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

   1 #undef Success
   2
   3 #include "mixer.h"
   4
   5 #include <assert.h>
   6 #include <epoxy/egl.h>
   7 #include <movit/effect_chain.h>
   8 #include <movit/effect_util.h>
   9 #include <movit/flat_input.h>
  10 #include <movit/image_format.h>
  11 #include <movit/init.h>
  12 #include <movit/resource_pool.h>
  13 #include <pthread.h>
  14 #include <stdint.h>
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <algorithm>
  18 #include <chrono>
  19 #include <condition_variable>
  20 #include <cstddef>
  21 #include <cstdint>
  22 #include <memory>
  23 #include <mutex>
  24 #include <ratio>
  25 #include <string>
  26 #include <thread>
  27 #include <utility>
  28 #include <vector>
  29
  30 #include "DeckLinkAPI.h"
  31 #include "LinuxCOM.h"
  32 #include "alsa_output.h"
  33 #include "basic_stats.h"
  34 #include "bmusb/bmusb.h"
  35 #include "bmusb/fake_capture.h"
  36 #ifdef HAVE_CEF
  37 #include "cef_capture.h"
  38 #endif
  39 #include "chroma_subsampler.h"
  40 #include "shared/context.h"
  41 #include "decklink_capture.h"
  42 #include "decklink_output.h"
  43 #include "defs.h"
  44 #include "shared/disk_space_estimator.h"
  45 #include "ffmpeg_capture.h"
  46 #include "flags.h"
  47 #include "image_input.h"
  48 #include "input_mapping.h"
  49 #include "shared/metrics.h"
  50 #include "mjpeg_encoder.h"
  51 #include "pbo_frame_allocator.h"
  52 #include "shared/ref_counted_gl_sync.h"
  53 #include "resampling_queue.h"
  54 #include "shared/timebase.h"
  55 #include "timecode_renderer.h"
  56 #include "v210_converter.h"
  57 #include "va_display_with_cleanup.h"
  58 #include "video_encoder.h"
  59
  60 #undef Status
  61 #include <google/protobuf/util/json_util.h>
  62 #include "json.pb.h"
  63
  64 class IDeckLink;
  65 class QOpenGLContext;
  66
  67 using namespace movit;
  68 using namespace std;
  69 using namespace std::chrono;
  70 using namespace std::placeholders;
  71 using namespace bmusb;
  72
  73 Mixer *global_mixer = nullptr;
  74
  75 namespace {
  76
  77 void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
  78 {
  79         if (interlaced) {
  80                 for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) {  // :-)
  81                         input_state->buffered_frames[card_index][frame_num] =
  82                                 input_state->buffered_frames[card_index][frame_num - 1];
  83                 }
  84                 input_state->buffered_frames[card_index][0] = { frame, field_num };
  85         } else {
  86                 for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
  87                         input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
  88                 }
  89         }
  90 }
  91
  92 void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width)
  93 {
  94         bool first;
  95         switch (userdata->pixel_format) {
  96         case PixelFormat_10BitYCbCr:
  97                 first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
  98                 break;
  99         case PixelFormat_8BitYCbCr:
 100                 first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
 101                 break;
 102         case PixelFormat_8BitBGRA:
 103                 first = userdata->tex_rgba[field] == 0;
 104                 break;
 105         case PixelFormat_8BitYCbCrPlanar:
 106                 first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0;
 107                 break;
 108         default:
 109                 assert(false);
 110         }
 111
 112         if (first ||
 113             width != userdata->last_width[field] ||
 114             height != userdata->last_height[field] ||
 115             cbcr_width != userdata->last_cbcr_width[field] ||
 116             cbcr_height != userdata->last_cbcr_height[field]) {
 117                 // We changed resolution since last use of this texture, so we need to create
 118                 // a new object. Note that this each card has its own PBOFrameAllocator,
 119                 // we don't need to worry about these flip-flopping between resolutions.
 120                 switch (userdata->pixel_format) {
 121                 case PixelFormat_10BitYCbCr:
 122                         glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
 123                         check_error();
 124                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 125                         check_error();
 126                         break;
 127                 case PixelFormat_8BitYCbCr: {
 128                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 129                         check_error();
 130                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 131                         check_error();
 132                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 133                         check_error();
 134                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 135                         check_error();
 136                         break;
 137                 }
 138                 case PixelFormat_8BitYCbCrPlanar: {
 139                         glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 140                         check_error();
 141                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 142                         check_error();
 143                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]);
 144                         check_error();
 145                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 146                         check_error();
 147                         glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]);
 148                         check_error();
 149                         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 150                         check_error();
 151                         break;
 152                 }
 153                 case PixelFormat_8BitBGRA:
 154                         glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 155                         check_error();
 156                         if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
 157                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 158                         } else {
 159                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 160                         }
 161                         check_error();
 162                         break;
 163                 default:
 164                         assert(false);
 165                 }
 166                 userdata->last_width[field] = width;
 167                 userdata->last_height[field] = height;
 168                 userdata->last_cbcr_width[field] = cbcr_width;
 169                 userdata->last_cbcr_height[field] = cbcr_height;
 170         }
 171         if (global_flags.ten_bit_input &&
 172             (first || v210_width != userdata->last_v210_width[field])) {
 173                 // Same as above; we need to recreate the texture.
 174                 glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
 175                 check_error();
 176                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 177                 check_error();
 178                 userdata->last_v210_width[field] = v210_width;
 179         }
 180 }
 181
 182 void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
 183 {
 184         if (interlaced_stride) {
 185                 stride *= 2;
 186         }
 187         if (global_flags.flush_pbos) {
 188                 glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
 189                 check_error();
 190         }
 191
 192         glBindTexture(GL_TEXTURE_2D, tex);
 193         check_error();
 194         if (interlaced_stride) {
 195                 glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
 196                 check_error();
 197         } else {
 198                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 199                 check_error();
 200         }
 201
 202         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
 203         check_error();
 204         glBindTexture(GL_TEXTURE_2D, 0);
 205         check_error();
 206         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 207         check_error();
 208 }
 209
 210 }  // namespace
 211
 212 void JitterHistory::register_metrics(const vector<pair<string, string>> &labels)
 213 {
 214         global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames);
 215         global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE);
 216 }
 217
 218 void JitterHistory::unregister_metrics(const vector<pair<string, string>> &labels)
 219 {
 220         global_metrics.remove("input_underestimated_jitter_frames", labels);
 221         global_metrics.remove("input_estimated_max_jitter_seconds", labels);
 222 }
 223
 224 void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames)
 225 {
 226         if (expected_timestamp > steady_clock::time_point::min()) {
 227                 expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 228                 double jitter_seconds = fabs(duration<double>(expected_timestamp - now).count());
 229                 history.push_back(orders.insert(jitter_seconds));
 230                 if (jitter_seconds > estimate_max_jitter()) {
 231                         ++metric_input_underestimated_jitter_frames;
 232                 }
 233
 234                 metric_input_estimated_max_jitter_seconds = estimate_max_jitter();
 235
 236                 if (history.size() > history_length) {
 237                         orders.erase(history.front());
 238                         history.pop_front();
 239                 }
 240                 assert(history.size() <= history_length);
 241         }
 242         expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE);
 243 }
 244
 245 double JitterHistory::estimate_max_jitter() const
 246 {
 247         if (orders.empty()) {
 248                 return 0.0;
 249         }
 250         size_t elem_idx = lrint((orders.size() - 1) * percentile);
 251         if (percentile <= 0.5) {
 252                 return *next(orders.begin(), elem_idx) * multiplier;
 253         } else {
 254                 return *prev(orders.end(), orders.size() - elem_idx) * multiplier;
 255         }
 256 }
 257
 258 void QueueLengthPolicy::register_metrics(const vector<pair<string, string>> &labels)
 259 {
 260         global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE);
 261 }
 262
 263 void QueueLengthPolicy::unregister_metrics(const vector<pair<string, string>> &labels)
 264 {
 265         global_metrics.remove("input_queue_safe_length_frames", labels);
 266 }
 267
 268 void QueueLengthPolicy::update_policy(steady_clock::time_point now,
 269                                       steady_clock::time_point expected_next_frame,
 270                                       int64_t input_frame_duration,
 271                                       int64_t master_frame_duration,
 272                                       double max_input_card_jitter_seconds,
 273                                       double max_master_card_jitter_seconds)
 274 {
 275         double input_frame_duration_seconds = input_frame_duration / double(TIMEBASE);
 276         double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE);
 277
 278         // Figure out when we can expect the next frame for this card, assuming
 279         // worst-case jitter (ie., the frame is maximally late).
