git.sesse.net Git - nageru/blob - mixer.cpp

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   7 #include <movit/effect_chain.h>
   8 #include <movit/effect_util.h>
   9 #include <movit/flat_input.h>
  10 #include <movit/image_format.h>
  11 #include <movit/init.h>
  12 #include <movit/resource_pool.h>
  13 #include <pthread.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <sys/resource.h>
  18 #include <algorithm>
  19 #include <chrono>
  20 #include <condition_variable>
  21 #include <cstddef>
  22 #include <cstdint>
  23 #include <memory>
  24 #include <mutex>
  25 #include <ratio>
  26 #include <string>
  27 #include <thread>
  28 #include <utility>
  29 #include <vector>
  30
  31 #include "DeckLinkAPI.h"
  32 #include "LinuxCOM.h"
  33 #include "alsa_output.h"
  34 #include "bmusb/bmusb.h"
  35 #include "bmusb/fake_capture.h"
  36 #include "chroma_subsampler.h"
  37 #include "context.h"
  38 #include "decklink_capture.h"
  39 #include "decklink_output.h"
  40 #include "defs.h"
  41 #include "disk_space_estimator.h"
  42 #include "ffmpeg_capture.h"
  43 #include "flags.h"
  44 #include "input_mapping.h"
  45 #include "pbo_frame_allocator.h"
  46 #include "ref_counted_gl_sync.h"
  47 #include "resampling_queue.h"
  48 #include "timebase.h"
  49 #include "timecode_renderer.h"
  50 #include "v210_converter.h"
  51 #include "video_encoder.h"
  52
  53 class IDeckLink;
  54 class QOpenGLContext;
  55
  56 using namespace movit;
  57 using namespace std;
  58 using namespace std::chrono;
  59 using namespace std::placeholders;
  60 using namespace bmusb;
  61
  62 Mixer *global_mixer = nullptr;
  63 bool uses_mlock = false;
  64
  65 namespace {
  66
  67 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  68 {
  69         if (interlaced) {
  70                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  71                         input_state->buffered_frames[card_index][frame_num] =
  72                                 input_state->buffered_frames[card_index][frame_num - 1];
  73                 }
  74                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  75         } else {
  76                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  77                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  78                 }
  79         }
  80 }
  81
  82 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned v210_width)
  83 {
  84         bool first;
  85         switch (userdata->pixel_format) {
  86         case bmusb::PixelFormat_10BitYCbCr:
  87                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  88                 break;
  89         case bmusb::PixelFormat_8BitYCbCr:
  90                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
  91                 break;
  92         case bmusb::PixelFormat_8BitBGRA:
  93                 first = userdata->tex_rgba[field] == 0;
  94                 break;
  95         default:
  96                 assert(false);
  97         }
  98
  99         if (first ||
 100             width != userdata->last_width[field] ||
 101             height != userdata->last_height[field]) {
 102                 // We changed resolution since last use of this texture, so we need to create
 103                 // a new object. Note that this each card has its own PBOFrameAllocator,
 104                 // we don't need to worry about these flip-flopping between resolutions.
 105                 switch (userdata->pixel_format) {
 106                 case bmusb::PixelFormat_10BitYCbCr:
 107                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 108                         check_error();
 109                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 110                         check_error();
 111                         break;
 112                 case bmusb::PixelFormat_8BitYCbCr: {
 113                         size_t cbcr_width = width / 2;
 114
 115                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 116                         check_error();
 117                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 118                         check_error();
 119                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 120                         check_error();
 121                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 122                         check_error();
 123                         break;
 124                 }
 125                 case bmusb::PixelFormat_8BitBGRA:
 126                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 127                         check_error();
 128                         if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
 129                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 130                         } else {
 131                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 132                         }
 133                         check_error();
 134                         break;
 135                 }
 136                 userdata->last_width[field] = width;
 137                 userdata->last_height[field] = height;
 138         }
 139         if (global_flags.ten_bit_input &&
 140             (first || v210_width != userdata->last_v210_width[field])) {
 141                 // Same as above; we need to recreate the texture.
 142                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 143                 check_error();
 144                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 145                 check_error();
 146                 userdata->last_v210_width[field] = v210_width;
 147         }
 148 }
 149
 150 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 151 {
 152         if (interlaced_stride) {
 153                 stride *= 2;
 154         }
 155         if (global_flags.flush_pbos) {
 156                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 157                 check_error();
 158         }
 159
 160         glBindTexture(GL_TEXTURE_2D, tex);
 161         check_error();
 162         if (interlaced_stride) {
 163                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 164                 check_error();
 165         } else {
 166                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 167                 check_error();
 168         }
 169
 170         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 171         check_error();
 172         glBindTexture(GL_TEXTURE_2D, 0);
 173         check_error();
 174         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 175         check_error();
 176 }
 177
 178 }  // namespace
 179
 180 void QueueLengthPolicy::update_policy(unsigned queue_length)
 181 {
 182         if (queue_length == 0) {  // Starvation.
 183                 if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
 184                         ++safe_queue_length;
 185                         fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
 186                                 card_index, safe_queue_length);
 187                 }
 188                 frames_with_at_least_one = 0;
 189                 been_at_safe_point_since_last_starvation = false;
 190                 return;
 191         }
 192         if (queue_length >= safe_queue_length) {
 193                 been_at_safe_point_since_last_starvation = true;
 194         }
 195         if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
 196                 --safe_queue_length;
 197                 fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
 198                         card_index, safe_queue_length);
 199                 frames_with_at_least_one = 0;
 200         }
 201 }
 202
 203 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 204         : httpd(),
 205           num_cards(num_cards),
 206           mixer_surface(create_surface(format)),
 207           h264_encoder_surface(create_surface(format)),
 208           decklink_output_surface(create_surface(format)),
 209           audio_mixer(num_cards)
 210 {
 211         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 212         check_error();
 213
 214         // This nearly always should be true.
