git.sesse.net Git - nageru/blob - nageru/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 "shared/metrics.h"
  16 #include "print_latency.h"
  17 #include "shared/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 bool DeckLinkOutput::set_device(IDeckLink *decklink)
  86 {
  87         if (decklink->QueryInterface(IID_IDeckLinkInput, (void**)&input) != S_OK) {
  88                 input = nullptr;
  89         }
  90         if (decklink->QueryInterface(IID_IDeckLinkOutput, (void**)&output) != S_OK) {
  91                 fprintf(stderr, "Warning: Card %u has no outputs\n", card_index);
  92                 return false;
  93         }
  94
  95         IDeckLinkDisplayModeIterator *mode_it;
  96         if (output->GetDisplayModeIterator(&mode_it) != S_OK) {
  97                 fprintf(stderr, "Warning: Failed to enumerate output display modes for card %u\n", card_index);
  98                 return false;
  99         }
 100
 101         video_modes.clear();
 102
 103         for (const auto &it : summarize_video_modes(mode_it, card_index)) {
 104                 if (it.second.width != width || it.second.height != height) {
 105                         continue;
 106                 }
 107
 108                 // We could support interlaced modes, but let's stay out of it for now,
 109                 // since we don't have interlaced stream output.
 110                 if (it.second.interlaced) {
 111                         continue;
 112                 }
 113
 114                 video_modes.insert(it);
 115         }
 116
 117         mode_it->Release();
 118
 119         // HDMI or SDI generally mean “both HDMI and SDI at the same time” on DeckLink cards
 120         // that support both; pick_default_video_connection() will generally pick one of those
 121         // if they exist. (--prefer-hdmi-input would also affect the selection despite the name
 122         // of the option, but since either generally means both, it's inconsequential.)
 123         // We're not very likely to need analog outputs, so we don't need a way to change
 124         // beyond that.
 125         video_connection = pick_default_video_connection(decklink, BMDDeckLinkVideoOutputConnections, card_index);
 126         return true;
 127 }
 128
 129 void DeckLinkOutput::start_output(uint32_t mode, int64_t base_pts)
 130 {
 131         assert(output);
 132         assert(!playback_initiated);
 133
 134         if (video_modes.empty()) {
 135                 fprintf(stderr, "ERROR: No matching output modes for %dx%d found\n", width, height);
 136                 abort();
 137         }
 138
 139         should_quit.unquit();
 140         playback_initiated = true;
 141         playback_started = false;
 142         this->base_pts = base_pts;
 143
 144         IDeckLinkConfiguration *config = nullptr;
 145         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
 146                 fprintf(stderr, "Failed to get configuration interface for output card\n");
 147                 abort();
 148         }
 149         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
 150                 fprintf(stderr, "Failed to set low latency output\n");
 151                 abort();
 152         }
 153         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
 154                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
 155                 abort();
 156         }
 157         if (config->SetFlag(bmdDeckLinkConfigOutput1080pAsPsF, true) != S_OK) {
 158                 fprintf(stderr, "Failed to set PsF flag for card\n");
 159                 abort();
 160         }
 161         if (config->SetFlag(bmdDeckLinkConfigSMPTELevelAOutput, true) != S_OK) {
 162                 // This affects at least some no-name SDI->HDMI converters.
 163                 // Warn, but don't die.
