git.sesse.net Git - nageru/blob - nageru/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 <algorithm>
  18 #include <chrono>
  19 #include <condition_variable>
  20 #include <cstddef>
  21 #include <cstdint>
  22 #include <memory>
  23 #include <mutex>
  24 #include <ratio>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29
  30 #include "DeckLinkAPI.h"
  31 #include "LinuxCOM.h"
  32 #include "alsa_output.h"
  33 #include "basic_stats.h"
  34 #include "bmusb/bmusb.h"
  35 #include "bmusb/fake_capture.h"
  36 #ifdef HAVE_CEF
  37 #include "cef_capture.h"
  38 #endif
  39 #include "chroma_subsampler.h"
  40 #include "shared/context.h"
  41 #include "decklink_capture.h"
  42 #include "decklink_output.h"
  43 #include "defs.h"
  44 #include "shared/disk_space_estimator.h"
  45 #include "ffmpeg_capture.h"
  46 #include "flags.h"
  47 #include "image_input.h"
  48 #include "input_mapping.h"
  49 #include "shared/metrics.h"
  50 #include "mjpeg_encoder.h"
  51 #include "pbo_frame_allocator.h"
  52 #include "shared/ref_counted_gl_sync.h"
  53 #include "resampling_queue.h"
  54 #include "shared/timebase.h"
  55 #include "timecode_renderer.h"
  56 #include "v210_converter.h"
  57 #include "va_display_with_cleanup.h"
  58 #include "video_encoder.h"
  59
  60 #undef Status
  61 #include <google/protobuf/util/json_util.h>
  62 #include "json.pb.h"
  63
  64 class IDeckLink;
  65 class QOpenGLContext;
  66
  67 using namespace movit;
  68 using namespace std;
  69 using namespace std::chrono;
  70 using namespace std::placeholders;
  71 using namespace bmusb;
  72
  73 Mixer *global_mixer = nullptr;
  74
  75 namespace {
  76
  77 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  78 {
  79         if (interlaced) {
  80                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  81                         input_state->buffered_frames[card_index][frame_num] =
  82                                 input_state->buffered_frames[card_index][frame_num - 1];
  83                 }
  84                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  85         } else {
  86                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  87                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  88                 }
  89         }
  90 }
  91
  92 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
  93 {
  94         bool first;
  95         switch (userdata->pixel_format) {
  96         case PixelFormat_10BitYCbCr:
  97                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  98                 break;
  99         case PixelFormat_8BitYCbCr:
 100                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
 101                 break;
 102         case PixelFormat_8BitBGRA:
 103                 first = userdata->tex_rgba[field] == 0;
 104                 break;
 105         case PixelFormat_8BitYCbCrPlanar:
 106                 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
 107                 break;
 108         default:
 109                 assert(false);
 110         }
 111
 112         if (first ||
 113             width != userdata->last_width[field] ||
 114             height != userdata->last_height[field] ||
 115             cbcr_width != userdata->last_cbcr_width[field] ||
 116             cbcr_height != userdata->last_cbcr_height[field]) {
 117                 // We changed resolution since last use of this texture, so we need to create
 118                 // a new object. Note that this each card has its own PBOFrameAllocator,
 119                 // we don't need to worry about these flip-flopping between resolutions.
 120                 switch (userdata->pixel_format) {
 121                 case PixelFormat_10BitYCbCr:
 122                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 123                         check_error();
 124                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 125                         check_error();
 126                         break;
 127                 case PixelFormat_8BitYCbCr: {
 128                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 129                         check_error();
 130                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 131                         check_error();
 132                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 133                         check_error();
 134                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 135                         check_error();
 136                         break;
 137                 }
 138                 case PixelFormat_8BitYCbCrPlanar: {
 139                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 140                         check_error();
 141                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 142                         check_error();
 143                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
 144                         check_error();
 145                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 146                         check_error();
 147                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
 148                         check_error();
 149                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 150                         check_error();
 151                         break;
 152                 }
 153                 case PixelFormat_8BitBGRA:
 154                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 155                         check_error();
 156                         // NOTE: sRGB may be disabled by sRGBSwitchingFlatInput.
 157                         glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 158                         check_error();
 159                         break;
 160                 default:
 161                         assert(false);
 162                 }
 163                 userdata->last_width[field] = width;
 164                 userdata->last_height[field] = height;
 165                 userdata->last_cbcr_width[field] = cbcr_width;
 166                 userdata->last_cbcr_height[field] = cbcr_height;
 167         }
 168         if (global_flags.ten_bit_input &&
 169             (first || v210_width != userdata->last_v210_width[field])) {
 170                 // Same as above; we need to recreate the texture.
 171                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 172                 check_error();
 173                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 174                 check_error();
 175                 userdata->last_v210_width[field] = v210_width;
 176         }
 177 }
 178
 179 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 180 {
 181         if (interlaced_stride) {
 182                 stride *= 2;
 183         }
 184         if (global_flags.flush_pbos) {
 185                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 186                 check_error();
 187         }
 188
 189         glBindTexture(GL_TEXTURE_2D, tex);
 190         check_error();
 191         if (interlaced_stride) {
 192                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 193                 check_error();
 194         } else {
 195                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 196                 check_error();
 197         }
 198
 199         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 200         check_error();
 201         glBindTexture(GL_TEXTURE_2D, 0);
 202         check_error();
 203         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 204         check_error();
 205 }
 206
 207 }  // namespace
 208
 209 void JitterHistory::register_metrics(const vector<pair<string, string>> &labels)
 210 {
 211         global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames);
 212         global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE);
 213 }
 214
 215 void JitterHistory::unregister_metrics(const vector<pair<string, string>> &labels)
 216 {
 217         global_metrics.remove("input_underestimated_jitter_frames", labels);
 218         global_metrics.remove("input_estimated_max_jitter_seconds", labels);
 219 }
 220
 221 void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames)
 222 {
 223         if (expected_timestamp > steady_clock::time_point::min()) {
 224                 expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 225                 double jitter_seconds = fabs(duration<double>(expected_timestamp - now).count());
 226                 history.push_back(orders.insert(jitter_seconds));
 227                 if (jitter_seconds > estimate_max_jitter()) {
 228                         ++metric_input_underestimated_jitter_frames;
 229                 }
 230
 231                 metric_input_estimated_max_jitter_seconds = estimate_max_jitter();
 232
 233                 if (history.size() > history_length) {
 234                         orders.erase(history.front());
 235                         history.pop_front();
 236                 }
 237                 assert(history.size() <= history_length);
 238         }
 239         expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 240 }
 241
 242 double JitterHistory::estimate_max_jitter() const
 243 {
 244         if (orders.empty()) {
 245                 return 0.0;
 246         }
 247         size_t elem_idx = lrint((orders.size() - 1) * percentile);
 248         if (percentile <= 0.5) {
 249                 return *next(orders.begin(), elem_idx) * multiplier;
 250         } else {
 251                 return *prev(orders.end(), orders.size() - elem_idx) * multiplier;
 252         }
 253 }
 254
 255 void QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
 256 {
 257         global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
 258 }
 259
 260 void QueueLengthPolicy::unregister_metrics(const vector<pair<string, string>> &labels)
 261 {
 262         global_metrics.remove("input_queue_safe_length_frames", labels);
 263 }
 264
 265 void QueueLengthPolicy::update_policy(steady_clock::time_point now,
 266                                       steady_clock::time_point expected_next_frame,
 267                                       int64_t input_frame_duration,
 268                                       int64_t master_frame_duration,
 269                                       double max_input_card_jitter_seconds,
 270                                       double max_master_card_jitter_seconds)
 271 {
 272         double input_frame_duration_seconds = input_frame_duration / double(TIMEBASE);
 273         double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
 274
 275         // Figure out when we can expect the next frame for this card, assuming
 276         // worst-case jitter (ie., the frame is maximally late).
