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.h>
   8 #include <movit/effect_chain.h>
   9 #include <movit/effect_util.h>
  10 #include <movit/flat_input.h>
  11 #include <movit/image_format.h>
  12 #include <movit/init.h>
  13 #include <movit/resource_pool.h>
  14 #include <pthread.h>
  15 #include <stdint.h>
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <algorithm>
  19 #include <chrono>
  20 #include <condition_variable>
  21 #include <cstddef>
  22 #include <cstdint>
  23 #include <memory>
  24 #include <mutex>
  25 #include <ratio>
  26 #include <string>
  27 #include <thread>
  28 #include <utility>
  29 #include <vector>
  30
  31 #include "DeckLinkAPI.h"
  32 #include "LinuxCOM.h"
  33 #include "alsa_output.h"
  34 #include "basic_stats.h"
  35 #include "bmusb/bmusb.h"
  36 #include "bmusb/fake_capture.h"
  37 #ifdef HAVE_CEF
  38 #include "cef_capture.h"
  39 #endif
  40 #include "chroma_subsampler.h"
  41 #include "shared/context.h"
  42 #include "decklink_capture.h"
  43 #include "decklink_output.h"
  44 #include "defs.h"
  45 #include "shared/disk_space_estimator.h"
  46 #include "ffmpeg_capture.h"
  47 #include "flags.h"
  48 #include "image_input.h"
  49 #include "input_mapping.h"
  50 #include "shared/metrics.h"
  51 #include "mjpeg_encoder.h"
  52 #include "pbo_frame_allocator.h"
  53 #include "shared/ref_counted_gl_sync.h"
  54 #include "resampling_queue.h"
  55 #include "shared/timebase.h"
  56 #include "timecode_renderer.h"
  57 #include "v210_converter.h"
  58 #include "va_display_with_cleanup.h"
  59 #include "video_encoder.h"
  60
  61 #undef Status
  62 #include <google/protobuf/util/json_util.h>
  63 #include "json.pb.h"
  64
  65 class IDeckLink;
  66 class QOpenGLContext;
  67
  68 using namespace movit;
  69 using namespace std;
  70 using namespace std::chrono;
  71 using namespace std::placeholders;
  72 using namespace bmusb;
  73
  74 Mixer *global_mixer = nullptr;
  75
  76 namespace {
  77
  78 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  79 {
  80         if (interlaced) {
  81                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  82                         input_state->buffered_frames[card_index][frame_num] =
  83                                 input_state->buffered_frames[card_index][frame_num - 1];
  84                 }
  85                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  86         } else {
  87                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  88                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  89                 }
  90         }
  91 }
  92
  93 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
  94 {
  95         bool first;
  96         switch (userdata->pixel_format) {
  97         case PixelFormat_10BitYCbCr:
  98                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  99                 break;
 100         case PixelFormat_8BitYCbCr:
 101                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
 102                 break;
 103         case PixelFormat_8BitBGRA:
 104                 first = userdata->tex_rgba[field] == 0;
 105                 break;
 106         case PixelFormat_8BitYCbCrPlanar:
 107                 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
 108                 break;
 109         default:
 110                 assert(false);
 111         }
 112
 113         if (first ||
 114             width != userdata->last_width[field] ||
 115             height != userdata->last_height[field] ||
 116             cbcr_width != userdata->last_cbcr_width[field] ||
 117             cbcr_height != userdata->last_cbcr_height[field]) {
 118                 // We changed resolution since last use of this texture, so we need to create
 119                 // a new object. Note that this each card has its own PBOFrameAllocator,
 120                 // we don't need to worry about these flip-flopping between resolutions.
 121                 switch (userdata->pixel_format) {
 122                 case PixelFormat_10BitYCbCr:
 123                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 124                         check_error();
 125                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 126                         check_error();
 127                         break;
 128                 case PixelFormat_8BitYCbCr: {
 129                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 130                         check_error();
 131                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 132                         check_error();
 133                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 134                         check_error();
 135                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 136                         check_error();
 137                         break;
 138                 }
 139                 case PixelFormat_8BitYCbCrPlanar: {
 140                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 141                         check_error();
 142                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 143                         check_error();
 144                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
 145                         check_error();
 146                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 147                         check_error();
 148                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
 149                         check_error();
 150                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 151                         check_error();
 152                         break;
 153                 }
 154                 case PixelFormat_8BitBGRA:
 155                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 156                         check_error();
 157                         // NOTE: sRGB may be disabled by sRGBSwitchingFlatInput.
 158                         glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 159                         check_error();
 160                         break;
 161                 default:
 162                         assert(false);
 163                 }
 164                 userdata->last_width[field] = width;
 165                 userdata->last_height[field] = height;
 166                 userdata->last_cbcr_width[field] = cbcr_width;
 167                 userdata->last_cbcr_height[field] = cbcr_height;
 168         }
 169         if (global_flags.ten_bit_input &&
 170             (first || v210_width != userdata->last_v210_width[field])) {
 171                 // Same as above; we need to recreate the texture.
 172                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 173                 check_error();
 174                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 175                 check_error();
 176                 userdata->last_v210_width[field] = v210_width;
 177         }
 178 }
 179
 180 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 181 {
 182         if (interlaced_stride) {
 183                 stride *= 2;
 184         }
 185         if (global_flags.flush_pbos) {
 186                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 187                 check_error();
 188         }
 189
 190         glBindTexture(GL_TEXTURE_2D, tex);
 191         check_error();
 192         if (interlaced_stride) {
 193                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 194                 check_error();
 195         } else {
 196                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 197                 check_error();
 198         }
 199
 200         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 201         check_error();
 202         glBindTexture(GL_TEXTURE_2D, 0);
 203         check_error();
 204         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 205         check_error();
 206 }
 207
 208 }  // namespace
 209
 210 void JitterHistory::register_metrics(const vector<pair<string, string>> &labels)
 211 {
 212         global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames);
 213         global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE);
 214 }
 215
 216 void JitterHistory::unregister_metrics(const vector<pair<string, string>> &labels)
 217 {
 218         global_metrics.remove("input_underestimated_jitter_frames", labels);
 219         global_metrics.remove("input_estimated_max_jitter_seconds", labels);
 220 }
 221
 222 void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames)
 223 {
 224         if (expected_timestamp > steady_clock::time_point::min()) {
 225                 expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 226                 double jitter_seconds = fabs(duration<double>(expected_timestamp - now).count());
 227                 history.push_back(orders.insert(jitter_seconds));
 228                 if (jitter_seconds > estimate_max_jitter()) {
 229                         ++metric_input_underestimated_jitter_frames;
 230                 }
 231
 232                 metric_input_estimated_max_jitter_seconds = estimate_max_jitter();
 233
 234                 if (history.size() > history_length) {
 235                         orders.erase(history.front());
 236                         history.pop_front();
 237                 }
 238                 assert(history.size() <= history_length);
 239         }
 240         expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 241 }
 242
 243 double JitterHistory::estimate_max_jitter() const
 244 {
 245         if (orders.empty()) {
 246                 return 0.0;
 247         }
 248         size_t elem_idx = lrint((orders.size() - 1) * percentile);
 249         if (percentile <= 0.5) {
 250                 return *next(orders.begin(), elem_idx) * multiplier;
 251         } else {
 252                 return *prev(orders.end(), orders.size() - elem_idx) * multiplier;
 253         }
 254 }
 255
 256 void QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
 257 {
 258         global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
 259 }
 260
 261 void QueueLengthPolicy::unregister_metrics(const vector<pair<string, string>> &labels)
 262 {
 263         global_metrics.remove("input_queue_safe_length_frames", labels);
 264 }
 265
 266 void QueueLengthPolicy::update_policy(steady_clock::time_point now,
 267                                       steady_clock::time_point expected_next_frame,
 268                                       int64_t input_frame_duration,
 269                                       int64_t master_frame_duration,
 270                                       double max_input_card_jitter_seconds,
 271                                       double max_master_card_jitter_seconds)
 272 {
 273         double input_frame_duration_seconds = input_frame_duration / double(TIMEBASE);
 274         double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
 275
 276         // Figure out when we can expect the next frame for this card, assuming
 277         // worst-case jitter (ie., the frame is maximally late).
