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

   1 #include <movit/effect_util.h>
   2 #include <movit/util.h>
   3 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
   4 #include <pthread.h>
   5 #include <unistd.h>
   6
   7 #include <mutex>
   8
   9 #include <epoxy/egl.h>
  10
  11 #include "chroma_subsampler.h"
  12 #include "decklink_output.h"
  13 #include "decklink_util.h"
  14 #include "flags.h"
  15 #include "metrics.h"
  16 #include "print_latency.h"
  17 #include "timebase.h"
  18 #include "v210_converter.h"
  19
  20 using namespace movit;
  21 using namespace std;
  22 using namespace std::chrono;
  23
  24 namespace {
  25
  26 // This class can be deleted during regular use, so make all the metrics static.
  27 once_flag decklink_metrics_inited;
  28 LatencyHistogram latency_histogram;
  29 atomic<int64_t> metric_decklink_output_width_pixels{-1};
  30 atomic<int64_t> metric_decklink_output_height_pixels{-1};
  31 atomic<int64_t> metric_decklink_output_frame_rate_den{-1};
  32 atomic<int64_t> metric_decklink_output_frame_rate_nom{-1};
  33 atomic<int64_t> metric_decklink_output_inflight_frames{0};
  34 atomic<int64_t> metric_decklink_output_color_mismatch_frames{0};
  35
  36 atomic<int64_t> metric_decklink_output_scheduled_frames_dropped{0};
  37 atomic<int64_t> metric_decklink_output_scheduled_frames_late{0};
  38 atomic<int64_t> metric_decklink_output_scheduled_frames_normal{0};
  39 atomic<int64_t> metric_decklink_output_scheduled_frames_preroll{0};
  40
  41 atomic<int64_t> metric_decklink_output_completed_frames_completed{0};
  42 atomic<int64_t> metric_decklink_output_completed_frames_dropped{0};
  43 atomic<int64_t> metric_decklink_output_completed_frames_flushed{0};
  44 atomic<int64_t> metric_decklink_output_completed_frames_late{0};
  45 atomic<int64_t> metric_decklink_output_completed_frames_unknown{0};
  46
  47 atomic<int64_t> metric_decklink_output_scheduled_samples{0};
  48
  49 Summary metric_decklink_output_margin_seconds;
  50
  51 }  // namespace
  52
  53 DeckLinkOutput::DeckLinkOutput(ResourcePool *resource_pool, QSurface *surface, unsigned width, unsigned height, unsigned card_index)
  54         : resource_pool(resource_pool), surface(surface), width(width), height(height), card_index(card_index)
  55 {
  56         chroma_subsampler.reset(new ChromaSubsampler(resource_pool));
  57
  58         call_once(decklink_metrics_inited, [](){
  59                 latency_histogram.init("decklink_output");
  60                 global_metrics.add("decklink_output_width_pixels", &metric_decklink_output_width_pixels, Metrics::TYPE_GAUGE);
  61                 global_metrics.add("decklink_output_height_pixels", &metric_decklink_output_height_pixels, Metrics::TYPE_GAUGE);
  62                 global_metrics.add("decklink_output_frame_rate_den", &metric_decklink_output_frame_rate_den, Metrics::TYPE_GAUGE);
  63                 global_metrics.add("decklink_output_frame_rate_nom", &metric_decklink_output_frame_rate_nom, Metrics::TYPE_GAUGE);
  64                 global_metrics.add("decklink_output_inflight_frames", &metric_decklink_output_inflight_frames, Metrics::TYPE_GAUGE);
  65                 global_metrics.add("decklink_output_color_mismatch_frames", &metric_decklink_output_color_mismatch_frames);
  66
  67                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "dropped" }}, &metric_decklink_output_scheduled_frames_dropped);
  68                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "late" }}, &metric_decklink_output_scheduled_frames_late);
  69                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "normal" }}, &metric_decklink_output_scheduled_frames_normal);
  70                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "preroll" }}, &metric_decklink_output_scheduled_frames_preroll);
  71
  72                 global_metrics.add("decklink_output_completed_frames", {{ "status", "completed" }}, &metric_decklink_output_completed_frames_completed);
  73                 global_metrics.add("decklink_output_completed_frames", {{ "status", "dropped" }}, &metric_decklink_output_completed_frames_dropped);
