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

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   7 #include <movit/effect_chain.h>
   8 #include <movit/effect_util.h>
   9 #include <movit/flat_input.h>
  10 #include <movit/image_format.h>
  11 #include <movit/init.h>
  12 #include <movit/resource_pool.h>
  13 #include <pthread.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <sys/resource.h>
  18 #include <algorithm>
  19 #include <chrono>
  20 #include <condition_variable>
  21 #include <cstddef>
  22 #include <cstdint>
  23 #include <memory>
  24 #include <mutex>
  25 #include <ratio>
  26 #include <string>
  27 #include <thread>
  28 #include <utility>
  29 #include <vector>
  30
  31 #include "DeckLinkAPI.h"
  32 #include "LinuxCOM.h"
  33 #include "alsa_output.h"
  34 #include "bmusb/bmusb.h"
  35 #include "bmusb/fake_capture.h"
  36 #include "chroma_subsampler.h"
  37 #include "context.h"
  38 #include "decklink_capture.h"
  39 #include "decklink_output.h"
  40 #include "defs.h"
  41 #include "disk_space_estimator.h"
  42 #include "ffmpeg_capture.h"
  43 #include "flags.h"
  44 #include "input_mapping.h"
  45 #include "metrics.h"
  46 #include "pbo_frame_allocator.h"
  47 #include "ref_counted_gl_sync.h"
  48 #include "resampling_queue.h"
  49 #include "timebase.h"
  50 #include "timecode_renderer.h"
  51 #include "v210_converter.h"
  52 #include "video_encoder.h"
  53
  54 class IDeckLink;
  55 class QOpenGLContext;
  56
  57 using namespace movit;
  58 using namespace std;
  59 using namespace std::chrono;
  60 using namespace std::placeholders;
  61 using namespace bmusb;
  62
  63 Mixer *global_mixer = nullptr;
  64 bool uses_mlock = false;
  65
  66 namespace {
  67
  68 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  69 {
  70         if (interlaced) {
  71                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  72                         input_state->buffered_frames[card_index][frame_num] =
  73                                 input_state->buffered_frames[card_index][frame_num - 1];
  74                 }
  75                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  76         } else {
  77                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  78                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  79                 }
  80         }
  81 }
  82
  83 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
  84 {
  85         bool first;
  86         switch (userdata->pixel_format) {
  87         case PixelFormat_10BitYCbCr:
  88                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  89                 break;
  90         case PixelFormat_8BitYCbCr:
  91                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
  92                 break;
  93         case PixelFormat_8BitBGRA:
  94                 first = userdata->tex_rgba[field] == 0;
  95                 break;
  96         case PixelFormat_8BitYCbCrPlanar:
  97                 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
  98                 break;
  99         default:
 100                 assert(false);
 101         }
 102
 103         if (first ||
 104             width != userdata->last_width[field] ||
 105             height != userdata->last_height[field] ||
 106             cbcr_width != userdata->last_cbcr_width[field] ||
 107             cbcr_height != userdata->last_cbcr_height[field]) {
 108                 // We changed resolution since last use of this texture, so we need to create
 109                 // a new object. Note that this each card has its own PBOFrameAllocator,
 110                 // we don't need to worry about these flip-flopping between resolutions.
 111                 switch (userdata->pixel_format) {
 112                 case PixelFormat_10BitYCbCr:
 113                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 114                         check_error();
 115                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 116                         check_error();
 117                         break;
 118                 case PixelFormat_8BitYCbCr: {
 119                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 120                         check_error();
 121                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 122                         check_error();
 123                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 124                         check_error();
 125                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 126                         check_error();
 127                         break;
 128                 }
 129                 case PixelFormat_8BitYCbCrPlanar: {
 130                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 131                         check_error();
 132                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 133                         check_error();
 134                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
 135                         check_error();
 136                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 137                         check_error();
 138                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
 139                         check_error();
 140                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 141                         check_error();
 142                         break;
 143                 }
 144                 case PixelFormat_8BitBGRA:
 145                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 146                         check_error();
 147                         if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
 148                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 149                         } else {
 150                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 151                         }
 152                         check_error();
 153                         break;
 154                 }
 155                 userdata->last_width[field] = width;
 156                 userdata->last_height[field] = height;
 157                 userdata->last_cbcr_width[field] = cbcr_width;
 158                 userdata->last_cbcr_height[field] = cbcr_height;
 159         }
 160         if (global_flags.ten_bit_input &&
 161             (first || v210_width != userdata->last_v210_width[field])) {
 162                 // Same as above; we need to recreate the texture.
 163                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 164                 check_error();
 165                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 166                 check_error();
 167                 userdata->last_v210_width[field] = v210_width;
 168         }
 169 }
 170
 171 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 172 {
 173         if (interlaced_stride) {
 174                 stride *= 2;
 175         }
 176         if (global_flags.flush_pbos) {
 177                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 178                 check_error();
 179         }
 180
 181         glBindTexture(GL_TEXTURE_2D, tex);
 182         check_error();
 183         if (interlaced_stride) {
 184                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 185                 check_error();
 186         } else {
 187                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 188                 check_error();
 189         }
 190
 191         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 192         check_error();
 193         glBindTexture(GL_TEXTURE_2D, 0);
 194         check_error();
 195         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 196         check_error();
 197 }
 198
 199 }  // namespace
 200
 201 void JitterHistory::register_metrics(const vector<pair<string, string>> &labels)
 202 {
 203         global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames);
 204         global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE);
 205 }
 206
 207 void JitterHistory::unregister_metrics(const vector<pair<string, string>> &labels)
 208 {
 209         global_metrics.remove("input_underestimated_jitter_frames", labels);
 210         global_metrics.remove("input_estimated_max_jitter_seconds", labels);
 211 }
 212
 213 void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames)
 214 {
 215         if (expected_timestamp > steady_clock::time_point::min()) {
 216                 expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 217                 double jitter_seconds = fabs(duration<double>(expected_timestamp - now).count());
 218                 history.push_back(orders.insert(jitter_seconds));
 219                 if (jitter_seconds > estimate_max_jitter()) {
 220                         ++metric_input_underestimated_jitter_frames;
 221                 }
 222
 223                 metric_input_estimated_max_jitter_seconds = estimate_max_jitter();
 224
 225                 if (history.size() > history_length) {
 226                         orders.erase(history.front());
 227                         history.pop_front();
 228                 }
 229                 assert(history.size() <= history_length);
 230         }
 231         expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 232 }
 233
 234 double JitterHistory::estimate_max_jitter() const
 235 {
 236         if (orders.empty()) {
 237                 return 0.0;
 238         }
 239         size_t elem_idx = lrint((orders.size() - 1) * percentile);
 240         if (percentile <= 0.5) {
 241                 return *next(orders.begin(), elem_idx) * multiplier;
 242         } else {
 243                 return *prev(orders.end(), elem_idx + 1) * multiplier;
 244         }
 245 }
 246
 247 void QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
 248 {
 249         global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
 250 }
 251
 252 void QueueLengthPolicy::unregister_metrics(const vector<pair<string, string>> &labels)
 253 {
 254         global_metrics.remove("input_queue_safe_length_frames", labels);
 255 }
 256
 257 void QueueLengthPolicy::update_policy(steady_clock::time_point now,
 258                                       steady_clock::time_point expected_next_frame,
 259                                       int64_t master_frame_duration,
 260                                       double max_input_card_jitter_seconds,
 261                                       double max_master_card_jitter_seconds)
 262 {
 263         double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
 264
 265         // Figure out when we can expect the next frame for this card, assuming
 266         // worst-case jitter (ie., the frame is maximally late).
