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 DeckLinkOutput::~DeckLinkOutput()
  86 {
  87         if (output != nullptr) {
  88                 output->Release();
  89         }
  90 }
  91
  92 bool DeckLinkOutput::set_device(IDeckLink *decklink, IDeckLinkInput *input_arg)
  93 {
  94         input = input_arg;
  95         if (decklink->QueryInterface(IID_IDeckLinkOutput, (void**)&output) != S_OK) {
  96                 fprintf(stderr, "Warning: Card %u has no outputs\n", card_index);
  97                 return false;
  98         }
  99
 100         IDeckLinkDisplayModeIterator *mode_it;
 101         if (output->GetDisplayModeIterator(&mode_it) != S_OK) {
 102                 fprintf(stderr, "Warning: Failed to enumerate output display modes for card %u\n", card_index);
 103                 return false;
 104         }
 105
 106         video_modes.clear();
 107
 108         for (const auto &it : summarize_video_modes(mode_it, card_index)) {
 109                 if (it.second.width != width || it.second.height != height) {
 110                         continue;
 111                 }
 112
 113                 // We could support interlaced modes, but let's stay out of it for now,
 114                 // since we don't have interlaced stream output.
 115                 if (it.second.interlaced) {
 116                         continue;
 117                 }
 118
 119                 video_modes.insert(it);
 120         }
 121
 122         mode_it->Release();
 123
 124         // HDMI or SDI generally mean “both HDMI and SDI at the same time” on DeckLink cards
 125         // that support both; pick_default_video_connection() will generally pick one of those
 126         // if they exist. (--prefer-hdmi-input would also affect the selection despite the name
 127         // of the option, but since either generally means both, it's inconsequential.)
 128         // We're not very likely to need analog outputs, so we don't need a way to change
 129         // beyond that.
 130         video_connection = pick_default_video_connection(decklink, BMDDeckLinkVideoOutputConnections, card_index);
 131         return true;
 132 }
 133
 134 void DeckLinkOutput::start_output(uint32_t mode, int64_t base_pts, bool is_master_card_arg)
 135 {
 136         assert(output);
 137         assert(!playback_initiated);
 138         this->is_master_card = is_master_card_arg;
 139
 140         if (video_modes.empty()) {
 141                 fprintf(stderr, "ERROR: No matching output modes for %dx%d found\n", width, height);
 142                 abort();
 143         }
 144
 145         should_quit.unquit();
 146         playback_initiated = true;
 147         playback_started = false;
 148         this->base_pts = base_pts;
 149
 150         IDeckLinkConfiguration *config = nullptr;
 151         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
 152                 fprintf(stderr, "Failed to get configuration interface for output card\n");
 153                 abort();
 154         }
 155         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
 156                 fprintf(stderr, "Failed to set low latency output\n");
 157                 abort();
 158         }
 159         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
 160                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
 161                 abort();
 162         }
 163         if (config->SetFlag(bmdDeckLinkConfigOutput1080pAsPsF, true) != S_OK) {
 164                 fprintf(stderr, "Failed to set PsF flag for card\n");
 165                 abort();
 166         }
 167         if (config->SetFlag(bmdDeckLinkConfigSMPTELevelAOutput, true) != S_OK) {
 168                 // This affects at least some no-name SDI->HDMI converters.
 169                 // Warn, but don't die.