  74                 global_metrics.add("decklink_output_completed_frames", {{ "status", "flushed" }}, &metric_decklink_output_completed_frames_flushed);
  75                 global_metrics.add("decklink_output_completed_frames", {{ "status", "late" }}, &metric_decklink_output_completed_frames_late);
  76                 global_metrics.add("decklink_output_completed_frames", {{ "status", "unknown" }}, &metric_decklink_output_completed_frames_unknown);
  77
  78                 global_metrics.add("decklink_output_scheduled_samples", &metric_decklink_output_scheduled_samples);
  79                 vector<double> quantiles{0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99};
  80                 metric_decklink_output_margin_seconds.init(quantiles, 60.0);
  81                 global_metrics.add("decklink_output_margin_seconds", &metric_decklink_output_margin_seconds);
  82         });
  83 }
  84
  85 void DeckLinkOutput::set_device(IDeckLink *decklink)
  86 {
  87         if (decklink->QueryInterface(IID_IDeckLinkOutput, (void**)&output) != S_OK) {
  88                 fprintf(stderr, "Card %u has no outputs\n", card_index);
  89                 exit(1);
  90         }
  91
  92         IDeckLinkDisplayModeIterator *mode_it;
  93         if (output->GetDisplayModeIterator(&mode_it) != S_OK) {
  94                 fprintf(stderr, "Failed to enumerate output display modes for card %u\n", card_index);
  95                 exit(1);
  96         }
  97
  98         video_modes.clear();
  99
 100         for (const auto &it : summarize_video_modes(mode_it, card_index)) {
 101                 if (it.second.width != width || it.second.height != height) {
 102                         continue;
 103                 }
 104
 105                 // We could support interlaced modes, but let's stay out of it for now,
 106                 // since we don't have interlaced stream output.
 107                 if (it.second.interlaced) {
 108                         continue;
 109                 }
 110
 111                 video_modes.insert(it);
 112         }
 113
 114         mode_it->Release();
 115
 116         // HDMI or SDI generally mean “both HDMI and SDI at the same time” on DeckLink cards
 117         // that support both; pick_default_video_connection() will generally pick one of those
 118         // if they exist. We're not very likely to need analog outputs, so we don't need a way
 119         // to change beyond that.
 120         video_connection = pick_default_video_connection(decklink, BMDDeckLinkVideoOutputConnections, card_index);
 121 }
 122
 123 void DeckLinkOutput::start_output(uint32_t mode, int64_t base_pts)
 124 {
 125         assert(output);
 126         assert(!playback_initiated);
 127
 128         if (video_modes.empty()) {
 129                 fprintf(stderr, "ERROR: No matching output modes for %dx%d found\n", width, height);
 130                 exit(1);
 131         }
 132
 133         should_quit.unquit();
 134         playback_initiated = true;
 135         playback_started = false;
 136         this->base_pts = base_pts;
 137
 138         IDeckLinkConfiguration *config = nullptr;
 139         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
 140                 fprintf(stderr, "Failed to get configuration interface for output card\n");
 141                 exit(1);
 142         }
 143         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
 144                 fprintf(stderr, "Failed to set low latency output\n");
 145                 exit(1);
 146         }
 147         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
 148                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
 149                 exit(1);
 150         }
 151         if (config->SetFlag(bmdDeckLinkConfigUse1080pNotPsF, true) != S_OK) {
 152                 fprintf(stderr, "Failed to set PsF flag for card\n");
 153                 exit(1);
 154         }
 155         if (config->SetFlag(bmdDeckLinkConfigSMPTELevelAOutput, true) != S_OK) {
 156                 // This affects at least some no-name SDI->HDMI converters.
 157                 // Warn, but don't die.