 280         double seconds_until_next_frame = max(duration<double>(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0);
 281
 282         // How many times are the master card expected to tick in that time?
 283         // We assume the master clock has worst-case jitter but not any rate
 284         // discrepancy, ie., it ticks as early as possible every time, but not
 285         // cumulatively.
 286         double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds;
 287
 288         // As a special case, if the master card ticks faster than the input card,
 289         // we expect the queue to drain by itself even without dropping. But if
 290         // the difference is small (e.g. 60 Hz master and 59.94 input), it would
 291         // go slowly enough that the effect wouldn't really be appreciable.
 292         // We account for this by looking at the situation five frames ahead,
 293         // assuming everything else is the same.
 294         double frames_allowed;
 295         if (master_frame_duration < input_frame_duration) {
 296                 frames_allowed = frames_needed + 5 * (input_frame_duration_seconds - master_frame_duration_seconds) / master_frame_duration_seconds;
 297         } else {
 298                 frames_allowed = frames_needed;
 299         }
 300
 301         safe_queue_length = max<int>(floor(frames_allowed), 0);
 302         metric_input_queue_safe_length_frames = safe_queue_length;
 303 }
 304
 305 Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 306         : httpd(),
 307           num_cards(num_cards),
 308           mixer_surface(create_surface(format)),
 309           h264_encoder_surface(create_surface(format)),
 310           decklink_output_surface(create_surface(format)),
 311           image_update_surface(create_surface(format))
 312 {
 313         memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation));
 314         CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
 315         check_error();
 316
 317         // This nearly always should be true.
 318         global_flags.can_disable_srgb_decoder =
 319                 epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") &&
 320                 epoxy_has_gl_extension("GL_ARB_sampler_objects");
 321
 322         // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
 323         // will be halved when sampling them, and we need to compensate here.
 324         movit_texel_subpixel_precision /= 2.0;
 325
 326         resource_pool.reset(new ResourcePool);
 327         for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
 328                 output_channel[i].parent = this;
 329                 output_channel[i].channel = i;
 330         }
 331
 332         ImageFormat inout_format;
 333         inout_format.color_space = COLORSPACE_sRGB;
 334         inout_format.gamma_curve = GAMMA_sRGB;
 335
 336         // Matches the 4:2:0 format created by the main chain.
 337         YCbCrFormat ycbcr_format;
 338         ycbcr_format.chroma_subsampling_x = 2;
 339         ycbcr_format.chroma_subsampling_y = 2;
 340         if (global_flags.ycbcr_rec709_coefficients) {
 341                 ycbcr_format.luma_coefficients = YCBCR_REC_709;
 342         } else {
 343                 ycbcr_format.luma_coefficients = YCBCR_REC_601;
 344         }
 345         ycbcr_format.full_range = false;
 346         ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
 347         ycbcr_format.cb_x_position = 0.0f;
 348         ycbcr_format.cr_x_position = 0.0f;
 349         ycbcr_format.cb_y_position = 0.5f;
 350         ycbcr_format.cr_y_position = 0.5f;
 351
 352         // Display chain; shows the live output produced by the main chain (or rather, a copy of it).
 353         display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get()));
 354         check_error();
 355         GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
 356         display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type);
 357         display_chain->add_input(display_input);
 358         display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
 359         display_chain->set_dither_bits(0);  // Don't bother.
 360         display_chain->finalize();
 361
 362         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));
 363         if (!global_flags.card_to_mjpeg_stream_export.empty()) {
 364                 mjpeg_encoder.reset(new MJPEGEncoder(&httpd, global_flags.va_display));
 365         }
 366
 367         // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
 368         theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
 369
 370         // Must be instantiated after the theme, as the theme decides the number of FFmpeg inputs.
 371         std::vector<FFmpegCapture *> video_inputs = theme->get_video_inputs();
 372         audio_mixer.reset(new AudioMixer(num_cards, video_inputs.size()));
 373
 374         httpd.add_endpoint("/channels", bind(&Mixer::get_channels_json, this), HTTPD::ALLOW_ALL_ORIGINS);
 375         for (int channel_idx = 0; channel_idx < theme->get_num_channels(); ++channel_idx) {
 376                 char url[256];
 377                 snprintf(url, sizeof(url), "/channels/%d/color", channel_idx + 2);
 378                 httpd.add_endpoint(url, bind(&Mixer::get_channel_color_http, this, unsigned(channel_idx + 2)), HTTPD::ALLOW_ALL_ORIGINS);
 379         }
 380
 381         // Start listening for clients only once VideoEncoder has written its header, if any.
 382         httpd.start(global_flags.http_port);
 383
 384         // First try initializing the then PCI devices, then USB, then
 385         // fill up with fake cards until we have the desired number of cards.
 386         unsigned num_pci_devices = 0;
 387         unsigned card_index = 0;
 388
 389         {
 390                 IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
 391                 if (decklink_iterator != nullptr) {
 392                         for ( ; card_index < num_cards; ++card_index) {
 393                                 IDeckLink *decklink;
 394                                 if (decklink_iterator->Next(&decklink) != S_OK) {
 395                                         break;
 396                                 }
 397
 398                                 DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
 399                                 DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
 400                                 if (!output->set_device(decklink)) {
 401                                         delete output;
 402                                         output = nullptr;
 403                                 }
 404                                 configure_card(card_index, capture, CardType::LIVE_CARD, output);
 405                                 ++num_pci_devices;
 406                         }
 407                         decklink_iterator->Release();
 408                         fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
 409                 } else {
 410                         fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
 411                 }
 412         }
 413
 414         unsigned num_usb_devices = BMUSBCapture::num_cards();
 415         for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
 416                 BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
 417                 capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
 418                 configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
 419         }
 420         fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 421
 422         unsigned num_fake_cards = 0;
 423         for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
 424                 FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 425                 configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
 426         }
 427
 428         if (num_fake_cards > 0) {
 429                 fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
 430         }
 431
 432         // Initialize all video inputs the theme asked for. Note that these are
 433         // all put _after_ the regular cards, which stop at <num_cards> - 1.
 434         for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) {
 435                 if (card_index >= MAX_VIDEO_CARDS) {
 436                         fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n");
 437                         abort();
 438                 }
 439                 configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr);
 440                 video_inputs[video_card_index]->set_card_index(card_index);
 441         }
 442         num_video_inputs = video_inputs.size();
 443
 444 #ifdef HAVE_CEF
 445         // Same, for HTML inputs.
 446         std::vector<CEFCapture *> html_inputs = theme->get_html_inputs();
 447         for (unsigned html_card_index = 0; html_card_index < html_inputs.size(); ++card_index, ++html_card_index) {
 448                 if (card_index >= MAX_VIDEO_CARDS) {
 449                         fprintf(stderr, "ERROR: Not enough card slots available for the HTML inputs the theme requested.\n");
 450                         abort();
 451                 }
 452                 configure_card(card_index, html_inputs[html_card_index], CardType::CEF_INPUT, /*output=*/nullptr);
 453                 html_inputs[html_card_index]->set_card_index(card_index);
 454         }
 455         num_html_inputs = html_inputs.size();
 456 #endif
 457
 458         BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
 459         BMUSBCapture::start_bm_thread();
 460
 461         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 462                 cards[card_index].queue_length_policy.reset(card_index);
 463         }
 464
 465         chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 466
 467         if (global_flags.ten_bit_input) {
 468                 if (!v210Converter::has_hardware_support()) {
 469                         fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
 470                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 471                         abort();
 472                 }
 473                 v210_converter.reset(new v210Converter());
 474
 475                 // These are all the widths listed in the Blackmagic SDK documentation
 476                 // (section 2.7.3, “Display Modes”).