 278         double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
 279
 280         // How many times are the master card expected to tick in that time?
 281         // We assume the master clock has worst-case jitter but not any rate
 282         // discrepancy, ie., it ticks as early as possible every time, but not
 283         // cumulatively.
 284         double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
 285
 286         // As a special case, if the master card ticks faster than the input card,
 287         // we expect the queue to drain by itself even without dropping. But if
 288         // the difference is small (e.g. 60 Hz master and 59.94 input), it would
 289         // go slowly enough that the effect wouldn't really be appreciable.
 290         // We account for this by looking at the situation five frames ahead,
 291         // assuming everything else is the same.
 292         double frames_allowed;
 293         if (master_frame_duration < input_frame_duration) {
 294                 frames_allowed = frames_needed + 5 * (input_frame_duration_seconds - master_frame_duration_seconds) / master_frame_duration_seconds;
 295         } else {
 296                 frames_allowed = frames_needed;
 297         }
 298
 299         safe_queue_length = max<int>(floor(frames_allowed), 0);
 300         metric_input_queue_safe_length_frames = safe_queue_length;
 301 }
 302
 303 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 304         : httpd(),
 305           num_cards(num_cards),
 306           mixer_surface(create_surface(format)),
 307           h264_encoder_surface(create_surface(format)),
 308           decklink_output_surface(create_surface(format)),
 309           image_update_surface(create_surface(format))
 310 {
 311         memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
 312         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 313         check_error();
 314
 315         if (!epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") ||
 316             !epoxy_has_gl_extension("GL_ARB_sampler_objects")) {
 317                 fprintf(stderr, "Nageru requires GL_EXT_texture_sRGB_decode and GL_ARB_sampler_objects to run.\n");
 318                 exit(1);
 319         }
 320
 321         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 322         // will be halved when sampling them, and we need to compensate here.
 323         movit_texel_subpixel_precision /= 2.0;
 324
 325         resource_pool.reset(new ResourcePool);
 326         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 327                 output_channel[i].parent = this;
 328                 output_channel[i].channel = i;
 329         }
 330
 331         ImageFormat inout_format;
 332         inout_format.color_space = COLORSPACE_sRGB;
 333         inout_format.gamma_curve = GAMMA_sRGB;
 334
 335         // Matches the 4:2:0 format created by the main chain.
 336         YCbCrFormat ycbcr_format;
 337         ycbcr_format.chroma_subsampling_x = 2;
 338         ycbcr_format.chroma_subsampling_y = 2;
 339         if (global_flags.ycbcr_rec709_coefficients) {
 340                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 341         } else {
 342                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 343         }
 344         ycbcr_format.full_range = false;
 345         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 346         ycbcr_format.cb_x_position = 0.0f;
 347         ycbcr_format.cr_x_position = 0.0f;
 348         ycbcr_format.cb_y_position = 0.5f;
 349         ycbcr_format.cr_y_position = 0.5f;
 350
 351         // Initialize the neutral colors to sane values.
 352         for (unsigned i = 0; i < MAX_VIDEO_CARDS; ++i) {
 353                 last_received_neutral_color[i] = RGBTriplet(1.0f, 1.0f, 1.0f);
 354         }
 355
 356         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 357         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 358         check_error();
 359         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 360         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 361         display_chain->add_input(display_input);
 362         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 363         display_chain->set_dither_bits(0);  // Don't bother.
 364         display_chain->finalize();
 365
 366         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));
 367         if (!global_flags.card_to_mjpeg_stream_export.empty()) {
 368                 mjpeg_encoder.reset(new MJPEGEncoder(&httpd, global_flags.va_display));
 369         }
 370
 371         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 372         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 373
 374         // Must be instantiated after the theme, as the theme decides the number of FFmpeg inputs.
 375         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 376         audio_mixer.reset(new AudioMixer(num_cards, video_inputs.size()));
 377
 378         httpd.add_endpoint("/channels", bind(&Mixer::get_channels_json, this), HTTPD::ALLOW_ALL_ORIGINS);
 379         for (int channel_idx = 0; channel_idx < theme->get_num_channels(); ++channel_idx) {
 380                 char url[256];
 381                 snprintf(url, sizeof(url), "/channels/%d/color", channel_idx + 2);
 382                 httpd.add_endpoint(url, bind(&Mixer::get_channel_color_http, this, unsigned(channel_idx + 2)), HTTPD::ALLOW_ALL_ORIGINS);
 383         }
 384
 385         // Start listening for clients only once VideoEncoder has written its header, if any.
 386         httpd.start(global_flags.http_port);
 387
 388         // First try initializing the then PCI devices, then USB, then
 389         // fill up with fake cards until we have the desired number of cards.
 390         unsigned num_pci_devices = 0;
 391         unsigned card_index = 0;
 392
 393         {
 394                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 395                 if (decklink_iterator != nullptr) {
 396                         for ( ; card_index < num_cards; ++card_index) {
 397                                 IDeckLink *decklink;
 398                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 399                                         break;
 400                                 }
 401
 402                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 403                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 404                                 if (!output->set_device(decklink)) {
 405                                         delete output;
 406                                         output = nullptr;
 407                                 }
 408                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 409                                 ++num_pci_devices;
 410                         }
 411                         decklink_iterator->Release();
 412                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 413                 } else {
 414                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 415                 }
 416         }
 417
 418         unsigned num_usb_devices = BMUSBCapture::num_cards();
 419         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 420                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 421                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 422                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 423         }
 424         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 425
 426         unsigned num_fake_cards = 0;
 427         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 428                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 429                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 430         }
 431
 432         if (num_fake_cards > 0) {
 433                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 434         }
 435
 436         // Initialize all video inputs the theme asked for. Note that these are
 437         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 438         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 439                 if (card_index >= MAX_VIDEO_CARDS) {
 440                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 441                         abort();
 442                 }
 443                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 444                 video_inputs[video_card_index]->set_card_index(card_index);
 445         }
 446         num_video_inputs = video_inputs.size();
 447
 448 #ifdef HAVE_CEF
 449         // Same, for HTML inputs.
 450         std::vector<CEFCapture *> html_inputs = theme->get_html_inputs();
 451         for (unsigned html_card_index = 0; html_card_index < html_inputs.size(); ++card_index, ++html_card_index) {
 452                 if (card_index >= MAX_VIDEO_CARDS) {
 453                         fprintf(stderr, "ERROR: Not enough card slots available for the HTML inputs the theme requested.\n");
 454                         abort();
 455                 }
 456                 configure_card(card_index, html_inputs[html_card_index], CardType::CEF_INPUT, /*output=*/nullptr);
 457                 html_inputs[html_card_index]->set_card_index(card_index);
 458         }
 459         num_html_inputs = html_inputs.size();
 460 #endif
 461
 462         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 463         BMUSBCapture::start_bm_thread();
 464
 465         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 466                 cards[card_index].queue_length_policy.reset(card_index);
 467         }
 468
 469         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 470
 471         if (global_flags.ten_bit_input) {
 472                 if (!v210Converter::has_hardware_support()) {
 473                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 474                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 475                         abort();
 476                 }
 477                 v210_converter.reset(new v210Converter());
 478
 479                 // These are all the widths listed in the Blackmagic SDK documentation
 480                 // (section 2.7.3, “Display Modes”).