 267         double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
 268
 269         // How many times are the master card expected to tick in that time?
 270         // We assume the master clock has worst-case jitter but not any rate
 271         // discrepancy, ie., it ticks as early as possible every time, but not
 272         // cumulatively.
 273         double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
 274
 275         // As a special case, if the master card ticks faster than the input card,
 276         // we expect the queue to drain by itself even without dropping. But if
 277         // the difference is small (e.g. 60 Hz master and 59.94 input), it would
 278         // go slowly enough that the effect wouldn't really be appreciable.
 279         // We account for this by looking at the situation five frames ahead,
 280         // assuming everything else is the same.
 281         double frames_allowed;
 282         if (max_master_card_jitter_seconds < max_input_card_jitter_seconds) {
 283                 frames_allowed = frames_needed + 5 * (max_input_card_jitter_seconds - max_master_card_jitter_seconds) / master_frame_duration_seconds;
 284         } else {
 285                 frames_allowed = frames_needed;
 286         }
 287
 288         safe_queue_length = max<int>(floor(frames_allowed), 0);
 289         metric_input_queue_safe_length_frames = safe_queue_length;
 290 }
 291
 292 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 293         : httpd(),
 294           num_cards(num_cards),
 295           mixer_surface(create_surface(format)),
 296           h264_encoder_surface(create_surface(format)),
 297           decklink_output_surface(create_surface(format)),
 298           audio_mixer(num_cards)
 299 {
 300         memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
 301         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 302         check_error();
 303
 304         // This nearly always should be true.
 305         global_flags.can_disable_srgb_decoder =
 306                 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
 307                 epoxy_has_gl_extension("GL_ARB_sampler_objects");
 308
 309         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 310         // will be halved when sampling them, and we need to compensate here.
 311         movit_texel_subpixel_precision /= 2.0;
 312
 313         resource_pool.reset(new ResourcePool);
 314         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 315                 output_channel[i].parent = this;
 316                 output_channel[i].channel = i;
 317         }
 318
 319         ImageFormat inout_format;
 320         inout_format.color_space = COLORSPACE_sRGB;
 321         inout_format.gamma_curve = GAMMA_sRGB;
 322
 323         // Matches the 4:2:0 format created by the main chain.
 324         YCbCrFormat ycbcr_format;
 325         ycbcr_format.chroma_subsampling_x = 2;
 326         ycbcr_format.chroma_subsampling_y = 2;
 327         if (global_flags.ycbcr_rec709_coefficients) {
 328                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 329         } else {
 330                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 331         }
 332         ycbcr_format.full_range = false;
 333         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 334         ycbcr_format.cb_x_position = 0.0f;
 335         ycbcr_format.cr_x_position = 0.0f;
 336         ycbcr_format.cb_y_position = 0.5f;
 337         ycbcr_format.cr_y_position = 0.5f;
 338
 339         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 340         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 341         check_error();
 342         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 343         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 344         display_chain->add_input(display_input);
 345         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 346         display_chain->set_dither_bits(0);  // Don't bother.
 347         display_chain->finalize();
 348
 349         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));
 350
 351         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 352         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 353
 354         // Start listening for clients only once VideoEncoder has written its header, if any.
 355         httpd.start(9095);
 356
 357         // First try initializing the then PCI devices, then USB, then
 358         // fill up with fake cards until we have the desired number of cards.
 359         unsigned num_pci_devices = 0;
 360         unsigned card_index = 0;
 361
 362         {
 363                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 364                 if (decklink_iterator != nullptr) {
 365                         for ( ; card_index < num_cards; ++card_index) {
 366                                 IDeckLink *decklink;
 367                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 368                                         break;
 369                                 }
 370
 371                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 372                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 373                                 output->set_device(decklink);
 374                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 375                                 ++num_pci_devices;
 376                         }
 377                         decklink_iterator->Release();
 378                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 379                 } else {
 380                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 381                 }
 382         }
 383
 384         unsigned num_usb_devices = BMUSBCapture::num_cards();
 385         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 386                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 387                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 388                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 389         }
 390         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 391
 392         unsigned num_fake_cards = 0;
 393         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 394                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 395                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 396         }
 397
 398         if (num_fake_cards > 0) {
 399                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 400         }
 401
 402         // Initialize all video inputs the theme asked for. Note that these are
 403         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 404         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 405         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 406                 if (card_index >= MAX_VIDEO_CARDS) {
 407                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 408                         exit(1);
 409                 }
 410                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 411                 video_inputs[video_card_index]->set_card_index(card_index);
 412         }
 413         num_video_inputs = video_inputs.size();
 414
 415         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 416         BMUSBCapture::start_bm_thread();
 417
 418         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 419                 cards[card_index].queue_length_policy.reset(card_index);
 420         }
 421
 422         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 423
 424         if (global_flags.ten_bit_input) {
 425                 if (!v210Converter::has_hardware_support()) {
 426                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 427                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 428                         exit(1);
 429                 }
 430                 v210_converter.reset(new v210Converter());
 431
 432                 // These are all the widths listed in the Blackmagic SDK documentation
 433                 // (section 2.7.3, “Display Modes”).
 434                 v210_converter->precompile_shader(720);
 435                 v210_converter->precompile_shader(1280);
 436                 v210_converter->precompile_shader(1920);
 437                 v210_converter->precompile_shader(2048);
 438                 v210_converter->precompile_shader(3840);
 439                 v210_converter->precompile_shader(4096);
 440         }
 441         if (global_flags.ten_bit_output) {
 442                 if (!v210Converter::has_hardware_support()) {
 443                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 444                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 445                         exit(1);
 446                 }
 447         }
 448
 449         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 450         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 451         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 452
 453         if (global_flags.enable_alsa_output) {
 454                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 455         }
 456         if (global_flags.output_card != -1) {
 457                 desired_output_card_index = global_flags.output_card;
 458                 set_output_card_internal(global_flags.output_card);
 459         }
 460
 461         metric_start_time_seconds = get_timestamp_for_metrics();
 462
 463         output_jitter_history.register_metrics({{ "card", "output" }});
 464         global_metrics.add("frames_output_total", &metric_frames_output_total);
 465         global_metrics.add("frames_output_dropped", &metric_frames_output_dropped);
 466         global_metrics.add("start_time_seconds", &metric_start_time_seconds, Metrics::TYPE_GAUGE);
 467         global_metrics.add("memory_used_bytes", &metrics_memory_used_bytes);
 468         global_metrics.add("memory_locked_limit_bytes", &metrics_memory_locked_limit_bytes);
 469 }
 470
 471 Mixer::~Mixer()
 472 {
 473         BMUSBCapture::stop_bm_thread();
 474
 475         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 476                 {
 477                         unique_lock<mutex> lock(card_mutex);
 478                         cards[card_index].should_quit = true;  // Unblock thread.