 477                 v210_converter->precompile_shader(720);
 478                 v210_converter->precompile_shader(1280);
 479                 v210_converter->precompile_shader(1920);
 480                 v210_converter->precompile_shader(2048);
 481                 v210_converter->precompile_shader(3840);
 482                 v210_converter->precompile_shader(4096);
 483         }
 484         if (global_flags.ten_bit_output) {
 485                 if (!v210Converter::has_hardware_support()) {
 486                         fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n");
 487                         fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
 488                         abort();
 489                 }
 490         }
 491
 492         timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
 493         display_timecode_in_stream = global_flags.display_timecode_in_stream;
 494         display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
 495
 496         if (global_flags.enable_alsa_output) {
 497                 alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
 498         }
 499         if (global_flags.output_card != -1) {
 500                 desired_output_card_index = global_flags.output_card;
 501                 set_output_card_internal(global_flags.output_card);
 502         }
 503
 504         output_jitter_history.register_metrics({{ "card", "output" }});
 505
 506         ImageInput::start_update_thread(image_update_surface);
 507 }
 508
 509 Mixer::~Mixer()
 510 {
 511         ImageInput::end_update_thread();
 512
 513         if (mjpeg_encoder != nullptr) {
 514                 mjpeg_encoder->stop();
 515         }
 516         httpd.stop();
 517         BMUSBCapture::stop_bm_thread();
 518
 519         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
 520                 cards[card_index].capture->stop_dequeue_thread();
 521                 if (cards[card_index].output) {
 522                         cards[card_index].output->end_output();
 523                         cards[card_index].output.reset();
 524                 }
 525         }
 526
 527         video_encoder.reset(nullptr);
 528 }
 529
 530 void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output)
 531 {
 532         printf("Configuring card %d...\n", card_index);
 533
 534         CaptureCard *card = &cards[card_index];
 535         if (card->capture != nullptr) {
 536                 card->capture->stop_dequeue_thread();
 537         }
 538         card->capture.reset(capture);
 539         card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE);
 540         card->is_cef_capture = (card_type == CardType::CEF_INPUT);
 541         card->may_have_dropped_last_frame = false;
 542         card->type = card_type;
 543         if (card->output.get() != output) {
 544                 card->output.reset(output);
 545         }
 546
 547         PixelFormat pixel_format;
 548         if (card_type == CardType::FFMPEG_INPUT) {
 549                 pixel_format = capture->get_current_pixel_format();
 550         } else if (card_type == CardType::CEF_INPUT) {
 551                 pixel_format = PixelFormat_8BitBGRA;
 552         } else if (global_flags.ten_bit_input) {
 553                 pixel_format = PixelFormat_10BitYCbCr;
 554         } else {
 555                 pixel_format = PixelFormat_8BitYCbCr;
 556         }
 557
 558         card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
 559         if (card->frame_allocator == nullptr) {
 560                 card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height, card_index, mjpeg_encoder.get()));  // 8 MB.
 561         }
 562         card->capture->set_video_frame_allocator(card->frame_allocator.get());
 563         if (card->surface == nullptr) {
 564                 card->surface = create_surface_with_same_format(mixer_surface);
 565         }
 566         while (!card->new_frames.empty()) card->new_frames.pop_front();
 567         card->last_timecode = -1;
 568         card->capture->set_pixel_format(pixel_format);
 569         card->capture->configure_card();
 570
 571         // NOTE: start_bm_capture() happens in thread_func().
 572
 573         DeviceSpec device;
 574         if (card_type == CardType::FFMPEG_INPUT) {
 575                 device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 576         } else {
 577                 device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 578         }
 579         audio_mixer->reset_resampler(device);
 580         audio_mixer->set_display_name(device, card->capture->get_description());
 581         audio_mixer->trigger_state_changed_callback();
 582
 583         // Unregister old metrics, if any.
 584         if (!card->labels.empty()) {
 585                 const vector<pair<string, string>> &labels = card->labels;
 586                 card->jitter_history.unregister_metrics(labels);
 587                 card->queue_length_policy.unregister_metrics(labels);
 588                 global_metrics.remove("input_received_frames", labels);
 589                 global_metrics.remove("input_dropped_frames_jitter", labels);
 590                 global_metrics.remove("input_dropped_frames_error", labels);
 591                 global_metrics.remove("input_dropped_frames_resets", labels);
 592                 global_metrics.remove("input_queue_length_frames", labels);
 593                 global_metrics.remove("input_queue_duped_frames", labels);
 594
 595                 global_metrics.remove("input_has_signal_bool", labels);
 596                 global_metrics.remove("input_is_connected_bool", labels);
 597                 global_metrics.remove("input_interlaced_bool", labels);
 598                 global_metrics.remove("input_width_pixels", labels);
 599                 global_metrics.remove("input_height_pixels", labels);
 600                 global_metrics.remove("input_frame_rate_nom", labels);
 601                 global_metrics.remove("input_frame_rate_den", labels);
 602                 global_metrics.remove("input_sample_rate_hz", labels);
 603         }
 604
 605         // Register metrics.
 606         vector<pair<string, string>> labels;
 607         char card_name[64];
 608         snprintf(card_name, sizeof(card_name), "%d", card_index);
 609         labels.emplace_back("card", card_name);
 610
 611         switch (card_type) {
 612         case CardType::LIVE_CARD:
 613                 labels.emplace_back("cardtype", "live");
 614                 break;
 615         case CardType::FAKE_CAPTURE:
 616                 labels.emplace_back("cardtype", "fake");
 617                 break;
 618         case CardType::FFMPEG_INPUT:
 619                 labels.emplace_back("cardtype", "ffmpeg");
 620                 break;
 621         case CardType::CEF_INPUT:
 622                 labels.emplace_back("cardtype", "cef");
 623                 break;
 624         default:
 625                 assert(false);
 626         }
 627         card->jitter_history.register_metrics(labels);
 628         card->queue_length_policy.register_metrics(labels);
 629         global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames);
 630         global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter);
 631         global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error);
 632         global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets);
 633         global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE);
 634         global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames);
 635
 636         global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE);
 637         global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE);
 638         global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE);
 639         global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE);
 640         global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE);
 641         global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE);
 642         global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE);
 643         global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE);
 644         card->labels = labels;
 645 }
 646
 647 void Mixer::set_output_card_internal(int card_index)
 648 {
 649         // We don't really need to take card_mutex, since we're in the mixer
 650         // thread and don't mess with any queues (which is the only thing that happens
 651         // from other threads), but it's probably the safest in the long run.
 652         unique_lock<mutex> lock(card_mutex);
 653         if (output_card_index != -1) {
 654                 // Switch the old card from output to input.
 655                 CaptureCard *old_card = &cards[output_card_index];
 656                 old_card->output->end_output();
 657
 658                 // Stop the fake card that we put into place.
 659                 // This needs to _not_ happen under the mutex, to avoid deadlock
 660                 // (delivering the last frame needs to take the mutex).
 661                 CaptureInterface *fake_capture = old_card->capture.get();
 662                 lock.unlock();
 663                 fake_capture->stop_dequeue_thread();
 664                 lock.lock();
 665                 old_card->capture = move(old_card->parked_capture);  // TODO: reset the metrics
 666                 old_card->is_fake_capture = false;
 667                 old_card->capture->start_bm_capture();
 668         }
 669         if (card_index != -1) {
 670                 CaptureCard *card = &cards[card_index];
 671                 CaptureInterface *capture = card->capture.get();
 672                 // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
 673                 // several seconds to complete (blocking on DisableVideoInput);
 674                 // see if we can maybe do it asynchronously.