 215         global_flags.can_disable_srgb_decoder = epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode");
 216
 217         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 218         // will be halved when sampling them, and we need to compensate here.
 219         movit_texel_subpixel_precision /= 2.0;
 220
 221         resource_pool.reset(new ResourcePool);
 222         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 223                 output_channel[i].parent = this;
 224                 output_channel[i].channel = i;
 225         }
 226
 227         ImageFormat inout_format;
 228         inout_format.color_space = COLORSPACE_sRGB;
 229         inout_format.gamma_curve = GAMMA_sRGB;
 230
 231         // Matches the 4:2:0 format created by the main chain.
 232         YCbCrFormat ycbcr_format;
 233         ycbcr_format.chroma_subsampling_x = 2;
 234         ycbcr_format.chroma_subsampling_y = 2;
 235         if (global_flags.ycbcr_rec709_coefficients) {
 236                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 237         } else {
 238                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 239         }
 240         ycbcr_format.full_range = false;
 241         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 242         ycbcr_format.cb_x_position = 0.0f;
 243         ycbcr_format.cr_x_position = 0.0f;
 244         ycbcr_format.cb_y_position = 0.5f;
 245         ycbcr_format.cr_y_position = 0.5f;
 246
 247         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 248         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 249         check_error();
 250         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 251         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 252         display_chain->add_input(display_input);
 253         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 254         display_chain->set_dither_bits(0);  // Don't bother.
 255         display_chain->finalize();
 256
 257         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));
 258
 259         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 260         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 261
 262         // Start listening for clients only once VideoEncoder has written its header, if any.
 263         httpd.start(9095);
 264
 265         // First try initializing the then PCI devices, then USB, then
 266         // fill up with fake cards until we have the desired number of cards.
 267         unsigned num_pci_devices = 0;
 268         unsigned card_index = 0;
 269
 270         {
 271                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 272                 if (decklink_iterator != nullptr) {
 273                         for ( ; card_index < num_cards; ++card_index) {
 274                                 IDeckLink *decklink;
 275                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 276                                         break;
 277                                 }
 278
 279                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 280                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 281                                 output->set_device(decklink);
 282                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 283                                 ++num_pci_devices;
 284                         }
 285                         decklink_iterator->Release();
 286                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 287                 } else {
 288                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 289                 }
 290         }
 291
 292         unsigned num_usb_devices = BMUSBCapture::num_cards();
 293         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 294                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 295                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 296                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 297         }
 298         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 299
 300         unsigned num_fake_cards = 0;
 301         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 302                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 303                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 304         }
 305
 306         if (num_fake_cards > 0) {
 307                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 308         }
 309
 310         // Initialize all video inputs the theme asked for. Note that these are
 311         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 312         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 313         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 314                 if (card_index >= MAX_VIDEO_CARDS) {
 315                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 316                         exit(1);
 317                 }
 318                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 319                 video_inputs[video_card_index]->set_card_index(card_index);
 320         }
 321         num_video_inputs = video_inputs.size();
 322
 323         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 324         BMUSBCapture::start_bm_thread();
 325
 326         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 327                 cards[card_index].queue_length_policy.reset(card_index);
 328         }
 329
 330         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 331
 332         if (global_flags.ten_bit_input) {
 333                 if (!v210Converter::has_hardware_support()) {
 334                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 335                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 336                         exit(1);
 337                 }
 338                 v210_converter.reset(new v210Converter());
 339
 340                 // These are all the widths listed in the Blackmagic SDK documentation
 341                 // (section 2.7.3, “Display Modes”).
 342                 v210_converter->precompile_shader(720);
 343                 v210_converter->precompile_shader(1280);
 344                 v210_converter->precompile_shader(1920);
 345                 v210_converter->precompile_shader(2048);
 346                 v210_converter->precompile_shader(3840);
 347                 v210_converter->precompile_shader(4096);
 348         }
 349         if (global_flags.ten_bit_output) {
 350                 if (!v210Converter::has_hardware_support()) {
 351                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 352                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 353                         exit(1);
 354                 }
 355         }
 356
 357         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 358         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 359         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 360
 361         if (global_flags.enable_alsa_output) {
 362                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 363         }
 364         if (global_flags.output_card != -1) {
 365                 desired_output_card_index = global_flags.output_card;
 366                 set_output_card_internal(global_flags.output_card);
 367         }
 368 }
 369
 370 Mixer::~Mixer()
 371 {
 372         BMUSBCapture::stop_bm_thread();
 373
 374         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 375                 {
 376                         unique_lock<mutex> lock(card_mutex);
 377                         cards[card_index].should_quit = true;  // Unblock thread.