 158                 fprintf(stderr, "WARNING: Failed to enable SMTPE Level A; resolutions like 1080p60 might have issues.\n");
 159         }
 160
 161         BMDDisplayModeSupport support;
 162         IDeckLinkDisplayMode *display_mode;
 163         BMDPixelFormat pixel_format = global_flags.ten_bit_output ? bmdFormat10BitYUV : bmdFormat8BitYUV;
 164         if (output->DoesSupportVideoMode(mode, pixel_format, bmdVideoOutputFlagDefault,
 165                                          &support, &display_mode) != S_OK) {
 166                 fprintf(stderr, "Couldn't ask for format support\n");
 167                 exit(1);
 168         }
 169
 170         if (support == bmdDisplayModeNotSupported) {
 171                 fprintf(stderr, "Requested display mode not supported\n");
 172                 exit(1);
 173         }
 174
 175         current_mode_flags = display_mode->GetFlags();
 176
 177         BMDTimeValue time_value;
 178         BMDTimeScale time_scale;
 179         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
 180                 fprintf(stderr, "Couldn't get frame rate\n");
 181                 exit(1);
 182         }
 183
 184         metric_decklink_output_width_pixels = width;
 185         metric_decklink_output_height_pixels = height;
 186         metric_decklink_output_frame_rate_nom = time_value;
 187         metric_decklink_output_frame_rate_den = time_scale;
 188
 189         frame_duration = time_value * TIMEBASE / time_scale;
 190
 191         display_mode->Release();
 192
 193         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
 194         if (result != S_OK) {
 195                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
 196                 exit(1);
 197         }
 198         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
 199                 fprintf(stderr, "Couldn't set callback\n");
 200                 exit(1);
 201         }
 202         assert(OUTPUT_FREQUENCY == 48000);
 203         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
 204                 fprintf(stderr, "Couldn't enable audio output\n");
 205                 exit(1);
 206         }
 207         if (output->BeginAudioPreroll() != S_OK) {
 208                 fprintf(stderr, "Couldn't begin audio preroll\n");
 209                 exit(1);
 210         }
 211
 212         present_thread = thread([this]{
 213                 QOpenGLContext *context = create_context(this->surface);
 214                 eglBindAPI(EGL_OPENGL_API);
 215                 if (!make_current(context, this->surface)) {
 216                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
 217                                 eglGetError());
 218                         exit(1);
 219                 }
 220                 present_thread_func();
 221                 delete_context(context);
 222         });
 223 }
 224
 225 void DeckLinkOutput::end_output()
 226 {
 227         if (!playback_initiated) {
 228                 return;
 229         }
 230
 231         should_quit.quit();
 232         frame_queues_changed.notify_all();
 233         present_thread.join();
 234         playback_initiated = false;
 235
 236         output->StopScheduledPlayback(0, nullptr, 0);
 237         output->DisableVideoOutput();
 238         output->DisableAudioOutput();
 239
 240         // Wait until all frames are accounted for, and free them.
 241         {
 242                 unique_lock<mutex> lock(frame_queue_mutex);
 243                 while (!(frame_freelist.empty() && num_frames_in_flight == 0)) {
 244                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
 245                         frame_freelist.pop();
 246                 }
 247         }
 248 }
 249
 250 void DeckLinkOutput::send_frame(GLuint y_tex, GLuint cbcr_tex, YCbCrLumaCoefficients output_ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, int64_t pts, int64_t duration)
 251 {
 252         assert(!should_quit.should_quit());
 253
 254         if ((current_mode_flags & bmdDisplayModeColorspaceRec601) && output_ycbcr_coefficients == YCBCR_REC_709) {
 255                 if (!last_frame_had_mode_mismatch) {
 256                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 257                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 258                 }
 259                 last_frame_had_mode_mismatch = true;
 260                 ++metric_decklink_output_color_mismatch_frames;
 261         } else if ((current_mode_flags & bmdDisplayModeColorspaceRec709) && output_ycbcr_coefficients == YCBCR_REC_601) {
 262                 if (!last_frame_had_mode_mismatch) {
 263                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 264                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 265                 }
 266                 last_frame_had_mode_mismatch = true;
 267                 ++metric_decklink_output_color_mismatch_frames;
 268         } else {
 269                 last_frame_had_mode_mismatch = false;
 270         }
 271
 272         unique_ptr<Frame> frame = get_frame();
 273         if (global_flags.ten_bit_output) {
 274                 chroma_subsampler->create_v210(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 275         } else {
 276                 chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 277         }
 278
 279         // Download the UYVY texture to the PBO.