 675                 lock.unlock();
 676                 capture->stop_dequeue_thread();
 677                 lock.lock();
 678                 card->parked_capture = move(card->capture);
 679                 CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
 680                 configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release());
 681                 card->queue_length_policy.reset(card_index);
 682                 card->capture->start_bm_capture();
 683                 desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
 684                 card->output->start_output(desired_output_video_mode, pts_int);
 685         }
 686         output_card_index = card_index;
 687         output_jitter_history.clear();
 688 }
 689
 690 namespace {
 691
 692 int unwrap_timecode(uint16_t current_wrapped, int last)
 693 {
 694         uint16_t last_wrapped = last & 0xffff;
 695         if (current_wrapped > last_wrapped) {
 696                 return (last & ~0xffff) | current_wrapped;
 697         } else {
 698                 return 0x10000 + ((last & ~0xffff) | current_wrapped);
 699         }
 700 }
 701
 702 DeviceSpec card_index_to_device(unsigned card_index, unsigned num_cards)
 703 {
 704         if (card_index >= num_cards) {
 705                 return DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards};
 706         } else {
 707                 return DeviceSpec{InputSourceType::CAPTURE_CARD, card_index};
 708         }
 709 }
 710
 711 }  // namespace
 712
 713 void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 714                      FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format,
 715                      FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
 716 {
 717         DeviceSpec device = card_index_to_device(card_index, num_cards);
 718         CaptureCard *card = &cards[card_index];
 719
 720         ++card->metric_input_received_frames;
 721         card->metric_input_has_signal_bool = video_format.has_signal;
 722         card->metric_input_is_connected_bool = video_format.is_connected;
 723         card->metric_input_interlaced_bool = video_format.interlaced;
 724         card->metric_input_width_pixels = video_format.width;
 725         card->metric_input_height_pixels = video_format.height;
 726         card->metric_input_frame_rate_nom = video_format.frame_rate_nom;
 727         card->metric_input_frame_rate_den = video_format.frame_rate_den;
 728         card->metric_input_sample_rate_hz = audio_format.sample_rate;
 729
 730         if (is_mode_scanning[card_index]) {
 731                 if (video_format.has_signal) {
 732                         // Found a stable signal, so stop scanning.
 733                         is_mode_scanning[card_index] = false;
 734                 } else {
 735                         static constexpr double switch_time_s = 0.1;  // Should be enough time for the signal to stabilize.
 736                         steady_clock::time_point now = steady_clock::now();
 737                         double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
 738                         if (sec_since_last_switch > switch_time_s) {
 739                                 // It isn't this mode; try the next one.
 740                                 mode_scanlist_index[card_index]++;
 741                                 mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
 742                                 cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
 743                                 last_mode_scan_change[card_index] = now;
 744                         }
 745                 }
 746         }
 747
 748         int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
 749         assert(frame_length > 0);
 750
 751         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;
 752         if (num_samples > OUTPUT_FREQUENCY / 10 && card->type != CardType::FFMPEG_INPUT) {
 753                 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",
 754                         spec_to_string(device).c_str(), int(audio_frame.len), int(audio_offset),
 755                         timecode, int(video_frame.len), int(video_offset), video_format.id);
 756                 if (video_frame.owner) {
 757                         video_frame.owner->release_frame(video_frame);
 758                 }
 759                 if (audio_frame.owner) {
 760                         audio_frame.owner->release_frame(audio_frame);
 761                 }
 762                 return;
 763         }
 764
 765         int dropped_frames = 0;
 766         if (card->last_timecode != -1) {
 767                 dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
 768         }
 769
 770         // Number of samples per frame if we need to insert silence.
 771         // (Could be nonintegral, but resampling will save us then.)
 772         const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;
 773
 774         if (dropped_frames > MAX_FPS * 2) {
 775                 fprintf(stderr, "%s lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
 776                         spec_to_string(device).c_str(), card->last_timecode, timecode);
 777                 audio_mixer->reset_resampler(device);
 778                 dropped_frames = 0;
 779                 ++card->metric_input_resets;
 780         } else if (dropped_frames > 0) {
 781                 // Insert silence as needed.
 782                 fprintf(stderr, "%s dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
 783                         spec_to_string(device).c_str(), dropped_frames, timecode);
 784                 card->metric_input_dropped_frames_error += dropped_frames;
 785
 786                 bool success;
 787                 do {
 788                         success = audio_mixer->add_silence(device, silence_samples, dropped_frames);
 789                 } while (!success);
 790         }
 791
 792         if (num_samples > 0) {
 793                 audio_mixer->add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, audio_frame.received_timestamp);
 794
 795                 // Audio for the MJPEG stream. We don't resample; audio that's not in 48 kHz
 796                 // just gets dropped for now.
 797                 //
 798                 // Only bother doing MJPEG encoding if there are any connected clients
 799                 // that want the stream.
 800                 if (httpd.get_num_connected_multicam_clients() > 0) {
 801                         vector<int32_t> converted_samples = convert_audio_to_fixed32(audio_frame.data + audio_offset, num_samples, audio_format, 2);
 802                         lock_guard<mutex> lock(card_mutex);
 803                         if (card->new_raw_audio.empty()) {
 804                                 card->new_raw_audio = move(converted_samples);
 805                         } else {
 806                                 // For raw audio, we don't really synchronize audio and video;
 807                                 // we just put the audio in frame by frame, and if a video frame is
 808                                 // dropped, we still keep the audio, which means it will be added
 809                                 // to the beginning of the next frame. It would probably be better
 810                                 // to move the audio pts earlier to show this, but most players can
 811                                 // live with some jitter, and in a lot of ways, it's much nicer for
 812                                 // Futatabi to have all audio locked to a video frame.
 813                                 card->new_raw_audio.insert(card->new_raw_audio.end(), converted_samples.begin(), converted_samples.end());
 814
 815                                 // Truncate to one second, just to be sure we don't have infinite buildup in case of weirdness.
 816                                 if (card->new_raw_audio.size() > OUTPUT_FREQUENCY * 2) {
 817                                         size_t excess_samples = card->new_raw_audio.size() - OUTPUT_FREQUENCY * 2;
 818                                         card->new_raw_audio.erase(card->new_raw_audio.begin(), card->new_raw_audio.begin() + excess_samples);
 819                                 }
 820                         }
 821                 }
 822         }
 823
 824         // Done with the audio, so release it.
 825         if (audio_frame.owner) {
 826                 audio_frame.owner->release_frame(audio_frame);
 827         }
 828
 829         card->last_timecode = timecode;
 830
 831         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
 832         if (card->type == CardType::FFMPEG_INPUT && userdata != nullptr) {
 833                 FFmpegCapture *ffmpeg_capture = static_cast<FFmpegCapture *>(card->capture.get());
 834                 userdata->has_last_subtitle = ffmpeg_capture->get_has_last_subtitle();
 835                 userdata->last_subtitle = ffmpeg_capture->get_last_subtitle();
 836         }
 837
 838         size_t cbcr_width, cbcr_height, cbcr_offset, y_offset;
 839         size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
 840         if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) {
 841                 // The calculation above is wrong for planar Y'CbCr, so just override it.
 842                 assert(card->type == CardType::FFMPEG_INPUT);
 843                 assert(video_offset == 0);
 844                 expected_length = video_frame.len;
 845
 846                 userdata->ycbcr_format = (static_cast<FFmpegCapture *>(card->capture.get()))->get_current_frame_ycbcr_format();
 847                 cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x;
 848                 cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y;
 849                 cbcr_offset = video_format.width * video_format.height;
 850                 y_offset = 0;
 851         } else {
 852                 // All the other Y'CbCr formats are 4:2:2.
 853                 cbcr_width = video_format.width / 2;
 854                 cbcr_height = video_format.height;
 855                 cbcr_offset = video_offset / 2;
 856                 y_offset = video_frame.size / 2 + video_offset / 2;
 857         }
 858         if (video_frame.len - video_offset == 0 ||
 859             video_frame.len - video_offset != expected_length) {
 860                 if (video_frame.len != 0) {
 861                         printf("%s: Dropping video frame with wrong length (%zu; expected %zu)\n",
 862                                 spec_to_string(device).c_str(), video_frame.len - video_offset, expected_length);
 863                 }
 864                 if (video_frame.owner) {
 865                         video_frame.owner->release_frame(video_frame);
 866                 }
 867
 868                 // Still send on the information that we _had_ a frame, even though it's corrupted,
 869                 // so that pts can go up accordingly.
 870                 {
 871                         lock_guard<mutex> lock(card_mutex);
 872                         CaptureCard::NewFrame new_frame;
 873                         new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
 874                         new_frame.length = frame_length;
 875                         new_frame.interlaced = false;
 876                         new_frame.dropped_frames = dropped_frames;
 877                         new_frame.received_timestamp = video_frame.received_timestamp;
 878                         card->new_frames.push_back(move(new_frame));
 879                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
 880                 }
 881                 card->new_frames_changed.notify_all();
 882                 return;
 883         }
 884
 885         unsigned num_fields = video_format.interlaced ? 2 : 1;
 886         steady_clock::time_point frame_upload_start;
 887         bool interlaced_stride = false;
 888         if (video_format.interlaced) {
 889                 // Send the two fields along as separate frames; the other side will need to add
 890                 // a deinterlacer to actually get this right.