 481                 v210_converter->precompile_shader(720);
 482                 v210_converter->precompile_shader(1280);
 483                 v210_converter->precompile_shader(1920);
 484                 v210_converter->precompile_shader(2048);
 485                 v210_converter->precompile_shader(3840);
 486                 v210_converter->precompile_shader(4096);
 487         }
 488         if (global_flags.ten_bit_output) {
 489                 if (!v210Converter::has_hardware_support()) {
 490                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 491                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 492                         abort();
 493                 }
 494         }
 495
 496         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 497         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 498         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 499
 500         if (global_flags.enable_alsa_output) {
 501                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 502         }
 503         if (global_flags.output_card != -1) {
 504                 desired_output_card_index = global_flags.output_card;
 505                 set_output_card_internal(global_flags.output_card);
 506         }
 507
 508         output_jitter_history.register_metrics({{ "card", "output" }});
 509
 510         ImageInput::start_update_thread(image_update_surface);
 511 }
 512
 513 Mixer::~Mixer()
 514 {
 515         ImageInput::end_update_thread();
 516
 517         if (mjpeg_encoder != nullptr) {
 518                 mjpeg_encoder->stop();
 519         }
 520         httpd.stop();
 521         BMUSBCapture::stop_bm_thread();
 522
 523         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 524                 cards[card_index].capture->stop_dequeue_thread();
 525                 if (cards[card_index].output) {
 526                         cards[card_index].output->end_output();
 527                         cards[card_index].output.reset();
 528                 }
 529         }
 530
 531         video_encoder.reset(nullptr);
 532 }
 533
 534 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 535 {
 536         printf("Configuring card %d...\n", card_index);
 537
 538         CaptureCard *card = &cards[card_index];
 539         if (card->capture != nullptr) {
 540                 card->capture->stop_dequeue_thread();
 541         }
 542         card->capture.reset(capture);
 543         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 544         card->is_cef_capture = (card_type == CardType::CEF_INPUT);
 545         card->may_have_dropped_last_frame = false;
 546         card->type = card_type;
 547         if (card->output.get() != output) {
 548                 card->output.reset(output);
 549         }
 550
 551         PixelFormat pixel_format;
 552         if (card_type == CardType::FFMPEG_INPUT) {
 553                 pixel_format = capture->get_current_pixel_format();
 554         } else if (card_type == CardType::CEF_INPUT) {
 555                 pixel_format = PixelFormat_8BitBGRA;
 556         } else if (global_flags.ten_bit_input) {
 557                 pixel_format = PixelFormat_10BitYCbCr;
 558         } else {
 559                 pixel_format = PixelFormat_8BitYCbCr;
 560         }
 561
 562         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 563         if (card->frame_allocator == nullptr) {
 564                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height, card_index, mjpeg_encoder.get()));  // 8 MB.
 565         }
 566         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 567         if (card->surface == nullptr) {
 568                 card->surface = create_surface_with_same_format(mixer_surface);
 569         }
 570         while (!card->new_frames.empty()) card->new_frames.pop_front();
 571         card->last_timecode = -1;
 572         card->capture->set_pixel_format(pixel_format);
 573         card->capture->configure_card();
 574
 575         // NOTE: start_bm_capture() happens in thread_func().
 576
 577         DeviceSpec device;
 578         if (card_type == CardType::FFMPEG_INPUT) {
 579                 device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 580         } else {
 581                 device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 582         }
 583         audio_mixer->reset_resampler(device);
 584         audio_mixer->set_display_name(device, card->capture->get_description());
 585         audio_mixer->trigger_state_changed_callback();
 586
 587         // Unregister old metrics, if any.
 588         if (!card->labels.empty()) {
 589                 const vector<pair<string, string>> &labels = card->labels;
 590                 card->jitter_history.unregister_metrics(labels);
 591                 card->queue_length_policy.unregister_metrics(labels);
 592                 global_metrics.remove("input_received_frames", labels);
 593                 global_metrics.remove("input_dropped_frames_jitter", labels);
 594                 global_metrics.remove("input_dropped_frames_error", labels);
 595                 global_metrics.remove("input_dropped_frames_resets", labels);
 596                 global_metrics.remove("input_queue_length_frames", labels);
 597                 global_metrics.remove("input_queue_duped_frames", labels);
 598
 599                 global_metrics.remove("input_has_signal_bool", labels);
 600                 global_metrics.remove("input_is_connected_bool", labels);
 601                 global_metrics.remove("input_interlaced_bool", labels);
 602                 global_metrics.remove("input_width_pixels", labels);
 603                 global_metrics.remove("input_height_pixels", labels);
 604                 global_metrics.remove("input_frame_rate_nom", labels);
 605                 global_metrics.remove("input_frame_rate_den", labels);
 606                 global_metrics.remove("input_sample_rate_hz", labels);
 607         }
 608
 609         // Register metrics.
 610         vector<pair<string, string>> labels;
 611         char card_name[64];
 612         snprintf(card_name, sizeof(card_name), "%d", card_index);
 613         labels.emplace_back("card", card_name);
 614
 615         switch (card_type) {
 616         case CardType::LIVE_CARD:
 617                 labels.emplace_back("cardtype", "live");
 618                 break;
 619         case CardType::FAKE_CAPTURE:
 620                 labels.emplace_back("cardtype", "fake");
 621                 break;
 622         case CardType::FFMPEG_INPUT:
 623                 labels.emplace_back("cardtype", "ffmpeg");
 624                 break;
 625         case CardType::CEF_INPUT:
 626                 labels.emplace_back("cardtype", "cef");
 627                 break;
 628         default:
 629                 assert(false);
 630         }
 631         card->jitter_history.register_metrics(labels);
 632         card->queue_length_policy.register_metrics(labels);
 633         global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
 634         global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
 635         global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
 636         global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
 637         global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
 638         global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
 639
 640         global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
 641         global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
 642         global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
 643         global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
 644         global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
 645         global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
 646         global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
 647         global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
 648         card->labels = labels;
 649 }
 650
 651 void Mixer::set_output_card_internal(int card_index)
 652 {
 653         // We don't really need to take card_mutex, since we're in the mixer
 654         // thread and don't mess with any queues (which is the only thing that happens
 655         // from other threads), but it's probably the safest in the long run.
 656         unique_lock<mutex> lock(card_mutex);
 657         if (output_card_index != -1) {
 658                 // Switch the old card from output to input.
 659                 CaptureCard *old_card = &cards[output_card_index];
 660                 old_card->output->end_output();
 661
 662                 // Stop the fake card that we put into place.
 663                 // This needs to _not_ happen under the mutex, to avoid deadlock
 664                 // (delivering the last frame needs to take the mutex).
 665                 CaptureInterface *fake_capture = old_card->capture.get();
 666                 lock.unlock();
 667                 fake_capture->stop_dequeue_thread();
 668                 lock.lock();
 669                 old_card->capture = move(old_card->parked_capture);  // TODO: reset the metrics
 670                 old_card->is_fake_capture = false;
 671                 old_card->capture->start_bm_capture();
 672         }
 673         if (card_index != -1) {
 674                 CaptureCard *card = &cards[card_index];
 675                 CaptureInterface *capture = card->capture.get();
 676                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 677                 // several seconds to complete (blocking on DisableVideoInput);
 678                 // see if we can maybe do it asynchronously.
 679                 lock.unlock();
 680                 capture->stop_dequeue_thread();
 681                 lock.lock();
 682                 card->parked_capture = move(card->capture);
 683                 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 684                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 685                 card->queue_length_policy.reset(card_index);
 686                 card->capture->start_bm_capture();
 687                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 688                 card->output->start_output(desired_output_video_mode, pts_int);
 689         }
 690         output_card_index = card_index;
 691         output_jitter_history.clear();
 692 }
 693
 694 namespace {
 695
 696 int unwrap_timecode(uint16_t current_wrapped, int last)
 697 {
 698         uint16_t last_wrapped = last & 0xffff;
 699         if (current_wrapped > last_wrapped) {
 700                 return (last & ~0xffff) | current_wrapped;
 701         } else {
 702                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 703         }
 704 }
 705
 706 DeviceSpec card_index_to_device(unsigned card_index, unsigned num_cards)
 707 {
 708         if (card_index >= num_cards) {
 709                 return DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 710         } else {
 711                 return DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 712         }
 713 }
 714
 715 }  // namespace
 716
 717 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 718                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 719                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 720 {
 721         DeviceSpec device = card_index_to_device(card_index, num_cards);
 722         CaptureCard *card = &cards[card_index];
 723
 724         ++card->metric_input_received_frames;
 725         card->metric_input_has_signal_bool = video_format.has_signal;
 726         card->metric_input_is_connected_bool = video_format.is_connected;
 727         card->metric_input_interlaced_bool = video_format.interlaced;
 728         card->metric_input_width_pixels = video_format.width;
 729         card->metric_input_height_pixels = video_format.height;
 730         card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
 731         card->metric_input_frame_rate_den = video_format.frame_rate_den;
 732         card->metric_input_sample_rate_hz = audio_format.sample_rate;
 733
 734         if (is_mode_scanning[card_index]) {
 735                 if (video_format.has_signal) {
 736                         // Found a stable signal, so stop scanning.