 479                         cards[card_index].new_frames_changed.notify_all();
 480                 }
 481                 cards[card_index].capture->stop_dequeue_thread();
 482                 if (cards[card_index].output) {
 483                         cards[card_index].output->end_output();
 484                         cards[card_index].output.reset();
 485                 }
 486         }
 487
 488         video_encoder.reset(nullptr);
 489 }
 490
 491 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 492 {
 493         printf("Configuring card %d...\n", card_index);
 494
 495         CaptureCard *card = &cards[card_index];
 496         if (card->capture != nullptr) {
 497                 card->capture->stop_dequeue_thread();
 498         }
 499         card->capture.reset(capture);
 500         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 501         card->type = card_type;
 502         if (card->output.get() != output) {
 503                 card->output.reset(output);
 504         }
 505
 506         PixelFormat pixel_format;
 507         if (card_type == CardType::FFMPEG_INPUT) {
 508                 pixel_format = capture->get_current_pixel_format();
 509         } else if (global_flags.ten_bit_input) {
 510                 pixel_format = PixelFormat_10BitYCbCr;
 511         } else {
 512                 pixel_format = PixelFormat_8BitYCbCr;
 513         }
 514
 515         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 516         if (card->frame_allocator == nullptr) {
 517                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height));  // 8 MB.
 518         }
 519         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 520         if (card->surface == nullptr) {
 521                 card->surface = create_surface_with_same_format(mixer_surface);
 522         }
 523         while (!card->new_frames.empty()) card->new_frames.pop_front();
 524         card->last_timecode = -1;
 525         card->capture->set_pixel_format(pixel_format);
 526         card->capture->configure_card();
 527
 528         // NOTE: start_bm_capture() happens in thread_func().
 529
 530         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 531         audio_mixer.reset_resampler(device);
 532         audio_mixer.set_display_name(device, card->capture->get_description());
 533         audio_mixer.trigger_state_changed_callback();
 534
 535         // Unregister old metrics, if any.
 536         if (!card->labels.empty()) {
 537                 const vector<pair<string, string>> &labels = card->labels;
 538                 card->jitter_history.unregister_metrics(labels);
 539                 card->queue_length_policy.unregister_metrics(labels);
 540                 global_metrics.remove("input_received_frames", labels);
 541                 global_metrics.remove("input_dropped_frames_jitter", labels);
 542                 global_metrics.remove("input_dropped_frames_error", labels);
 543                 global_metrics.remove("input_dropped_frames_resets", labels);
 544                 global_metrics.remove("input_queue_length_frames", labels);
 545                 global_metrics.remove("input_queue_duped_frames", labels);
 546
 547                 global_metrics.remove("input_has_signal_bool", labels);
 548                 global_metrics.remove("input_is_connected_bool", labels);
 549                 global_metrics.remove("input_interlaced_bool", labels);
 550                 global_metrics.remove("input_width_pixels", labels);
 551                 global_metrics.remove("input_height_pixels", labels);
 552                 global_metrics.remove("input_frame_rate_nom", labels);
 553                 global_metrics.remove("input_frame_rate_den", labels);
 554                 global_metrics.remove("input_sample_rate_hz", labels);
 555         }
 556
 557         // Register metrics.
 558         vector<pair<string, string>> labels;
 559         char card_name[64];
 560         snprintf(card_name, sizeof(card_name), "%d", card_index);
 561         labels.emplace_back("card", card_name);
 562
 563         switch (card_type) {
 564         case CardType::LIVE_CARD:
 565                 labels.emplace_back("cardtype", "live");
 566                 break;
 567         case CardType::FAKE_CAPTURE:
 568                 labels.emplace_back("cardtype", "fake");
 569                 break;
 570         case CardType::FFMPEG_INPUT:
 571                 labels.emplace_back("cardtype", "ffmpeg");
 572                 break;
 573         default:
 574                 assert(false);
 575         }
 576         card->jitter_history.register_metrics(labels);
 577         card->queue_length_policy.register_metrics(labels);
 578         global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
 579         global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
 580         global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
 581         global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
 582         global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
 583         global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
 584
 585         global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
 586         global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
 587         global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
 588         global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
 589         global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
 590         global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
 591         global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
 592         global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
 593         card->labels = labels;
 594 }
 595
 596 void Mixer::set_output_card_internal(int card_index)
 597 {
 598         // We don't really need to take card_mutex, since we're in the mixer
 599         // thread and don't mess with any queues (which is the only thing that happens
 600         // from other threads), but it's probably the safest in the long run.
 601         unique_lock<mutex> lock(card_mutex);
 602         if (output_card_index != -1) {
 603                 // Switch the old card from output to input.
 604                 CaptureCard *old_card = &cards[output_card_index];
 605                 old_card->output->end_output();
 606
 607                 // Stop the fake card that we put into place.
 608                 // This needs to _not_ happen under the mutex, to avoid deadlock
 609                 // (delivering the last frame needs to take the mutex).
 610                 CaptureInterface *fake_capture = old_card->capture.get();
 611                 lock.unlock();
 612                 fake_capture->stop_dequeue_thread();
 613                 lock.lock();
 614                 old_card->capture = move(old_card->parked_capture);  // TODO: reset the metrics
 615                 old_card->is_fake_capture = false;
 616                 old_card->capture->start_bm_capture();
 617         }
 618         if (card_index != -1) {
 619                 CaptureCard *card = &cards[card_index];
 620                 CaptureInterface *capture = card->capture.get();
 621                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 622                 // several seconds to complete (blocking on DisableVideoInput);
 623                 // see if we can maybe do it asynchronously.