 891                 assert(video_format.height % 2 == 0);
 892                 video_format.height /= 2;
 893                 cbcr_height /= 2;
 894                 assert(frame_length % 2 == 0);
 895                 frame_length /= 2;
 896                 num_fields = 2;
 897                 if (video_format.second_field_start == 1) {
 898                         interlaced_stride = true;
 899                 }
 900                 frame_upload_start = steady_clock::now();
 901         }
 902         assert(userdata != nullptr);
 903         userdata->last_interlaced = video_format.interlaced;
 904         userdata->last_has_signal = video_format.has_signal;
 905         userdata->last_is_connected = video_format.is_connected;
 906         userdata->last_frame_rate_nom = video_format.frame_rate_nom;
 907         userdata->last_frame_rate_den = video_format.frame_rate_den;
 908         RefCountedFrame frame(video_frame);
 909
 910         // Upload the textures.
 911         for (unsigned field = 0; field < num_fields; ++field) {
 912                 // Put the actual texture upload in a lambda that is executed in the main thread.
 913                 // It is entirely possible to do this in the same thread (and it might even be
 914                 // faster, depending on the GPU and driver), but it appears to be trickling
 915                 // driver bugs very easily.
 916                 //
 917                 // Note that this means we must hold on to the actual frame data in <userdata>
 918                 // until the upload command is run, but we hold on to <frame> much longer than that
 919                 // (in fact, all the way until we no longer use the texture in rendering).
 920                 auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() {
 921                         unsigned field_start_line;
 922                         if (field == 1) {
 923                                 field_start_line = video_format.second_field_start;
 924                         } else {
 925                                 field_start_line = video_format.extra_lines_top;
 926                         }
 927
 928                         // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used.
 929                         size_t v210_width = video_format.stride / sizeof(uint32_t);
 930                         ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width);
 931
 932                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
 933                         check_error();
 934
 935                         switch (userdata->pixel_format) {
 936                         case PixelFormat_10BitYCbCr: {
 937                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 938                                 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);
 939                                 v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
 940                                 break;
 941                         }
 942                         case PixelFormat_8BitYCbCr: {
 943                                 size_t field_y_start = y_offset + video_format.width * field_start_line;
 944                                 size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 945
 946                                 // Make up our own strides, since we are interleaving.
 947                                 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);
 948                                 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);
 949                                 break;
 950                         }
 951                         case PixelFormat_8BitYCbCrPlanar: {
 952                                 assert(field_start_line == 0);  // We don't really support interlaced here.
 953                                 size_t field_y_start = y_offset;
 954                                 size_t field_cb_start = cbcr_offset;
 955                                 size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height;
 956
 957                                 // Make up our own strides, since we are interleaving.
 958                                 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);
 959                                 upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start);
 960                                 upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start);
 961                                 break;
 962                         }
 963                         case PixelFormat_8BitBGRA: {
 964                                 size_t field_start = video_offset + video_format.stride * field_start_line;
 965                                 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);
 966                                 // These could be asked to deliver mipmaps at any time.
 967                                 glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]);
 968                                 check_error();
 969                                 glGenerateMipmap(GL_TEXTURE_2D);
 970                                 check_error();
 971                                 glBindTexture(GL_TEXTURE_2D, 0);
 972                                 check_error();
 973                                 break;
 974                         }
 975                         default:
 976                                 assert(false);
 977                         }
 978
 979                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 980                         check_error();
 981                 };
 982
 983                 if (field == 1) {
 984                         // Don't upload the second field as fast as we can; wait until
 985                         // the field time has approximately passed. (Otherwise, we could
 986                         // get timing jitter against the other sources, and possibly also
 987                         // against the video display, although the latter is not as critical.)
 988                         // This requires our system clock to be reasonably close to the
 989                         // video clock, but that's not an unreasonable assumption.
 990                         steady_clock::time_point second_field_start = frame_upload_start +
 991                                 nanoseconds(frame_length * 1000000000 / TIMEBASE);
 992                         this_thread::sleep_until(second_field_start);
 993                 }
 994
 995                 {
 996                         lock_guard<mutex> lock(card_mutex);
 997                         CaptureCard::NewFrame new_frame;
 998                         new_frame.frame = frame;
 999                         new_frame.length = frame_length;
1000                         new_frame.field = field;
1001                         new_frame.interlaced = video_format.interlaced;
1002                         new_frame.upload_func = upload_func;
1003                         new_frame.dropped_frames = dropped_frames;
1004                         new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
1005                         new_frame.video_format = video_format;
1006                         new_frame.y_offset = y_offset;
1007                         new_frame.cbcr_offset = cbcr_offset;
1008                         card->new_frames.push_back(move(new_frame));
1009                         card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
1010                         card->may_have_dropped_last_frame = false;
1011                 }
1012                 card->new_frames_changed.notify_all();
1013         }
1014 }
1015
1016 void Mixer::bm_hotplug_add(libusb_device *dev)
1017 {
1018         lock_guard<mutex> lock(hotplug_mutex);
1019         hotplugged_cards.push_back(dev);
1020 }
1021
1022 void Mixer::bm_hotplug_remove(unsigned card_index)
1023 {
1024         cards[card_index].new_frames_changed.notify_all();
1025 }
1026
1027 void Mixer::thread_func()
1028 {
1029         pthread_setname_np(pthread_self(), "Mixer_OpenGL");
1030
1031         eglBindAPI(EGL_OPENGL_API);
1032         QOpenGLContext *context = create_context(mixer_surface);
1033         if (!make_current(context, mixer_surface)) {
1034                 printf("oops\n");
1035                 abort();
1036         }
1037
1038         // Start the actual capture. (We don't want to do it before we're actually ready
1039         // to process output frames.)
1040         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1041                 if (int(card_index) != output_card_index) {
1042                         cards[card_index].capture->start_bm_capture();
1043                 }
1044         }
1045
1046         BasicStats basic_stats(/*verbose=*/true, /*use_opengl=*/true);
1047         int stats_dropped_frames = 0;
1048
1049         while (!should_quit) {
1050                 if (desired_output_card_index != output_card_index) {
1051                         set_output_card_internal(desired_output_card_index);
1052                 }
1053                 if (output_card_index != -1 &&
1054                     desired_output_video_mode != output_video_mode) {
1055                         DeckLinkOutput *output = cards[output_card_index].output.get();
1056                         output->end_output();
1057                         desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
1058                         output->start_output(desired_output_video_mode, pts_int);
1059                 }
1060
1061                 CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
1062                 bool has_new_frame[MAX_VIDEO_CARDS] = { false };
1063
1064                 bool master_card_is_output;
1065                 unsigned master_card_index;
1066                 if (output_card_index != -1) {
1067                         master_card_is_output = true;
1068                         master_card_index = output_card_index;
1069                 } else {
1070                         master_card_is_output = false;
1071                         master_card_index = theme->map_signal(master_clock_channel);
1072                         assert(master_card_index < num_cards + num_video_inputs);
1073                 }
1074
1075                 vector<int32_t> raw_audio[MAX_VIDEO_CARDS];  // For MJPEG encoding.
1076                 OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame, raw_audio);
1077                 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);
1078                 stats_dropped_frames += output_frame_info.dropped_frames;
1079
1080                 handle_hotplugged_cards();
1081
1082                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1083                         DeviceSpec device = card_index_to_device(card_index, num_cards);
1084                         if (card_index == master_card_index || !has_new_frame[card_index]) {
1085                                 continue;
1086                         }
1087                         if (new_frames[card_index].frame->len == 0) {
1088                                 ++new_frames[card_index].dropped_frames;
1089                         }
1090                         if (new_frames[card_index].dropped_frames > 0) {
1091                                 printf("%s dropped %d frames before this\n",
1092                                         spec_to_string(device).c_str(), int(new_frames[card_index].dropped_frames));
1093                         }
1094                 }
1095
1096                 // If the first card is reporting a corrupted or otherwise dropped frame,
1097                 // just increase the pts (skipping over this frame) and don't try to compute anything new.
