length /= 2;
 654                 num_fields = 2;
 655                 if (video_format.second_field_start == 1) {
 656                         interlaced_stride = true;
 657                 }
 658                 frame_upload_start = steady_clock::now();
 659         }
 660         userdata->last_interlaced = video_format.interlaced;
 661         userdata->last_has_signal = video_format.has_signal;
 662         userdata->last_is_connected = video_format.is_connected;
 663         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 664         userdata->last_frame_rate_den = video_format.frame_rate_den;
 665         RefCountedFrame frame(video_frame);
 666
 667         // Upload the textures.
 668         for (unsigned field = 0; field < num_fields; ++field) {
 669                 // Put the actual texture upload in a lambda that is executed in the main thread.
 670                 // It is entirely possible to do this in the same thread (and it might even be
 671                 // faster, depending on the GPU and driver), but it appears to be trickling
 672                 // driver bugs very easily.
 673                 //
 674                 // Note that this means we must hold on to the actual frame data in <userdata>
 675                 // until the upload command is run, but we hold on to <frame> much longer than that
 676                 // (in fact, all the way until we no longer use the texture in rendering).
 677                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
 678                         unsigned field_start_line;
 679                         if (field == 1) {
 680                                 field_start_line = video_format.second_field_start;
 681                         } else {
 682                                 field_start_line = video_format.extra_lines_top;
 683                         }
 684
 685                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 686                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 687                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
 688
 689                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 690                         check_error();
 691
 692                         switch (userdata->pixel_format) {
 693                         case PixelFormat_10BitYCbCr: {
 694                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 695                                 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);
 696                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 697                                 break;
 698                         }
 699                         case PixelFormat_8BitYCbCr: {
 700                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 701                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 702
 703                                 // Make up our own strides, since we are interleaving.
 704                                 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);
 705                                 upload_texture(userdata->tex_cbcr[field], cbcr_width, cbcr_height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
 706                                 break;
 707                         }
 708                         case PixelFormat_8BitYCbCrPlanar: {
 709                                 assert(field_start_line == 0);  // We don't really support interlaced here.
 710                                 size_t field_y_start = y_offset;
 711                                 size_t field_cb_start = cbcr_offset;
 712                                 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
 713
 714                                 // Make up our own strides, since we are interleaving.
 715                                 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);
 716                                 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
 717                                 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
 718                                 break;
 719                         }
 720                         case PixelFormat_8BitBGRA: {
 721                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 722                                 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);
 723                                 // These could be asked to deliver mipmaps at any time.
 724                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 725                                 check_error();
 726                                 glGenerateMipmap(GL_TEXTURE_2D);
 727                                 check_error();
 728                                 glBindTexture(GL_TEXTURE_2D, 0);
 729                                 check_error();
 730                                 break;
 731                         }
 732                         default:
 733                                 assert(false);
 734                         }
 735
 736                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 737                         check_error();
 738                 };
 739
 740                 if (field == 1) {
 741                         // Don't upload the second field as fast as we can; wait until
 742                         // the field time has approximately passed. (Otherwise, we could
 743                         // get timing jitter against the other sources, and possibly also
 744                         // against the video display, although the latter is not as critical.)
 745                         // This requires our system clock to be reasonably close to the
 746                         // video clock, but that's not an unreasonable assumption.