 280         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 281         check_error();
 282
 283         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 284         check_error();
 285
 286         if (global_flags.ten_bit_output) {
 287                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 288                 check_error();
 289                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, BUFFER_OFFSET(0));
 290                 check_error();
 291         } else {
 292                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 293                 check_error();
 294                 glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 295                 check_error();
 296         }
 297
 298         glBindTexture(GL_TEXTURE_2D, 0);
 299         check_error();
 300         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 301         check_error();
 302
 303         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 304         check_error();
 305
 306         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 307         check_error();
 308         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 309         check_error();
 310
 311         frame->input_frames = input_frames;
 312         frame->received_ts = find_received_timestamp(input_frames);
 313         frame->pts = pts;
 314         frame->duration = duration;
 315
 316         {
 317                 unique_lock<mutex> lock(frame_queue_mutex);
 318                 pending_video_frames.push(move(frame));
 319         }
 320         frame_queues_changed.notify_all();
 321 }
 322
 323 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 324 {
 325         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 326         for (size_t i = 0; i < samples.size(); ++i) {
 327                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 328         }
 329
 330         uint32_t frames_written;
 331         HRESULT result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 332                 pts, TIMEBASE, &frames_written);
 333         if (result != S_OK) {
 334                 fprintf(stderr, "ScheduleAudioSamples(pts=%ld) failed (result=0x%08x)\n", pts, result);
 335         } else {
 336                 if (frames_written != samples.size() / 2) {
 337                         fprintf(stderr, "ScheduleAudioSamples() returned short write (%u/%ld)\n", frames_written, samples.size() / 2);
 338                 }
 339         }
 340         metric_decklink_output_scheduled_samples += samples.size() / 2;
 341 }
 342
 343 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
 344 {
 345         assert(!should_quit.should_quit());
 346
 347         *dropped_frames = 0;
 348         *frame_duration = this->frame_duration;
 349
 350         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 351         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 352         BMDTimeValue target_time = pts - buffer;
 353
 354         // While prerolling, we send out frames as quickly as we can.
 355         if (target_time < base_pts) {
 356                 *is_preroll = true;
 357                 ++metric_decklink_output_scheduled_frames_preroll;
 358                 return;
 359         }
 360
 361         *is_preroll = !playback_started;
 362
 363         if (!playback_started) {
 364                 if (output->EndAudioPreroll() != S_OK) {
 365                         fprintf(stderr, "Could not end audio preroll\n");
 366                         exit(1);  // TODO
 367                 }
 368                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 369                         fprintf(stderr, "Could not start playback\n");
 370                         exit(1);  // TODO
 371                 }
 372                 playback_started = true;
 373         }
 374
 375         BMDTimeValue stream_frame_time;
 376         double playback_speed;
 377         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 378
 379         *frame_timestamp = steady_clock::now() +
 380                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 381
 382         metric_decklink_output_margin_seconds.count_event(
 383                 (target_time - stream_frame_time) / double(TIMEBASE));
 384
 385         // If we're ahead of time, wait for the frame to (approximately) start.
 386         if (stream_frame_time < target_time) {
 387                 should_quit.sleep_until(*frame_timestamp);
 388                 ++metric_decklink_output_scheduled_frames_normal;
 389                 return;
 390         }
 391
 392         // If we overshot the previous frame by just a little,
 393         // fire off one immediately.
 394         if (stream_frame_time < target_time + max_overshoot) {
 395                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 396                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 397                 ++metric_decklink_output_scheduled_frames_late;
 398                 return;
 399         }
 400
 401         // Oops, we missed by more than one frame. Return immediately,
 402         // but drop so that we catch up.