 170                 fprintf(stderr, "WARNING: Failed to enable SMTPE Level A; resolutions like 1080p60 might have issues.\n");
 171         }
 172
 173         BMDDisplayModeSupport support;
 174         IDeckLinkDisplayMode *display_mode;
 175         BMDPixelFormat pixel_format = global_flags.bit_depth > 8 ? bmdFormat10BitYUV : bmdFormat8BitYUV;
 176         if (output->DoesSupportVideoMode(mode, pixel_format, bmdVideoOutputFlagDefault,
 177                                          &support, &display_mode) != S_OK) {
 178                 fprintf(stderr, "Couldn't ask for format support\n");
 179                 abort();
 180         }
 181
 182         if (support == bmdDisplayModeNotSupported) {
 183                 fprintf(stderr, "Requested display mode not supported\n");
 184                 abort();
 185         }
 186
 187         current_mode_flags = display_mode->GetFlags();
 188
 189         BMDTimeValue time_value;
 190         BMDTimeScale time_scale;
 191         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
 192                 fprintf(stderr, "Couldn't get frame rate\n");
 193                 abort();
 194         }
 195
 196         metric_decklink_output_width_pixels = width;
 197         metric_decklink_output_height_pixels = height;
 198         metric_decklink_output_frame_rate_nom = time_value;
 199         metric_decklink_output_frame_rate_den = time_scale;
 200
 201         frame_duration = time_value * TIMEBASE / time_scale;
 202
 203         display_mode->Release();
 204
 205         if (input != nullptr) {
 206                 if (input->DisableVideoInput() != S_OK) {
 207                         fprintf(stderr, "Warning: Failed to disable video input for card %d\n", card_index);
 208                 }
 209                 if (input->DisableAudioInput() != S_OK) {
 210                         fprintf(stderr, "Warning: Failed to disable audio input for card %d\n", card_index);
 211                 }
 212         }
 213
 214         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
 215         if (result != S_OK) {
 216                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
 217                 abort();
 218         }
 219         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
 220                 fprintf(stderr, "Couldn't set callback\n");
 221                 abort();
 222         }
 223         assert(OUTPUT_FREQUENCY == 48000);
 224         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
 225                 fprintf(stderr, "Couldn't enable audio output\n");
 226                 abort();
 227         }
 228         if (is_master_card) {
 229                 if (output->BeginAudioPreroll() != S_OK) {
 230                         fprintf(stderr, "Couldn't begin audio preroll\n");
 231                         abort();
 232                 }
 233         } else {
 234                 playback_started = true;
 235         }
 236
 237         present_thread = thread([this]{
 238                 QOpenGLContext *context = create_context(this->surface);
 239                 eglBindAPI(EGL_OPENGL_API);
 240                 if (!make_current(context, this->surface)) {
 241                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
 242                                 eglGetError());
 243                         abort();
 244                 }
 245                 present_thread_func();
 246                 delete_context(context);
 247         });
 248 }
 249
 250 void DeckLinkOutput::end_output()
 251 {
 252         if (!playback_initiated) {
 253                 return;
 254         }
 255
 256         should_quit.quit();
 257         frame_queues_changed.notify_all();
 258         present_thread.join();
 259         playback_initiated = false;
 260
 261         if (is_master_card) {
 262                 output->StopScheduledPlayback(0, nullptr, 0);
 263         }
 264         output->DisableVideoOutput();
 265         output->DisableAudioOutput();
 266
 267         // Wait until all frames are accounted for, and free them.
 268         {
 269                 unique_lock<mutex> lock(frame_queue_mutex);
 270                 while (!(frame_freelist.empty() && scheduled_frames.empty())) {
 271                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
 272                         frame_freelist.pop();
 273                 }
 274         }
 275 }
 276
 277 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)
 278 {
 279         assert(!should_quit.should_quit());
 280
 281         if ((current_mode_flags & bmdDisplayModeColorspaceRec601) && output_ycbcr_coefficients == YCBCR_REC_709) {
 282                 if (!last_frame_had_mode_mismatch) {
 283                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 284                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 285                 }
 286                 last_frame_had_mode_mismatch = true;
 287                 ++metric_decklink_output_color_mismatch_frames;
 288         } else if ((current_mode_flags & bmdDisplayModeColorspaceRec709) && output_ycbcr_coefficients == YCBCR_REC_601) {
 289                 if (!last_frame_had_mode_mismatch) {
 290                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 291                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 292                 }
 293                 last_frame_had_mode_mismatch = true;
 294                 ++metric_decklink_output_color_mismatch_frames;
 295         } else {
 296                 last_frame_had_mode_mismatch = false;
 297         }
 298
 299         unique_ptr<Frame> frame = get_frame();
 300         if (global_flags.bit_depth > 8) {
 301                 chroma_subsampler->create_v210(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 302         } else {
 303                 chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 304         }
 305
 306         // Download the UYVY texture to the PBO.
 307         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 308         check_error();
 309
 310         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 311         check_error();
 312
 313         if (global_flags.bit_depth > 8) {
 314                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 315                 check_error();
 316                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, BUFFER_OFFSET(0));
 317                 check_error();
 318         } else {
 319                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 320                 check_error();
 321                 glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 322                 check_error();
 323         }
 324
 325         glBindTexture(GL_TEXTURE_2D, 0);
 326         check_error();
 327         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 328         check_error();
 329
 330         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 331         check_error();
 332
 333         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 334         check_error();
 335         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 336         check_error();
 337
 338         frame->input_frames = input_frames;
 339         frame->received_ts = find_received_timestamp(input_frames);
 340         frame->pts = pts;
 341         frame->duration = duration;
 342
 343         {
 344                 lock_guard<mutex> lock(frame_queue_mutex);
 345                 pending_video_frames.push(move(frame));
 346         }
 347         frame_queues_changed.notify_all();
 348 }
 349
 350 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 351 {
 352         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 353         for (size_t i = 0; i < samples.size(); ++i) {
 354                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 355         }
 356
 357         uint32_t frames_written;
 358         HRESULT result;
 359         if (is_master_card) {
 360                 result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 361                         pts, TIMEBASE, &frames_written);
 362         } else {
 363                 result = output->WriteAudioSamplesSync(int_samples.get(), samples.size() / 2,
 364                         &frames_written);
 365         }
 366         if (result != S_OK) {
 367                 fprintf(stderr, "write audio to DeckLink (pts=%" PRId64 ") failed (result=0x%08x)\n", pts, result);
 368         } else {
 369                 // Non-master card is not really synchronized on audio at all, so we don't warn on it.