 378                         cards[card_index].new_frames_changed.notify_all();
 379                 }
 380                 cards[card_index].capture->stop_dequeue_thread();
 381                 if (cards[card_index].output) {
 382                         cards[card_index].output->end_output();
 383                         cards[card_index].output.reset();
 384                 }
 385         }
 386
 387         video_encoder.reset(nullptr);
 388 }
 389
 390 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 391 {
 392         printf("Configuring card %d...\n", card_index);
 393
 394         CaptureCard *card = &cards[card_index];
 395         if (card->capture != nullptr) {
 396                 card->capture->stop_dequeue_thread();
 397         }
 398         card->capture.reset(capture);
 399         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 400         if (card->output.get() != output) {
 401                 card->output.reset(output);
 402         }
 403
 404         bmusb::PixelFormat pixel_format;
 405         if (card_type == CardType::FFMPEG_INPUT) {
 406                 pixel_format = bmusb::PixelFormat_8BitBGRA;
 407         } else if (global_flags.ten_bit_input) {
 408                 pixel_format = PixelFormat_10BitYCbCr;
 409         } else {
 410                 pixel_format = PixelFormat_8BitYCbCr;
 411         }
 412
 413         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 414         if (card->frame_allocator == nullptr) {
 415                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 416         }
 417         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 418         if (card->surface == nullptr) {
 419                 card->surface = create_surface_with_same_format(mixer_surface);
 420         }
 421         while (!card->new_frames.empty()) card->new_frames.pop_front();
 422         card->last_timecode = -1;
 423         card->capture->set_pixel_format(pixel_format);
 424         card->capture->configure_card();
 425
 426         // NOTE: start_bm_capture() happens in thread_func().
 427
 428         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 429         audio_mixer.reset_resampler(device);
 430         audio_mixer.set_display_name(device, card->capture->get_description());
 431         audio_mixer.trigger_state_changed_callback();
 432 }
 433
 434 void Mixer::set_output_card_internal(int card_index)
 435 {
 436         // We don't really need to take card_mutex, since we're in the mixer
 437         // thread and don't mess with any queues (which is the only thing that happens
 438         // from other threads), but it's probably the safest in the long run.
 439         unique_lock<mutex> lock(card_mutex);
 440         if (output_card_index != -1) {
 441                 // Switch the old card from output to input.
 442                 CaptureCard *old_card = &cards[output_card_index];
 443                 old_card->output->end_output();
 444
 445                 // Stop the fake card that we put into place.
 446                 // This needs to _not_ happen under the mutex, to avoid deadlock
 447                 // (delivering the last frame needs to take the mutex).
 448                 bmusb::CaptureInterface *fake_capture = old_card->capture.get();
 449                 lock.unlock();
 450                 fake_capture->stop_dequeue_thread();
 451                 lock.lock();
 452                 old_card->capture = move(old_card->parked_capture);
 453                 old_card->is_fake_capture = false;
 454                 old_card->capture->start_bm_capture();
 455         }
 456         if (card_index != -1) {
 457                 CaptureCard *card = &cards[card_index];
 458                 bmusb::CaptureInterface *capture = card->capture.get();
 459                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 460                 // several seconds to complete (blocking on DisableVideoInput);
 461                 // see if we can maybe do it asynchronously.
 462                 lock.unlock();
 463                 capture->stop_dequeue_thread();
 464                 lock.lock();
 465                 card->parked_capture = move(card->capture);
 466                 bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 467                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 468                 card->queue_length_policy.reset(card_index);
 469                 card->capture->start_bm_capture();
 470                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 471                 card->output->start_output(desired_output_video_mode, pts_int);
 472         }
 473         output_card_index = card_index;
 474 }
 475
 476 namespace {
 477
 478 int unwrap_timecode(uint16_t current_wrapped, int last)
 479 {
 480         uint16_t last_wrapped = last & 0xffff;
 481         if (current_wrapped > last_wrapped) {
 482                 return (last & ~0xffff) | current_wrapped;
 483         } else {
 484                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 485         }
 486 }
 487
 488 }  // namespace
 489
 490 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 491                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 492                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 493 {
 494         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 495         CaptureCard *card = &cards[card_index];
 496
 497         if (is_mode_scanning[card_index]) {
 498                 if (video_format.has_signal) {
 499                         // Found a stable signal, so stop scanning.
 500                         is_mode_scanning[card_index] = false;
 501                 } else {
 502                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 503                         steady_clock::time_point now = steady_clock::now();
 504                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 505                         if (sec_since_last_switch > switch_time_s) {
 506                                 // It isn't this mode; try the next one.
 507                                 mode_scanlist_index[card_index]++;
 508                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 509                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 510                                 last_mode_scan_change[card_index] = now;
 511                         }
 512                 }
 513         }
 514
 515         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 516         assert(frame_length > 0);
 517
 518         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;
 519         if (num_samples > OUTPUT_FREQUENCY / 10) {
 520                 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",
 521                         card_index, int(audio_frame.len), int(audio_offset),
 522                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 523                 if (video_frame.owner) {
 524                         video_frame.owner->release_frame(video_frame);
 525                 }
 526                 if (audio_frame.owner) {
 527                         audio_frame.owner->release_frame(audio_frame);
 528                 }
 529                 return;
 530         }
 531
 532         int dropped_frames = 0;
 533         if (card->last_timecode != -1) {
 534                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 535         }
 536
 537         // Number of samples per frame if we need to insert silence.
 538         // (Could be nonintegral, but resampling will save us then.)
 539         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 540
 541         if (dropped_frames > MAX_FPS * 2) {
 542                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 543                         card_index, card->last_timecode, timecode);
 544                 audio_mixer.reset_resampler(device);
 545                 dropped_frames = 0;
 546         } else if (dropped_frames > 0) {
 547                 // Insert silence as needed.
 548                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 549                         card_index, dropped_frames, timecode);
 550
 551                 bool success;
 552                 do {
 553                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 554                 } while (!success);
 555         }
 556
 557         audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 558
 559         // Done with the audio, so release it.