1098                 if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
1099                         ++stats_dropped_frames;
1100                         pts_int += new_frames[master_card_index].length;
1101                         continue;
1102                 }
1103
1104                 for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1105                         if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
1106                                 continue;
1107
1108                         CaptureCard::NewFrame *new_frame = &new_frames[card_index];
1109                         assert(new_frame->frame != nullptr);
1110                         insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
1111                         check_error();
1112
1113                         // The new texture might need uploading before use.
1114                         if (new_frame->upload_func) {
1115                                 new_frame->upload_func();
1116                                 new_frame->upload_func = nullptr;
1117                         }
1118
1119                         if (new_frame->frame->data_copy != nullptr) {
1120                                 int mjpeg_card_index = mjpeg_encoder->get_mjpeg_stream_for_card(card_index);
1121                                 if (mjpeg_card_index != -1) {
1122                                         mjpeg_encoder->upload_frame(pts_int, mjpeg_card_index, new_frame->frame, new_frame->video_format, new_frame->y_offset, new_frame->cbcr_offset, move(raw_audio[card_index]));
1123                                 }
1124                         }
1125                 }
1126
1127                 int64_t frame_duration = output_frame_info.frame_duration;
1128                 render_one_frame(frame_duration);
1129                 {
1130                         lock_guard<mutex> lock(frame_num_mutex);
1131                         ++frame_num;
1132                 }
1133                 frame_num_updated.notify_all();
1134                 pts_int += frame_duration;
1135
1136                 basic_stats.update(frame_num, stats_dropped_frames);
1137                 // if (frame_num % 100 == 0) chain->print_phase_timing();
1138
1139                 if (should_cut.exchange(false)) {  // Test and clear.
1140                         video_encoder->do_cut(frame_num);
1141                 }
1142
1143 #if 0
1144                 // Reset every 100 frames, so that local variations in frame times
1145                 // (especially for the first few frames, when the shaders are
1146                 // compiled etc.) don't make it hard to measure for the entire
1147                 // remaining duration of the program.
1148                 if (frame == 10000) {
1149                         frame = 0;
1150                         start = now;
1151                 }
1152 #endif
1153                 check_error();
1154         }
1155
1156         resource_pool->clean_context();
1157 }
1158
1159 bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
1160 {
1161         if (output_card_index != -1) {
1162                 // The output card (ie., cards[output_card_index].output) is the master clock,
1163                 // so no input card (ie., cards[card_index].capture) is.
1164                 return false;
1165         }
1166         return (card_index == master_card_index);
1167 }
1168
1169 void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length)
1170 {
1171         // Count the number of frames in the queue, including any frames
1172         // we dropped. It's hard to know exactly how we should deal with
1173         // dropped (corrupted) input frames; they don't help our goal of
1174         // avoiding starvation, but they still add to the problem of latency.
1175         // Since dropped frames is going to mean a bump in the signal anyway,
1176         // we err on the side of having more stable latency instead.
1177         unsigned queue_length = 0;
1178         for (const CaptureCard::NewFrame &frame : card->new_frames) {
1179                 queue_length += frame.dropped_frames + 1;
1180         }
1181
1182         // If needed, drop frames until the queue is below the safe limit.
1183         // We prefer to drop from the head, because all else being equal,
1184         // we'd like more recent frames (less latency).
1185         unsigned dropped_frames = 0;
1186         while (queue_length > safe_queue_length) {
1187                 assert(!card->new_frames.empty());
1188                 assert(queue_length > card->new_frames.front().dropped_frames);
1189                 queue_length -= card->new_frames.front().dropped_frames;
1190
1191                 if (queue_length <= safe_queue_length) {
1192                         // No need to drop anything.
1193                         break;
1194                 }
1195
1196                 card->new_frames.pop_front();
1197                 card->new_frames_changed.notify_all();
1198                 --queue_length;
1199                 ++dropped_frames;
1200
1201                 if (queue_length == 0 && card->is_cef_capture) {
1202                         card->may_have_dropped_last_frame = true;
1203                 }
1204         }
1205
1206         card->metric_input_dropped_frames_jitter += dropped_frames;
1207         card->metric_input_queue_length_frames = queue_length;
1208
1209 #if 0
1210         if (dropped_frames > 0) {
1211                 fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
1212                         card_index, dropped_frames);
1213         }
1214 #endif
1215 }
1216
1217 pair<string, string> Mixer::get_channels_json()
1218 {
1219         Channels ret;
1220         for (int channel_idx = 2; channel_idx < theme->get_num_channels(); ++channel_idx) {
1221                 Channel *channel = ret.add_channel();
1222                 channel->set_index(channel_idx);
1223                 channel->set_name(theme->get_channel_name(channel_idx));
1224                 channel->set_color(theme->get_channel_color(channel_idx));
1225         }
1226         string contents;
1227         google::protobuf::util::MessageToJsonString(ret, &contents);  // Ignore any errors.
1228         return make_pair(contents, "text/json");
1229 }
1230
1231 pair<string, string> Mixer::get_channel_color_http(unsigned channel_idx)
1232 {
1233         return make_pair(theme->get_channel_color(channel_idx), "text/plain");
1234 }
1235
1236 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])
1237 {
1238         OutputFrameInfo output_frame_info;
1239 start:
1240         unique_lock<mutex> lock(card_mutex, defer_lock);
1241         if (master_card_is_output) {
1242                 // Clocked to the output, so wait for it to be ready for the next frame.
1243                 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);
1244                 lock.lock();
1245         } else {
1246                 // Wait for the master card to have a new frame.
1247                 // TODO: Add a timeout.
1248                 output_frame_info.is_preroll = false;
1249                 lock.lock();
1250                 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(); });
1251         }
1252
1253         if (master_card_is_output) {
1254                 handle_hotplugged_cards();
1255         } else if (cards[master_card_index].new_frames.empty()) {
1256                 // We were woken up, but not due to a new frame. Deal with it
1257                 // and then restart.
1258                 assert(cards[master_card_index].capture->get_disconnected());
1259                 handle_hotplugged_cards();
1260                 lock.unlock();
1261                 goto start;
1262         }
1263
1264         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1265                 CaptureCard *card = &cards[card_index];
1266                 if (card->new_frames.empty()) {  // Starvation.
1267                         ++card->metric_input_duped_frames;
1268 #ifdef HAVE_CEF
1269                         if (card->is_cef_capture && card->may_have_dropped_last_frame) {
1270                                 // Unlike other sources, CEF is not guaranteed to send us a steady
1271                                 // stream of frames, so we'll have to ask it to repaint the frame
1272                                 // we dropped. (may_have_dropped_last_frame is set whenever we
1273                                 // trim the queue completely away, and cleared when we actually
1274                                 // get a new frame.)
1275                                 ((CEFCapture *)card->capture.get())->request_new_frame();
1276                         }
1277 #endif
1278                 } else {
1279                         new_frames[card_index] = move(card->new_frames.front());
1280                         has_new_frame[card_index] = true;
1281                         card->new_frames.pop_front();
1282                         card->new_frames_changed.notify_all();
1283                 }
1284
1285                 raw_audio[card_index] = move(card->new_raw_audio);
1286         }
1287
1288         if (!master_card_is_output) {
1289                 output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp;
1290                 output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
1291                 output_frame_info.frame_duration = new_frames[master_card_index].length;
1292         }
1293
1294         if (!output_frame_info.is_preroll) {
1295                 output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames);
1296         }
1297
1298         for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) {
1299                 CaptureCard *card = &cards[card_index];
1300                 if (has_new_frame[card_index] &&
1301                     !input_card_is_master_clock(card_index, master_card_index) &&
1302                     !output_frame_info.is_preroll) {
1303                         card->queue_length_policy.update_policy(
1304                                 output_frame_info.frame_timestamp,
1305                                 card->jitter_history.get_expected_next_frame(),
1306                                 new_frames[master_card_index].length,
1307                                 output_frame_info.frame_duration,
1308                                 card->jitter_history.estimate_max_jitter(),
1309                                 output_jitter_history.estimate_max_jitter());
1310                         trim_queue(card, min<int>(global_flags.max_input_queue_frames,
1311                                                   card->queue_length_policy.get_safe_queue_length()));
1312                 }
1313         }
1314
1315         // This might get off by a fractional sample when changing master card
1316         // between ones with different frame rates, but that's fine.