 747                         steady_clock::time_point second_field_start = frame_upload_start +
 748                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 749                         this_thread::sleep_until(second_field_start);
 750                 }
 751
 752                 {
 753                         unique_lock<mutex> lock(card_mutex);
 754                         CaptureCard::NewFrame new_frame;
 755                         new_frame.frame = frame;
 756                         new_frame.length = frame_length;
 757                         new_frame.field = field;
 758                         new_frame.interlaced = video_format.interlaced;
 759                         new_frame.upload_func = upload_func;
 760                         new_frame.dropped_frames = dropped_frames;
 761                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 762                         card->new_frames.push_back(move(new_frame));
 763                         card->new_frames_changed.notify_all();
 764                 }
 765         }
 766 }
 767
 768 void Mixer::bm_hotplug_add(libusb_device *dev)
 769 {
 770         lock_guard<mutex> lock(hotplug_mutex);
 771         hotplugged_cards.push_back(dev);
 772 }
 773
 774 void Mixer::bm_hotplug_remove(unsigned card_index)
 775 {
 776         cards[card_index].new_frames_changed.notify_all();
 777 }
 778
 779 void Mixer::thread_func()
 780 {
 781         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 782
 783         eglBindAPI(EGL_OPENGL_API);
 784         QOpenGLContext *context = create_context(mixer_surface);
 785         if (!make_current(context, mixer_surface)) {
 786                 printf("oops\n");
 787                 exit(1);
 788         }
 789
 790         // Start the actual capture. (We don't want to do it before we're actually ready
 791         // to process output frames.)
 792         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 793                 if (int(card_index) != output_card_index) {
 794                         cards[card_index].capture->start_bm_capture();
 795                 }
 796         }
 797
 798         steady_clock::time_point start, now;
 799         start = steady_clock::now();
 800
 801         int stats_dropped_frames = 0;
 802
 803         while (!should_quit) {
 804                 if (desired_output_card_index != output_card_index) {
 805                         set_output_card_internal(desired_output_card_index);
 806                 }
 807                 if (output_card_index != -1 &&
 808                     desired_output_video_mode != output_video_mode) {
 809                         DeckLinkOutput *output = cards[output_card_index].output.get();
 810                         output->end_output();
 811                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 812                         output->start_output(desired_output_video_mode, pts_int);
 813                 }
 814
 815                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 816                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 817
 818                 bool master_card_is_output;
 819                 unsigned master_card_index;
 820                 if (output_card_index != -1) {
 821                         master_card_is_output = true;
 822                         master_card_index = output_card_index;
 823                 } else {
 824                         master_card_is_output = false;
 825                         master_card_index = theme->map_signal(master_clock_channel);
 826                         assert(master_card_index < num_cards);
 827                 }
 828
 829                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 830                 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);
 831                 stats_dropped_frames += output_frame_info.dropped_frames;
 832
 833                 handle_hotplugged_cards();
 834
 835                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 836                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 837                                 continue;
 838                         }
 839                         if (new_frames[card_index].frame->len == 0) {
 840                                 ++new_frames[card_index].dropped_frames;
 841                         }
 842                         if (new_frames[card_index].dropped_frames > 0) {
 843                                 printf("Card %u dropped %d frames before this\n",
 844                                         card_index, int(new_frames[card_index].dropped_frames));
 845                         }
 846                 }
 847
 848                 // If the first card is reporting a corrupted or otherwise dropped frame,
 849                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 850                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 851                         ++stats_dropped_frames;
 852                         pts_int += new_frames[master_card_index].length;
 853                         continue;
 854                 }
 855
 856                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 857                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 858                                 continue;
 859
 860                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 861                         assert(new_frame->frame != nullptr);
 862                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 863                         check_error();
 864
 865                         // The new texture might need uploading before use.
 866                         if (new_frame->upload_func) {
 867                                 new_frame->upload_func();
 868                                 new_frame->upload_func = nullptr;
 869                         }
 870                 }
 871
 872                 int64_t frame_duration = output_frame_info.frame_duration;
 873                 render_one_frame(frame_duration);
 874                 ++frame_num;
 875                 pts_int += frame_duration;
 876
 877                 now = steady_clock::now();
 878                 double elapsed = duration<double>(now - start).count();
 879
 880                 metrics_num_frames = frame_num;
 881                 metrics_dropped_frames = stats_dropped_frames;
 882                 metrics_uptime = elapsed;
 883
 884                 if (frame_num % 100 == 0) {
 885                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 886                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 887                                 1e3 * elapsed / frame_num);
 888                 //      chain->print_phase_timing();
 889
 890                         // Check our memory usage, to see if we are close to our mlockall()
 891                         // limit (if at all set).