 277         double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
 278
 279         // How many times are the master card expected to tick in that time?
 280         // We assume the master clock has worst-case jitter but not any rate
 281         // discrepancy, ie., it ticks as early as possible every time, but not
 282         // cumulatively.
 283         double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
 284
 285         // As a special case, if the master card ticks faster than the input card,
 286         // we expect the queue to drain by itself even without dropping. But if
 287         // the difference is small (e.g. 60 Hz master and 59.94 input), it would
 288         // go slowly enough that the effect wouldn't really be appreciable.
 289         // We account for this by looking at the situation five frames ahead,
 290         // assuming everything else is the same.
 291         double frames_allowed;
 292         if (master_frame_duration < input_frame_duration) {
 293                 frames_allowed = frames_needed + 5 * (input_frame_duration_seconds - master_frame_duration_seconds) / master_frame_duration_seconds;
 294         } else {
 295                 frames_allowed = frames_needed;
 296         }
 297
 298         safe_queue_length = max<int>(floor(frames_allowed), 0);
 299         metric_input_queue_safe_length_frames = safe_queue_length;
 300 }
 301
 302 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 303         : httpd(),
 304           num_cards(num_cards),
 305           mixer_surface(create_surface(format)),
 306           h264_encoder_surface(create_surface(format)),
 307           decklink_output_surface(create_surface(format)),
 308           image_update_surface(create_surface(format))
 309 {
 310         memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
 311         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 312         check_error();
 313
 314         if (!epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") ||
 315             !epoxy_has_gl_extension("GL_ARB_sampler_objects")) {
 316                 fprintf(stderr, "Nageru requires GL_EXT_texture_sRGB_decode and GL_ARB_sampler_objects to run.\n");
 317                 exit(1);
 318         }
 319
 320         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 321         // will be halved when sampling them, and we need to compensate here.
 322         movit_texel_subpixel_precision /= 2.0;
 323
 324         resource_pool.reset(new ResourcePool);
 325         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 326                 output_channel[i].parent = this;
 327                 output_channel[i].channel = i;
 328         }
 329
 330         ImageFormat inout_format;
 331         inout_format.color_space = COLORSPACE_sRGB;
 332         inout_format.gamma_curve = GAMMA_sRGB;
 333
 334         // Matches the 4:2:0 format created by the main chain.
 335         YCbCrFormat ycbcr_format;
 336         ycbcr_format.chroma_subsampling_x = 2;
 337         ycbcr_format.chroma_subsampling_y = 2;
 338         if (global_flags.ycbcr_rec709_coefficients) {
 339                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 340         } else {
 341                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 342         }
 343         ycbcr_format.full_range = false;
 344         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 345         ycbcr_format.cb_x_position = 0.0f;
 346         ycbcr_format.cr_x_position = 0.0f;
 347         ycbcr_format.cb_y_position = 0.5f;
 348         ycbcr_format.cr_y_position = 0.5f;
 349
 350         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 351         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 352         check_error();
 353         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 354         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 355         display_chain->add_input(display_input);
 356         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 357         display_chain->set_dither_bits(0);  // Don't bother.
 358         display_chain->finalize();
 359
 360         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));
 361         if (!global_flags.card_to_mjpeg_stream_export.empty()) {
 362                 mjpeg_encoder.reset(new MJPEGEncoder(&httpd, global_flags.va_display));
 363         }
 364
 365         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 366         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 367
 368         // Must be instantiated after the theme, as the theme decides the number of FFmpeg inputs.
 369         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 370         audio_mixer.reset(new AudioMixer(num_cards, video_inputs.size()));
 371
 372         httpd.add_endpoint("/channels", bind(&Mixer::get_channels_json, this), HTTPD::ALLOW_ALL_ORIGINS);
 373         for (int channel_idx = 0; channel_idx < theme->get_num_channels(); ++channel_idx) {
 374                 char url[256];
 375                 snprintf(url, sizeof(url), "/channels/%d/color", channel_idx + 2);
 376                 httpd.add_endpoint(url, bind(&Mixer::get_channel_color_http, this, unsigned(channel_idx + 2)), HTTPD::ALLOW_ALL_ORIGINS);
 377         }
 378
 379         // Start listening for clients only once VideoEncoder has written its header, if any.
 380         httpd.start(global_flags.http_port);
 381
 382         // First try initializing the then PCI devices, then USB, then
 383         // fill up with fake cards until we have the desired number of cards.
 384         unsigned num_pci_devices = 0;
 385         unsigned card_index = 0;
 386
 387         {
 388                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 389                 if (decklink_iterator != nullptr) {
 390                         for ( ; card_index < num_cards; ++card_index) {
 391                                 IDeckLink *decklink;
 392                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 393                                         break;
 394                                 }
 395
 396                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 397                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 398                                 if (!output->set_device(decklink)) {
 399                                         delete output;
 400                                         output = nullptr;
 401                                 }
 402                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 403                                 ++num_pci_devices;
 404                         }
 405                         decklink_iterator->Release();
 406                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 407                 } else {
 408                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 409                 }
 410         }
 411
 412         unsigned num_usb_devices = BMUSBCapture::num_cards();
 413         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 414                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 415                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 416                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 417         }
 418         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 419
 420         unsigned num_fake_cards = 0;
 421         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 422                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 423                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 424         }
 425
 426         if (num_fake_cards > 0) {
 427                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 428         }
 429
 430         // Initialize all video inputs the theme asked for. Note that these are
 431         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 432         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 433                 if (card_index >= MAX_VIDEO_CARDS) {
 434                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 435                         abort();
 436                 }
 437                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 438                 video_inputs[video_card_index]->set_card_index(card_index);
 439         }
 440         num_video_inputs = video_inputs.size();
 441
 442 #ifdef HAVE_CEF
 443         // Same, for HTML inputs.
 444         std::vector<CEFCapture *> html_inputs = theme->get_html_inputs();
 445         for (unsigned html_card_index = 0; html_card_index < html_inputs.size(); ++card_index, ++html_card_index) {
 446                 if (card_index >= MAX_VIDEO_CARDS) {
 447                         fprintf(stderr, "ERROR: Not enough card slots available for the HTML inputs the theme requested.\n");
 448                         abort();
 449                 }
 450                 configure_card(card_index, html_inputs[html_card_index], CardType::CEF_INPUT, /*output=*/nullptr);
 451                 html_inputs[html_card_index]->set_card_index(card_index);
 452         }
 453         num_html_inputs = html_inputs.size();
 454 #endif
 455
 456         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 457         BMUSBCapture::start_bm_thread();
 458
 459         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 460                 cards[card_index].queue_length_policy.reset(card_index);
 461         }
 462
 463         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 464
 465         if (global_flags.ten_bit_input) {
 466                 if (!v210Converter::has_hardware_support()) {
 467                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 468                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 469                         abort();
 470                 }
 471                 v210_converter.reset(new v210Converter());
 472
 473                 // These are all the widths listed in the Blackmagic SDK documentation
 474                 // (section 2.7.3, “Display Modes”).