1317         int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples;
1318         output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE;
1319         fractional_samples = num_samples_times_timebase % TIMEBASE;
1320         assert(output_frame_info.num_samples >= 0);
1321
1322         return output_frame_info;
1323 }
1324
1325 void Mixer::handle_hotplugged_cards()
1326 {
1327         // Check for cards that have been disconnected since last frame.
1328         for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1329                 CaptureCard *card = &cards[card_index];
1330                 if (card->capture->get_disconnected()) {
1331                         fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
1332                         FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
1333                         configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr);
1334                         card->queue_length_policy.reset(card_index);
1335                         card->capture->start_bm_capture();
1336                 }
1337         }
1338
1339         // Check for cards that have been connected since last frame.
1340         vector<libusb_device *> hotplugged_cards_copy;
1341         {
1342                 lock_guard<mutex> lock(hotplug_mutex);
1343                 swap(hotplugged_cards, hotplugged_cards_copy);
1344         }
1345         for (libusb_device *new_dev : hotplugged_cards_copy) {
1346                 // Look for a fake capture card where we can stick this in.
1347                 int free_card_index = -1;
1348                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1349                         if (cards[card_index].is_fake_capture) {
1350                                 free_card_index = card_index;
1351                                 break;
1352                         }
1353                 }
1354
1355                 if (free_card_index == -1) {
1356                         fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
1357                         libusb_unref_device(new_dev);
1358                 } else {
1359                         // BMUSBCapture takes ownership.
1360                         fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
1361                         CaptureCard *card = &cards[free_card_index];
1362                         BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
1363                         configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr);
1364                         card->queue_length_policy.reset(free_card_index);
1365                         capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
1366                         capture->start_bm_capture();
1367                 }
1368         }
1369 }
1370
1371
1372 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)
1373 {
1374         // Resample the audio as needed, including from previously dropped frames.
1375         assert(num_cards > 0);
1376         for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
1377                 const bool dropped_frame = (frame_num != dropped_frames);
1378                 {
1379                         // Signal to the audio thread to process this frame.
1380                         // Note that if the frame is a dropped frame, we signal that
1381                         // we don't want to use this frame as base for adjusting
1382                         // the resampler rate. The reason for this is that the timing
1383                         // of these frames is often way too late; they typically don't
1384                         // “arrive” before we synthesize them. Thus, we could end up
1385                         // in a situation where we have inserted e.g. five audio frames
1386                         // into the queue before we then start pulling five of them
1387                         // back out. This makes ResamplingQueue overestimate the delay,
1388                         // causing undue resampler changes. (We _do_ use the last,
1389                         // non-dropped frame; perhaps we should just discard that as well,
1390                         // since dropped frames are expected to be rare, and it might be
1391                         // better to just wait until we have a slightly more normal situation).
1392                         lock_guard<mutex> lock(audio_mutex);
1393                         bool adjust_rate = !dropped_frame && !is_preroll;
1394                         audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
1395                         audio_task_queue_changed.notify_one();
1396                 }
1397                 if (dropped_frame) {
1398                         // For dropped frames, increase the pts. Note that if the format changed
1399                         // in the meantime, we have no way of detecting that; we just have to
1400                         // assume the frame length is always the same.
1401                         pts_int += length_per_frame;
1402                 }
1403         }
1404 }
1405
1406 void Mixer::render_one_frame(int64_t duration)
1407 {
1408         // Determine the time code for this frame before we start rendering.
1409         string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
1410         if (display_timecode_on_stdout) {
1411                 printf("Timecode: '%s'\n", timecode_text.c_str());
1412         }
1413
1414         // Update Y'CbCr settings for all cards.
1415         {
1416                 lock_guard<mutex> lock(card_mutex);
1417                 for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
1418                         YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index];
1419                         input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto;
1420                         input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients;
1421                         input_state.full_range[card_index] = interpretation->full_range;
1422                 }
1423         }
1424
1425         // Get the main chain from the theme, and set its state immediately.
1426         Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
1427         EffectChain *chain = theme_main_chain.chain;
1428         theme_main_chain.setup_chain();
1429         //theme_main_chain.chain->enable_phase_timing(true);
1430
1431         // If HDMI/SDI output is active and the user has requested auto mode,
1432         // its mode overrides the existing Y'CbCr setting for the chain.
1433         YCbCrLumaCoefficients ycbcr_output_coefficients;
1434         if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) {
1435                 ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients();
1436         } else {
1437                 ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601;
1438         }
1439
1440         // TODO: Reduce the duplication against theme.cpp.
1441         YCbCrFormat output_ycbcr_format;
1442         output_ycbcr_format.chroma_subsampling_x = 1;
1443         output_ycbcr_format.chroma_subsampling_y = 1;
1444         output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients;
1445         output_ycbcr_format.full_range = false;
1446         output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth;
1447         chain->change_ycbcr_output_format(output_ycbcr_format);
1448
1449         // Render main chain. If we're using zerocopy Quick Sync encoding
1450         // (the default case), we take an extra copy of the created outputs,
1451         // so that we can display it back to the screen later (it's less memory
1452         // bandwidth than writing and reading back an RGBA texture, even at 16-bit).
1453         // Ideally, we'd like to avoid taking copies and just use the main textures
1454         // for display as well, but they're just views into VA-API memory and must be
1455         // unmapped during encoding, so we can't use them for display, unfortunately.
1456         GLuint y_tex, cbcr_full_tex, cbcr_tex;
1457         GLuint y_copy_tex, cbcr_copy_tex = 0;
1458         GLuint y_display_tex, cbcr_display_tex;
1459         GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8;
1460         GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8;
1461         const bool is_zerocopy = video_encoder->is_zerocopy();
1462         if (is_zerocopy) {
1463                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1464                 y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1465                 cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1466
1467                 y_display_tex = y_copy_tex;
1468                 cbcr_display_tex = cbcr_copy_tex;
1469
1470                 // y_tex and cbcr_tex will be given by VideoEncoder.
1471         } else {
1472                 cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height);
1473                 y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height);
1474                 cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2);
1475
1476                 y_display_tex = y_tex;
1477                 cbcr_display_tex = cbcr_tex;
1478         }
1479
1480         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1481         bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
1482         assert(got_frame);
1483
1484         GLuint fbo;
1485         if (is_zerocopy) {
1486                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex);
1487         } else {
1488                 fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex);
1489         }
1490         check_error();
1491         chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
1492
1493         if (display_timecode_in_stream) {
1494                 // Render the timecode on top.
1495                 timecode_renderer->render_timecode(fbo, timecode_text);
1496         }
1497
1498         resource_pool->release_fbo(fbo);
1499
1500         if (is_zerocopy) {
1501                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex);
1502         } else {
1503                 chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
1504         }
1505         if (output_card_index != -1) {
1506                 cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration);
1507         }
1508         resource_pool->release_2d_texture(cbcr_full_tex);
1509
1510         // Set the right state for the Y' and CbCr textures we use for display.
1511         glBindFramebuffer(GL_FRAMEBUFFER, 0);
1512         glBindTexture(GL_TEXTURE_2D, y_display_tex);
1513         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1514         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1515         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1516
1517         glBindTexture(GL_TEXTURE_2D, cbcr_display_tex);
1518         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
1519         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
1520         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
1521
1522         RefCountedGLsync fence = video_encoder->end_frame();
1523
1524         // The live frame pieces the Y'CbCr texture copies back into RGB and displays them.
1525         // It owns y_display_tex and cbcr_display_tex now (whichever textures they are).
1526         DisplayFrame live_frame;
1527         live_frame.chain = display_chain.get();
1528         live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{
1529                 display_input->set_texture_num(0, y_display_tex);
1530                 display_input->set_texture_num(1, cbcr_display_tex);
1531         };
1532         live_frame.ready_fence = fence;
1533         live_frame.input_frames = {};
1534         live_frame.temp_textures = { y_display_tex, cbcr_display_tex };
1535         output_channel[OUTPUT_LIVE].output_frame(move(live_frame));
1536
1537         // Set up preview and any additional channels.