 892                         rusage used;
 893                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 894                                 perror("getrusage(RUSAGE_SELF)");
 895                                 assert(false);
 896                         }
 897
 898                         if (uses_mlock) {
 899                                 rlimit limit;
 900                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 901                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 902                                         assert(false);
 903                                 }
 904
 905                                 if (limit.rlim_cur == 0) {
 906                                         printf(", using %ld MB memory (locked)",
 907                                                 long(used.ru_maxrss / 1024));
 908                                 } else {
 909                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 910                                                 long(used.ru_maxrss / 1024),
 911                                                 long(limit.rlim_cur / 1048576),
 912                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 913                                 }
 914                         } else {
 915                                 printf(", using %ld MB memory (not locked)",
 916                                         long(used.ru_maxrss / 1024));
 917                         }
 918
 919                         printf("\n");
 920                 }
 921
 922
 923                 if (should_cut.exchange(false)) {  // Test and clear.
 924                         video_encoder->do_cut(frame_num);
 925                 }
 926
 927 #if 0
 928                 // Reset every 100 frames, so that local variations in frame times
 929                 // (especially for the first few frames, when the shaders are
 930                 // compiled etc.) don't make it hard to measure for the entire
 931                 // remaining duration of the program.
 932                 if (frame == 10000) {
 933                         frame = 0;
 934                         start = now;
 935                 }
 936 #endif
 937                 check_error();
 938         }
 939
 940         resource_pool->clean_context();
 941 }
 942
 943 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 944 {
 945         if (output_card_index != -1) {
 946                 // The output card (ie., cards[output_card_index].output) is the master clock,
 947                 // so no input card (ie., cards[card_index].capture) is.
 948                 return false;
 949         }
 950         return (card_index == master_card_index);
 951 }
 952
 953 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 954 {
 955         // Count the number of frames in the queue, including any frames
 956         // we dropped. It's hard to know exactly how we should deal with
 957         // dropped (corrupted) input frames; they don't help our goal of
 958         // avoiding starvation, but they still add to the problem of latency.
 959         // Since dropped frames is going to mean a bump in the signal anyway,
 960         // we err on the side of having more stable latency instead.
 961         unsigned queue_length = 0;
 962         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 963                 queue_length += frame.dropped_frames + 1;
 964         }
 965         card->queue_length_policy.update_policy(queue_length);
 966
 967         // If needed, drop frames until the queue is below the safe limit.
 968         // We prefer to drop from the head, because all else being equal,
 969         // we'd like more recent frames (less latency).
 970         unsigned dropped_frames = 0;
 971         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 972                 assert(!card->new_frames.empty());
 973                 assert(queue_length > card->new_frames.front().dropped_frames);
 974                 queue_length -= card->new_frames.front().dropped_frames;
 975
 976                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 977                         // No need to drop anything.
 978                         break;
 979                 }
 980
 981                 card->new_frames.pop_front();
 982                 card->new_frames_changed.notify_all();
 983                 --queue_length;
 984                 ++dropped_frames;
 985         }
 986
 987 #if 0
 988         if (dropped_frames > 0) {
 989                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 990                         card_index, dropped_frames);
 991         }
 992 #endif
 993 }
 994
 995
 996 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])
 997 {
 998         OutputFrameInfo output_frame_info;
 999 start:
1000         unique_lock<mutex> lock(card_mutex, defer_lock);
1001         if (master_card_is_output) {
1002                 // Clocked to the output, so wait for it to be ready for the next frame.
1003                 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);
1004                 lock.lock();
1005         } else {
1006                 // Wait for the master card to have a new frame.
1007                 // TODO: Add a timeout.
1008                 output_frame_info.is_preroll = false;
1009                 lock.lock();
1010                 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(); });
1011         }
1012
1013         if (master_card_is_output) {
1014                 handle_hotplugged_cards();
1015         } else if (cards[master_card_index].new_frames.empty()) {
1016                 // We were woken up, but not due to a new frame. Deal with it
1017                 // and then restart.