 624                 lock.unlock();
 625                 capture->stop_dequeue_thread();
 626                 lock.lock();
 627                 card->parked_capture = move(card->capture);
 628                 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 629                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 630                 card->queue_length_policy.reset(card_index);
 631                 card->capture->start_bm_capture();
 632                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 633                 card->output->start_output(desired_output_video_mode, pts_int);
 634         }
 635         output_card_index = card_index;
 636         output_jitter_history.clear();
 637 }
 638
 639 namespace {
 640
 641 int unwrap_timecode(uint16_t current_wrapped, int last)
 642 {
 643         uint16_t last_wrapped = last & 0xffff;
 644         if (current_wrapped > last_wrapped) {
 645                 return (last & ~0xffff) | current_wrapped;
 646         } else {
 647                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 648         }
 649 }
 650
 651 }  // namespace
 652
 653 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 654                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 655                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 656 {
 657         DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
 658         CaptureCard *card = &cards[card_index];
 659
 660         ++card->metric_input_received_frames;
 661         card->metric_input_has_signal_bool = video_format.has_signal;
 662         card->metric_input_is_connected_bool = video_format.is_connected;
 663         card->metric_input_interlaced_bool = video_format.interlaced;
 664         card->metric_input_width_pixels = video_format.width;
 665         card->metric_input_height_pixels = video_format.height;
 666         card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
 667         card->metric_input_frame_rate_den = video_format.frame_rate_den;
 668         card->metric_input_sample_rate_hz = audio_format.sample_rate;
 669
 670         if (is_mode_scanning[card_index]) {
 671                 if (video_format.has_signal) {
 672                         // Found a stable signal, so stop scanning.
 673                         is_mode_scanning[card_index] = false;
 674                 } else {
 675                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 676                         steady_clock::time_point now = steady_clock::now();
 677                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 678                         if (sec_since_last_switch > switch_time_s) {
 679                                 // It isn't this mode; try the next one.
 680                                 mode_scanlist_index[card_index]++;
 681                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 682                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 683                                 last_mode_scan_change[card_index] = now;
 684                         }
 685                 }
 686         }
 687
 688         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 689         assert(frame_length > 0);
 690
 691         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;
 692         if (num_samples > OUTPUT_FREQUENCY / 10) {
 693                 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",
 694                         card_index, int(audio_frame.len), int(audio_offset),
 695                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 696                 if (video_frame.owner) {
 697                         video_frame.owner->release_frame(video_frame);
 698                 }
 699                 if (audio_frame.owner) {
 700                         audio_frame.owner->release_frame(audio_frame);
 701                 }
 702                 return;
 703         }
 704
 705         int dropped_frames = 0;
 706         if (card->last_timecode != -1) {
 707                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 708         }
 709
 710         // Number of samples per frame if we need to insert silence.
 711         // (Could be nonintegral, but resampling will save us then.)
 712         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 713
 714         if (dropped_frames > MAX_FPS * 2) {
 715                 fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 716                         card_index, card->last_timecode, timecode);
 717                 audio_mixer.reset_resampler(device);
 718                 dropped_frames = 0;
 719                 ++card->metric_input_resets;
 720         } else if (dropped_frames > 0) {
 721                 // Insert silence as needed.
 722                 fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 723                         card_index, dropped_frames, timecode);
 724                 card->metric_input_dropped_frames_error += dropped_frames;
 725
 726                 bool success;
 727                 do {
 728                         success = audio_mixer.add_silence(device, silence_samples, dropped_frames, frame_length);
 729                 } while (!success);
 730         }
 731
 732         if (num_samples > 0) {
 733                 audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 734         }
 735
 736         // Done with the audio, so release it.
 737         if (audio_frame.owner) {
 738                 audio_frame.owner->release_frame(audio_frame);
 739         }
 740
 741         card->last_timecode = timecode;
 742
 743         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 744
 745         size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
 746         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 747         if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
 748                 // The calculation above is wrong for planar Y'CbCr, so just override it.
 749                 assert(card->type == CardType::FFMPEG_INPUT);
 750                 assert(video_offset == 0);
 751                 expected_length = video_frame.len;
 752
 753                 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
 754                 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
 755                 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
 756                 cbcr_offset = video_format.width * video_format.height;
 757                 y_offset = 0;
 758         } else {
 759                 // All the other Y'CbCr formats are 4:2:2.
 760                 cbcr_width = video_format.width / 2;
 761                 cbcr_height = video_format.height;
 762                 cbcr_offset = video_offset / 2;
 763                 y_offset = video_frame.size / 2 + video_offset / 2;
 764         }
 765         if (video_frame.len - video_offset == 0 ||
 766             video_frame.len - video_offset != expected_length) {
 767                 if (video_frame.len != 0) {
 768                         printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n",
 769                                 card_index, video_frame.len - video_offset, expected_length);
 770                 }
 771                 if (video_frame.owner) {
 772                         video_frame.owner->release_frame(video_frame);
 773                 }
 774
 775                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 776                 // so that pts can go up accordingly.
 777                 {
 778                         unique_lock<mutex> lock(card_mutex);
 779                         CaptureCard::NewFrame new_frame;
 780                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 781                         new_frame.length = frame_length;
 782                         new_frame.interlaced = false;
 783                         new_frame.dropped_frames = dropped_frames;
 784                         new_frame.received_timestamp = video_frame.received_timestamp;
 785                         card->new_frames.push_back(move(new_frame));
 786                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
 787                 }
 788                 card->new_frames_changed.notify_all();
 789                 return;
 790         }
 791
 792         unsigned num_fields = video_format.interlaced ? 2 : 1;
 793         steady_clock::time_point frame_upload_start;
 794         bool interlaced_stride = false;
 795         if (video_format.interlaced) {
 796                 // Send the two fields along as separate frames; the other side will need to add
 797                 // a deinterlacer to actually get this right.
 798                 assert(video_format.height % 2 == 0);
 799                 video_format.height /= 2;
 800                 cbcr_height /= 2;
 801                 assert(frame_length % 2 == 0);
 802                 frame_length /= 2;
 803                 num_fields = 2;
 804                 if (video_format.second_field_start == 1) {
 805                         interlaced_stride = true;
 806                 }
 807                 frame_upload_start = steady_clock::now();
 808         }
 809         userdata->last_interlaced = video_format.interlaced;
 810         userdata->last_has_signal = video_format.has_signal;
 811         userdata->last_is_connected = video_format.is_connected;
 812         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 813         userdata->last_frame_rate_den = video_format.frame_rate_den;
 814         RefCountedFrame frame(video_frame);
 815
 816         // Upload the textures.
 817         for (unsigned field = 0; field < num_fields; ++field) {
 818                 // Put the actual texture upload in a lambda that is executed in the main thread.
 819                 // It is entirely possible to do this in the same thread (and it might even be
 820                 // faster, depending on the GPU and driver), but it appears to be trickling
 821                 // driver bugs very easily.
 822                 //
 823                 // Note that this means we must hold on to the actual frame data in <userdata>
 824                 // until the upload command is run, but we hold on to <frame> much longer than that
 825                 // (in fact, all the way until we no longer use the texture in rendering).
 826                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
 827                         unsigned field_start_line;
 828                         if (field == 1) {
 829                                 field_start_line = video_format.second_field_start;
 830                         } else {
 831                                 field_start_line = video_format.extra_lines_top;
 832                         }
 833
 834                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 835                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 836                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
 837
 838                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 839                         check_error();
 840
 841                         switch (userdata->pixel_format) {
 842                         case PixelFormat_10BitYCbCr: {
 843                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 844                                 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);
 845                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 846                                 break;
 847                         }
 848                         case PixelFormat_8BitYCbCr: {
 849                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 850                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 851
 852                                 // Make up our own strides, since we are interleaving.
