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