 403         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 404         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
 405         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
 406         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 407         metric_decklink_output_scheduled_frames_dropped += *dropped_frames;
 408         ++metric_decklink_output_scheduled_frames_normal;
 409 }
 410
 411 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
 412 {
 413         if (video_modes.count(mode)) {
 414                 return mode;
 415         }
 416
 417         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
 418         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 0 }, { 30000, 1001 } }) {
 419                 for (const auto &it : video_modes) {
 420                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
 421                                 return it.first;
 422                         }
 423                 }
 424         }
 425
 426         uint32_t best_mode = 0;
 427         double best_fps = 0.0;
 428         for (const auto &it : video_modes) {
 429                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
 430                 if (fps > best_fps) {
 431                         best_mode = it.first;
 432                         best_fps = fps;
 433                 }
 434         }
 435         return best_mode;
 436 }
 437
 438 YCbCrLumaCoefficients DeckLinkOutput::preferred_ycbcr_coefficients() const
 439 {
 440         if (current_mode_flags & bmdDisplayModeColorspaceRec601) {
 441                 return YCBCR_REC_601;
 442         } else {
 443                 // Don't bother checking bmdDisplayModeColorspaceRec709;
 444                 // if none is set, 709 is a good default anyway.
 445                 return YCBCR_REC_709;
 446         }
 447 }
 448
 449 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 450 {
 451         Frame *frame = static_cast<Frame *>(completedFrame);
 452         switch (result) {
 453         case bmdOutputFrameCompleted:
 454                 ++metric_decklink_output_completed_frames_completed;
 455                 break;
 456         case bmdOutputFrameDisplayedLate:
 457                 fprintf(stderr, "Output frame displayed late (pts=%ld)\n", frame->pts);
 458                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 459                 ++metric_decklink_output_completed_frames_late;
 460                 break;
 461         case bmdOutputFrameDropped:
 462                 fprintf(stderr, "Output frame was dropped (pts=%ld)\n", frame->pts);
 463                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 464                 ++metric_decklink_output_completed_frames_dropped;
 465                 break;
 466         case bmdOutputFrameFlushed:
 467                 fprintf(stderr, "Output frame was flushed (pts=%ld)\n", frame->pts);
 468                 ++metric_decklink_output_completed_frames_flushed;
 469                 break;
 470         default:
 471                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 472                 ++metric_decklink_output_completed_frames_unknown;
 473                 break;
 474         }
 475
 476         static int frameno = 0;
 477         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &frameno, &latency_histogram);
 478
 479         {
 480                 lock_guard<mutex> lock(frame_queue_mutex);
 481                 frame_freelist.push(unique_ptr<Frame>(frame));
 482                 --num_frames_in_flight;
 483                 --metric_decklink_output_inflight_frames;
 484         }
 485
 486         return S_OK;
 487 }
 488
 489 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 490 {
 491         printf("playback stopped!\n");
 492         return S_OK;
 493 }
 494
 495 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 496 {
 497         lock_guard<mutex> lock(frame_queue_mutex);
 498
 499         if (!frame_freelist.empty()) {
 500                 unique_ptr<Frame> frame = move(frame_freelist.front());
 501                 frame_freelist.pop();
 502                 return frame;
 503         }
 504
 505         unique_ptr<Frame> frame(new Frame);
 506
 507         size_t stride;
 508         if (global_flags.ten_bit_output) {
 509                 stride = v210Converter::get_v210_stride(width);
 510                 GLint v210_width = stride / sizeof(uint32_t);
 511                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGB10_A2, v210_width, height);
 512
 513                 // We need valid texture state, or NVIDIA won't allow us to write to the texture.