 853                                 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);
 854                                 upload_texture(userdata->tex_cbcr[field], cbcr_width, cbcr_height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
 855                                 break;
 856                         }
 857                         case PixelFormat_8BitYCbCrPlanar: {
 858                                 assert(field_start_line == 0);  // We don't really support interlaced here.
 859                                 size_t field_y_start = y_offset;
 860                                 size_t field_cb_start = cbcr_offset;
 861                                 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
 862
 863                                 // Make up our own strides, since we are interleaving.
 864                                 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);
 865                                 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
 866                                 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
 867                                 break;
 868                         }
 869                         case PixelFormat_8BitBGRA: {
 870                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 871                                 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);
 872                                 // These could be asked to deliver mipmaps at any time.
 873                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 874                                 check_error();
 875                                 glGenerateMipmap(GL_TEXTURE_2D);
 876                                 check_error();
 877                                 glBindTexture(GL_TEXTURE_2D, 0);
 878                                 check_error();
 879                                 break;
 880                         }
 881                         default:
 882                                 assert(false);
 883                         }
 884
 885                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 886                         check_error();
 887                 };
 888
 889                 if (field == 1) {
 890                         // Don't upload the second field as fast as we can; wait until
 891                         // the field time has approximately passed. (Otherwise, we could
 892                         // get timing jitter against the other sources, and possibly also
 893                         // against the video display, although the latter is not as critical.)
 894                         // This requires our system clock to be reasonably close to the
 895                         // video clock, but that's not an unreasonable assumption.
 896                         steady_clock::time_point second_field_start = frame_upload_start +
 897                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 898                         this_thread::sleep_until(second_field_start);
 899                 }
 900
 901                 {
 902                         unique_lock<mutex> lock(card_mutex);
 903                         CaptureCard::NewFrame new_frame;
 904                         new_frame.frame = frame;
 905                         new_frame.length = frame_length;
 906                         new_frame.field = field;
 907                         new_frame.interlaced = video_format.interlaced;
 908                         new_frame.upload_func = upload_func;
 909                         new_frame.dropped_frames = dropped_frames;
 910                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
 911                         card->new_frames.push_back(move(new_frame));
 912                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
 913                 }
 914                 card->new_frames_changed.notify_all();
 915         }
 916 }
 917
 918 void Mixer::bm_hotplug_add(libusb_device *dev)
 919 {
 920         lock_guard<mutex> lock(hotplug_mutex);
 921         hotplugged_cards.push_back(dev);
 922 }
 923
 924 void Mixer::bm_hotplug_remove(unsigned card_index)
 925 {
 926         cards[card_index].new_frames_changed.notify_all();
 927 }
 928
 929 void Mixer::thread_func()
 930 {
 931         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
 932
 933         eglBindAPI(EGL_OPENGL_API);
 934         QOpenGLContext *context = create_context(mixer_surface);
 935         if (!make_current(context, mixer_surface)) {
 936                 printf("oops\n");
 937                 exit(1);
 938         }
 939
 940         // Start the actual capture. (We don't want to do it before we're actually ready
 941         // to process output frames.)
 942         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 943                 if (int(card_index) != output_card_index) {
 944                         cards[card_index].capture->start_bm_capture();
 945                 }
 946         }
 947
 948         steady_clock::time_point start, now;
 949         start = steady_clock::now();
 950
 951         int stats_dropped_frames = 0;
 952
 953         while (!should_quit) {
 954                 if (desired_output_card_index != output_card_index) {
 955                         set_output_card_internal(desired_output_card_index);
 956                 }
 957                 if (output_card_index != -1 &&
 958                     desired_output_video_mode != output_video_mode) {
 959                         DeckLinkOutput *output = cards[output_card_index].output.get();
 960                         output->end_output();
 961                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
 962                         output->start_output(desired_output_video_mode, pts_int);
 963                 }
 964
 965                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
 966                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 967
 968                 bool master_card_is_output;
 969                 unsigned master_card_index;
 970                 if (output_card_index != -1) {
 971                         master_card_is_output = true;
 972                         master_card_index = output_card_index;
 973                 } else {
 974                         master_card_is_output = false;
 975                         master_card_index = theme->map_signal(master_clock_channel);
 976                         assert(master_card_index < num_cards);
 977                 }
 978
 979                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
 980                 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);
 981                 stats_dropped_frames += output_frame_info.dropped_frames;
 982
 983                 handle_hotplugged_cards();
 984
 985                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
 986                         if (card_index == master_card_index || !has_new_frame[card_index]) {
 987                                 continue;
 988                         }
 989                         if (new_frames[card_index].frame->len == 0) {
 990                                 ++new_frames[card_index].dropped_frames;
 991                         }
 992                         if (new_frames[card_index].dropped_frames > 0) {
 993                                 printf("Card %u dropped %d frames before this\n",
 994                                         card_index, int(new_frames[card_index].dropped_frames));
 995                         }
 996                 }
 997
 998                 // If the first card is reporting a corrupted or otherwise dropped frame,
 999                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
1000                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
1001                         ++stats_dropped_frames;
1002                         pts_int += new_frames[master_card_index].length;
1003                         continue;
1004                 }
1005
1006                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1007                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1008                                 continue;
1009
1010                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1011                         assert(new_frame->frame != nullptr);
1012                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1013                         check_error();
1014
1015                         // The new texture might need uploading before use.
1016                         if (new_frame->upload_func) {
1017                                 new_frame->upload_func();
1018                                 new_frame->upload_func = nullptr;
1019                         }
1020                 }
1021
1022                 int64_t frame_duration = output_frame_info.frame_duration;
1023                 render_one_frame(frame_duration);
1024                 ++frame_num;
1025                 pts_int += frame_duration;
1026
1027                 now = steady_clock::now();
1028                 double elapsed = duration<double>(now - start).count();
1029
1030                 metric_frames_output_total = frame_num;
1031                 metric_frames_output_dropped = stats_dropped_frames;
1032
1033                 if (frame_num % 100 == 0) {
1034                         printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
1035                                 frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
1036                                 1e3 * elapsed / frame_num);
1037                 //      chain->print_phase_timing();
1038
1039                         // Check our memory usage, to see if we are close to our mlockall()
1040                         // limit (if at all set).