 560         if (audio_frame.owner) {
 561                 audio_frame.owner->release_frame(audio_frame);
 562         }
 563
 564         card->last_timecode = timecode;
 565
 566         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 567         if (video_frame.len - video_offset == 0 ||
 568             video_frame.len - video_offset != expected_length) {
 569                 if (video_frame.len != 0) {
 570                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 571                                 card_index, video_frame.len - video_offset, expected_length);
 572                 }
 573                 if (video_frame.owner) {
 574                         video_frame.owner->release_frame(video_frame);
 575                 }
 576
 577                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 578                 // so that pts can go up accordingly.
 579                 {
 580                         unique_lock<mutex> lock(card_mutex);
 581                         CaptureCard::NewFrame new_frame;
 582                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 583                         new_frame.length = frame_length;
 584                         new_frame.interlaced = false;
 585                         new_frame.dropped_frames = dropped_frames;
 586                         new_frame.received_timestamp = video_frame.received_timestamp;
 587                         card->new_frames.push_back(move(new_frame));
 588                         card->new_frames_changed.notify_all();
 589                 }
 590                 return;
 591         }
 592
 593         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 594
 595         unsigned num_fields = video_format.interlaced ? 2 : 1;
 596         steady_clock::time_point frame_upload_start;
 597         bool interlaced_stride = false;
 598         if (video_format.interlaced) {
 599                 // Send the two fields along as separate frames; the other side will need to add
 600                 // a deinterlacer to actually get this right.
 601                 assert(video_format.height % 2 == 0);
 602                 video_format.height /= 2;
 603                 assert(frame_length % 2 == 0);
 604                 frame_length /= 2;
 605                 num_fields = 2;
 606                 if (video_format.second_field_start == 1) {
 607                         interlaced_stride = true;
 608                 }
 609                 frame_upload_start = steady_clock::now();
 610         }
 611         userdata->last_interlaced = video_format.interlaced;
 612         userdata->last_has_signal = video_format.has_signal;
 613         userdata->last_is_connected = video_format.is_connected;
 614         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 615         userdata->last_frame_rate_den = video_format.frame_rate_den;
 616         RefCountedFrame frame(video_frame);
 617
 618         // Upload the textures.
 619         const size_t cbcr_width = video_format.width / 2;
 620         const size_t cbcr_offset = video_offset / 2;
 621         const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 622
 623         for (unsigned field = 0; field < num_fields; ++field) {
 624                 // Put the actual texture upload in a lambda that is executed in the main thread.
 625                 // It is entirely possible to do this in the same thread (and it might even be
 626                 // faster, depending on the GPU and driver), but it appears to be trickling
 627                 // driver bugs very easily.
 628                 //
 629                 // Note that this means we must hold on to the actual frame data in <userdata>
 630                 // until the upload command is run, but we hold on to <frame> much longer than that
 631                 // (in fact, all the way until we no longer use the texture in rendering).
 632                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
 633                         unsigned field_start_line;
 634                         if (field == 1) {
 635                                 field_start_line = video_format.second_field_start;
 636                         } else {
 637                                 field_start_line = video_format.extra_lines_top;
 638                         }
 639
 640                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 641                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 642                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 643
 644                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 645                         check_error();
 646
 647                         switch (userdata->pixel_format) {
 648                         case bmusb::PixelFormat_10BitYCbCr: {
 649                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 650                                 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);
 651                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 652                                 break;
 653                         }
 654                         case bmusb::PixelFormat_8BitYCbCr: {
 655                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 656                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 657
 658                                 // Make up our own strides, since we are interleaving.
 659                                 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);
 660                                 upload_texture(userdata->tex_cbcr[field], cbcr_width, video_format.height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
 661                                 break;
 662                         }
 663                         case bmusb::PixelFormat_8BitBGRA: {
 664                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 665                                 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);
 666                                 break;
 667                         }
 668                         default:
 669                                 assert(false);
 670                         }
 671
 672                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 673                         check_error();
 674                 };
 675
 676                 if (field == 1) {
 677                         // Don't upload the second field as fast as we can; wait until
 678                         // the field time has approximately passed. (Otherwise, we could
 679                         // get timing jitter against the other sources, and possibly also
 680                         // against the video display, although the latter is not as critical.)
 681                         // This requires our system clock to be reasonably close to the
 682                         // video clock, but that's not an unreasonable assumption.
 683                         steady_clock::time_point second_field_start = frame_upload_start +
 684                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 685                         this_thread::sleep_until(second_field_start);
 686                 }
 687
 688                 {
 689                         unique_lock<mutex> lock(card_mutex);
 690                         CaptureCard::NewFrame new_frame;
 691                         new_frame.frame = frame;
 692                         new_frame.length = frame_length;
 693                         new_frame.field = field;
 694                         new_frame.interlaced = video_format.interlaced;
 695                         new_frame.upload_func = upload_func;
 696                         new_frame.dropped_frames = dropped_frames;
 697                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 698                         card->new_frames.push_back(move(new_frame));
 699                         card->new_frames_changed.notify_all();
 700                 }
 701         }
 702 }
 703
 704 void Mixer::bm_hotplug_add(libusb_device *dev)
 705 {
 706         lock_guard<mutex> lock(hotplug_mutex);
 707         hotplugged_cards.push_back(dev);
 708 }
 709
 710 void Mixer::bm_hotplug_remove(unsigned card_index)
 711 {
 712         cards[card_index].new_frames_changed.notify_all();
 713 }
 714
 715 void Mixer::thread_func()
 716 {
 717         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 718
 719         eglBindAPI(EGL_OPENGL_API);
 720         QOpenGLContext *context = create_context(mixer_surface);
 721         if (!make_current(context, mixer_surface)) {
 722                 printf("oops\n");
 723                 exit(1);
 724         }
 725
 726         // Start the actual capture. (We don't want to do it before we're actually ready
 727         // to process output frames.)