 475                 v210_converter->precompile_shader(720);
 476                 v210_converter->precompile_shader(1280);
 477                 v210_converter->precompile_shader(1920);
 478                 v210_converter->precompile_shader(2048);
 479                 v210_converter->precompile_shader(3840);
 480                 v210_converter->precompile_shader(4096);
 481         }
 482         if (global_flags.ten_bit_output) {
 483                 if (!v210Converter::has_hardware_support()) {
 484                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 485                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 486                         abort();
 487                 }
 488         }
 489
 490         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 491         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 492         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 493
 494         if (global_flags.enable_alsa_output) {
 495                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 496         }
 497         if (global_flags.output_card != -1) {
 498                 desired_output_card_index = global_flags.output_card;
 499                 set_output_card_internal(global_flags.output_card);
 500         }
 501
 502         output_jitter_history.register_metrics({{ "card", "output" }});
 503
 504         ImageInput::start_update_thread(image_update_surface);
 505 }
 506
 507 Mixer::~Mixer()
 508 {
 509         ImageInput::end_update_thread();
 510
 511         if (mjpeg_encoder != nullptr) {
 512                 mjpeg_encoder->stop();
 513         }
 514         httpd.stop();
 515         BMUSBCapture::stop_bm_thread();
 516
 517         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 518                 cards[card_index].capture->stop_dequeue_thread();
 519                 if (cards[card_index].output) {
 520                         cards[card_index].output->end_output();
 521                         cards[card_index].output.reset();
 522                 }
 523         }
 524
 525         video_encoder.reset(nullptr);
 526 }
 527
 528 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 529 {
 530         printf("Configuring card %d...\n", card_index);
 531
 532         CaptureCard *card = &cards[card_index];
 533         if (card->capture != nullptr) {
 534                 card->capture->stop_dequeue_thread();
 535         }
 536         card->capture.reset(capture);
 537         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 538         card->is_cef_capture = (card_type == CardType::CEF_INPUT);
 539         card->may_have_dropped_last_frame = false;
 540         card->type = card_type;
 541         if (card->output.get() != output) {
 542                 card->output.reset(output);
 543         }
 544
 545         PixelFormat pixel_format;
 546         if (card_type == CardType::FFMPEG_INPUT) {
 547                 pixel_format = capture->get_current_pixel_format();
 548         } else if (card_type == CardType::CEF_INPUT) {
 549                 pixel_format = PixelFormat_8BitBGRA;
 550         } else if (global_flags.ten_bit_input) {
 551                 pixel_format = PixelFormat_10BitYCbCr;
 552         } else {
 553                 pixel_format = PixelFormat_8BitYCbCr;
 554         }
 555
 556         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 557         if (card->frame_allocator == nullptr) {
 558                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height, card_index, mjpeg_encoder.get()));  // 8 MB.
 559         }
 560         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 561         if (card->surface == nullptr) {
 562                 card->surface = create_surface_with_same_format(mixer_surface);
 563         }
 564         while (!card->new_frames.empty()) card->new_frames.pop_front();
 565         card->last_timecode = -1;
 566         card->capture->set_pixel_format(pixel_format);
 567         card->capture->configure_card();
 568
 569         // NOTE: start_bm_capture() happens in thread_func().
 570
 571         DeviceSpec device;
 572         if (card_type == CardType::FFMPEG_INPUT) {
 573                 device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 574         } else {
 575                 device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 576         }
 577         audio_mixer->reset_resampler(device);
 578         audio_mixer->set_display_name(device, card->capture->get_description());
 579         audio_mixer->trigger_state_changed_callback();
 580
 581         // Unregister old metrics, if any.
 582         if (!card->labels.empty()) {
 583                 const vector<pair<string, string>> &labels = card->labels;
 584                 card->jitter_history.unregister_metrics(labels);
 585                 card->queue_length_policy.unregister_metrics(labels);
 586                 global_metrics.remove("input_received_frames", labels);
 587                 global_metrics.remove("input_dropped_frames_jitter", labels);
 588                 global_metrics.remove("input_dropped_frames_error", labels);
 589                 global_metrics.remove("input_dropped_frames_resets", labels);
 590                 global_metrics.remove("input_queue_length_frames", labels);
 591                 global_metrics.remove("input_queue_duped_frames", labels);
 592
 593                 global_metrics.remove("input_has_signal_bool", labels);
 594                 global_metrics.remove("input_is_connected_bool", labels);
 595                 global_metrics.remove("input_interlaced_bool", labels);
 596                 global_metrics.remove("input_width_pixels", labels);
 597                 global_metrics.remove("input_height_pixels", labels);
 598                 global_metrics.remove("input_frame_rate_nom", labels);
 599                 global_metrics.remove("input_frame_rate_den", labels);
 600                 global_metrics.remove("input_sample_rate_hz", labels);
 601         }
 602
 603         // Register metrics.
 604         vector<pair<string, string>> labels;
 605         char card_name[64];
 606         snprintf(card_name, sizeof(card_name), "%d", card_index);
 607         labels.emplace_back("card", card_name);
 608
 609         switch (card_type) {
 610         case CardType::LIVE_CARD:
 611                 labels.emplace_back("cardtype", "live");
 612                 break;
 613         case CardType::FAKE_CAPTURE:
 614                 labels.emplace_back("cardtype", "fake");
 615                 break;
 616         case CardType::FFMPEG_INPUT:
 617                 labels.emplace_back("cardtype", "ffmpeg");
 618                 break;
 619         case CardType::CEF_INPUT:
 620                 labels.emplace_back("cardtype", "cef");
 621                 break;
 622         default:
 623                 assert(false);
 624         }
 625         card->jitter_history.register_metrics(labels);
 626         card->queue_length_policy.register_metrics(labels);
 627         global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
 628         global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
 629         global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
 630         global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
 631         global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
 632         global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
 633
 634         global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
 635         global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
 636         global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
 637         global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
 638         global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
 639         global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
 640         global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
 641         global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
 642         card->labels = labels;
 643 }
 644
 645 void Mixer::set_output_card_internal(int card_index)
 646 {
 647         // We don't really need to take card_mutex, since we're in the mixer
 648         // thread and don't mess with any queues (which is the only thing that happens
 649         // from other threads), but it's probably the safest in the long run.
 650         unique_lock<mutex> lock(card_mutex);
 651         if (output_card_index != -1) {
 652                 // Switch the old card from output to input.
 653                 CaptureCard *old_card = &cards[output_card_index];
 654                 old_card->output->end_output();
 655
 656                 // Stop the fake card that we put into place.
 657                 // This needs to _not_ happen under the mutex, to avoid deadlock
 658                 // (delivering the last frame needs to take the mutex).
 659                 CaptureInterface *fake_capture = old_card->capture.get();
 660                 lock.unlock();
 661                 fake_capture->stop_dequeue_thread();
 662                 lock.lock();
 663                 old_card->capture = move(old_card->parked_capture);  // TODO: reset the metrics
 664                 old_card->is_fake_capture = false;
 665                 old_card->capture->start_bm_capture();
 666         }
 667         if (card_index != -1) {
 668                 CaptureCard *card = &cards[card_index];
 669                 CaptureInterface *capture = card->capture.get();
 670                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 671                 // several seconds to complete (blocking on DisableVideoInput);
 672                 // see if we can maybe do it asynchronously.