 370                 if (frames_written != samples.size() / 2 && is_master_card) {
 371                         fprintf(stderr, "write audio to DeckLink returned short write (%u/%zu)\n", frames_written, samples.size() / 2);
 372                 }
 373         }
 374         metric_decklink_output_scheduled_samples += samples.size() / 2;
 375 }
 376
 377 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
 378 {
 379         assert(!should_quit.should_quit());
 380
 381         *dropped_frames = 0;
 382         *frame_duration = this->frame_duration;
 383
 384         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 385         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 386         BMDTimeValue target_time = pts - buffer;
 387
 388         // While prerolling, we send out frames as quickly as we can.
 389         if (target_time < base_pts) {
 390                 *is_preroll = true;
 391                 ++metric_decklink_output_scheduled_frames_preroll;
 392                 return;
 393         }
 394
 395         *is_preroll = !playback_started;
 396
 397         if (!playback_started) {
 398                 if (output->EndAudioPreroll() != S_OK) {
 399                         fprintf(stderr, "Could not end audio preroll\n");
 400                         abort();  // TODO
 401                 }
 402                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 403                         fprintf(stderr, "Could not start playback\n");
 404                         abort();  // TODO
 405                 }
 406                 playback_started = true;
 407         }
 408
 409         BMDTimeValue stream_frame_time;
 410         double playback_speed;
 411         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 412
 413         *frame_timestamp = steady_clock::now() +
 414                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 415
 416         metric_decklink_output_margin_seconds.count_event(
 417                 (target_time - stream_frame_time) / double(TIMEBASE));
 418
 419         // If we're ahead of time, wait for the frame to (approximately) start.
 420         if (stream_frame_time < target_time) {
 421                 should_quit.sleep_until(*frame_timestamp);
 422                 ++metric_decklink_output_scheduled_frames_normal;
 423                 return;
 424         }
 425
 426         // If we overshot the previous frame by just a little,
 427         // fire off one immediately.
 428         if (stream_frame_time < target_time + max_overshoot) {
 429                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 430                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 431                 ++metric_decklink_output_scheduled_frames_late;
 432                 return;
 433         }
 434
 435         // Oops, we missed by more than one frame. Return immediately,
 436         // but drop so that we catch up.
 437         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 438         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
 439         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
 440         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 441         metric_decklink_output_scheduled_frames_dropped += *dropped_frames;
 442         ++metric_decklink_output_scheduled_frames_normal;
 443 }
 444
 445 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
 446 {
 447         if (video_modes.count(mode)) {
 448                 return mode;
 449         }
 450
 451         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
 452         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 1 }, { 30000, 1001 } }) {
 453                 for (const auto &it : video_modes) {
 454                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
 455                                 return it.first;
 456                         }
 457                 }
 458         }
 459
 460         uint32_t best_mode = 0;
 461         double best_fps = 0.0;
 462         for (const auto &it : video_modes) {
 463                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
 464                 if (fps > best_fps) {
 465                         best_mode = it.first;
 466                         best_fps = fps;
 467                 }
 468         }
 469         return best_mode;
 470 }
 471
 472 YCbCrLumaCoefficients DeckLinkOutput::preferred_ycbcr_coefficients() const
 473 {
 474         if (current_mode_flags & bmdDisplayModeColorspaceRec601) {
 475                 return YCBCR_REC_601;
 476         } else {
 477                 // Don't bother checking bmdDisplayModeColorspaceRec709;
 478                 // if none is set, 709 is a good default anyway.