1018                 assert(cards[master_card_index].capture->get_disconnected());
1019                 handle_hotplugged_cards();
1020                 goto start;
1021         }
1022
1023         if (!master_card_is_output) {
1024                 output_frame_info.frame_timestamp =
1025                         cards[master_card_index].new_frames.front().received_timestamp;
1026         }
1027
1028         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1029                 CaptureCard *card = &cards[card_index];
1030                 if (input_card_is_master_clock(card_index, master_card_index)) {
1031                         // We don't use the queue length policy for the master card,
1032                         // but we will if it stops being the master. Thus, clear out
1033                         // the policy in case we switch in the future.
1034                         card->queue_length_policy.reset(card_index);
1035                         assert(!card->new_frames.empty());
1036                 } else {
1037                         trim_queue(card, card_index);
1038                 }
1039                 if (!card->new_frames.empty()) {
1040                         new_frames[card_index] = move(card->new_frames.front());
1041                         has_new_frame[card_index] = true;
1042                         card->new_frames.pop_front();
1043                         card->new_frames_changed.notify_all();
1044                 }
1045         }
1046
1047         if (!master_card_is_output) {
1048                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1049                 output_frame_info.frame_duration = new_frames[master_card_index].length;
1050         }
1051
1052         // This might get off by a fractional sample when changing master card
1053         // between ones with different frame rates, but that's fine.
1054         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1055         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1056         fractional_samples = num_samples_times_timebase % TIMEBASE;
1057         assert(output_frame_info.num_samples >= 0);
1058
1059         return output_frame_info;
1060 }
1061
1062 void Mixer::handle_hotplugged_cards()
1063 {
1064         // Check for cards that have been disconnected since last frame.
1065         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1066                 CaptureCard *card = &cards[card_index];
1067                 if (card->capture->get_disconnected()) {
1068                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1069                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1070                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1071                         card->queue_length_policy.reset(card_index);
1072                         card->capture->start_bm_capture();
1073                 }
1074         }
1075
1076         // Check for cards that have been connected since last frame.
1077         vector<libusb_device *> hotplugged_cards_copy;
1078         {
1079                 lock_guard<mutex> lock(hotplug_mutex);
1080                 swap(hotplugged_cards, hotplugged_cards_copy);
1081         }
1082         for (libusb_device *new_dev : hotplugged_cards_copy) {
1083                 // Look for a fake capture card where we can stick this in.
1084                 int free_card_index = -1;
1085                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1086                         if (cards[card_index].is_fake_capture) {
1087                                 free_card_index = card_index;
1088                                 break;
1089                         }
1090                 }
1091
1092                 if (free_card_index == -1) {
1093                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1094                         libusb_unref_device(new_dev);
1095                 } else {
1096                         // BMUSBCapture takes ownership.
1097                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1098                         CaptureCard *card = &cards[free_card_index];
1099                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1100                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1101                         card->queue_length_policy.reset(free_card_index);
1102                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1103                         capture->start_bm_capture();
1104                 }
1105         }
1106 }
1107
1108
1109 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)
1110 {
1111         // Resample the audio as needed, including from previously dropped frames.
1112         assert(num_cards > 0);
1113         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1114                 const bool dropped_frame = (frame_num != dropped_frames);
1115                 {
1116                         // Signal to the audio thread to process this frame.
1117                         // Note that if the frame is a dropped frame, we signal that
1118                         // we don't want to use this frame as base for adjusting
1119                         // the resampler rate. The reason for this is that the timing
1120                         // of these frames is often way too late; they typically don't
1121                         // “arrive” before we synthesize them. Thus, we could end up
1122                         // in a situation where we have inserted e.g. five audio frames
1123                         // into the queue before we then start pulling five of them
1124                         // back out. This makes ResamplingQueue overestimate the delay,
1125                         // causing undue resampler changes. (We _do_ use the last,
1126                         // non-dropped frame; perhaps we should just discard that as well,
1127                         // since dropped frames are expected to be rare, and it might be
1128                         // better to just wait until we have a slightly more normal situation).