 728         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 729                 if (int(card_index) != output_card_index) {
 730                         cards[card_index].capture->start_bm_capture();
 731                 }
 732         }
 733
 734         steady_clock::time_point start, now;
 735         start = steady_clock::now();
 736
 737         int stats_dropped_frames = 0;
 738
 739         while (!should_quit) {
 740                 if (desired_output_card_index != output_card_index) {
 741                         set_output_card_internal(desired_output_card_index);
 742                 }
 743                 if (output_card_index != -1 &&
 744                     desired_output_video_mode != output_video_mode) {
 745                         DeckLinkOutput *output = cards[output_card_index].output.get();
 746                         output->end_output();
 747                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 748                         output->start_output(desired_output_video_mode, pts_int);
 749                 }
 750
 751                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 752                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 753
 754                 bool master_card_is_output;
 755                 unsigned master_card_index;
 756                 if (output_card_index != -1) {
 757                         master_card_is_output = true;
 758                         master_card_index = output_card_index;
 759                 } else {
 760                         master_card_is_output = false;
 761                         master_card_index = theme->map_signal(master_clock_channel);
 762                         assert(master_card_index < num_cards);
 763                 }
 764
 765                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 766                 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);
 767                 stats_dropped_frames += output_frame_info.dropped_frames;
 768
 769                 handle_hotplugged_cards();
 770
 771                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 772                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 773                                 continue;
 774                         }
 775                         if (new_frames[card_index].frame->len == 0) {
 776                                 ++new_frames[card_index].dropped_frames;
 777                         }
 778                         if (new_frames[card_index].dropped_frames > 0) {
 779                                 printf("Card %u dropped %d frames before this\n",
 780                                         card_index, int(new_frames[card_index].dropped_frames));
 781                         }
 782                 }
 783
 784                 // If the first card is reporting a corrupted or otherwise dropped frame,
 785                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
 786                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
 787                         ++stats_dropped_frames;
 788                         pts_int += new_frames[master_card_index].length;
 789                         continue;
 790                 }
 791
 792                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 793                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
 794                                 continue;
 795
 796                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
 797                         assert(new_frame->frame != nullptr);
 798                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
 799                         check_error();
 800
 801                         // The new texture might need uploading before use.
 802                         if (new_frame->upload_func) {
 803                                 new_frame->upload_func();
 804                                 new_frame->upload_func = nullptr;
 805                         }
 806                 }
 807
 808                 int64_t frame_duration = output_frame_info.frame_duration;
 809                 render_one_frame(frame_duration);
 810                 ++frame_num;
 811                 pts_int += frame_duration;
 812
 813                 now = steady_clock::now();
 814                 double elapsed = duration<double>(now - start).count();
 815                 if (frame_num % 100 == 0) {
 816                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
 817                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
 818                                 1e3 * elapsed / frame_num);
 819                 //      chain->print_phase_timing();
 820
 821                         // Check our memory usage, to see if we are close to our mlockall()
 822                         // limit (if at all set).
 823                         rusage used;
 824                         if (getrusage(RUSAGE_SELF, &used) == -1) {
 825                                 perror("getrusage(RUSAGE_SELF)");
 826                                 assert(false);
 827                         }
 828
 829                         if (uses_mlock) {
 830                                 rlimit limit;
 831                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
 832                                         perror("getrlimit(RLIMIT_MEMLOCK)");
 833                                         assert(false);
 834                                 }
 835
 836                                 if (limit.rlim_cur == 0) {
 837                                         printf(", using %ld MB memory (locked)",
 838                                                 long(used.ru_maxrss / 1024));
 839                                 } else {
 840                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
 841                                                 long(used.ru_maxrss / 1024),
 842                                                 long(limit.rlim_cur / 1048576),
 843                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
 844                                 }
 845                         } else {
 846                                 printf(", using %ld MB memory (not locked)",
 847                                         long(used.ru_maxrss / 1024));
 848                         }
 849
 850                         printf("\n");
 851                 }
 852
 853
 854                 if (should_cut.exchange(false)) {  // Test and clear.
 855                         video_encoder->do_cut(frame_num);
 856                 }
 857
 858 #if 0
 859                 // Reset every 100 frames, so that local variations in frame times
 860                 // (especially for the first few frames, when the shaders are
 861                 // compiled etc.) don't make it hard to measure for the entire
 862                 // remaining duration of the program.
 863                 if (frame == 10000) {
 864                         frame = 0;
 865                         start = now;
 866                 }
 867 #endif
 868                 check_error();
 869         }
 870
 871         resource_pool->clean_context();
 872 }
 873
 874 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 875 {
 876         if (output_card_index != -1) {
 877                 // The output card (ie., cards[output_card_index].output) is the master clock,
 878                 // so no input card (ie., cards[card_index].capture) is.
 879                 return false;
 880         }
 881         return (card_index == master_card_index);
 882 }
 883
 884 void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
 885 {
 886         // Count the number of frames in the queue, including any frames
 887         // we dropped. It's hard to know exactly how we should deal with
 888         // dropped (corrupted) input frames; they don't help our goal of
 889         // avoiding starvation, but they still add to the problem of latency.
 890         // Since dropped frames is going to mean a bump in the signal anyway,
 891         // we err on the side of having more stable latency instead.
 892         unsigned queue_length = 0;
 893         for (const CaptureCard::NewFrame &frame : card->new_frames) {
 894                 queue_length += frame.dropped_frames + 1;
 895         }
 896         card->queue_length_policy.update_policy(queue_length);
 897
 898         // If needed, drop frames until the queue is below the safe limit.