 164                 fprintf(stderr, "WARNING: Failed to enable SMTPE Level A; resolutions like 1080p60 might have issues.\n");
 165         }
 166
 167         BMDDisplayModeSupport support;
 168         IDeckLinkDisplayMode *display_mode;
 169         BMDPixelFormat pixel_format = global_flags.ten_bit_output ? bmdFormat10BitYUV : bmdFormat8BitYUV;
 170         if (output->DoesSupportVideoMode(mode, pixel_format, bmdVideoOutputFlagDefault,
 171                                          &support, &display_mode) != S_OK) {
 172                 fprintf(stderr, "Couldn't ask for format support\n");
 173                 abort();
 174         }
 175
 176         if (support == bmdDisplayModeNotSupported) {
 177                 fprintf(stderr, "Requested display mode not supported\n");
 178                 abort();
 179         }
 180
 181         current_mode_flags = display_mode->GetFlags();
 182
 183         BMDTimeValue time_value;
 184         BMDTimeScale time_scale;
 185         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
 186                 fprintf(stderr, "Couldn't get frame rate\n");
 187                 abort();
 188         }
 189
 190         metric_decklink_output_width_pixels = width;
 191         metric_decklink_output_height_pixels = height;
 192         metric_decklink_output_frame_rate_nom = time_value;
 193         metric_decklink_output_frame_rate_den = time_scale;
 194
 195         frame_duration = time_value * TIMEBASE / time_scale;
 196
 197         display_mode->Release();
 198
 199         if (input != nullptr) {
 200                 if (input->DisableVideoInput() != S_OK) {
 201                         fprintf(stderr, "Warning: Failed to disable video input for card %d\n", card_index);
 202                 }
 203                 if (input->DisableAudioInput() != S_OK) {
 204                         fprintf(stderr, "Warning: Failed to disable audio input for card %d\n", card_index);
 205                 }
 206         }
 207
 208         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
 209         if (result != S_OK) {
 210                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
 211                 abort();
 212         }
 213         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
 214                 fprintf(stderr, "Couldn't set callback\n");
 215                 abort();
 216         }
 217         assert(OUTPUT_FREQUENCY == 48000);
 218         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
 219                 fprintf(stderr, "Couldn't enable audio output\n");
 220                 abort();
 221         }
 222         if (output->BeginAudioPreroll() != S_OK) {
 223                 fprintf(stderr, "Couldn't begin audio preroll\n");
 224                 abort();
 225         }
 226
 227         present_thread = thread([this]{
 228                 QOpenGLContext *context = create_context(this->surface);
 229                 eglBindAPI(EGL_OPENGL_API);
 230                 if (!make_current(context, this->surface)) {
 231                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
 232                                 eglGetError());
 233                         abort();
 234                 }
 235                 present_thread_func();
 236                 delete_context(context);
 237         });
 238 }
 239
 240 void DeckLinkOutput::end_output()
 241 {
 242         if (!playback_initiated) {
 243                 return;
 244         }
 245
 246         should_quit.quit();
 247         frame_queues_changed.notify_all();
 248         present_thread.join();
 249         playback_initiated = false;
 250
 251         output->StopScheduledPlayback(0, nullptr, 0);
 252         output->DisableVideoOutput();
 253         output->DisableAudioOutput();
 254
 255         // Wait until all frames are accounted for, and free them.
 256         {
 257                 unique_lock<mutex> lock(frame_queue_mutex);
 258                 while (!(frame_freelist.empty() && num_frames_in_flight == 0)) {
 259                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
 260                         frame_freelist.pop();
 261                 }
 262         }
 263
 264         if (input != nullptr) {
 265                 input->Release();
 266                 input = nullptr;
 267         }
 268         if (output != nullptr) {
 269                 output->Release();
 270                 output = nullptr;
 271         }
 272 }
 273
 274 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)
 275 {
 276         assert(!should_quit.should_quit());
 277
 278         if ((current_mode_flags & bmdDisplayModeColorspaceRec601) && output_ycbcr_coefficients == YCBCR_REC_709) {
 279                 if (!last_frame_had_mode_mismatch) {
 280                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 281                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 282                 }
 283                 last_frame_had_mode_mismatch = true;
 284                 ++metric_decklink_output_color_mismatch_frames;
 285         } else if ((current_mode_flags & bmdDisplayModeColorspaceRec709) && output_ycbcr_coefficients == YCBCR_REC_601) {
 286                 if (!last_frame_had_mode_mismatch) {
 287                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 288                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 289                 }
 290                 last_frame_had_mode_mismatch = true;
 291                 ++metric_decklink_output_color_mismatch_frames;
 292         } else {
 293                 last_frame_had_mode_mismatch = false;
 294         }
 295
 296         unique_ptr<Frame> frame = get_frame();
 297         if (global_flags.ten_bit_output) {
 298                 chroma_subsampler->create_v210(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 299         } else {
 300                 chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 301         }
 302
 303         // Download the UYVY texture to the PBO.