1129                         unique_lock<mutex> lock(audio_mutex);
1130                         bool adjust_rate = !dropped_frame && !is_preroll;
1131                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1132                         audio_task_queue_changed.notify_one();
1133                 }
1134                 if (dropped_frame) {
1135                         // For dropped frames, increase the pts. Note that if the format changed
1136                         // in the meantime, we have no way of detecting that; we just have to
1137                         // assume the frame length is always the same.
1138                         pts_int += length_per_frame;
1139                 }
1140         }
1141 }
1142
1143 void Mixer::render_one_frame(int64_t duration)
1144 {
1145         // Determine the time code for this frame before we start rendering.
1146         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1147         if (display_timecode_on_stdout) {
1148                 printf("Timecode: '%s'\n", timecode_text.c_str());
1149         }
1150
1151         // Update Y'CbCr settings for all cards.
1152         {
1153                 unique_lock<mutex> lock(card_mutex);
1154                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1155                         YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1156                         input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1157                         input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1158                         input_state.full_range[card_index] = interpretation->full_range;
1159                 }
1160         }
1161
1162         // Get the main chain from the theme, and set its state immediately.
1163         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1164         EffectChain *chain = theme_main_chain.chain;
1165         theme_main_chain.setup_chain();
1166         //theme_main_chain.chain->enable_phase_timing(true);
1167
1168         // The theme can't (or at least shouldn't!) call connect_signal() on
1169         // each FFmpeg input, so we'll do it here.
1170         for (const pair<LiveInputWrapper *, FFmpegCapture *> &conn : theme->get_signal_connections()) {
1171                 conn.first->connect_signal_raw(conn.second->get_card_index());
1172         }
1173
1174         // If HDMI/SDI output is active and the user has requested auto mode,
1175         // its mode overrides the existing Y'CbCr setting for the chain.
1176         YCbCrLumaCoefficients ycbcr_output_coefficients;
1177         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1178                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1179         } else {
1180                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1181         }
1182
1183         // TODO: Reduce the duplication against theme.cpp.
1184         YCbCrFormat output_ycbcr_format;
1185         output_ycbcr_format.chroma_subsampling_x = 1;
1186         output_ycbcr_format.chroma_subsampling_y = 1;
1187         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1188         output_ycbcr_format.full_range = false;
1189         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1190         chain->change_ycbcr_output_format(output_ycbcr_format);
1191
1192         // Render main chain. If we're using zerocopy Quick Sync encoding
1193         // (the default case), we take an extra copy of the created outputs,
1194         // so that we can display it back to the screen later (it's less memory
1195         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1196         // Ideally, we'd like to avoid taking copies and just use the main textures
1197         // for display as well, but they're just views into VA-API memory and must be
1198         // unmapped during encoding, so we can't use them for display, unfortunately.
1199         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1200         GLuint y_copy_tex, cbcr_copy_tex = 0;
1201         GLuint y_display_tex, cbcr_display_tex;
1202         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1203         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1204         const bool is_zerocopy = video_encoder->is_zerocopy();
1205         if (is_zerocopy) {
1206                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1207                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1208                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1209
1210                 y_display_tex = y_copy_tex;
1211                 cbcr_display_tex = cbcr_copy_tex;
1212
1213                 // y_tex and cbcr_tex will be given by VideoEncoder.
1214         } else {
1215                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1216                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1217                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1218
1219                 y_display_tex = y_tex;
1220                 cbcr_display_tex = cbcr_tex;
1221         }
1222
1223         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1224         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1225         assert(got_frame);
1226
1227         GLuint fbo;
1228         if (is_zerocopy) {
1229                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1230         } else {
1231                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1232         }
1233         check_error();
1234         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1235
1236         if (display_timecode_in_stream) {
1237                 // Render the timecode on top.
1238                 timecode_renderer->render_timecode(fbo, timecode_text);
1239         }
1240
1241         resource_pool->release_fbo(fbo);
1242
1243         if (is_zerocopy) {
1244                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1245         } else {
1246                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1247         }
1248         if (output_card_index != -1) {
1249                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1250         }
1251         resource_pool->release_2d_texture(cbcr_full_tex);
1252
1253         // Set the right state for the Y' and CbCr textures we use for display.