 899         // We prefer to drop from the head, because all else being equal,
 900         // we'd like more recent frames (less latency).
 901         unsigned dropped_frames = 0;
 902         while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
 903                 assert(!card->new_frames.empty());
 904                 assert(queue_length > card->new_frames.front().dropped_frames);
 905                 queue_length -= card->new_frames.front().dropped_frames;
 906
 907                 if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
 908                         // No need to drop anything.
 909                         break;
 910                 }
 911
 912                 card->new_frames.pop_front();
 913                 card->new_frames_changed.notify_all();
 914                 --queue_length;
 915                 ++dropped_frames;
 916         }
 917
 918 #if 0
 919         if (dropped_frames > 0) {
 920                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
 921                         card_index, dropped_frames);
 922         }
 923 #endif
 924 }
 925
 926
 927 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])
 928 {
 929         OutputFrameInfo output_frame_info;
 930 start:
 931         unique_lock<mutex> lock(card_mutex, defer_lock);
 932         if (master_card_is_output) {
 933                 // Clocked to the output, so wait for it to be ready for the next frame.
 934                 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);
 935                 lock.lock();
 936         } else {
 937                 // Wait for the master card to have a new frame.
 938                 // TODO: Add a timeout.
 939                 output_frame_info.is_preroll = false;
 940                 lock.lock();
 941                 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(); });
 942         }
 943
 944         if (master_card_is_output) {
 945                 handle_hotplugged_cards();
 946         } else if (cards[master_card_index].new_frames.empty()) {
 947                 // We were woken up, but not due to a new frame. Deal with it
 948                 // and then restart.
 949                 assert(cards[master_card_index].capture->get_disconnected());
 950                 handle_hotplugged_cards();
 951                 goto start;
 952         }
 953
 954         if (!master_card_is_output) {
 955                 output_frame_info.frame_timestamp =
 956                         cards[master_card_index].new_frames.front().received_timestamp;
 957         }
 958
 959         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 960                 CaptureCard *card = &cards[card_index];
 961                 if (input_card_is_master_clock(card_index, master_card_index)) {
 962                         // We don't use the queue length policy for the master card,
 963                         // but we will if it stops being the master. Thus, clear out
 964                         // the policy in case we switch in the future.
 965                         card->queue_length_policy.reset(card_index);
 966                         assert(!card->new_frames.empty());
 967                 } else {
 968                         trim_queue(card, card_index);
 969                 }
 970                 if (!card->new_frames.empty()) {
 971                         new_frames[card_index] = move(card->new_frames.front());
 972                         has_new_frame[card_index] = true;
 973                         card->new_frames.pop_front();
 974                         card->new_frames_changed.notify_all();
 975                 }
 976         }
 977
 978         if (!master_card_is_output) {
 979                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
 980                 output_frame_info.frame_duration = new_frames[master_card_index].length;
 981         }
 982
 983         // This might get off by a fractional sample when changing master card
 984         // between ones with different frame rates, but that's fine.
 985         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
 986         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
 987         fractional_samples = num_samples_times_timebase % TIMEBASE;
 988         assert(output_frame_info.num_samples >= 0);
 989
 990         return output_frame_info;
 991 }
 992
 993 void Mixer::handle_hotplugged_cards()
 994 {
 995         // Check for cards that have been disconnected since last frame.
 996         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
 997                 CaptureCard *card = &cards[card_index];
 998                 if (card->capture->get_disconnected()) {
 999                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1000                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1001                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1002                         card->queue_length_policy.reset(card_index);
1003                         card->capture->start_bm_capture();
1004                 }
1005         }
1006
1007         // Check for cards that have been connected since last frame.
1008         vector<libusb_device *> hotplugged_cards_copy;
1009         {
1010                 lock_guard<mutex> lock(hotplug_mutex);
1011                 swap(hotplugged_cards, hotplugged_cards_copy);
1012         }
1013         for (libusb_device *new_dev : hotplugged_cards_copy) {
1014                 // Look for a fake capture card where we can stick this in.
1015                 int free_card_index = -1;
1016                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1017                         if (cards[card_index].is_fake_capture) {
1018                                 free_card_index = card_index;
1019                                 break;
1020                         }
1021                 }
1022
1023                 if (free_card_index == -1) {
1024                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1025                         libusb_unref_device(new_dev);
1026                 } else {
1027                         // BMUSBCapture takes ownership.
1028                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1029                         CaptureCard *card = &cards[free_card_index];
1030                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1031                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1032                         card->queue_length_policy.reset(free_card_index);
1033                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1034                         capture->start_bm_capture();
1035                 }
1036         }
1037 }
1038
1039
1040 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)
1041 {
1042         // Resample the audio as needed, including from previously dropped frames.
1043         assert(num_cards > 0);
1044         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1045                 const bool dropped_frame = (frame_num != dropped_frames);
1046                 {
1047                         // Signal to the audio thread to process this frame.
1048                         // Note that if the frame is a dropped frame, we signal that
1049                         // we don't want to use this frame as base for adjusting
1050                         // the resampler rate. The reason for this is that the timing
1051                         // of these frames is often way too late; they typically don't
1052                         // “arrive” before we synthesize them. Thus, we could end up
1053                         // in a situation where we have inserted e.g. five audio frames
1054                         // into the queue before we then start pulling five of them
1055                         // back out. This makes ResamplingQueue overestimate the delay,
1056                         // causing undue resampler changes. (We _do_ use the last,
1057                         // non-dropped frame; perhaps we should just discard that as well,
1058                         // since dropped frames are expected to be rare, and it might be
1059                         // better to just wait until we have a slightly more normal situation).