 304         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 305         check_error();
 306
 307         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 308         check_error();
 309
 310         if (global_flags.ten_bit_output) {
 311                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 312                 check_error();
 313                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, BUFFER_OFFSET(0));
 314                 check_error();
 315         } else {
 316                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 317                 check_error();
 318                 glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 319                 check_error();
 320         }
 321
 322         glBindTexture(GL_TEXTURE_2D, 0);
 323         check_error();
 324         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 325         check_error();
 326
 327         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 328         check_error();
 329
 330         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 331         check_error();
 332         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 333         check_error();
 334
 335         frame->input_frames = input_frames;
 336         frame->received_ts = find_received_timestamp(input_frames);
 337         frame->pts = pts;
 338         frame->duration = duration;
 339
 340         {
 341                 lock_guard<mutex> lock(frame_queue_mutex);
 342                 pending_video_frames.push(move(frame));
 343         }
 344         frame_queues_changed.notify_all();
 345 }
 346
 347 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 348 {
 349         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 350         for (size_t i = 0; i < samples.size(); ++i) {
 351                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 352         }
 353
 354         uint32_t frames_written;
 355         HRESULT result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 356                 pts, TIMEBASE, &frames_written);
 357         if (result != S_OK) {
 358                 fprintf(stderr, "ScheduleAudioSamples(pts=%" PRId64 ") failed (result=0x%08x)\n", pts, result);
 359         } else {
 360                 if (frames_written != samples.size() / 2) {
 361                         fprintf(stderr, "ScheduleAudioSamples() returned short write (%u/%zu)\n", frames_written, samples.size() / 2);
 362                 }
 363         }
 364         metric_decklink_output_scheduled_samples += samples.size() / 2;
 365 }
 366
 367 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
 368 {
 369         assert(!should_quit.should_quit());
 370
 371         *dropped_frames = 0;
 372         *frame_duration = this->frame_duration;
 373
 374         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 375         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 376         BMDTimeValue target_time = pts - buffer;
 377
 378         // While prerolling, we send out frames as quickly as we can.
 379         if (target_time < base_pts) {
 380                 *is_preroll = true;
 381                 ++metric_decklink_output_scheduled_frames_preroll;
 382                 return;
 383         }
 384
 385         *is_preroll = !playback_started;
 386
 387         if (!playback_started) {
 388                 if (output->EndAudioPreroll() != S_OK) {
 389                         fprintf(stderr, "Could not end audio preroll\n");
 390                         abort();  // TODO
 391                 }
 392                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 393                         fprintf(stderr, "Could not start playback\n");
 394                         abort();  // TODO
 395                 }
 396                 playback_started = true;
 397         }
 398
 399         BMDTimeValue stream_frame_time;
 400         double playback_speed;
 401         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 402
 403         *frame_timestamp = steady_clock::now() +
 404                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 405
 406         metric_decklink_output_margin_seconds.count_event(
 407                 (target_time - stream_frame_time) / double(TIMEBASE));
 408
 409         // If we're ahead of time, wait for the frame to (approximately) start.
 410         if (stream_frame_time < target_time) {
 411                 should_quit.sleep_until(*frame_timestamp);
 412                 ++metric_decklink_output_scheduled_frames_normal;
 413                 return;
 414         }
 415
 416         // If we overshot the previous frame by just a little,
 417         // fire off one immediately.
 418         if (stream_frame_time < target_time + max_overshoot) {
 419                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 420                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 421                 ++metric_decklink_output_scheduled_frames_late;
 422                 return;
 423         }
 424
 425         // Oops, we missed by more than one frame. Return immediately,
 426         // but drop so that we catch up.