 673                 lock.unlock();
 674                 capture->stop_dequeue_thread();
 675                 lock.lock();
 676                 card->parked_capture = move(card->capture);
 677                 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 678                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 679                 card->queue_length_policy.reset(card_index);
 680                 card->capture->start_bm_capture();
 681                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 682                 card->output->start_output(desired_output_video_mode, pts_int);
 683         }
 684         output_card_index = card_index;
 685         output_jitter_history.clear();
 686 }
 687
 688 namespace {
 689
 690 int unwrap_timecode(uint16_t current_wrapped, int last)
 691 {
 692         uint16_t last_wrapped = last & 0xffff;
 693         if (current_wrapped > last_wrapped) {
 694                 return (last & ~0xffff) | current_wrapped;
 695         } else {
 696                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 697         }
 698 }
 699
 700 DeviceSpec card_index_to_device(unsigned card_index, unsigned num_cards)
 701 {
 702         if (card_index >= num_cards) {
 703                 return DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 704         } else {
 705                 return DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 706         }
 707 }
 708
 709 }  // namespace
 710
 711 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 712                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 713                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 714 {
 715         DeviceSpec device = card_index_to_device(card_index, num_cards);
 716         CaptureCard *card = &cards[card_index];
 717
 718         ++card->metric_input_received_frames;
 719         card->metric_input_has_signal_bool = video_format.has_signal;
 720         card->metric_input_is_connected_bool = video_format.is_connected;
 721         card->metric_input_interlaced_bool = video_format.interlaced;
 722         card->metric_input_width_pixels = video_format.width;
 723         card->metric_input_height_pixels = video_format.height;
 724         card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
 725         card->metric_input_frame_rate_den = video_format.frame_rate_den;
 726         card->metric_input_sample_rate_hz = audio_format.sample_rate;
 727
 728         if (is_mode_scanning[card_index]) {
 729                 if (video_format.has_signal) {
 730                         // Found a stable signal, so stop scanning.
 731                         is_mode_scanning[card_index] = false;
 732                 } else {
 733                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 734                         steady_clock::time_point now = steady_clock::now();
 735                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 736                         if (sec_since_last_switch > switch_time_s) {
 737                                 // It isn't this mode; try the next one.
 738                                 mode_scanlist_index[card_index]++;
 739                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 740                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 741                                 last_mode_scan_change[card_index] = now;
 742                         }
 743                 }
 744         }
 745
 746         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 747         assert(frame_length > 0);
 748
 749         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;
 750         if (num_samples > OUTPUT_FREQUENCY / 10 && card->type != CardType::FFMPEG_INPUT) {
 751                 printf("%s: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
 752                         spec_to_string(device).c_str(), int(audio_frame.len), int(audio_offset),
 753                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 754                 if (video_frame.owner) {
 755                         video_frame.owner->release_frame(video_frame);
 756                 }
 757                 if (audio_frame.owner) {
 758                         audio_frame.owner->release_frame(audio_frame);
 759                 }
 760                 return;
 761         }
 762
 763         int dropped_frames = 0;
 764         if (card->last_timecode != -1) {
 765                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 766         }
 767
 768         // Number of samples per frame if we need to insert silence.
 769         // (Could be nonintegral, but resampling will save us then.)
 770         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 771
 772         if (dropped_frames > MAX_FPS * 2) {
 773                 fprintf(stderr, "%s lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 774                         spec_to_string(device).c_str(), card->last_timecode, timecode);
 775                 audio_mixer->reset_resampler(device);
 776                 dropped_frames = 0;
 777                 ++card->metric_input_resets;
 778         } else if (dropped_frames > 0) {
 779                 // Insert silence as needed.
 780                 fprintf(stderr, "%s dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 781                         spec_to_string(device).c_str(), dropped_frames, timecode);
 782                 card->metric_input_dropped_frames_error += dropped_frames;
 783
 784                 bool success;
 785                 do {
 786                         success = audio_mixer->add_silence(device, silence_samples, dropped_frames);
 787                 } while (!success);
 788         }
 789
 790         if (num_samples > 0) {
 791                 audio_mixer->add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, audio_frame.received_timestamp);
 792
 793                 // Audio for the MJPEG stream. We don't resample; audio that's not in 48 kHz
 794                 // just gets dropped for now.
 795                 //
 796                 // Only bother doing MJPEG encoding if there are any connected clients
 797                 // that want the stream.
 798                 if (httpd.get_num_connected_multicam_clients() > 0) {
 799                         vector<int32_t> converted_samples = convert_audio_to_fixed32(audio_frame.data + audio_offset, num_samples, audio_format, 2);
 800                         lock_guard<mutex> lock(card_mutex);
 801                         if (card->new_raw_audio.empty()) {
 802                                 card->new_raw_audio = move(converted_samples);
 803                         } else {
 804                                 // For raw audio, we don't really synchronize audio and video;
 805                                 // we just put the audio in frame by frame, and if a video frame is
 806                                 // dropped, we still keep the audio, which means it will be added
 807                                 // to the beginning of the next frame. It would probably be better
 808                                 // to move the audio pts earlier to show this, but most players can
 809                                 // live with some jitter, and in a lot of ways, it's much nicer for
 810                                 // Futatabi to have all audio locked to a video frame.
 811                                 card->new_raw_audio.insert(card->new_raw_audio.end(), converted_samples.begin(), converted_samples.end());
 812
 813                                 // Truncate to one second, just to be sure we don't have infinite buildup in case of weirdness.
 814                                 if (card->new_raw_audio.size() > OUTPUT_FREQUENCY * 2) {
 815                                         size_t excess_samples = card->new_raw_audio.size() - OUTPUT_FREQUENCY * 2;
 816                                         card->new_raw_audio.erase(card->new_raw_audio.begin(), card->new_raw_audio.begin() + excess_samples);
 817                                 }
 818                         }
 819                 }
 820         }
 821
 822         // Done with the audio, so release it.
 823         if (audio_frame.owner) {
 824                 audio_frame.owner->release_frame(audio_frame);
 825         }
 826
 827         card->last_timecode = timecode;
 828
 829         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 830         if (card->type == CardType::FFMPEG_INPUT && userdata != nullptr) {
 831                 FFmpegCapture *ffmpeg_capture = static_cast<FFmpegCapture *>(card->capture.get());
 832                 userdata->has_last_subtitle = ffmpeg_capture->get_has_last_subtitle();
 833                 userdata->last_subtitle = ffmpeg_capture->get_last_subtitle();
 834         }
 835
 836         size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
 837         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 838         if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
 839                 // The calculation above is wrong for planar Y'CbCr, so just override it.
 840                 assert(card->type == CardType::FFMPEG_INPUT);
 841                 assert(video_offset == 0);
 842                 expected_length = video_frame.len;
 843
 844                 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
 845                 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
 846                 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
 847                 cbcr_offset = video_format.width * video_format.height;
 848                 y_offset = 0;
 849         } else {
 850                 // All the other Y'CbCr formats are 4:2:2.
 851                 cbcr_width = video_format.width / 2;
 852                 cbcr_height = video_format.height;
 853                 cbcr_offset = video_offset / 2;
 854                 y_offset = video_frame.size / 2 + video_offset / 2;
 855         }
 856         if (video_frame.len - video_offset == 0 ||
 857             video_frame.len - video_offset != expected_length) {
 858                 if (video_frame.len != 0) {
 859                         printf("%s: Dropping video frame with wrong length (%zu; expected %zu)\n",
 860                                 spec_to_string(device).c_str(), video_frame.len - video_offset, expected_length);
 861                 }
 862                 if (video_frame.owner) {
 863                         video_frame.owner->release_frame(video_frame);
 864                 }
 865
 866                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 867                 // so that pts can go up accordingly.
 868                 {
 869                         lock_guard<mutex> lock(card_mutex);
 870                         CaptureCard::NewFrame new_frame;
 871                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 872                         new_frame.length = frame_length;
 873                         new_frame.interlaced = false;
 874                         new_frame.dropped_frames = dropped_frames;
 875                         new_frame.received_timestamp = video_frame.received_timestamp;
 876                         card->new_frames.push_back(move(new_frame));
 877                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
 878                 }
 879                 card->new_frames_changed.notify_all();
 880                 return;
 881         }
 882
 883         unsigned num_fields = video_format.interlaced ? 2 : 1;
 884         steady_clock::time_point frame_upload_start;
 885         bool interlaced_stride = false;
 886         if (video_format.interlaced) {
 887                 // Send the two fields along as separate frames; the other side will need to add
 888                 // a deinterlacer to actually get this right.