1060                         unique_lock<mutex> lock(audio_mutex);
1061                         bool adjust_rate = !dropped_frame && !is_preroll;
1062                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1063                         audio_task_queue_changed.notify_one();
1064                 }
1065                 if (dropped_frame) {
1066                         // For dropped frames, increase the pts. Note that if the format changed
1067                         // in the meantime, we have no way of detecting that; we just have to
1068                         // assume the frame length is always the same.
1069                         pts_int += length_per_frame;
1070                 }
1071         }
1072 }
1073
1074 void Mixer::render_one_frame(int64_t duration)
1075 {
1076         // Determine the time code for this frame before we start rendering.
1077         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1078         if (display_timecode_on_stdout) {
1079                 printf("Timecode: '%s'\n", timecode_text.c_str());
1080         }
1081
1082         // Get the main chain from the theme, and set its state immediately.
1083         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1084         EffectChain *chain = theme_main_chain.chain;
1085         theme_main_chain.setup_chain();
1086         //theme_main_chain.chain->enable_phase_timing(true);
1087
1088         // The theme can't (or at least shouldn't!) call connect_signal() on
1089         // each FFmpeg input, so we'll do it here.
1090         for (const pair<LiveInputWrapper *, FFmpegCapture *> &conn : theme->get_signal_connections()) {
1091                 conn.first->connect_signal_raw(conn.second->get_card_index());
1092         }
1093
1094         // If HDMI/SDI output is active and the user has requested auto mode,
1095         // its mode overrides the existing Y'CbCr setting for the chain.
1096         YCbCrLumaCoefficients ycbcr_output_coefficients;
1097         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1098                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1099         } else {
1100                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1101         }
1102
1103         // TODO: Reduce the duplication against theme.cpp.
1104         YCbCrFormat output_ycbcr_format;
1105         output_ycbcr_format.chroma_subsampling_x = 1;
1106         output_ycbcr_format.chroma_subsampling_y = 1;
1107         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1108         output_ycbcr_format.full_range = false;
1109         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1110         chain->change_ycbcr_output_format(output_ycbcr_format);
1111
1112         // Render main chain. If we're using zerocopy Quick Sync encoding
1113         // (the default case), we take an extra copy of the created outputs,
1114         // so that we can display it back to the screen later (it's less memory
1115         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1116         // Ideally, we'd like to avoid taking copies and just use the main textures
1117         // for display as well, but they're just views into VA-API memory and must be
1118         // unmapped during encoding, so we can't use them for display, unfortunately.
1119         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1120         GLuint y_copy_tex, cbcr_copy_tex = 0;
1121         GLuint y_display_tex, cbcr_display_tex;
1122         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1123         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1124         const bool is_zerocopy = video_encoder->is_zerocopy();
1125         if (is_zerocopy) {
1126                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1127                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1128                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1129
1130                 y_display_tex = y_copy_tex;
1131                 cbcr_display_tex = cbcr_copy_tex;
1132
1133                 // y_tex and cbcr_tex will be given by VideoEncoder.
1134         } else {
1135                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1136                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1137                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1138
1139                 y_display_tex = y_tex;
1140                 cbcr_display_tex = cbcr_tex;
1141         }
1142
1143         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1144         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1145         assert(got_frame);
1146
1147         GLuint fbo;
1148         if (is_zerocopy) {
1149                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1150         } else {
1151                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1152         }
1153         check_error();
1154         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1155
1156         if (display_timecode_in_stream) {
1157                 // Render the timecode on top.
1158                 timecode_renderer->render_timecode(fbo, timecode_text);
1159         }
1160
1161         resource_pool->release_fbo(fbo);
1162
1163         if (is_zerocopy) {
1164                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1165         } else {
1166                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1167         }
1168         if (output_card_index != -1) {
1169                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1170         }
1171         resource_pool->release_2d_texture(cbcr_full_tex);
1172
1173         // Set the right state for the Y' and CbCr textures we use for display.
1174         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1175         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1176         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1177         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1178         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1179
1180         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1181         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1182         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1183         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1184
1185         RefCountedGLsync fence = video_encoder->end_frame();
1186
1187         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1188         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1189         DisplayFrame live_frame;
1190         live_frame.chain = display_chain.get();
1191         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1192                 display_input->set_texture_num(0, y_display_tex);
1193                 display_input->set_texture_num(1, cbcr_display_tex);
1194         };
1195         live_frame.ready_fence = fence;
1196         live_frame.input_frames = {};
1197         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1198         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1199
1200         // Set up preview and any additional channels.
1201         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1202                 DisplayFrame display_frame;
1203                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1204                 display_frame.chain = chain.chain;
1205                 display_frame.setup_chain = chain.setup_chain;
1206                 display_frame.ready_fence = fence;
1207                 display_frame.input_frames = chain.input_frames;
1208                 display_frame.temp_textures = {};
1209                 output_channel[i].output_frame(display_frame);
1210         }
1211 }
1212
1213 void Mixer::audio_thread_func()
1214 {
1215         pthread_setname_np(pthread_self(), "Mixer_Audio");
1216
1217         while (!should_quit) {
1218                 AudioTask task;
1219
1220                 {
1221                         unique_lock<mutex> lock(audio_mutex);
1222                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1223                         if (should_quit) {
1224                                 return;
1225                         }
1226                         task = audio_task_queue.front();
1227                         audio_task_queue.pop();
1228                 }
1229
1230                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1231                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1232                 vector<float> samples_out = audio_mixer.get_output(
1233                         task.frame_timestamp,
1234                         task.num_samples,
1235                         rate_adjustment_policy);
1236
1237                 // Send the samples to the sound card, then add them to the output.