 427         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 428         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
 429         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
 430         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 431         metric_decklink_output_scheduled_frames_dropped += *dropped_frames;
 432         ++metric_decklink_output_scheduled_frames_normal;
 433 }
 434
 435 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
 436 {
 437         if (video_modes.count(mode)) {
 438                 return mode;
 439         }
 440
 441         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
 442         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 1 }, { 30000, 1001 } }) {
 443                 for (const auto &it : video_modes) {
 444                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
 445                                 return it.first;
 446                         }
 447                 }
 448         }
 449
 450         uint32_t best_mode = 0;
 451         double best_fps = 0.0;
 452         for (const auto &it : video_modes) {
 453                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
 454                 if (fps > best_fps) {
 455                         best_mode = it.first;
 456                         best_fps = fps;
 457                 }
 458         }
 459         return best_mode;
 460 }
 461
 462 YCbCrLumaCoefficients DeckLinkOutput::preferred_ycbcr_coefficients() const
 463 {
 464         if (current_mode_flags & bmdDisplayModeColorspaceRec601) {
 465                 return YCBCR_REC_601;
 466         } else {
 467                 // Don't bother checking bmdDisplayModeColorspaceRec709;
 468                 // if none is set, 709 is a good default anyway.
 469                 return YCBCR_REC_709;
 470         }
 471 }
 472
 473 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 474 {
 475         Frame *frame = static_cast<Frame *>(completedFrame);
 476         switch (result) {
 477         case bmdOutputFrameCompleted:
 478                 ++metric_decklink_output_completed_frames_completed;
 479                 break;
 480         case bmdOutputFrameDisplayedLate:
 481                 fprintf(stderr, "Output frame displayed late (pts=%" PRId64 ")\n", frame->pts);
 482                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 483                 ++metric_decklink_output_completed_frames_late;
 484                 break;
 485         case bmdOutputFrameDropped:
 486                 fprintf(stderr, "Output frame was dropped (pts=%" PRId64 ")\n", frame->pts);
 487                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 488                 ++metric_decklink_output_completed_frames_dropped;
 489                 break;
 490         case bmdOutputFrameFlushed:
 491                 fprintf(stderr, "Output frame was flushed (pts=%" PRId64 ")\n", frame->pts);
 492                 ++metric_decklink_output_completed_frames_flushed;
 493                 break;
 494         default:
 495                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 496                 ++metric_decklink_output_completed_frames_unknown;
 497                 break;
 498         }
 499
 500         static int frameno = 0;
 501         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &frameno, &latency_histogram);
 502
 503         {
 504                 lock_guard<mutex> lock(frame_queue_mutex);
 505                 frame_freelist.push(unique_ptr<Frame>(frame));
 506                 --num_frames_in_flight;
 507                 --metric_decklink_output_inflight_frames;
 508         }
 509
 510         return S_OK;
 511 }
 512
 513 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 514 {
 515         printf("playback stopped!\n");
 516         return S_OK;
 517 }
 518
 519 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 520 {
 521         lock_guard<mutex> lock(frame_queue_mutex);
 522
 523         if (!frame_freelist.empty()) {
 524                 unique_ptr<Frame> frame = move(frame_freelist.front());
 525                 frame_freelist.pop();
 526                 return frame;
 527         }
 528
 529         unique_ptr<Frame> frame(new Frame);
 530
 531         size_t stride;
 532         if (global_flags.ten_bit_output) {
 533                 stride = v210Converter::get_v210_stride(width);
 534                 GLint v210_width = stride / sizeof(uint32_t);
 535                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGB10_A2, v210_width, height);
 536
 537                 // We need valid texture state, or NVIDIA won't allow us to write to the texture.
 538                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 539                 check_error();
 540                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 541                 check_error();
 542         } else {
 543                 stride = width * 2;
 544                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 545         }
 546
 547         glGenBuffers(1, &frame->pbo);
 548         check_error();
 549         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 550         check_error();
 551         glBufferStorage(GL_PIXEL_PACK_BUFFER, stride * height, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 552         check_error();
 553         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, stride * height, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 554         check_error();
 555         frame->uyvy_ptr_local.reset(new uint8_t[stride * height]);
 556         frame->resource_pool = resource_pool;
 557
 558         return frame;
 559 }
 560
 561 void DeckLinkOutput::present_thread_func()
 562 {
 563         pthread_setname_np(pthread_self(), "DeckLinkOutput");
 564         for ( ;; ) {
 565                 unique_ptr<Frame> frame;
 566                 {
 567                         unique_lock<mutex> lock(frame_queue_mutex);
 568                         frame_queues_changed.wait(lock, [this]{
 569                                 return should_quit.should_quit() || !pending_video_frames.empty();
 570                         });
 571                         if (should_quit.should_quit()) {
 572                                 return;
 573                         }
 574                         frame = move(pending_video_frames.front());
 575                         pending_video_frames.pop();
 576                         ++num_frames_in_flight;
 577                         ++metric_decklink_output_inflight_frames;
 578                 }
 579
 580                 for ( ;; ) {
 581                         int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
 582                         if (err == GL_TIMEOUT_EXPIRED) {
 583                                 // NVIDIA likes to busy-wait; yield instead.