 737                         is_mode_scanning[card_index] = false;
 738                 } else {
 739                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 740                         steady_clock::time_point now = steady_clock::now();
 741                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 742                         if (sec_since_last_switch > switch_time_s) {
 743                                 // It isn't this mode; try the next one.
 744                                 mode_scanlist_index[card_index]++;
 745                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 746                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 747                                 last_mode_scan_change[card_index] = now;
 748                         }
 749                 }
 750         }
 751
 752         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 753         assert(frame_length > 0);
 754
 755         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;
 756         if (num_samples > OUTPUT_FREQUENCY / 10 && card->type != CardType::FFMPEG_INPUT) {
 757                 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",
 758                         spec_to_string(device).c_str(), int(audio_frame.len), int(audio_offset),
 759                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 760                 if (video_frame.owner) {
 761                         video_frame.owner->release_frame(video_frame);
 762                 }
 763                 if (audio_frame.owner) {
 764                         audio_frame.owner->release_frame(audio_frame);
 765                 }
 766                 return;
 767         }
 768
 769         int dropped_frames = 0;
 770         if (card->last_timecode != -1) {
 771                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 772         }
 773
 774         // Number of samples per frame if we need to insert silence.
 775         // (Could be nonintegral, but resampling will save us then.)
 776         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 777
 778         if (dropped_frames > MAX_FPS * 2) {
 779                 fprintf(stderr, "%s lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 780                         spec_to_string(device).c_str(), card->last_timecode, timecode);
 781                 audio_mixer->reset_resampler(device);
 782                 dropped_frames = 0;
 783                 ++card->metric_input_resets;
 784         } else if (dropped_frames > 0) {
 785                 // Insert silence as needed.
 786                 fprintf(stderr, "%s dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 787                         spec_to_string(device).c_str(), dropped_frames, timecode);
 788                 card->metric_input_dropped_frames_error += dropped_frames;
 789
 790                 bool success;
 791                 do {
 792                         success = audio_mixer->add_silence(device, silence_samples, dropped_frames);
 793                 } while (!success);
 794         }
 795
 796         if (num_samples > 0) {
 797                 audio_mixer->add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, audio_frame.received_timestamp);
 798
 799                 // Audio for the MJPEG stream. We don't resample; audio that's not in 48 kHz
 800                 // just gets dropped for now.
 801                 //
 802                 // Only bother doing MJPEG encoding if there are any connected clients
 803                 // that want the stream.
 804                 if (httpd.get_num_connected_multicam_clients() > 0 ||
 805                     httpd.get_num_connected_siphon_clients(card_index) > 0) {
 806                         vector<int32_t> converted_samples = convert_audio_to_fixed32(audio_frame.data + audio_offset, num_samples, audio_format, 2);
 807                         lock_guard<mutex> lock(card_mutex);
 808                         if (card->new_raw_audio.empty()) {
 809                                 card->new_raw_audio = move(converted_samples);
 810                         } else {
 811                                 // For raw audio, we don't really synchronize audio and video;
 812                                 // we just put the audio in frame by frame, and if a video frame is
 813                                 // dropped, we still keep the audio, which means it will be added
 814                                 // to the beginning of the next frame. It would probably be better
 815                                 // to move the audio pts earlier to show this, but most players can
 816                                 // live with some jitter, and in a lot of ways, it's much nicer for
 817                                 // Futatabi to have all audio locked to a video frame.
 818                                 card->new_raw_audio.insert(card->new_raw_audio.end(), converted_samples.begin(), converted_samples.end());
 819
 820                                 // Truncate to one second, just to be sure we don't have infinite buildup in case of weirdness.
 821                                 if (card->new_raw_audio.size() > OUTPUT_FREQUENCY * 2) {
 822                                         size_t excess_samples = card->new_raw_audio.size() - OUTPUT_FREQUENCY * 2;
 823                                         card->new_raw_audio.erase(card->new_raw_audio.begin(), card->new_raw_audio.begin() + excess_samples);
 824                                 }
 825                         }
 826                 }
 827         }
 828
 829         // Done with the audio, so release it.
 830         if (audio_frame.owner) {
 831                 audio_frame.owner->release_frame(audio_frame);
 832         }
 833
 834         card->last_timecode = timecode;
 835
 836         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 837         if (card->type == CardType::FFMPEG_INPUT && userdata != nullptr) {
 838                 FFmpegCapture *ffmpeg_capture = static_cast<FFmpegCapture *>(card->capture.get());
 839                 userdata->has_last_subtitle = ffmpeg_capture->get_has_last_subtitle();
 840                 userdata->last_subtitle = ffmpeg_capture->get_last_subtitle();
 841         }
 842
 843         size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
 844         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 845         if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
 846                 // The calculation above is wrong for planar Y'CbCr, so just override it.
 847                 assert(card->type == CardType::FFMPEG_INPUT);
 848                 assert(video_offset == 0);
 849                 expected_length = video_frame.len;
 850
 851                 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
 852                 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
 853                 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
 854                 cbcr_offset = video_format.width * video_format.height;
 855                 y_offset = 0;
 856         } else {
 857                 // All the other Y'CbCr formats are 4:2:2.
 858                 cbcr_width = video_format.width / 2;
 859                 cbcr_height = video_format.height;
 860                 cbcr_offset = video_offset / 2;
 861                 y_offset = video_frame.size / 2 + video_offset / 2;
 862         }
 863         if (video_frame.len - video_offset == 0 ||
 864             video_frame.len - video_offset != expected_length) {
 865                 if (video_frame.len != 0) {
 866                         printf("%s: Dropping video frame with wrong length (%zu; expected %zu)\n",
 867                                 spec_to_string(device).c_str(), video_frame.len - video_offset, expected_length);
 868                 }
 869                 if (video_frame.owner) {
 870                         video_frame.owner->release_frame(video_frame);
 871                 }
 872
 873                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 874                 // so that pts can go up accordingly.
 875                 {
 876                         lock_guard<mutex> lock(card_mutex);
 877                         CaptureCard::NewFrame new_frame;
 878                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 879                         new_frame.length = frame_length;
 880                         new_frame.interlaced = false;
 881                         new_frame.dropped_frames = dropped_frames;
 882                         new_frame.received_timestamp = video_frame.received_timestamp;
 883                         card->new_frames.push_back(move(new_frame));
 884                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
 885                 }
 886                 card->new_frames_changed.notify_all();
 887                 return;
 888         }
 889
 890         unsigned num_fields = video_format.interlaced ? 2 : 1;
 891         steady_clock::time_point frame_upload_start;
 892         bool interlaced_stride = false;
 893         if (video_format.interlaced) {
 894                 // Send the two fields along as separate frames; the other side will need to add
 895                 // a deinterlacer to actually get this right.