1254         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1255         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1256         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1257         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1258         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1259
1260         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1261         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1262         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1263         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1264
1265         RefCountedGLsync fence = video_encoder->end_frame();
1266
1267         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1268         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1269         DisplayFrame live_frame;
1270         live_frame.chain = display_chain.get();
1271         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1272                 display_input->set_texture_num(0, y_display_tex);
1273                 display_input->set_texture_num(1, cbcr_display_tex);
1274         };
1275         live_frame.ready_fence = fence;
1276         live_frame.input_frames = {};
1277         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1278         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1279
1280         // Set up preview and any additional channels.
1281         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1282                 DisplayFrame display_frame;
1283                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1284                 display_frame.chain = chain.chain;
1285                 display_frame.setup_chain = chain.setup_chain;
1286                 display_frame.ready_fence = fence;
1287                 display_frame.input_frames = chain.input_frames;
1288                 display_frame.temp_textures = {};
1289                 output_channel[i].output_frame(display_frame);
1290         }
1291 }
1292
1293 void Mixer::audio_thread_func()
1294 {
1295         pthread_setname_np(pthread_self(), "Mixer_Audio");
1296
1297         while (!should_quit) {
1298                 AudioTask task;
1299
1300                 {
1301                         unique_lock<mutex> lock(audio_mutex);
1302                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1303                         if (should_quit) {
1304                                 return;
1305                         }
1306                         task = audio_task_queue.front();
1307                         audio_task_queue.pop();
1308                 }
1309
1310                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1311                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1312                 vector<float> samples_out = audio_mixer.get_output(
1313                         task.frame_timestamp,
1314                         task.num_samples,
1315                         rate_adjustment_policy);
1316
1317                 // Send the samples to the sound card, then add them to the output.
1318                 if (alsa) {
1319                         alsa->write(samples_out);
1320                 }
1321                 if (output_card_index != -1) {
1322                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1323                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1324                 }
1325                 video_encoder->add_audio(task.pts_int, move(samples_out));
1326         }
1327 }
1328
1329 void Mixer::release_display_frame(DisplayFrame *frame)
1330 {
1331         for (GLuint texnum : frame->temp_textures) {
1332                 resource_pool->release_2d_texture(texnum);
1333         }
1334         frame->temp_textures.clear();
1335         frame->ready_fence.reset();
1336         frame->input_frames.clear();
1337 }
1338
1339 void Mixer::start()
1340 {
1341         mixer_thread = thread(&Mixer::thread_func, this);
1342         audio_thread = thread(&Mixer::audio_thread_func, this);
1343 }
1344
1345 void Mixer::quit()
1346 {
1347         should_quit = true;
1348         audio_task_queue_changed.notify_one();
1349         mixer_thread.join();
1350         audio_thread.join();
1351 }
1352
1353 void Mixer::transition_clicked(int transition_num)
1354 {
1355         theme->transition_clicked(transition_num, pts());
1356 }
1357
1358 void Mixer::channel_clicked(int preview_num)
1359 {
1360         theme->channel_clicked(preview_num);
1361 }
1362
1363 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1364 {
1365         unique_lock<mutex> lock(card_mutex);
1366         return ycbcr_interpretation[card_index];
1367 }
1368
1369 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1370 {
1371         unique_lock<mutex> lock(card_mutex);
1372         ycbcr_interpretation[card_index] = interpretation;
1373 }
1374
1375 void Mixer::start_mode_scanning(unsigned card_index)
1376 {
1377         assert(card_index < num_cards);
1378         if (is_mode_scanning[card_index]) {
1379                 return;
1380         }
1381         is_mode_scanning[card_index] = true;
1382         mode_scanlist[card_index].clear();
1383         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1384                 mode_scanlist[card_index].push_back(mode.first);
1385         }
1386         assert(!mode_scanlist[card_index].empty());
1387         mode_scanlist_index[card_index] = 0;
1388         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1389         last_mode_scan_change[card_index] = steady_clock::now();
1390 }
1391
1392 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1393 {
1394         assert(desired_output_card_index != -1);
1395         unique_lock<mutex> lock(card_mutex);
1396         return cards[desired_output_card_index].output->get_available_video_modes();
1397 }
1398
1399 Mixer::OutputChannel::~OutputChannel()
1400 {
1401         if (has_current_frame) {
1402                 parent->release_display_frame(&current_frame);
1403         }
1404         if (has_ready_frame) {
1405                 parent->release_display_frame(&ready_frame);
1406         }
1407 }
1408
1409 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1410 {
1411         // Store this frame for display. Remove the ready frame if any
1412         // (it was seemingly never used).