 889                 assert(video_format.height % 2 == 0);
 890                 video_format.height /= 2;
 891                 cbcr_height /= 2;
 892                 assert(frame_length % 2 == 0);
 893                 frame_length /= 2;
 894                 num_fields = 2;
 895                 if (video_format.second_field_start == 1) {
 896                         interlaced_stride = true;
 897                 }
 898                 frame_upload_start = steady_clock::now();
 899         }
 900         assert(userdata != nullptr);
 901         userdata->last_interlaced = video_format.interlaced;
 902         userdata->last_has_signal = video_format.has_signal;
 903         userdata->last_is_connected = video_format.is_connected;
 904         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 905         userdata->last_frame_rate_den = video_format.frame_rate_den;
 906         RefCountedFrame frame(video_frame);
 907
 908         // Upload the textures.
 909         for (unsigned field = 0; field < num_fields; ++field) {
 910                 // Put the actual texture upload in a lambda that is executed in the main thread.
 911                 // It is entirely possible to do this in the same thread (and it might even be
 912                 // faster, depending on the GPU and driver), but it appears to be trickling
 913                 // driver bugs very easily.
 914                 //
 915                 // Note that this means we must hold on to the actual frame data in <userdata>
 916                 // until the upload command is run, but we hold on to <frame> much longer than that
 917                 // (in fact, all the way until we no longer use the texture in rendering).
 918                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
 919                         unsigned field_start_line;
 920                         if (field == 1) {
 921                                 field_start_line = video_format.second_field_start;
 922                         } else {
 923                                 field_start_line = video_format.extra_lines_top;
 924                         }
 925
 926                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 927                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 928                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
 929
 930                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 931                         check_error();
 932
 933                         switch (userdata->pixel_format) {
 934                         case PixelFormat_10BitYCbCr: {
 935                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 936                                 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);
 937                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 938                                 break;
 939                         }
 940                         case PixelFormat_8BitYCbCr: {
 941                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 942                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 943
 944                                 // Make up our own strides, since we are interleaving.
 945                                 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);
 946                                 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);
 947                                 break;
 948                         }
 949                         case PixelFormat_8BitYCbCrPlanar: {
 950                                 assert(field_start_line == 0);  // We don't really support interlaced here.
 951                                 size_t field_y_start = y_offset;
 952                                 size_t field_cb_start = cbcr_offset;
 953                                 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
 954
 955                                 // Make up our own strides, since we are interleaving.
 956                                 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);
 957                                 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
 958                                 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
 959                                 break;
 960                         }
 961                         case PixelFormat_8BitBGRA: {
 962                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 963                                 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);
 964                                 // These could be asked to deliver mipmaps at any time.
 965                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 966                                 check_error();
 967                                 glGenerateMipmap(GL_TEXTURE_2D);
 968                                 check_error();
 969                                 glBindTexture(GL_TEXTURE_2D, 0);
 970                                 check_error();
 971                                 break;
 972                         }
 973                         default:
 974                                 assert(false);
 975                         }
 976
 977                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 978                         check_error();
 979                 };
 980
 981                 if (field == 1) {
 982                         // Don't upload the second field as fast as we can; wait until
 983                         // the field time has approximately passed. (Otherwise, we could
 984                         // get timing jitter against the other sources, and possibly also
 985                         // against the video display, although the latter is not as critical.)
 986                         // This requires our system clock to be reasonably close to the
 987                         // video clock, but that's not an unreasonable assumption.
 988                         steady_clock::time_point second_field_start = frame_upload_start +
 989                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 990                         this_thread::sleep_until(second_field_start);
 991                 }
 992
 993                 {
 994                         lock_guard<mutex> lock(card_mutex);
 995                         CaptureCard::NewFrame new_frame;
 996                         new_frame.frame = frame;
 997                         new_frame.length = frame_length;
 998                         new_frame.field = field;
 999                         new_frame.interlaced = video_format.interlaced;
1000                         new_frame.upload_func = upload_func;
1001                         new_frame.dropped_frames = dropped_frames;
1002                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
1003                         new_frame.video_format = video_format;
1004                         new_frame.y_offset = y_offset;
1005                         new_frame.cbcr_offset = cbcr_offset;
1006                         card->new_frames.push_back(move(new_frame));
1007                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
1008                         card->may_have_dropped_last_frame = false;
1009                 }
1010                 card->new_frames_changed.notify_all();
1011         }
1012 }
1013
1014 void Mixer::bm_hotplug_add(libusb_device *dev)
1015 {
1016         lock_guard<mutex> lock(hotplug_mutex);
1017         hotplugged_cards.push_back(dev);
1018 }
1019
1020 void Mixer::bm_hotplug_remove(unsigned card_index)
1021 {
1022         cards[card_index].new_frames_changed.notify_all();
1023 }
1024
1025 void Mixer::thread_func()
1026 {
1027         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
1028
1029         eglBindAPI(EGL_OPENGL_API);
1030         QOpenGLContext *context = create_context(mixer_surface);
1031         if (!make_current(context, mixer_surface)) {
1032                 printf("oops\n");
1033                 abort();
1034         }
1035
1036         // Start the actual capture. (We don't want to do it before we're actually ready
1037         // to process output frames.)
1038         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1039                 if (int(card_index) != output_card_index) {
1040                         cards[card_index].capture->start_bm_capture();
1041                 }
1042         }
1043
1044         BasicStats basic_stats(/*verbose=*/true, /*use_opengl=*/true);
1045         int stats_dropped_frames = 0;
1046
1047         while (!should_quit) {
1048                 if (desired_output_card_index != output_card_index) {
1049                         set_output_card_internal(desired_output_card_index);
1050                 }
1051                 if (output_card_index != -1 &&
1052                     desired_output_video_mode != output_video_mode) {
1053                         DeckLinkOutput *output = cards[output_card_index].output.get();
1054                         output->end_output();
1055                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
1056                         output->start_output(desired_output_video_mode, pts_int);
1057                 }
1058
1059                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
1060                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
1061
1062                 bool master_card_is_output;
1063                 unsigned master_card_index;
1064                 if (output_card_index != -1) {
1065                         master_card_is_output = true;
1066                         master_card_index = output_card_index;
1067                 } else {
1068                         master_card_is_output = false;
1069                         master_card_index = theme->map_signal(master_clock_channel);
1070                         assert(master_card_index < num_cards + num_video_inputs);
1071                 }
1072
1073                 vector<int32_t> raw_audio[MAX_VIDEO_CARDS];  // For MJPEG encoding.
1074                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame, raw_audio);
1075                 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);
1076                 stats_dropped_frames += output_frame_info.dropped_frames;
1077
1078                 handle_hotplugged_cards();
1079
1080                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1081                         DeviceSpec device = card_index_to_device(card_index, num_cards);
1082                         if (card_index == master_card_index || !has_new_frame[card_index]) {
1083                                 continue;
1084                         }
1085                         if (new_frames[card_index].frame->len == 0) {
1086                                 ++new_frames[card_index].dropped_frames;
1087                         }
1088                         if (new_frames[card_index].dropped_frames > 0) {
1089                                 printf("%s dropped %d frames before this\n",
1090                                         spec_to_string(device).c_str(), int(new_frames[card_index].dropped_frames));
1091                         }
1092                 }
1093
1094                 // If the first card is reporting a corrupted or otherwise dropped frame,
1095                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
1096                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
1097                         ++stats_dropped_frames;
1098                         pts_int += new_frames[master_card_index].length;
1099                         continue;
1100                 }
1101
1102                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1103                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1104                                 continue;
1105
1106                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1107                         assert(new_frame->frame != nullptr);
1108                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1109                         check_error();
1110
1111                         // The new texture might need uploading before use.