 896                 assert(video_format.height % 2 == 0);
 897                 video_format.height /= 2;
 898                 cbcr_height /= 2;
 899                 assert(frame_length % 2 == 0);
 900                 frame_length /= 2;
 901                 num_fields = 2;
 902                 if (video_format.second_field_start == 1) {
 903                         interlaced_stride = true;
 904                 }
 905                 frame_upload_start = steady_clock::now();
 906         }
 907         assert(userdata != nullptr);
 908         userdata->last_interlaced = video_format.interlaced;
 909         userdata->last_has_signal = video_format.has_signal;
 910         userdata->last_is_connected = video_format.is_connected;
 911         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 912         userdata->last_frame_rate_den = video_format.frame_rate_den;
 913         RefCountedFrame frame(video_frame);
 914
 915         // Upload the textures.
 916         for (unsigned field = 0; field < num_fields; ++field) {
 917                 // Put the actual texture upload in a lambda that is executed in the main thread.
 918                 // It is entirely possible to do this in the same thread (and it might even be
 919                 // faster, depending on the GPU and driver), but it appears to be trickling
 920                 // driver bugs very easily.
 921                 //
 922                 // Note that this means we must hold on to the actual frame data in <userdata>
 923                 // until the upload command is run, but we hold on to <frame> much longer than that
 924                 // (in fact, all the way until we no longer use the texture in rendering).
 925                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
 926                         unsigned field_start_line;
 927                         if (field == 1) {
 928                                 field_start_line = video_format.second_field_start;
 929                         } else {
 930                                 field_start_line = video_format.extra_lines_top;
 931                         }
 932
 933                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 934                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 935                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
 936
 937                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 938                         check_error();
 939
 940                         switch (userdata->pixel_format) {
 941                         case PixelFormat_10BitYCbCr: {
 942                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 943                                 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);
 944                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 945                                 break;
 946                         }
 947                         case PixelFormat_8BitYCbCr: {
 948                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 949                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 950
 951                                 // Make up our own strides, since we are interleaving.
 952                                 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);
 953                                 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);
 954                                 break;
 955                         }
 956                         case PixelFormat_8BitYCbCrPlanar: {
 957                                 assert(field_start_line == 0);  // We don't really support interlaced here.
 958                                 size_t field_y_start = y_offset;
 959                                 size_t field_cb_start = cbcr_offset;
 960                                 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
 961
 962                                 // Make up our own strides, since we are interleaving.
 963                                 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);
 964                                 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
 965                                 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
 966                                 break;
 967                         }
 968                         case PixelFormat_8BitBGRA: {
 969                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 970                                 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);
 971                                 // These could be asked to deliver mipmaps at any time.
 972                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 973                                 check_error();
 974                                 glGenerateMipmap(GL_TEXTURE_2D);
 975                                 check_error();
 976                                 glBindTexture(GL_TEXTURE_2D, 0);
 977                                 check_error();
 978                                 break;
 979                         }
 980                         default:
 981                                 assert(false);
 982                         }
 983
 984                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 985                         check_error();
 986                 };
 987
 988                 if (field == 1) {
 989                         // Don't upload the second field as fast as we can; wait until
 990                         // the field time has approximately passed. (Otherwise, we could
 991                         // get timing jitter against the other sources, and possibly also
 992                         // against the video display, although the latter is not as critical.)
 993                         // This requires our system clock to be reasonably close to the
 994                         // video clock, but that's not an unreasonable assumption.
 995                         steady_clock::time_point second_field_start = frame_upload_start +
 996                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 997                         this_thread::sleep_until(second_field_start);
 998                 }
 999
1000                 {
1001                         lock_guard<mutex> lock(card_mutex);
1002                         CaptureCard::NewFrame new_frame;
1003                         new_frame.frame = frame;
1004                         new_frame.length = frame_length;
1005                         new_frame.field = field;
1006                         new_frame.interlaced = video_format.interlaced;
1007                         new_frame.upload_func = upload_func;
1008                         new_frame.dropped_frames = dropped_frames;
1009                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
1010                         new_frame.video_format = video_format;
1011                         new_frame.y_offset = y_offset;
1012                         new_frame.cbcr_offset = cbcr_offset;
1013                         if (card->type == CardType::FFMPEG_INPUT) {
1014                                 FFmpegCapture *ffmpeg_capture = static_cast<FFmpegCapture *>(card->capture.get());
1015                                 new_frame.neutral_color = ffmpeg_capture->get_last_neutral_color();
1016                         }
1017                         card->new_frames.push_back(move(new_frame));
1018                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
1019                         card->may_have_dropped_last_frame = false;
1020                 }
1021                 card->new_frames_changed.notify_all();
1022         }
1023 }
1024
1025 void Mixer::bm_hotplug_add(libusb_device *dev)
1026 {
1027         lock_guard<mutex> lock(hotplug_mutex);
1028         hotplugged_cards.push_back(dev);
1029 }
1030
1031 void Mixer::bm_hotplug_remove(unsigned card_index)
1032 {
1033         cards[card_index].new_frames_changed.notify_all();
1034 }
1035
1036 void Mixer::thread_func()
1037 {
1038         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
1039
1040         eglBindAPI(EGL_OPENGL_API);
1041         QOpenGLContext *context = create_context(mixer_surface);
1042         if (!make_current(context, mixer_surface)) {
1043                 printf("oops\n");
1044                 abort();
1045         }
1046
1047         // Start the actual capture. (We don't want to do it before we're actually ready
1048         // to process output frames.)
1049         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1050                 if (int(card_index) != output_card_index) {
1051                         cards[card_index].capture->start_bm_capture();
1052                 }
1053         }
1054
1055         BasicStats basic_stats(/*verbose=*/true, /*use_opengl=*/true);
1056         int stats_dropped_frames = 0;
1057
1058         while (!should_quit) {
1059                 if (desired_output_card_index != output_card_index) {
1060                         set_output_card_internal(desired_output_card_index);
1061                 }
1062                 if (output_card_index != -1 &&
1063                     desired_output_video_mode != output_video_mode) {
1064                         DeckLinkOutput *output = cards[output_card_index].output.get();
1065                         output->end_output();
1066                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
1067                         output->start_output(desired_output_video_mode, pts_int);
1068                 }
1069
1070                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
1071                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
1072
1073                 bool master_card_is_output;
1074                 unsigned master_card_index;
1075                 if (output_card_index != -1) {
1076                         master_card_is_output = true;
1077                         master_card_index = output_card_index;
1078                 } else {
1079                         master_card_is_output = false;
1080                         master_card_index = theme->map_signal(master_clock_channel);
1081                         assert(master_card_index < num_cards + num_video_inputs);
1082                 }
1083
1084                 vector<int32_t> raw_audio[MAX_VIDEO_CARDS];  // For MJPEG encoding.
1085                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame, raw_audio);
1086                 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);
1087                 stats_dropped_frames += output_frame_info.dropped_frames;
1088
1089                 handle_hotplugged_cards();
1090
1091                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1092                         DeviceSpec device = card_index_to_device(card_index, num_cards);
1093                         if (card_index == master_card_index || !has_new_frame[card_index]) {
1094                                 continue;
1095                         }
1096                         if (new_frames[card_index].frame->len == 0) {
1097                                 ++new_frames[card_index].dropped_frames;
1098                         }
1099                         if (new_frames[card_index].dropped_frames > 0) {
1100                                 printf("%s dropped %d frames before this\n",
1101                                         spec_to_string(device).c_str(), int(new_frames[card_index].dropped_frames));
1102                         }
1103                 }
1104
1105                 // If the first card is reporting a corrupted or otherwise dropped frame,
1106                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
1107                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
1108                         ++stats_dropped_frames;
1109                         pts_int += new_frames[master_card_index].length;
1110                         continue;
1111                 }
1112
1113                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1114                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1115                                 continue;
1116
1117                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1118                         assert(new_frame->frame != nullptr);
1119                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1120                         check_error();
1121
1122                         // The new texture might need uploading before use.