 584                                 this_thread::sleep_for(milliseconds(1));
 585                         } else {
 586                                 break;
 587                         }
 588                 }
 589                 check_error();
 590                 frame->fence.reset();
 591
 592                 if (global_flags.ten_bit_output) {
 593                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, v210Converter::get_v210_stride(width) * height);
 594                 } else {
 595                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 596                 }
 597
 598                 // Release any input frames we needed to render this frame.
 599                 frame->input_frames.clear();
 600
 601                 BMDTimeValue pts = frame->pts;
 602                 BMDTimeValue duration = frame->duration;
 603                 HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 604                 if (res == S_OK) {
 605                         frame.release();  // Owned by the driver now.
 606                 } else {
 607                         fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 608
 609                         lock_guard<mutex> lock(frame_queue_mutex);
 610                         frame_freelist.push(move(frame));
 611                         --num_frames_in_flight;
 612                         --metric_decklink_output_inflight_frames;
 613                 }
 614         }
 615 }
 616
 617 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 618 {
 619         return E_NOINTERFACE;
 620 }
 621
 622 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 623 {
 624         return refcount.fetch_add(1) + 1;
 625 }
 626
 627 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 628 {
 629         int new_ref = refcount.fetch_sub(1) - 1;
 630         if (new_ref == 0)
 631                 delete this;
 632         return new_ref;
 633 }
 634
 635 DeckLinkOutput::Frame::~Frame()
 636 {
 637         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 638         check_error();
 639         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 640         check_error();
 641         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 642         check_error();
 643         glDeleteBuffers(1, &pbo);
 644         check_error();
 645         resource_pool->release_2d_texture(uyvy_tex);
 646         check_error();
 647 }
 648
 649 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 650 {
 651         return E_NOINTERFACE;
 652 }
 653
 654 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 655 {
 656         return refcount.fetch_add(1) + 1;
 657 }
 658
 659 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 660 {
 661         int new_ref = refcount.fetch_sub(1) - 1;
 662         if (new_ref == 0)
 663                 delete this;
 664         return new_ref;
 665 }
 666
 667 long DeckLinkOutput::Frame::GetWidth()
 668 {
 669         return global_flags.width;
 670 }
 671
 672 long DeckLinkOutput::Frame::GetHeight()
 673 {
 674         return global_flags.height;
 675 }
 676
 677 long DeckLinkOutput::Frame::GetRowBytes()
 678 {
 679         if (global_flags.ten_bit_output) {
 680                 return v210Converter::get_v210_stride(global_flags.width);
 681         } else {
 682                 return global_flags.width * 2;
 683         }
 684 }
 685
 686 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 687 {
 688         if (global_flags.ten_bit_output) {
 689                 return bmdFormat10BitYUV;
 690         } else {
 691                 return bmdFormat8BitYUV;
 692         }
 693 }
 694
 695 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 696 {
 697         return bmdFrameFlagDefault;
 698 }
 699
 700 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 701 {
 702         *buffer = uyvy_ptr_local.get();
 703         return S_OK;
 704 }
 705
 706 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 707 {
 708         fprintf(stderr, "STUB: GetTimecode()\n");
 709         return E_NOTIMPL;
 710 }
 711
 712 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 713 {
 714         fprintf(stderr, "STUB: GetAncillaryData()\n");
 715         return E_NOTIMPL;
 716 }