1112                         if (new_frame->upload_func) {
1113                                 new_frame->upload_func();
1114                                 new_frame->upload_func = nullptr;
1115                         }
1116
1117                         if (new_frame->frame->data_copy != nullptr) {
1118                                 int mjpeg_card_index = mjpeg_encoder->get_mjpeg_stream_for_card(card_index);
1119                                 if (mjpeg_card_index != -1) {
1120                                         mjpeg_encoder->upload_frame(pts_int, mjpeg_card_index, new_frame->frame, new_frame->video_format, new_frame->y_offset, new_frame->cbcr_offset, move(raw_audio[card_index]));
1121                                 }
1122                         }
1123                 }
1124
1125                 int64_t frame_duration = output_frame_info.frame_duration;
1126                 render_one_frame(frame_duration);
1127                 {
1128                         lock_guard<mutex> lock(frame_num_mutex);
1129                         ++frame_num;
1130                 }
1131                 frame_num_updated.notify_all();
1132                 pts_int += frame_duration;
1133
1134                 basic_stats.update(frame_num, stats_dropped_frames);
1135                 // if (frame_num % 100 == 0) chain->print_phase_timing();
1136
1137                 if (should_cut.exchange(false)) {  // Test and clear.
1138                         video_encoder->do_cut(frame_num);
1139                 }
1140
1141 #if 0
1142                 // Reset every 100 frames, so that local variations in frame times
1143                 // (especially for the first few frames, when the shaders are
1144                 // compiled etc.) don't make it hard to measure for the entire
1145                 // remaining duration of the program.
1146                 if (frame == 10000) {
1147                         frame = 0;
1148                         start = now;
1149                 }
1150 #endif
1151                 check_error();
1152         }
1153
1154         resource_pool->clean_context();
1155 }
1156
1157 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1158 {
1159         if (output_card_index != -1) {
1160                 // The output card (ie., cards[output_card_index].output) is the master clock,
1161                 // so no input card (ie., cards[card_index].capture) is.
1162                 return false;
1163         }
1164         return (card_index == master_card_index);
1165 }
1166
1167 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1168 {
1169         // Count the number of frames in the queue, including any frames
1170         // we dropped. It's hard to know exactly how we should deal with
1171         // dropped (corrupted) input frames; they don't help our goal of
1172         // avoiding starvation, but they still add to the problem of latency.
1173         // Since dropped frames is going to mean a bump in the signal anyway,
1174         // we err on the side of having more stable latency instead.
1175         unsigned queue_length = 0;
1176         for (const CaptureCard::NewFrame &frame : card->new_frames) {
1177                 queue_length += frame.dropped_frames + 1;
1178         }
1179
1180         // If needed, drop frames until the queue is below the safe limit.
1181         // We prefer to drop from the head, because all else being equal,
1182         // we'd like more recent frames (less latency).
1183         unsigned dropped_frames = 0;
1184         while (queue_length > safe_queue_length) {
1185                 assert(!card->new_frames.empty());
1186                 assert(queue_length > card->new_frames.front().dropped_frames);
1187                 queue_length -= card->new_frames.front().dropped_frames;
1188
1189                 if (queue_length <= safe_queue_length) {
1190                         // No need to drop anything.
1191                         break;
1192                 }
1193
1194                 card->new_frames.pop_front();
1195                 card->new_frames_changed.notify_all();
1196                 --queue_length;
1197                 ++dropped_frames;
1198
1199                 if (queue_length == 0 && card->is_cef_capture) {
1200                         card->may_have_dropped_last_frame = true;
1201                 }
1202         }
1203
1204         card->metric_input_dropped_frames_jitter += dropped_frames;
1205         card->metric_input_queue_length_frames = queue_length;
1206
1207 #if 0
1208         if (dropped_frames > 0) {
1209                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1210                         card_index, dropped_frames);
1211         }
1212 #endif
1213 }
1214
1215 pair<string, string> Mixer::get_channels_json()
1216 {
1217         Channels ret;
1218         for (int channel_idx = 2; channel_idx < theme->get_num_channels(); ++channel_idx) {
1219                 Channel *channel = ret.add_channel();
1220                 channel->set_index(channel_idx);
1221                 channel->set_name(theme->get_channel_name(channel_idx));
1222                 channel->set_color(theme->get_channel_color(channel_idx));
1223         }
1224         string contents;
1225         google::protobuf::util::MessageToJsonString(ret, &contents);  // Ignore any errors.
1226         return make_pair(contents, "text/json");
1227 }
1228
1229 pair<string, string> Mixer::get_channel_color_http(unsigned channel_idx)
1230 {
1231         return make_pair(theme->get_channel_color(channel_idx), "text/plain");
1232 }
1233
1234 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], vector<int32_t> raw_audio[MAX_VIDEO_CARDS])
1235 {
1236         OutputFrameInfo output_frame_info;
1237 start:
1238         unique_lock<mutex> lock(card_mutex, defer_lock);
1239         if (master_card_is_output) {
1240                 // Clocked to the output, so wait for it to be ready for the next frame.
1241                 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);
1242                 lock.lock();
1243         } else {
1244                 // Wait for the master card to have a new frame.
1245                 // TODO: Add a timeout.
1246                 output_frame_info.is_preroll = false;
1247                 lock.lock();
1248                 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(); });
1249         }
1250
1251         if (master_card_is_output) {
1252                 handle_hotplugged_cards();
1253         } else if (cards[master_card_index].new_frames.empty()) {
1254                 // We were woken up, but not due to a new frame. Deal with it
1255                 // and then restart.
1256                 assert(cards[master_card_index].capture->get_disconnected());
1257                 handle_hotplugged_cards();
1258                 lock.unlock();
1259                 goto start;
1260         }
1261
1262         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1263                 CaptureCard *card = &cards[card_index];
1264                 if (card->new_frames.empty()) {  // Starvation.
1265                         ++card->metric_input_duped_frames;
1266 #ifdef HAVE_CEF
1267                         if (card->is_cef_capture && card->may_have_dropped_last_frame) {
1268                                 // Unlike other sources, CEF is not guaranteed to send us a steady
1269                                 // stream of frames, so we'll have to ask it to repaint the frame
1270                                 // we dropped. (may_have_dropped_last_frame is set whenever we
1271                                 // trim the queue completely away, and cleared when we actually
1272                                 // get a new frame.)
1273                                 ((CEFCapture *)card->capture.get())->request_new_frame();
1274                         }
1275 #endif
1276                 } else {
1277                         new_frames[card_index] = move(card->new_frames.front());
1278                         has_new_frame[card_index] = true;
1279                         card->new_frames.pop_front();
1280                         card->new_frames_changed.notify_all();
1281                 }
1282
1283                 raw_audio[card_index] = move(card->new_raw_audio);
1284         }
1285
1286         if (!master_card_is_output) {
1287                 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1288                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1289                 output_frame_info.frame_duration = new_frames[master_card_index].length;
1290         }
1291
1292         if (!output_frame_info.is_preroll) {
1293                 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1294         }
1295
1296         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1297                 CaptureCard *card = &cards[card_index];
1298                 if (has_new_frame[card_index] &&
1299                     !input_card_is_master_clock(card_index, master_card_index) &&
1300                     !output_frame_info.is_preroll) {
1301                         card->queue_length_policy.update_policy(
1302                                 output_frame_info.frame_timestamp,
1303                                 card->jitter_history.get_expected_next_frame(),
1304                                 new_frames[master_card_index].length,
1305                                 output_frame_info.frame_duration,
1306                                 card->jitter_history.estimate_max_jitter(),
1307                                 output_jitter_history.estimate_max_jitter());
1308                         trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1309                                                   card->queue_length_policy.get_safe_queue_length()));
1310                 }
1311         }
1312
1313         // This might get off by a fractional sample when changing master card
1314         // between ones with different frame rates, but that's fine.