1041                         rusage used;
1042                         if (getrusage(RUSAGE_SELF, &used) == -1) {
1043                                 perror("getrusage(RUSAGE_SELF)");
1044                                 assert(false);
1045                         }
1046
1047                         if (uses_mlock) {
1048                                 rlimit limit;
1049                                 if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
1050                                         perror("getrlimit(RLIMIT_MEMLOCK)");
1051                                         assert(false);
1052                                 }
1053
1054                                 if (limit.rlim_cur == 0) {
1055                                         printf(", using %ld MB memory (locked)",
1056                                                 long(used.ru_maxrss / 1024));
1057                                 } else {
1058                                         printf(", using %ld / %ld MB lockable memory (%.1f%%)",
1059                                                 long(used.ru_maxrss / 1024),
1060                                                 long(limit.rlim_cur / 1048576),
1061                                                 float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
1062                                 }
1063                                 metrics_memory_locked_limit_bytes = limit.rlim_cur;
1064                         } else {
1065                                 printf(", using %ld MB memory (not locked)",
1066                                         long(used.ru_maxrss / 1024));
1067                                 metrics_memory_locked_limit_bytes = 0.0 / 0.0;
1068                         }
1069
1070                         printf("\n");
1071
1072                         metrics_memory_used_bytes = used.ru_maxrss * 1024;
1073                 }
1074
1075
1076                 if (should_cut.exchange(false)) {  // Test and clear.
1077                         video_encoder->do_cut(frame_num);
1078                 }
1079
1080 #if 0
1081                 // Reset every 100 frames, so that local variations in frame times
1082                 // (especially for the first few frames, when the shaders are
1083                 // compiled etc.) don't make it hard to measure for the entire
1084                 // remaining duration of the program.
1085                 if (frame == 10000) {
1086                         frame = 0;
1087                         start = now;
1088                 }
1089 #endif
1090                 check_error();
1091         }
1092
1093         resource_pool->clean_context();
1094 }
1095
1096 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1097 {
1098         if (output_card_index != -1) {
1099                 // The output card (ie., cards[output_card_index].output) is the master clock,
1100                 // so no input card (ie., cards[card_index].capture) is.
1101                 return false;
1102         }
1103         return (card_index == master_card_index);
1104 }
1105
1106 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1107 {
1108         // Count the number of frames in the queue, including any frames
1109         // we dropped. It's hard to know exactly how we should deal with
1110         // dropped (corrupted) input frames; they don't help our goal of
1111         // avoiding starvation, but they still add to the problem of latency.
1112         // Since dropped frames is going to mean a bump in the signal anyway,
1113         // we err on the side of having more stable latency instead.
1114         unsigned queue_length = 0;
1115         for (const CaptureCard::NewFrame &frame : card->new_frames) {
1116                 queue_length += frame.dropped_frames + 1;
1117         }
1118
1119         // If needed, drop frames until the queue is below the safe limit.
1120         // We prefer to drop from the head, because all else being equal,
1121         // we'd like more recent frames (less latency).
1122         unsigned dropped_frames = 0;
1123         while (queue_length > safe_queue_length) {
1124                 assert(!card->new_frames.empty());
1125                 assert(queue_length > card->new_frames.front().dropped_frames);
1126                 queue_length -= card->new_frames.front().dropped_frames;
1127
1128                 if (queue_length <= safe_queue_length) {
1129                         // No need to drop anything.
1130                         break;
1131                 }
1132
1133                 card->new_frames.pop_front();
1134                 card->new_frames_changed.notify_all();
1135                 --queue_length;
1136                 ++dropped_frames;
1137         }
1138
1139         card->metric_input_dropped_frames_jitter += dropped_frames;
1140         card->metric_input_queue_length_frames = queue_length;
1141
1142 #if 0
1143         if (dropped_frames > 0) {
1144                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1145                         card_index, dropped_frames);
1146         }
1147 #endif
1148 }
1149
1150
1151 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])
1152 {
1153         OutputFrameInfo output_frame_info;
1154 start:
1155         unique_lock<mutex> lock(card_mutex, defer_lock);
1156         if (master_card_is_output) {
1157                 // Clocked to the output, so wait for it to be ready for the next frame.
1158                 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);
1159                 lock.lock();
1160         } else {
1161                 // Wait for the master card to have a new frame.
1162                 // TODO: Add a timeout.
1163                 output_frame_info.is_preroll = false;
1164                 lock.lock();
1165                 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(); });
1166         }
1167
1168         if (master_card_is_output) {
1169                 handle_hotplugged_cards();
1170         } else if (cards[master_card_index].new_frames.empty()) {
1171                 // We were woken up, but not due to a new frame. Deal with it
1172                 // and then restart.
1173                 assert(cards[master_card_index].capture->get_disconnected());
1174                 handle_hotplugged_cards();
1175                 lock.unlock();
1176                 goto start;
1177         }
1178
1179         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1180                 CaptureCard *card = &cards[card_index];
1181                 if (card->new_frames.empty()) {  // Starvation.
1182                         ++card->metric_input_duped_frames;
1183                 } else {
1184                         new_frames[card_index] = move(card->new_frames.front());
1185                         has_new_frame[card_index] = true;
1186                         card->new_frames.pop_front();
1187                         card->new_frames_changed.notify_all();
1188                 }
1189         }
1190
1191         if (!master_card_is_output) {
1192                 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1193                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1194                 output_frame_info.frame_duration = new_frames[master_card_index].length;
1195         }
1196
1197         if (!output_frame_info.is_preroll) {
1198                 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1199         }
1200
1201         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) {
1202                 CaptureCard *card = &cards[card_index];
1203                 if (has_new_frame[card_index] &&
1204                     !input_card_is_master_clock(card_index, master_card_index) &&
1205                     !output_frame_info.is_preroll) {
1206                         card->queue_length_policy.update_policy(
1207                                 output_frame_info.frame_timestamp,
1208                                 card->jitter_history.get_expected_next_frame(),
1209                                 output_frame_info.frame_duration,
1210                                 card->jitter_history.estimate_max_jitter(),
1211                                 output_jitter_history.estimate_max_jitter());
1212                         trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1213                                                   card->queue_length_policy.get_safe_queue_length()));
1214                 }
1215         }
1216
1217         // This might get off by a fractional sample when changing master card
1218         // between ones with different frame rates, but that's fine.
1219         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1220         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1221         fractional_samples = num_samples_times_timebase % TIMEBASE;
1222         assert(output_frame_info.num_samples >= 0);
1223
1224         return output_frame_info;
1225 }
1226
1227 void Mixer::handle_hotplugged_cards()
1228 {
1229         // Check for cards that have been disconnected since last frame.
1230         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1231                 CaptureCard *card = &cards[card_index];
1232                 if (card->capture->get_disconnected()) {
1233                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1234                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1235                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1236                         card->queue_length_policy.reset(card_index);
1237                         card->capture->start_bm_capture();
1238                 }
1239         }
1240
1241         // Check for cards that have been connected since last frame.
1242         vector<libusb_device *> hotplugged_cards_copy;
1243         {
1244                 lock_guard<mutex> lock(hotplug_mutex);
1245                 swap(hotplugged_cards, hotplugged_cards_copy);
1246         }
1247         for (libusb_device *new_dev : hotplugged_cards_copy) {
1248                 // Look for a fake capture card where we can stick this in.