1123                         if (new_frame->upload_func) {
1124                                 new_frame->upload_func();
1125                                 new_frame->upload_func = nullptr;
1126                         }
1127
1128                         // Only set the white balance if it actually changed. This means that the user
1129                         // is free to override the white balance in a video with no white balance information
1130                         // actually set (ie. r=g=b=1 all the time), or one where the white point is wrong,
1131                         // but frame-to-frame decisions will be heeded. We do this pretty much as late
1132                         // as possible (ie., after picking out the frame from the buffer), so that we are sure
1133                         // that the change takes effect on exactly the right frame.
1134                         if (fabs(new_frame->neutral_color.r - last_received_neutral_color[card_index].r) > 1e-3 ||
1135                             fabs(new_frame->neutral_color.g - last_received_neutral_color[card_index].g) > 1e-3 ||
1136                             fabs(new_frame->neutral_color.b - last_received_neutral_color[card_index].b) > 1e-3) {
1137                                 theme->set_wb_for_signal(card_index, new_frame->neutral_color.r, new_frame->neutral_color.g, new_frame->neutral_color.b);
1138                                 last_received_neutral_color[card_index] = new_frame->neutral_color;
1139                         }
1140
1141                         if (new_frame->frame->data_copy != nullptr && mjpeg_encoder->should_encode_mjpeg_for_card(card_index)) {
1142                                 RGBTriplet neutral_color = theme->get_white_balance_for_signal(card_index);
1143                                 mjpeg_encoder->upload_frame(pts_int, card_index, new_frame->frame, new_frame->video_format, new_frame->y_offset, new_frame->cbcr_offset, move(raw_audio[card_index]), neutral_color);
1144                         }
1145
1146                 }
1147
1148                 int64_t frame_duration = output_frame_info.frame_duration;
1149                 render_one_frame(frame_duration);
1150                 {
1151                         lock_guard<mutex> lock(frame_num_mutex);
1152                         ++frame_num;
1153                 }
1154                 frame_num_updated.notify_all();
1155                 pts_int += frame_duration;
1156
1157                 basic_stats.update(frame_num, stats_dropped_frames);
1158                 // if (frame_num % 100 == 0) chain->print_phase_timing();
1159
1160                 if (should_cut.exchange(false)) {  // Test and clear.
1161                         video_encoder->do_cut(frame_num);
1162                 }
1163
1164 #if 0
1165                 // Reset every 100 frames, so that local variations in frame times
1166                 // (especially for the first few frames, when the shaders are
1167                 // compiled etc.) don't make it hard to measure for the entire
1168                 // remaining duration of the program.
1169                 if (frame == 10000) {
1170                         frame = 0;
1171                         start = now;
1172                 }
1173 #endif
1174                 check_error();
1175         }
1176
1177         resource_pool->clean_context();
1178 }
1179
1180 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1181 {
1182         if (output_card_index != -1) {
1183                 // The output card (ie., cards[output_card_index].output) is the master clock,
1184                 // so no input card (ie., cards[card_index].capture) is.
1185                 return false;
1186         }
1187         return (card_index == master_card_index);
1188 }
1189
1190 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1191 {
1192         // Count the number of frames in the queue, including any frames
1193         // we dropped. It's hard to know exactly how we should deal with
1194         // dropped (corrupted) input frames; they don't help our goal of
1195         // avoiding starvation, but they still add to the problem of latency.
1196         // Since dropped frames is going to mean a bump in the signal anyway,
1197         // we err on the side of having more stable latency instead.
1198         unsigned queue_length = 0;
1199         for (const CaptureCard::NewFrame &frame : card->new_frames) {
1200                 queue_length += frame.dropped_frames + 1;
1201         }
1202
1203         // If needed, drop frames until the queue is below the safe limit.
1204         // We prefer to drop from the head, because all else being equal,
1205         // we'd like more recent frames (less latency).
1206         unsigned dropped_frames = 0;
1207         while (queue_length > safe_queue_length) {
1208                 assert(!card->new_frames.empty());
1209                 assert(queue_length > card->new_frames.front().dropped_frames);
1210                 queue_length -= card->new_frames.front().dropped_frames;
1211
1212                 if (queue_length <= safe_queue_length) {
1213                         // No need to drop anything.
1214                         break;
1215                 }
1216
1217                 card->new_frames.pop_front();
1218                 card->new_frames_changed.notify_all();
1219                 --queue_length;
1220                 ++dropped_frames;
1221
1222                 if (queue_length == 0 && card->is_cef_capture) {
1223                         card->may_have_dropped_last_frame = true;
1224                 }
1225         }
1226
1227         card->metric_input_dropped_frames_jitter += dropped_frames;
1228         card->metric_input_queue_length_frames = queue_length;
1229
1230 #if 0
1231         if (dropped_frames > 0) {
1232                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1233                         card_index, dropped_frames);
1234         }
1235 #endif
1236 }
1237
1238 pair<string, string> Mixer::get_channels_json()
1239 {
1240         Channels ret;
1241         for (int channel_idx = 2; channel_idx < theme->get_num_channels(); ++channel_idx) {
1242                 Channel *channel = ret.add_channel();
1243                 channel->set_index(channel_idx);
1244                 channel->set_name(theme->get_channel_name(channel_idx));
1245                 channel->set_color(theme->get_channel_color(channel_idx));
1246         }
1247         string contents;
1248         google::protobuf::util::MessageToJsonString(ret, &contents);  // Ignore any errors.
1249         return make_pair(contents, "text/json");
1250 }
1251
1252 pair<string, string> Mixer::get_channel_color_http(unsigned channel_idx)
1253 {
1254         return make_pair(theme->get_channel_color(channel_idx), "text/plain");
1255 }
1256
1257 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])
1258 {
1259         OutputFrameInfo output_frame_info;
1260 start:
1261         unique_lock<mutex> lock(card_mutex, defer_lock);
1262         if (master_card_is_output) {
1263                 // Clocked to the output, so wait for it to be ready for the next frame.
1264                 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);
1265                 lock.lock();
1266         } else {
1267                 // Wait for the master card to have a new frame.
1268                 // TODO: Add a timeout.
1269                 output_frame_info.is_preroll = false;
1270                 lock.lock();
1271                 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(); });
1272         }
1273
1274         if (master_card_is_output) {
1275                 handle_hotplugged_cards();
1276         } else if (cards[master_card_index].new_frames.empty()) {
1277                 // We were woken up, but not due to a new frame. Deal with it
1278                 // and then restart.
1279                 assert(cards[master_card_index].capture->get_disconnected());
1280                 handle_hotplugged_cards();
1281                 lock.unlock();
1282                 goto start;
1283         }
1284
1285         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1286                 CaptureCard *card = &cards[card_index];
1287                 if (card->new_frames.empty()) {  // Starvation.
1288                         ++card->metric_input_duped_frames;
1289 #ifdef HAVE_CEF
1290                         if (card->is_cef_capture && card->may_have_dropped_last_frame) {
1291                                 // Unlike other sources, CEF is not guaranteed to send us a steady
1292                                 // stream of frames, so we'll have to ask it to repaint the frame
1293                                 // we dropped. (may_have_dropped_last_frame is set whenever we
1294                                 // trim the queue completely away, and cleared when we actually
1295                                 // get a new frame.)
1296                                 ((CEFCapture *)card->capture.get())->request_new_frame(/*ignore_if_locked=*/true);
1297                         }
1298 #endif
1299                 } else {
1300                         new_frames[card_index] = move(card->new_frames.front());
1301                         has_new_frame[card_index] = true;
1302                         card->new_frames.pop_front();
1303                         card->new_frames_changed.notify_all();
1304                 }
1305
1306                 raw_audio[card_index] = move(card->new_raw_audio);
1307         }
1308
1309         if (!master_card_is_output) {
1310                 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1311                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1312                 output_frame_info.frame_duration = new_frames[master_card_index].length;
1313         }
1314
1315         if (!output_frame_info.is_preroll) {
1316                 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1317         }
1318
1319         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1320                 CaptureCard *card = &cards[card_index];
1321                 if (has_new_frame[card_index] &&
1322                     !input_card_is_master_clock(card_index, master_card_index) &&
1323                     !output_frame_info.is_preroll) {
1324                         card->queue_length_policy.update_policy(
1325                                 output_frame_info.frame_timestamp,
1326                                 card->jitter_history.get_expected_next_frame(),
1327                                 new_frames[master_card_index].length,
1328                                 output_frame_info.frame_duration,
1329                                 card->jitter_history.estimate_max_jitter(),
1330                                 output_jitter_history.estimate_max_jitter());
1331                         trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1332                                                   card->queue_length_policy.get_safe_queue_length()));
1333                 }
1334         }
1335
1336         // This might get off by a fractional sample when changing master card
1337         // between ones with different frame rates, but that's fine.