1315         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1316         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1317         fractional_samples = num_samples_times_timebase % TIMEBASE;
1318         assert(output_frame_info.num_samples >= 0);
1319
1320         return output_frame_info;
1321 }
1322
1323 void Mixer::handle_hotplugged_cards()
1324 {
1325         // Check for cards that have been disconnected since last frame.
1326         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1327                 CaptureCard *card = &cards[card_index];
1328                 if (card->capture->get_disconnected()) {
1329                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1330                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1331                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1332                         card->queue_length_policy.reset(card_index);
1333                         card->capture->start_bm_capture();
1334                 }
1335         }
1336
1337         // Check for cards that have been connected since last frame.
1338         vector<libusb_device *> hotplugged_cards_copy;
1339         {
1340                 lock_guard<mutex> lock(hotplug_mutex);
1341                 swap(hotplugged_cards, hotplugged_cards_copy);
1342         }
1343         for (libusb_device *new_dev : hotplugged_cards_copy) {
1344                 // Look for a fake capture card where we can stick this in.
1345                 int free_card_index = -1;
1346                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1347                         if (cards[card_index].is_fake_capture) {
1348                                 free_card_index = card_index;
1349                                 break;
1350                         }
1351                 }
1352
1353                 if (free_card_index == -1) {
1354                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1355                         libusb_unref_device(new_dev);
1356                 } else {
1357                         // BMUSBCapture takes ownership.
1358                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1359                         CaptureCard *card = &cards[free_card_index];
1360                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1361                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1362                         card->queue_length_policy.reset(free_card_index);
1363                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1364                         capture->start_bm_capture();
1365                 }
1366         }
1367 }
1368
1369
1370 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)
1371 {
1372         // Resample the audio as needed, including from previously dropped frames.
1373         assert(num_cards > 0);
1374         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1375                 const bool dropped_frame = (frame_num != dropped_frames);
1376                 {
1377                         // Signal to the audio thread to process this frame.
1378                         // Note that if the frame is a dropped frame, we signal that
1379                         // we don't want to use this frame as base for adjusting
1380                         // the resampler rate. The reason for this is that the timing
1381                         // of these frames is often way too late; they typically don't
1382                         // “arrive” before we synthesize them. Thus, we could end up
1383                         // in a situation where we have inserted e.g. five audio frames
1384                         // into the queue before we then start pulling five of them
1385                         // back out. This makes ResamplingQueue overestimate the delay,
1386                         // causing undue resampler changes. (We _do_ use the last,
1387                         // non-dropped frame; perhaps we should just discard that as well,
1388                         // since dropped frames are expected to be rare, and it might be
1389                         // better to just wait until we have a slightly more normal situation).
1390                         lock_guard<mutex> lock(audio_mutex);
1391                         bool adjust_rate = !dropped_frame && !is_preroll;
1392                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1393                         audio_task_queue_changed.notify_one();
1394                 }
1395                 if (dropped_frame) {
1396                         // For dropped frames, increase the pts. Note that if the format changed
1397                         // in the meantime, we have no way of detecting that; we just have to
1398                         // assume the frame length is always the same.
1399                         pts_int += length_per_frame;
1400                 }
1401         }
1402 }
1403
1404 void Mixer::render_one_frame(int64_t duration)
1405 {
1406         // Determine the time code for this frame before we start rendering.
1407         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1408         if (display_timecode_on_stdout) {
1409                 printf("Timecode: '%s'\n", timecode_text.c_str());
1410         }
1411
1412         // Update Y'CbCr settings for all cards.
1413         {
1414                 lock_guard<mutex> lock(card_mutex);
1415                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1416                         YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1417                         input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1418                         input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1419                         input_state.full_range[card_index] = interpretation->full_range;
1420                 }
1421         }
1422
1423         // Get the main chain from the theme, and set its state immediately.
1424         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1425         EffectChain *chain = theme_main_chain.chain;
1426         theme_main_chain.setup_chain();
1427         //theme_main_chain.chain->enable_phase_timing(true);
1428
1429         // If HDMI/SDI output is active and the user has requested auto mode,
1430         // its mode overrides the existing Y'CbCr setting for the chain.
1431         YCbCrLumaCoefficients ycbcr_output_coefficients;
1432         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1433                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1434         } else {
1435                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1436         }
1437
1438         // TODO: Reduce the duplication against theme.cpp.
1439         YCbCrFormat output_ycbcr_format;
1440         output_ycbcr_format.chroma_subsampling_x = 1;
1441         output_ycbcr_format.chroma_subsampling_y = 1;
1442         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1443         output_ycbcr_format.full_range = false;
1444         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1445         chain->change_ycbcr_output_format(output_ycbcr_format);
1446
1447         // Render main chain. If we're using zerocopy Quick Sync encoding
1448         // (the default case), we take an extra copy of the created outputs,
1449         // so that we can display it back to the screen later (it's less memory
1450         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1451         // Ideally, we'd like to avoid taking copies and just use the main textures
1452         // for display as well, but they're just views into VA-API memory and must be
1453         // unmapped during encoding, so we can't use them for display, unfortunately.
1454         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1455         GLuint y_copy_tex, cbcr_copy_tex = 0;
1456         GLuint y_display_tex, cbcr_display_tex;
1457         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1458         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1459         const bool is_zerocopy = video_encoder->is_zerocopy();
1460         if (is_zerocopy) {
1461                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1462                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1463                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1464
1465                 y_display_tex = y_copy_tex;
1466                 cbcr_display_tex = cbcr_copy_tex;
1467
1468                 // y_tex and cbcr_tex will be given by VideoEncoder.
1469         } else {
1470                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1471                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1472                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1473
1474                 y_display_tex = y_tex;
1475                 cbcr_display_tex = cbcr_tex;
1476         }
1477
1478         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1479         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1480         assert(got_frame);
1481
1482         GLuint fbo;
1483         if (is_zerocopy) {
1484                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1485         } else {
1486                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1487         }
1488         check_error();
1489         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1490
1491         if (display_timecode_in_stream) {
1492                 // Render the timecode on top.
1493                 timecode_renderer->render_timecode(fbo, timecode_text);
1494         }
1495
1496         resource_pool->release_fbo(fbo);
1497
1498         if (is_zerocopy) {
1499                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1500         } else {
1501                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1502         }
1503         if (output_card_index != -1) {
1504                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1505         }
1506         resource_pool->release_2d_texture(cbcr_full_tex);
1507
1508         // Set the right state for the Y' and CbCr textures we use for display.
1509         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1510         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1511         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1512         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1513         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1514
1515         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1516         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1517         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1518         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1519
1520         RefCountedGLsync fence = video_encoder->end_frame();
1521
1522         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1523         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1524         DisplayFrame live_frame;
1525         live_frame.chain = display_chain.get();
1526         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1527                 display_input->set_texture_num(0, y_display_tex);
1528                 display_input->set_texture_num(1, cbcr_display_tex);
1529         };
1530         live_frame.ready_fence = fence;
1531         live_frame.input_frames = {};
1532         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1533         output_channel[OUTPUT_LIVE].output_frame(move(live_frame));
1534
1535         // Set up preview and any additional channels.