1538         for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
1539                 DisplayFrame display_frame;
1540                 Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state);  // FIXME: dimensions
1541                 display_frame.chain = move(chain.chain);
1542                 display_frame.setup_chain = move(chain.setup_chain);
1543                 display_frame.ready_fence = fence;
1544                 display_frame.input_frames = move(chain.input_frames);
1545                 display_frame.temp_textures = {};
1546                 output_channel[i].output_frame(move(display_frame));
1547         }
1548 }
1549
1550 void Mixer::audio_thread_func()
1551 {
1552         pthread_setname_np(pthread_self(), "Mixer_Audio");
1553
1554         while (!should_quit) {
1555                 AudioTask task;
1556
1557                 {
1558                         unique_lock<mutex> lock(audio_mutex);
1559                         audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
1560                         if (should_quit) {
1561                                 return;
1562                         }
1563                         task = audio_task_queue.front();
1564                         audio_task_queue.pop();
1565                 }
1566
1567                 ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
1568                         task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
1569                 vector<float> samples_out = audio_mixer->get_output(
1570                         task.frame_timestamp,
1571                         task.num_samples,
1572                         rate_adjustment_policy);
1573
1574                 // Send the samples to the sound card, then add them to the output.
1575                 if (alsa) {
1576                         alsa->write(samples_out);
1577                 }
1578                 if (output_card_index != -1) {
1579                         const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
1580                         cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
1581                 }
1582                 video_encoder->add_audio(task.pts_int, move(samples_out));
1583         }
1584 }
1585
1586 void Mixer::release_display_frame(DisplayFrame *frame)
1587 {
1588         for (GLuint texnum : frame->temp_textures) {
1589                 resource_pool->release_2d_texture(texnum);
1590         }
1591         frame->temp_textures.clear();
1592         frame->ready_fence.reset();
1593         frame->input_frames.clear();
1594 }
1595
1596 void Mixer::start()
1597 {
1598         mixer_thread = thread(&Mixer::thread_func, this);
1599         audio_thread = thread(&Mixer::audio_thread_func, this);
1600 }
1601
1602 void Mixer::quit()
1603 {
1604         should_quit = true;
1605         audio_task_queue_changed.notify_one();
1606         mixer_thread.join();
1607         audio_thread.join();
1608 }
1609
1610 void Mixer::transition_clicked(int transition_num)
1611 {
1612         theme->transition_clicked(transition_num, pts());
1613 }
1614
1615 void Mixer::channel_clicked(int preview_num)
1616 {
1617         theme->channel_clicked(preview_num);
1618 }
1619
1620 YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const
1621 {
1622         lock_guard<mutex> lock(card_mutex);
1623         return ycbcr_interpretation[card_index];
1624 }
1625
1626 void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation)
1627 {
1628         lock_guard<mutex> lock(card_mutex);
1629         ycbcr_interpretation[card_index] = interpretation;
1630 }
1631
1632 void Mixer::start_mode_scanning(unsigned card_index)
1633 {
1634         assert(card_index < num_cards);
1635         if (is_mode_scanning[card_index]) {
1636                 return;
1637         }
1638         is_mode_scanning[card_index] = true;
1639         mode_scanlist[card_index].clear();
1640         for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
1641                 mode_scanlist[card_index].push_back(mode.first);
1642         }
1643         assert(!mode_scanlist[card_index].empty());
1644         mode_scanlist_index[card_index] = 0;
1645         cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
1646         last_mode_scan_change[card_index] = steady_clock::now();
1647 }
1648
1649 map<uint32_t, VideoMode> Mixer::get_available_output_video_modes() const
1650 {
1651         assert(desired_output_card_index != -1);
1652         lock_guard<mutex> lock(card_mutex);
1653         return cards[desired_output_card_index].output->get_available_video_modes();
1654 }
1655
1656 string Mixer::get_ffmpeg_filename(unsigned card_index) const
1657 {
1658         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1659         return ((FFmpegCapture *)(cards[card_index].capture.get()))->get_filename();
1660 }
1661
1662 void Mixer::set_ffmpeg_filename(unsigned card_index, const string &filename) {
1663         assert(card_index >= num_cards && card_index < num_cards + num_video_inputs);
1664         ((FFmpegCapture *)(cards[card_index].capture.get()))->change_filename(filename);
1665 }
1666
1667 void Mixer::wait_for_next_frame()
1668 {
1669         unique_lock<mutex> lock(frame_num_mutex);
1670         unsigned old_frame_num = frame_num;
1671         frame_num_updated.wait_for(lock, seconds(1),  // Timeout is just in case.
1672                 [old_frame_num, this]{ return this->frame_num > old_frame_num; });
1673 }
1674
1675 Mixer::OutputChannel::~OutputChannel()
1676 {
1677         if (has_current_frame) {
1678                 parent->release_display_frame(&current_frame);
1679         }
1680         if (has_ready_frame) {
1681                 parent->release_display_frame(&ready_frame);
1682         }
1683 }
1684
1685 void Mixer::OutputChannel::output_frame(DisplayFrame &&frame)
1686 {
1687         // Store this frame for display. Remove the ready frame if any
1688         // (it was seemingly never used).
1689         {
1690                 lock_guard<mutex> lock(frame_mutex);
1691                 if (has_ready_frame) {
1692                         parent->release_display_frame(&ready_frame);
1693                 }
1694                 ready_frame = move(frame);
1695                 has_ready_frame = true;
1696
1697                 // Call the callbacks under the mutex (they should be short),
1698                 // so that we don't race against a callback removal.
1699                 for (const auto &key_and_callback : new_frame_ready_callbacks) {
1700                         key_and_callback.second();
1701                 }
1702         }
1703
1704         // Reduce the number of callbacks by filtering duplicates. The reason
1705         // why we bother doing this is that Qt seemingly can get into a state
1706         // where its builds up an essentially unbounded queue of signals,
1707         // consuming more and more memory, and there's no good way of collapsing
1708         // user-defined signals or limiting the length of the queue.
1709         if (transition_names_updated_callback) {
1710                 vector<string> transition_names = global_mixer->get_transition_names();
1711                 bool changed = false;
1712                 if (transition_names.size() != last_transition_names.size()) {
1713                         changed = true;
1714                 } else {
1715                         for (unsigned i = 0; i < transition_names.size(); ++i) {
1716                                 if (transition_names[i] != last_transition_names[i]) {
1717                                         changed = true;
1718                                         break;
1719                                 }
1720                         }
1721                 }
1722                 if (changed) {
1723                         transition_names_updated_callback(transition_names);
1724                         last_transition_names = transition_names;
1725                 }
1726         }
1727         if (name_updated_callback) {
1728                 string name = global_mixer->get_channel_name(channel);
1729                 if (name != last_name) {
1730                         name_updated_callback(name);
1731                         last_name = name;
1732                 }
1733         }
1734         if (color_updated_callback) {
1735                 string color = global_mixer->get_channel_color(channel);
1736                 if (color != last_color) {
1737                         color_updated_callback(color);
1738                         last_color = color;
1739                 }
1740         }
1741 }
1742
1743 bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
1744 {
1745         lock_guard<mutex> lock(frame_mutex);
1746         if (!has_current_frame && !has_ready_frame) {
1747                 return false;
1748         }
1749
1750         if (has_current_frame && has_ready_frame) {
1751                 // We have a new ready frame. Toss the current one.
1752                 parent->release_display_frame(&current_frame);
1753                 has_current_frame = false;
1754         }
1755         if (has_ready_frame) {
1756                 assert(!has_current_frame);
1757                 current_frame = move(ready_frame);
1758                 ready_frame.ready_fence.reset();  // Drop the refcount.
1759                 ready_frame.input_frames.clear();  // Drop the refcounts.
1760                 has_current_frame = true;
1761                 has_ready_frame = false;
1762         }
1763
1764         *frame = current_frame;
1765         return true;
1766 }
1767
1768 void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback)
1769 {
1770         lock_guard<mutex> lock(frame_mutex);
1771         new_frame_ready_callbacks[key] = callback;
1772 }
1773
1774 void Mixer::OutputChannel::remove_frame_ready_callback(void *key)
1775 {
1776         lock_guard<mutex> lock(frame_mutex);
1777         new_frame_ready_callbacks.erase(key);
1778 }
1779
1780 void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
1781 {
1782         transition_names_updated_callback = callback;
1783 }
1784
1785 void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
1786 {
1787         name_updated_callback = callback;
1788 }
1789
1790 void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
1791 {
1792         color_updated_callback = callback;
1793 }
1794
1795 mutex RefCountedGLsync::fence_lock;