1249                 int free_card_index = -1;
1250                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1251                         if (cards[card_index].is_fake_capture) {
1252                                 free_card_index = card_index;
1253                                 break;
1254                         }
1255                 }
1256
1257                 if (free_card_index == -1) {
1258                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1259                         libusb_unref_device(new_dev);
1260                 } else {
1261                         // BMUSBCapture takes ownership.
1262                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1263                         CaptureCard *card = &cards[free_card_index];
1264                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1265                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1266                         card->queue_length_policy.reset(free_card_index);
1267                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1268                         capture->start_bm_capture();
1269                 }
1270         }
1271 }
1272
1273
1274 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)
1275 {
1276         // Resample the audio as needed, including from previously dropped frames.
1277         assert(num_cards > 0);
1278         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1279                 const bool dropped_frame = (frame_num != dropped_frames);
1280                 {
1281                         // Signal to the audio thread to process this frame.
1282                         // Note that if the frame is a dropped frame, we signal that
1283                         // we don't want to use this frame as base for adjusting
1284                         // the resampler rate. The reason for this is that the timing
1285                         // of these frames is often way too late; they typically don't
1286                         // “arrive” before we synthesize them. Thus, we could end up
1287                         // in a situation where we have inserted e.g. five audio frames
1288                         // into the queue before we then start pulling five of them
1289                         // back out. This makes ResamplingQueue overestimate the delay,
1290                         // causing undue resampler changes. (We _do_ use the last,
1291                         // non-dropped frame; perhaps we should just discard that as well,
1292                         // since dropped frames are expected to be rare, and it might be
1293                         // better to just wait until we have a slightly more normal situation).
1294                         unique_lock<mutex> lock(audio_mutex);
1295                         bool adjust_rate = !dropped_frame && !is_preroll;
1296                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1297                         audio_task_queue_changed.notify_one();
1298                 }
1299                 if (dropped_frame) {
1300                         // For dropped frames, increase the pts. Note that if the format changed
1301                         // in the meantime, we have no way of detecting that; we just have to
1302                         // assume the frame length is always the same.
1303                         pts_int += length_per_frame;
1304                 }
1305         }
1306 }
1307
1308 void Mixer::render_one_frame(int64_t duration)
1309 {
1310         // Determine the time code for this frame before we start rendering.
1311         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1312         if (display_timecode_on_stdout) {
1313                 printf("Timecode: '%s'\n", timecode_text.c_str());
1314         }
1315
1316         // Update Y'CbCr settings for all cards.
1317         {
1318                 unique_lock<mutex> lock(card_mutex);
1319                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1320                         YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1321                         input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1322                         input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1323                         input_state.full_range[card_index] = interpretation->full_range;
1324                 }
1325         }
1326
1327         // Get the main chain from the theme, and set its state immediately.
1328         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1329         EffectChain *chain = theme_main_chain.chain;
1330         theme_main_chain.setup_chain();
1331         //theme_main_chain.chain->enable_phase_timing(true);
1332
1333         // The theme can't (or at least shouldn't!) call connect_signal() on
1334         // each FFmpeg input, so we'll do it here.
1335         for (const pair<LiveInputWrapper *, FFmpegCapture *> &conn : theme->get_signal_connections()) {
1336                 conn.first->connect_signal_raw(conn.second->get_card_index(), input_state);
1337         }
1338
1339         // If HDMI/SDI output is active and the user has requested auto mode,
1340         // its mode overrides the existing Y'CbCr setting for the chain.
1341         YCbCrLumaCoefficients ycbcr_output_coefficients;
1342         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1343                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1344         } else {
1345                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1346         }
1347
1348         // TODO: Reduce the duplication against theme.cpp.
1349         YCbCrFormat output_ycbcr_format;
1350         output_ycbcr_format.chroma_subsampling_x = 1;
1351         output_ycbcr_format.chroma_subsampling_y = 1;
1352         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1353         output_ycbcr_format.full_range = false;
1354         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1355         chain->change_ycbcr_output_format(output_ycbcr_format);
1356
1357         // Render main chain. If we're using zerocopy Quick Sync encoding
1358         // (the default case), we take an extra copy of the created outputs,
1359         // so that we can display it back to the screen later (it's less memory
1360         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1361         // Ideally, we'd like to avoid taking copies and just use the main textures
1362         // for display as well, but they're just views into VA-API memory and must be
1363         // unmapped during encoding, so we can't use them for display, unfortunately.
1364         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1365         GLuint y_copy_tex, cbcr_copy_tex = 0;
1366         GLuint y_display_tex, cbcr_display_tex;
1367         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1368         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1369         const bool is_zerocopy = video_encoder->is_zerocopy();
1370         if (is_zerocopy) {
1371                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1372                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1373                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1374
1375                 y_display_tex = y_copy_tex;
1376                 cbcr_display_tex = cbcr_copy_tex;
1377
1378                 // y_tex and cbcr_tex will be given by VideoEncoder.
1379         } else {
1380                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1381                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1382                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1383
1384                 y_display_tex = y_tex;
1385                 cbcr_display_tex = cbcr_tex;
1386         }
1387
1388         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1389         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1390         assert(got_frame);
1391
1392         GLuint fbo;
1393         if (is_zerocopy) {
1394                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1395         } else {
1396                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1397         }
1398         check_error();
1399         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1400
1401         if (display_timecode_in_stream) {
1402                 // Render the timecode on top.
1403                 timecode_renderer->render_timecode(fbo, timecode_text);
1404         }
1405
1406         resource_pool->release_fbo(fbo);
1407
1408         if (is_zerocopy) {
1409                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1410         } else {
1411                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1412         }
1413         if (output_card_index != -1) {
1414                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1415         }
1416         resource_pool->release_2d_texture(cbcr_full_tex);
1417
1418         // Set the right state for the Y' and CbCr textures we use for display.
1419         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1420         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1421         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1422         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1423         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1424
1425         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1426         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1427         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1428         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1429
1430         RefCountedGLsync fence = video_encoder->end_frame();
1431
1432         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1433         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1434         DisplayFrame live_frame;
1435         live_frame.chain = display_chain.get();
1436         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1437                 display_input->set_texture_num(0, y_display_tex);
1438                 display_input->set_texture_num(1, cbcr_display_tex);
1439         };
1440         live_frame.ready_fence = fence;
1441         live_frame.input_frames = {};
1442         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1443         output_channel[OUTPUT_LIVE].output_frame(live_frame);
1444
1445         // Set up preview and any additional channels.