1536         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1537                 DisplayFrame display_frame;
1538                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1539                 display_frame.chain = move(chain.chain);
1540                 display_frame.setup_chain = move(chain.setup_chain);
1541                 display_frame.ready_fence = fence;
1542                 display_frame.input_frames = move(chain.input_frames);
1543                 display_frame.temp_textures = {};
1544                 output_channel[i].output_frame(move(display_frame));
1545         }
1546 }
1547
1548 void Mixer::audio_thread_func()
1549 {
1550         pthread_setname_np(pthread_self(), "Mixer_Audio");
1551
1552         while (!should_quit) {
1553                 AudioTask task;
1554
1555                 {
1556                         unique_lock<mutex> lock(audio_mutex);
1557                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1558                         if (should_quit) {
1559                                 return;
1560                         }
1561                         task = audio_task_queue.front();
1562                         audio_task_queue.pop();
1563                 }
1564
1565                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1566                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1567                 vector<float> samples_out = audio_mixer->get_output(
1568                         task.frame_timestamp,
1569                         task.num_samples,
1570                         rate_adjustment_policy);
1571
1572                 // Send the samples to the sound card, then add them to the output.
1573                 if (alsa) {
1574                         alsa->write(samples_out);
1575                 }
1576                 if (output_card_index != -1) {
1577                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1578                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1579                 }
1580                 video_encoder->add_audio(task.pts_int, move(samples_out));
1581         }
1582 }
1583
1584 void Mixer::release_display_frame(DisplayFrame *frame)
1585 {
1586         for (GLuint texnum : frame->temp_textures) {
1587                 resource_pool->release_2d_texture(texnum);
1588         }
1589         frame->temp_textures.clear();
1590         frame->ready_fence.reset();
1591         frame->input_frames.clear();
1592 }
1593
1594 void Mixer::start()
1595 {
1596         mixer_thread = thread(&Mixer::thread_func, this);
1597         audio_thread = thread(&Mixer::audio_thread_func, this);
1598 }
1599
1600 void Mixer::quit()
1601 {
1602         should_quit = true;
1603         audio_task_queue_changed.notify_one();
1604         mixer_thread.join();
1605         audio_thread.join();
1606 }
1607
1608 void Mixer::transition_clicked(int transition_num)
1609 {
1610         theme->transition_clicked(transition_num, pts());
1611 }
1612
1613 void Mixer::channel_clicked(int preview_num)
1614 {
1615         theme->channel_clicked(preview_num);
1616 }
1617
1618 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1619 {
1620         lock_guard<mutex> lock(card_mutex);
1621         return ycbcr_interpretation[card_index];
1622 }
1623
1624 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1625 {
1626         lock_guard<mutex> lock(card_mutex);
1627         ycbcr_interpretation[card_index] = interpretation;
1628 }
1629
1630 void Mixer::start_mode_scanning(unsigned card_index)
1631 {
1632         assert(card_index < num_cards);
1633         if (is_mode_scanning[card_index]) {
1634                 return;
1635         }
1636         is_mode_scanning[card_index] = true;
1637         mode_scanlist[card_index].clear();
1638         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1639                 mode_scanlist[card_index].push_back(mode.first);
1640         }
1641         assert(!mode_scanlist[card_index].empty());
1642         mode_scanlist_index[card_index] = 0;
1643         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1644         last_mode_scan_change[card_index] = steady_clock::now();
1645 }
1646
1647 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1648 {
1649         assert(desired_output_card_index != -1);
1650         lock_guard<mutex> lock(card_mutex);
1651         return cards[desired_output_card_index].output->get_available_video_modes();
1652 }
1653
1654 string Mixer::get_ffmpeg_filename(unsigned card_index) const
1655 {
1656         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1657         return ((FFmpegCapture *)(cards[card_index].capture.get()))->get_filename();
1658 }
1659
1660 void Mixer::set_ffmpeg_filename(unsigned card_index, const string &filename) {
1661         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1662         ((FFmpegCapture *)(cards[card_index].capture.get()))->change_filename(filename);
1663 }
1664
1665 void Mixer::wait_for_next_frame()
1666 {
1667         unique_lock<mutex> lock(frame_num_mutex);
1668         unsigned old_frame_num = frame_num;
1669         frame_num_updated.wait_for(lock, seconds(1),  // Timeout is just in case.
1670                 [old_frame_num, this]{ return this->frame_num > old_frame_num; });
1671 }
1672
1673 Mixer::OutputChannel::~OutputChannel()
1674 {
1675         if (has_current_frame) {
1676                 parent->release_display_frame(&current_frame);
1677         }
1678         if (has_ready_frame) {
1679                 parent->release_display_frame(&ready_frame);
1680         }
1681 }
1682
1683 void Mixer::OutputChannel::output_frame(DisplayFrame &&frame)
1684 {
1685         // Store this frame for display. Remove the ready frame if any
1686         // (it was seemingly never used).
1687         {
1688                 lock_guard<mutex> lock(frame_mutex);
1689                 if (has_ready_frame) {
1690                         parent->release_display_frame(&ready_frame);
1691                 }
1692                 ready_frame = move(frame);
1693                 has_ready_frame = true;
1694
1695                 // Call the callbacks under the mutex (they should be short),
1696                 // so that we don't race against a callback removal.
1697                 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1698                         key_and_callback.second();
1699                 }
1700         }
1701
1702         // Reduce the number of callbacks by filtering duplicates. The reason
1703         // why we bother doing this is that Qt seemingly can get into a state
1704         // where its builds up an essentially unbounded queue of signals,
1705         // consuming more and more memory, and there's no good way of collapsing
1706         // user-defined signals or limiting the length of the queue.
1707         if (transition_names_updated_callback) {
1708                 vector<string> transition_names = global_mixer->get_transition_names();
1709                 bool changed = false;
1710                 if (transition_names.size() != last_transition_names.size()) {
1711                         changed = true;
1712                 } else {
1713                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1714                                 if (transition_names[i] != last_transition_names[i]) {
1715                                         changed = true;
1716                                         break;
1717                                 }
1718                         }
1719                 }
1720                 if (changed) {
1721                         transition_names_updated_callback(transition_names);
1722                         last_transition_names = transition_names;
1723                 }
1724         }
1725         if (name_updated_callback) {
1726                 string name = global_mixer->get_channel_name(channel);
1727                 if (name != last_name) {
1728                         name_updated_callback(name);
1729                         last_name = name;
1730                 }
1731         }
1732         if (color_updated_callback) {
1733                 string color = global_mixer->get_channel_color(channel);
1734                 if (color != last_color) {
1735                         color_updated_callback(color);
1736                         last_color = color;
1737                 }
1738         }
1739 }
1740
1741 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1742 {
1743         lock_guard<mutex> lock(frame_mutex);
1744         if (!has_current_frame && !has_ready_frame) {
1745                 return false;
1746         }
1747
1748         if (has_current_frame && has_ready_frame) {
1749                 // We have a new ready frame. Toss the current one.
1750                 parent->release_display_frame(&current_frame);
1751                 has_current_frame = false;
1752         }
1753         if (has_ready_frame) {
1754                 assert(!has_current_frame);
1755                 current_frame = move(ready_frame);
1756                 ready_frame.ready_fence.reset();  // Drop the refcount.
1757                 ready_frame.input_frames.clear();  // Drop the refcounts.
1758                 has_current_frame = true;
1759                 has_ready_frame = false;
1760         }
1761
1762         *frame = current_frame;
1763         return true;
1764 }
1765
1766 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1767 {
1768         lock_guard<mutex> lock(frame_mutex);
1769         new_frame_ready_callbacks[key] = callback;
1770 }
1771
1772 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1773 {
1774         lock_guard<mutex> lock(frame_mutex);
1775         new_frame_ready_callbacks.erase(key);
1776 }
1777
1778 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1779 {
1780         transition_names_updated_callback = callback;
1781 }
1782
1783 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1784 {
1785         name_updated_callback = callback;
1786 }
1787
1788 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1789 {
1790         color_updated_callback = callback;
1791 }
1792
1793 mutex RefCountedGLsync::fence_lock;