1338         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1339         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1340         fractional_samples = num_samples_times_timebase % TIMEBASE;
1341         assert(output_frame_info.num_samples >= 0);
1342
1343         return output_frame_info;
1344 }
1345
1346 void Mixer::handle_hotplugged_cards()
1347 {
1348         // Check for cards that have been disconnected since last frame.
1349         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1350                 CaptureCard *card = &cards[card_index];
1351                 if (card->capture->get_disconnected()) {
1352                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1353                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1354                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1355                         card->queue_length_policy.reset(card_index);
1356                         card->capture->start_bm_capture();
1357                 }
1358         }
1359
1360         // Check for cards that have been connected since last frame.
1361         vector<libusb_device *> hotplugged_cards_copy;
1362         {
1363                 lock_guard<mutex> lock(hotplug_mutex);
1364                 swap(hotplugged_cards, hotplugged_cards_copy);
1365         }
1366         for (libusb_device *new_dev : hotplugged_cards_copy) {
1367                 // Look for a fake capture card where we can stick this in.
1368                 int free_card_index = -1;
1369                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1370                         if (cards[card_index].is_fake_capture) {
1371                                 free_card_index = card_index;
1372                                 break;
1373                         }
1374                 }
1375
1376                 if (free_card_index == -1) {
1377                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1378                         libusb_unref_device(new_dev);
1379                 } else {
1380                         // BMUSBCapture takes ownership.
1381                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1382                         CaptureCard *card = &cards[free_card_index];
1383                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1384                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1385                         card->queue_length_policy.reset(free_card_index);
1386                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1387                         capture->start_bm_capture();
1388                 }
1389         }
1390 }
1391
1392
1393 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)
1394 {
1395         // Resample the audio as needed, including from previously dropped frames.
1396         assert(num_cards > 0);
1397         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1398                 const bool dropped_frame = (frame_num != dropped_frames);
1399                 {
1400                         // Signal to the audio thread to process this frame.
1401                         // Note that if the frame is a dropped frame, we signal that
1402                         // we don't want to use this frame as base for adjusting
1403                         // the resampler rate. The reason for this is that the timing
1404                         // of these frames is often way too late; they typically don't
1405                         // “arrive” before we synthesize them. Thus, we could end up
1406                         // in a situation where we have inserted e.g. five audio frames
1407                         // into the queue before we then start pulling five of them
1408                         // back out. This makes ResamplingQueue overestimate the delay,
1409                         // causing undue resampler changes. (We _do_ use the last,
1410                         // non-dropped frame; perhaps we should just discard that as well,
1411                         // since dropped frames are expected to be rare, and it might be
1412                         // better to just wait until we have a slightly more normal situation).
1413                         lock_guard<mutex> lock(audio_mutex);
1414                         bool adjust_rate = !dropped_frame && !is_preroll;
1415                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1416                         audio_task_queue_changed.notify_one();
1417                 }
1418                 if (dropped_frame) {
1419                         // For dropped frames, increase the pts. Note that if the format changed
1420                         // in the meantime, we have no way of detecting that; we just have to
1421                         // assume the frame length is always the same.
1422                         pts_int += length_per_frame;
1423                 }
1424         }
1425 }
1426
1427 void Mixer::render_one_frame(int64_t duration)
1428 {
1429         // Determine the time code for this frame before we start rendering.
1430         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1431         if (display_timecode_on_stdout) {
1432                 printf("Timecode: '%s'\n", timecode_text.c_str());
1433         }
1434
1435         // Update Y'CbCr settings for all cards.
1436         {
1437                 lock_guard<mutex> lock(card_mutex);
1438                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1439                         YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1440                         input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1441                         input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1442                         input_state.full_range[card_index] = interpretation->full_range;
1443                 }
1444         }
1445
1446         // Get the main chain from the theme, and set its state immediately.
1447         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1448         EffectChain *chain = theme_main_chain.chain;
1449         theme_main_chain.setup_chain();
1450         //theme_main_chain.chain->enable_phase_timing(true);
1451
1452         // If HDMI/SDI output is active and the user has requested auto mode,
1453         // its mode overrides the existing Y'CbCr setting for the chain.
1454         YCbCrLumaCoefficients ycbcr_output_coefficients;
1455         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1456                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1457         } else {
1458                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1459         }
1460
1461         // TODO: Reduce the duplication against theme.cpp.
1462         YCbCrFormat output_ycbcr_format;
1463         output_ycbcr_format.chroma_subsampling_x = 1;
1464         output_ycbcr_format.chroma_subsampling_y = 1;
1465         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1466         output_ycbcr_format.full_range = false;
1467         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1468         chain->change_ycbcr_output_format(output_ycbcr_format);
1469
1470         // Render main chain. If we're using zerocopy Quick Sync encoding
1471         // (the default case), we take an extra copy of the created outputs,
1472         // so that we can display it back to the screen later (it's less memory
1473         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1474         // Ideally, we'd like to avoid taking copies and just use the main textures
1475         // for display as well, but they're just views into VA-API memory and must be
1476         // unmapped during encoding, so we can't use them for display, unfortunately.
1477         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1478         GLuint y_copy_tex, cbcr_copy_tex = 0;
1479         GLuint y_display_tex, cbcr_display_tex;
1480         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1481         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1482         const bool is_zerocopy = video_encoder->is_zerocopy();
1483         if (is_zerocopy) {
1484                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1485                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1486                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1487
1488                 y_display_tex = y_copy_tex;
1489                 cbcr_display_tex = cbcr_copy_tex;
1490
1491                 // y_tex and cbcr_tex will be given by VideoEncoder.
1492         } else {
1493                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1494                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1495                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1496
1497                 y_display_tex = y_tex;
1498                 cbcr_display_tex = cbcr_tex;
1499         }
1500
1501         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1502         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1503         assert(got_frame);
1504
1505         GLuint fbo;
1506         if (is_zerocopy) {
1507                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1508         } else {
1509                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1510         }
1511         check_error();
1512         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1513
1514         if (display_timecode_in_stream) {
1515                 // Render the timecode on top.
1516                 timecode_renderer->render_timecode(fbo, timecode_text);
1517         }
1518
1519         resource_pool->release_fbo(fbo);
1520
1521         if (is_zerocopy) {
1522                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1523         } else {
1524                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1525         }
1526         if (output_card_index != -1) {
1527                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1528         }
1529         resource_pool->release_2d_texture(cbcr_full_tex);
1530
1531         // Set the right state for the Y' and CbCr textures we use for display.
1532         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1533         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1534         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1535         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1536         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1537
1538         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1539         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1540         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1541         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1542
1543         RefCountedGLsync fence = video_encoder->end_frame();
1544
1545         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1546         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1547         DisplayFrame live_frame;
1548         live_frame.chain = display_chain.get();
1549         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1550                 display_input->set_texture_num(0, y_display_tex);
1551                 display_input->set_texture_num(1, cbcr_display_tex);
1552         };
1553         live_frame.ready_fence = fence;
1554         live_frame.input_frames = {};
1555         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1556         output_channel[OUTPUT_LIVE].output_frame(move(live_frame));
1557
1558         // Set up preview and any additional channels.