1446         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1447                 DisplayFrame display_frame;
1448                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1449                 display_frame.chain = chain.chain;
1450                 display_frame.setup_chain = chain.setup_chain;
1451                 display_frame.ready_fence = fence;
1452                 display_frame.input_frames = chain.input_frames;
1453                 display_frame.temp_textures = {};
1454                 output_channel[i].output_frame(display_frame);
1455         }
1456 }
1457
1458 void Mixer::audio_thread_func()
1459 {
1460         pthread_setname_np(pthread_self(), "Mixer_Audio");
1461
1462         while (!should_quit) {
1463                 AudioTask task;
1464
1465                 {
1466                         unique_lock<mutex> lock(audio_mutex);
1467                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1468                         if (should_quit) {
1469                                 return;
1470                         }
1471                         task = audio_task_queue.front();
1472                         audio_task_queue.pop();
1473                 }
1474
1475                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1476                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1477                 vector<float> samples_out = audio_mixer.get_output(
1478                         task.frame_timestamp,
1479                         task.num_samples,
1480                         rate_adjustment_policy);
1481
1482                 // Send the samples to the sound card, then add them to the output.
1483                 if (alsa) {
1484                         alsa->write(samples_out);
1485                 }
1486                 if (output_card_index != -1) {
1487                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1488                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1489                 }
1490                 video_encoder->add_audio(task.pts_int, move(samples_out));
1491         }
1492 }
1493
1494 void Mixer::release_display_frame(DisplayFrame *frame)
1495 {
1496         for (GLuint texnum : frame->temp_textures) {
1497                 resource_pool->release_2d_texture(texnum);
1498         }
1499         frame->temp_textures.clear();
1500         frame->ready_fence.reset();
1501         frame->input_frames.clear();
1502 }
1503
1504 void Mixer::start()
1505 {
1506         mixer_thread = thread(&Mixer::thread_func, this);
1507         audio_thread = thread(&Mixer::audio_thread_func, this);
1508 }
1509
1510 void Mixer::quit()
1511 {
1512         should_quit = true;
1513         audio_task_queue_changed.notify_one();
1514         mixer_thread.join();
1515         audio_thread.join();
1516 }
1517
1518 void Mixer::transition_clicked(int transition_num)
1519 {
1520         theme->transition_clicked(transition_num, pts());
1521 }
1522
1523 void Mixer::channel_clicked(int preview_num)
1524 {
1525         theme->channel_clicked(preview_num);
1526 }
1527
1528 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1529 {
1530         unique_lock<mutex> lock(card_mutex);
1531         return ycbcr_interpretation[card_index];
1532 }
1533
1534 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1535 {
1536         unique_lock<mutex> lock(card_mutex);
1537         ycbcr_interpretation[card_index] = interpretation;
1538 }
1539
1540 void Mixer::start_mode_scanning(unsigned card_index)
1541 {
1542         assert(card_index < num_cards);
1543         if (is_mode_scanning[card_index]) {
1544                 return;
1545         }
1546         is_mode_scanning[card_index] = true;
1547         mode_scanlist[card_index].clear();
1548         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1549                 mode_scanlist[card_index].push_back(mode.first);
1550         }
1551         assert(!mode_scanlist[card_index].empty());
1552         mode_scanlist_index[card_index] = 0;
1553         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1554         last_mode_scan_change[card_index] = steady_clock::now();
1555 }
1556
1557 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1558 {
1559         assert(desired_output_card_index != -1);
1560         unique_lock<mutex> lock(card_mutex);
1561         return cards[desired_output_card_index].output->get_available_video_modes();
1562 }
1563
1564 Mixer::OutputChannel::~OutputChannel()
1565 {
1566         if (has_current_frame) {
1567                 parent->release_display_frame(&current_frame);
1568         }
1569         if (has_ready_frame) {
1570                 parent->release_display_frame(&ready_frame);
1571         }
1572 }
1573
1574 void Mixer::OutputChannel::output_frame(DisplayFrame frame)
1575 {
1576         // Store this frame for display. Remove the ready frame if any
1577         // (it was seemingly never used).
1578         {
1579                 unique_lock<mutex> lock(frame_mutex);
1580                 if (has_ready_frame) {
1581                         parent->release_display_frame(&ready_frame);
1582                 }
1583                 ready_frame = frame;
1584                 has_ready_frame = true;
1585
1586                 // Call the callbacks under the mutex (they should be short),
1587                 // so that we don't race against a callback removal.
1588                 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1589                         key_and_callback.second();
1590                 }
1591         }
1592
1593         // Reduce the number of callbacks by filtering duplicates. The reason
1594         // why we bother doing this is that Qt seemingly can get into a state
1595         // where its builds up an essentially unbounded queue of signals,
1596         // consuming more and more memory, and there's no good way of collapsing
1597         // user-defined signals or limiting the length of the queue.
1598         if (transition_names_updated_callback) {
1599                 vector<string> transition_names = global_mixer->get_transition_names();
1600                 bool changed = false;
1601                 if (transition_names.size() != last_transition_names.size()) {
1602                         changed = true;
1603                 } else {
1604                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1605                                 if (transition_names[i] != last_transition_names[i]) {
1606                                         changed = true;
1607                                         break;
1608                                 }
1609                         }
1610                 }
1611                 if (changed) {
1612                         transition_names_updated_callback(transition_names);
1613                         last_transition_names = transition_names;
1614                 }
1615         }
1616         if (name_updated_callback) {
1617                 string name = global_mixer->get_channel_name(channel);
1618                 if (name != last_name) {
1619                         name_updated_callback(name);
1620                         last_name = name;
1621                 }
1622         }
1623         if (color_updated_callback) {
1624                 string color = global_mixer->get_channel_color(channel);
1625                 if (color != last_color) {
1626                         color_updated_callback(color);
1627                         last_color = color;
1628                 }
1629         }
1630 }
1631
1632 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1633 {
1634         unique_lock<mutex> lock(frame_mutex);
1635         if (!has_current_frame && !has_ready_frame) {
1636                 return false;
1637         }
1638
1639         if (has_current_frame && has_ready_frame) {
1640                 // We have a new ready frame. Toss the current one.
1641                 parent->release_display_frame(&current_frame);
1642                 has_current_frame = false;
1643         }
1644         if (has_ready_frame) {
1645                 assert(!has_current_frame);
1646                 current_frame = ready_frame;
1647                 ready_frame.ready_fence.reset();  // Drop the refcount.
1648                 ready_frame.input_frames.clear();  // Drop the refcounts.
1649                 has_current_frame = true;
1650                 has_ready_frame = false;
1651         }
1652
1653         *frame = current_frame;
1654         return true;
1655 }
1656
1657 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1658 {
1659         unique_lock<mutex> lock(frame_mutex);
1660         new_frame_ready_callbacks[key] = callback;
1661 }
1662
1663 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1664 {
1665         unique_lock<mutex> lock(frame_mutex);
1666         new_frame_ready_callbacks.erase(key);
1667 }
1668
1669 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1670 {
1671         transition_names_updated_callback = callback;
1672 }
1673
1674 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1675 {
1676         name_updated_callback = callback;
1677 }
1678
1679 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1680 {
1681         color_updated_callback = callback;
1682 }
1683
1684 mutex RefCountedGLsync::fence_lock;