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