1238                 if (alsa) {
1239                         alsa->write(samples_out);
1240                 }
1241                 if (output_card_index != -1) {
1242                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1243                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1244                 }
1245                 video_encoder->add_audio(task.pts_int, move(samples_out));
1246         }
1247 }
1248
1249 void Mixer::release_display_frame(DisplayFrame *frame)
1250 {
1251         for (GLuint texnum : frame->temp_textures) {
1252                 resource_pool->release_2d_texture(texnum);
1253         }
1254         frame->temp_textures.clear();
1255         frame->ready_fence.reset();
1256         frame->input_frames.clear();
1257 }
1258
1259 void Mixer::start()
1260 {
1261         mixer_thread = thread(&Mixer::thread_func, this);
1262         audio_thread = thread(&Mixer::audio_thread_func, this);
1263 }
1264
1265 void Mixer::quit()
1266 {
1267         should_quit = true;
1268         audio_task_queue_changed.notify_one();
1269         mixer_thread.join();
1270         audio_thread.join();
1271 }
1272
1273 void Mixer::transition_clicked(int transition_num)
1274 {
1275         theme->transition_clicked(transition_num, pts());
1276 }
1277
1278 void Mixer::channel_clicked(int preview_num)
1279 {
1280         theme->channel_clicked(preview_num);
1281 }
1282
1283 void Mixer::start_mode_scanning(unsigned card_index)
1284 {
1285         assert(card_index < num_cards);
1286         if (is_mode_scanning[card_index]) {
1287                 return;
1288         }
1289         is_mode_scanning[card_index] = true;
1290         mode_scanlist[card_index].clear();
1291         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1292                 mode_scanlist[card_index].push_back(mode.first);
1293         }
1294         assert(!mode_scanlist[card_index].empty());
1295         mode_scanlist_index[card_index] = 0;
1296         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1297         last_mode_scan_change[card_index] = steady_clock::now();
1298 }
1299
1300 map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
1301 {
1302         assert(desired_output_card_index != -1);
1303         unique_lock<mutex> lock(card_mutex);
1304         return cards[desired_output_card_index].output->get_available_video_modes();
1305 }
1306
1307 Mixer::OutputChannel::~OutputChannel()
1308 {
1309         if (has_current_frame) {
1310                 parent->release_display_frame(&current_frame);
1311         }
1312         if (has_ready_frame) {
1313                 parent->release_display_frame(&ready_frame);
1314         }
1315 }
1316
1317 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1318 {
1319         // Store this frame for display. Remove the ready frame if any
1320         // (it was seemingly never used).
1321         {
1322                 unique_lock<mutex> lock(frame_mutex);
1323                 if (has_ready_frame) {
1324                         parent->release_display_frame(&ready_frame);
1325                 }
1326                 ready_frame = frame;
1327                 has_ready_frame = true;
1328         }
1329
1330         if (new_frame_ready_callback) {
1331                 new_frame_ready_callback();
1332         }
1333
1334         // Reduce the number of callbacks by filtering duplicates. The reason
1335         // why we bother doing this is that Qt seemingly can get into a state
1336         // where its builds up an essentially unbounded queue of signals,
1337         // consuming more and more memory, and there's no good way of collapsing
1338         // user-defined signals or limiting the length of the queue.
1339         if (transition_names_updated_callback) {
1340                 vector<string> transition_names = global_mixer->get_transition_names();
1341                 bool changed = false;
1342                 if (transition_names.size() != last_transition_names.size()) {
1343                         changed = true;
1344                 } else {
1345                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1346                                 if (transition_names[i] != last_transition_names[i]) {
1347                                         changed = true;
1348                                         break;
1349                                 }
1350                         }
1351                 }
1352                 if (changed) {
1353                         transition_names_updated_callback(transition_names);
1354                         last_transition_names = transition_names;
1355                 }
1356         }
1357         if (name_updated_callback) {
1358                 string name = global_mixer->get_channel_name(channel);
1359                 if (name != last_name) {
1360                         name_updated_callback(name);
1361                         last_name = name;
1362                 }
1363         }
1364         if (color_updated_callback) {
1365                 string color = global_mixer->get_channel_color(channel);
1366                 if (color != last_color) {
1367                         color_updated_callback(color);
1368                         last_color = color;
1369                 }
1370         }
1371 }
1372
1373 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1374 {
1375         unique_lock<mutex> lock(frame_mutex);
1376         if (!has_current_frame && !has_ready_frame) {
1377                 return false;
1378         }
1379
1380         if (has_current_frame && has_ready_frame) {
1381                 // We have a new ready frame. Toss the current one.
1382                 parent->release_display_frame(&current_frame);
1383                 has_current_frame = false;
1384         }
1385         if (has_ready_frame) {
1386                 assert(!has_current_frame);
1387                 current_frame = ready_frame;
1388                 ready_frame.ready_fence.reset();  // Drop the refcount.
1389                 ready_frame.input_frames.clear();  // Drop the refcounts.
1390                 has_current_frame = true;
1391                 has_ready_frame = false;
1392         }
1393
1394         *frame = current_frame;
1395         return true;
1396 }
1397
1398 void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
1399 {
1400         new_frame_ready_callback = callback;
1401 }
1402
1403 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1404 {
1405         transition_names_updated_callback = callback;
1406 }
1407
1408 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1409 {
1410         name_updated_callback = callback;
1411 }
1412
1413 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1414 {
1415         color_updated_callback = callback;
1416 }
1417
1418 mutex RefCountedGLsync::fence_lock;