1559         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1560                 DisplayFrame display_frame;
1561                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1562                 display_frame.chain = move(chain.chain);
1563                 display_frame.setup_chain = move(chain.setup_chain);
1564                 display_frame.ready_fence = fence;
1565                 display_frame.input_frames = move(chain.input_frames);
1566                 display_frame.temp_textures = {};
1567                 output_channel[i].output_frame(move(display_frame));
1568         }
1569 }
1570
1571 void Mixer::audio_thread_func()
1572 {
1573         pthread_setname_np(pthread_self(), "Mixer_Audio");
1574
1575         while (!should_quit) {
1576                 AudioTask task;
1577
1578                 {
1579                         unique_lock<mutex> lock(audio_mutex);
1580                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1581                         if (should_quit) {
1582                                 return;
1583                         }
1584                         task = audio_task_queue.front();
1585                         audio_task_queue.pop();
1586                 }
1587
1588                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1589                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1590                 vector<float> samples_out = audio_mixer->get_output(
1591                         task.frame_timestamp,
1592                         task.num_samples,
1593                         rate_adjustment_policy);
1594
1595                 // Send the samples to the sound card, then add them to the output.
1596                 if (alsa) {
1597                         alsa->write(samples_out);
1598                 }
1599                 if (output_card_index != -1) {
1600                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1601                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1602                 }
1603                 video_encoder->add_audio(task.pts_int, move(samples_out));
1604         }
1605 }
1606
1607 void Mixer::release_display_frame(DisplayFrame *frame)
1608 {
1609         for (GLuint texnum : frame->temp_textures) {
1610                 resource_pool->release_2d_texture(texnum);
1611         }
1612         frame->temp_textures.clear();
1613         frame->ready_fence.reset();
1614         frame->input_frames.clear();
1615 }
1616
1617 void Mixer::start()
1618 {
1619         mixer_thread = thread(&Mixer::thread_func, this);
1620         audio_thread = thread(&Mixer::audio_thread_func, this);
1621 }
1622
1623 void Mixer::quit()
1624 {
1625         should_quit = true;
1626         audio_task_queue_changed.notify_one();
1627         mixer_thread.join();
1628         audio_thread.join();
1629 }
1630
1631 void Mixer::transition_clicked(int transition_num)
1632 {
1633         theme->transition_clicked(transition_num, pts());
1634 }
1635
1636 void Mixer::channel_clicked(int preview_num)
1637 {
1638         theme->channel_clicked(preview_num);
1639 }
1640
1641 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1642 {
1643         lock_guard<mutex> lock(card_mutex);
1644         return ycbcr_interpretation[card_index];
1645 }
1646
1647 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1648 {
1649         lock_guard<mutex> lock(card_mutex);
1650         ycbcr_interpretation[card_index] = interpretation;
1651 }
1652
1653 void Mixer::start_mode_scanning(unsigned card_index)
1654 {
1655         assert(card_index < num_cards);
1656         if (is_mode_scanning[card_index]) {
1657                 return;
1658         }
1659         is_mode_scanning[card_index] = true;
1660         mode_scanlist[card_index].clear();
1661         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1662                 mode_scanlist[card_index].push_back(mode.first);
1663         }
1664         assert(!mode_scanlist[card_index].empty());
1665         mode_scanlist_index[card_index] = 0;
1666         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1667         last_mode_scan_change[card_index] = steady_clock::now();
1668 }
1669
1670 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1671 {
1672         assert(desired_output_card_index != -1);
1673         lock_guard<mutex> lock(card_mutex);
1674         return cards[desired_output_card_index].output->get_available_video_modes();
1675 }
1676
1677 string Mixer::get_ffmpeg_filename(unsigned card_index) const
1678 {
1679         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1680         return ((FFmpegCapture *)(cards[card_index].capture.get()))->get_filename();
1681 }
1682
1683 void Mixer::set_ffmpeg_filename(unsigned card_index, const string &filename) {
1684         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1685         ((FFmpegCapture *)(cards[card_index].capture.get()))->change_filename(filename);
1686 }
1687
1688 void Mixer::wait_for_next_frame()
1689 {
1690         unique_lock<mutex> lock(frame_num_mutex);
1691         unsigned old_frame_num = frame_num;
1692         frame_num_updated.wait_for(lock, seconds(1),  // Timeout is just in case.
1693                 [old_frame_num, this]{ return this->frame_num > old_frame_num; });
1694 }
1695
1696 Mixer::OutputChannel::~OutputChannel()
1697 {
1698         if (has_current_frame) {
1699                 parent->release_display_frame(&current_frame);
1700         }
1701         if (has_ready_frame) {
1702                 parent->release_display_frame(&ready_frame);
1703         }
1704 }
1705
1706 void Mixer::OutputChannel::output_frame(DisplayFrame &&frame)
1707 {
1708         // Store this frame for display. Remove the ready frame if any
1709         // (it was seemingly never used).
1710         {
1711                 lock_guard<mutex> lock(frame_mutex);
1712                 if (has_ready_frame) {
1713                         parent->release_display_frame(&ready_frame);
1714                 }
1715                 ready_frame = move(frame);
1716                 has_ready_frame = true;
1717
1718                 // Call the callbacks under the mutex (they should be short),
1719                 // so that we don't race against a callback removal.
1720                 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1721                         key_and_callback.second();
1722                 }
1723         }
1724
1725         // Reduce the number of callbacks by filtering duplicates. The reason
1726         // why we bother doing this is that Qt seemingly can get into a state
1727         // where its builds up an essentially unbounded queue of signals,
1728         // consuming more and more memory, and there's no good way of collapsing
1729         // user-defined signals or limiting the length of the queue.
1730         if (transition_names_updated_callback) {
1731                 vector<string> transition_names = global_mixer->get_transition_names();
1732                 bool changed = false;
1733                 if (transition_names.size() != last_transition_names.size()) {
1734                         changed = true;
1735                 } else {
1736                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1737                                 if (transition_names[i] != last_transition_names[i]) {
1738                                         changed = true;
1739                                         break;
1740                                 }
1741                         }
1742                 }
1743                 if (changed) {
1744                         transition_names_updated_callback(transition_names);
1745                         last_transition_names = transition_names;
1746                 }
1747         }
1748         if (name_updated_callback) {
1749                 string name = global_mixer->get_channel_name(channel);
1750                 if (name != last_name) {
1751                         name_updated_callback(name);
1752                         last_name = name;
1753                 }
1754         }
1755         if (color_updated_callback) {
1756                 string color = global_mixer->get_channel_color(channel);
1757                 if (color != last_color) {
1758                         color_updated_callback(color);
1759                         last_color = color;
1760                 }
1761         }
1762 }
1763
1764 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1765 {
1766         lock_guard<mutex> lock(frame_mutex);
1767         if (!has_current_frame && !has_ready_frame) {
1768                 return false;
1769         }
1770
1771         if (has_current_frame && has_ready_frame) {
1772                 // We have a new ready frame. Toss the current one.
1773                 parent->release_display_frame(&current_frame);
1774                 has_current_frame = false;
1775         }
1776         if (has_ready_frame) {
1777                 assert(!has_current_frame);
1778                 current_frame = move(ready_frame);
1779                 ready_frame.ready_fence.reset();  // Drop the refcount.
1780                 ready_frame.input_frames.clear();  // Drop the refcounts.
1781                 has_current_frame = true;
1782                 has_ready_frame = false;
1783         }
1784
1785         *frame = current_frame;
1786         return true;
1787 }
1788
1789 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1790 {
1791         lock_guard<mutex> lock(frame_mutex);
1792         new_frame_ready_callbacks[key] = callback;
1793 }
1794
1795 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1796 {
1797         lock_guard<mutex> lock(frame_mutex);
1798         new_frame_ready_callbacks.erase(key);
1799 }
1800
1801 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1802 {
1803         transition_names_updated_callback = callback;
1804 }
1805
1806 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1807 {
1808         name_updated_callback = callback;
1809 }
1810
1811 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1812 {
1813         color_updated_callback = callback;
1814 }
1815
1816 mutex RefCountedGLsync::fence_lock;