1413         {
1414                 unique_lock<mutex> lock(frame_mutex);
1415                 if (has_ready_frame) {
1416                         parent->release_display_frame(&ready_frame);
1417                 }
1418                 ready_frame = frame;
1419                 has_ready_frame = true;
1420
1421                 // Call the callbacks under the mutex (they should be short),
1422                 // so that we don't race against a callback removal.
1423                 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1424                         key_and_callback.second();
1425                 }
1426         }
1427
1428         // Reduce the number of callbacks by filtering duplicates. The reason
1429         // why we bother doing this is that Qt seemingly can get into a state
1430         // where its builds up an essentially unbounded queue of signals,
1431         // consuming more and more memory, and there's no good way of collapsing
1432         // user-defined signals or limiting the length of the queue.
1433         if (transition_names_updated_callback) {
1434                 vector<string> transition_names = global_mixer->get_transition_names();
1435                 bool changed = false;
1436                 if (transition_names.size() != last_transition_names.size()) {
1437                         changed = true;
1438                 } else {
1439                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1440                                 if (transition_names[i] != last_transition_names[i]) {
1441                                         changed = true;
1442                                         break;
1443                                 }
1444                         }
1445                 }
1446                 if (changed) {
1447                         transition_names_updated_callback(transition_names);
1448                         last_transition_names = transition_names;
1449                 }
1450         }
1451         if (name_updated_callback) {
1452                 string name = global_mixer->get_channel_name(channel);
1453                 if (name != last_name) {
1454                         name_updated_callback(name);
1455                         last_name = name;
1456                 }
1457         }
1458         if (color_updated_callback) {
1459                 string color = global_mixer->get_channel_color(channel);
1460                 if (color != last_color) {
1461                         color_updated_callback(color);
1462                         last_color = color;
1463                 }
1464         }
1465 }
1466
1467 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1468 {
1469         unique_lock<mutex> lock(frame_mutex);
1470         if (!has_current_frame && !has_ready_frame) {
1471                 return false;
1472         }
1473
1474         if (has_current_frame && has_ready_frame) {
1475                 // We have a new ready frame. Toss the current one.
1476                 parent->release_display_frame(&current_frame);
1477                 has_current_frame = false;
1478         }
1479         if (has_ready_frame) {
1480                 assert(!has_current_frame);
1481                 current_frame = ready_frame;
1482                 ready_frame.ready_fence.reset();  // Drop the refcount.
1483                 ready_frame.input_frames.clear();  // Drop the refcounts.
1484                 has_current_frame = true;
1485                 has_ready_frame = false;
1486         }
1487
1488         *frame = current_frame;
1489         return true;
1490 }
1491
1492 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1493 {
1494         unique_lock<mutex> lock(frame_mutex);
1495         new_frame_ready_callbacks[key] = callback;
1496 }
1497
1498 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1499 {
1500         unique_lock<mutex> lock(frame_mutex);
1501         new_frame_ready_callbacks.erase(key);
1502 }
1503
1504 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1505 {
1506         transition_names_updated_callback = callback;
1507 }
1508
1509 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1510 {
1511         name_updated_callback = callback;
1512 }
1513
1514 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1515 {
1516         color_updated_callback = callback;
1517 }
1518
1519 mutex RefCountedGLsync::fence_lock;