 514                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 515                 check_error();
 516                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 517                 check_error();
 518         } else {
 519                 stride = width * 2;
 520                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 521         }
 522
 523         glGenBuffers(1, &frame->pbo);
 524         check_error();
 525         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 526         check_error();
 527         glBufferStorage(GL_PIXEL_PACK_BUFFER, stride * height, NULL, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 528         check_error();
 529         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, stride * height, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 530         check_error();
 531         frame->uyvy_ptr_local.reset(new uint8_t[stride * height]);
 532         frame->resource_pool = resource_pool;
 533
 534         return frame;
 535 }
 536
 537 void DeckLinkOutput::present_thread_func()
 538 {
 539         pthread_setname_np(pthread_self(), "DeckLinkOutput");
 540         for ( ;; ) {
 541                 unique_ptr<Frame> frame;
 542                 {
 543                         unique_lock<mutex> lock(frame_queue_mutex);
 544                         frame_queues_changed.wait(lock, [this]{
 545                                 return should_quit.should_quit() || !pending_video_frames.empty();
 546                         });
 547                         if (should_quit.should_quit()) {
 548                                 return;
 549                         }
 550                         frame = move(pending_video_frames.front());
 551                         pending_video_frames.pop();
 552                         ++num_frames_in_flight;
 553                         ++metric_decklink_output_inflight_frames;
 554                 }
 555
 556                 for ( ;; ) {
 557                         int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
 558                         if (err == GL_TIMEOUT_EXPIRED) {
 559                                 // NVIDIA likes to busy-wait; yield instead.
 560                                 this_thread::sleep_for(milliseconds(1));
 561                         } else {
 562                                 break;
 563                         }
 564                 }
 565                 check_error();
 566                 frame->fence.reset();
 567
 568                 if (global_flags.ten_bit_output) {
 569                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, v210Converter::get_v210_stride(width) * height);
 570                 } else {
 571                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 572                 }
 573
 574                 // Release any input frames we needed to render this frame.
 575                 frame->input_frames.clear();
 576
 577                 BMDTimeValue pts = frame->pts;
 578                 BMDTimeValue duration = frame->duration;
 579                 HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 580                 if (res == S_OK) {
 581                         frame.release();  // Owned by the driver now.
 582                 } else {
 583                         fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 584
 585                         lock_guard<mutex> lock(frame_queue_mutex);
 586                         frame_freelist.push(move(frame));
 587                         --num_frames_in_flight;
 588                         --metric_decklink_output_inflight_frames;
 589                 }
 590         }
 591 }
 592
 593 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 594 {
 595         return E_NOINTERFACE;
 596 }
 597
 598 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 599 {
 600         return refcount.fetch_add(1) + 1;
 601 }
 602
 603 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 604 {
 605         int new_ref = refcount.fetch_sub(1) - 1;
 606         if (new_ref == 0)
 607                 delete this;
 608         return new_ref;
 609 }
 610
 611 DeckLinkOutput::Frame::~Frame()
 612 {
 613         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 614         check_error();
 615         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 616         check_error();
 617         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 618         check_error();
 619         glDeleteBuffers(1, &pbo);
 620         check_error();
 621         resource_pool->release_2d_texture(uyvy_tex);
 622         check_error();
 623 }
 624
 625 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 626 {
 627         return E_NOINTERFACE;
 628 }
 629
 630 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 631 {
 632         return refcount.fetch_add(1) + 1;
 633 }
 634
 635 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 636 {
 637         int new_ref = refcount.fetch_sub(1) - 1;
 638         if (new_ref == 0)
 639                 delete this;
 640         return new_ref;
 641 }
 642
 643 long DeckLinkOutput::Frame::GetWidth()
 644 {
 645         return global_flags.width;
 646 }
 647
 648 long DeckLinkOutput::Frame::GetHeight()
 649 {
 650         return global_flags.height;
 651 }
 652
 653 long DeckLinkOutput::Frame::GetRowBytes()
 654 {
 655         if (global_flags.ten_bit_output) {
 656                 return v210Converter::get_v210_stride(global_flags.width);
 657         } else {
 658                 return global_flags.width * 2;
 659         }
 660 }
 661
 662 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 663 {
 664         if (global_flags.ten_bit_output) {
 665                 return bmdFormat10BitYUV;
 666         } else {
 667                 return bmdFormat8BitYUV;
 668         }
 669 }
 670
 671 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 672 {
 673         return bmdFrameFlagDefault;
 674 }
 675
 676 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 677 {
 678         *buffer = uyvy_ptr_local.get();
 679         return S_OK;
 680 }
 681
 682 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 683 {
 684         fprintf(stderr, "STUB: GetTimecode()\n");
 685         return E_NOTIMPL;
 686 }
 687
 688 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 689 {
 690         fprintf(stderr, "STUB: GetAncillaryData()\n");
 691         return E_NOTIMPL;
 692 }