 479                 return YCBCR_REC_709;
 480         }
 481 }
 482
 483 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 484 {
 485         Frame *frame = static_cast<Frame *>(completedFrame);
 486         switch (result) {
 487         case bmdOutputFrameCompleted:
 488                 ++metric_decklink_output_completed_frames_completed;
 489                 break;
 490         case bmdOutputFrameDisplayedLate:
 491                 fprintf(stderr, "Output frame displayed late (pts=%" PRId64 ")\n", frame->pts);
 492                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 493                 ++metric_decklink_output_completed_frames_late;
 494                 break;
 495         case bmdOutputFrameDropped:
 496                 fprintf(stderr, "Output frame was dropped (pts=%" PRId64 ")\n", frame->pts);
 497                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 498                 ++metric_decklink_output_completed_frames_dropped;
 499                 break;
 500         case bmdOutputFrameFlushed:
 501                 fprintf(stderr, "Output frame was flushed (pts=%" PRId64 ")\n", frame->pts);
 502                 ++metric_decklink_output_completed_frames_flushed;
 503                 break;
 504         default:
 505                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 506                 ++metric_decklink_output_completed_frames_unknown;
 507                 break;
 508         }
 509
 510         static int frameno = 0;
 511         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &frameno, &latency_histogram);
 512
 513         {
 514                 lock_guard<mutex> lock(frame_queue_mutex);
 515                 frame_freelist.push(unique_ptr<Frame>(frame));
 516                 assert(scheduled_frames.count(frame));
 517                 scheduled_frames.erase(frame);
 518                 --metric_decklink_output_inflight_frames;
 519         }
 520
 521         return S_OK;
 522 }
 523
 524 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 525 {
 526         printf("playback stopped!\n");
 527         return S_OK;
 528 }
 529
 530 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 531 {
 532         lock_guard<mutex> lock(frame_queue_mutex);
 533
 534         if (!frame_freelist.empty()) {
 535                 unique_ptr<Frame> frame = move(frame_freelist.front());
 536                 frame_freelist.pop();
 537                 return frame;
 538         }
 539
 540         unique_ptr<Frame> frame(new Frame);
 541
 542         size_t stride;
 543         if (global_flags.bit_depth > 8) {
 544                 stride = v210Converter::get_v210_stride(width);
 545                 GLint v210_width = stride / sizeof(uint32_t);
 546                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGB10_A2, v210_width, height);
 547
 548                 // We need valid texture state, or NVIDIA won't allow us to write to the texture.
 549                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 550                 check_error();
 551                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 552                 check_error();
 553         } else {
 554                 stride = width * 2;
 555                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 556         }
 557
 558         glGenBuffers(1, &frame->pbo);
 559         check_error();
 560         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 561         check_error();
 562         glBufferStorage(GL_PIXEL_PACK_BUFFER, stride * height, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 563         check_error();
 564         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, stride * height, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 565         check_error();
 566         frame->uyvy_ptr_local.reset(new uint8_t[stride * height]);
 567         frame->resource_pool = resource_pool;
 568
 569         return frame;
 570 }
 571
 572 void DeckLinkOutput::present_thread_func()
 573 {
 574         pthread_setname_np(pthread_self(), "DeckLinkOutput");
 575         for ( ;; ) {
 576                 unique_ptr<Frame> frame;
 577                 {
 578                         unique_lock<mutex> lock(frame_queue_mutex);
 579                         frame_queues_changed.wait(lock, [this]{
 580                                 return should_quit.should_quit() || !pending_video_frames.empty();
 581                         });
 582                         if (should_quit.should_quit()) {
 583                                 return;
 584                         }
 585                         frame = move(pending_video_frames.front());
 586                         pending_video_frames.pop();
 587                 }
 588
 589                 for ( ;; ) {
 590                         int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
 591                         if (err == GL_TIMEOUT_EXPIRED) {
 592                                 // NVIDIA likes to busy-wait; yield instead.
 593                                 this_thread::sleep_for(milliseconds(1));
 594                         } else {
 595                                 break;
 596                         }
 597                 }
 598                 check_error();
 599                 frame->fence.reset();
 600
 601                 if (global_flags.bit_depth > 8) {
 602                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, v210Converter::get_v210_stride(width) * height);
 603                 } else {
 604                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 605                 }
 606
 607                 // Release any input frames we needed to render this frame.
 608                 frame->input_frames.clear();
 609
 610                 if (is_master_card) {
 611                         BMDTimeValue pts = frame->pts;
 612                         BMDTimeValue duration = frame->duration;
 613                         HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 614                         lock_guard<mutex> lock(frame_queue_mutex);
 615                         if (res == S_OK) {
 616                                 scheduled_frames.insert(frame.release());  // Owned by the driver now.
 617                                 ++metric_decklink_output_inflight_frames;
 618                         } else {
 619                                 fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 620
 621                                 frame_freelist.push(move(frame));
 622                         }
 623                 } else {
 624                         HRESULT res = output->DisplayVideoFrameSync(frame.get());
 625                         if (res != S_OK) {
 626                                 fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 627                         }
 628                         frame_freelist.push(move(frame));
 629                 }
 630         }
 631 }
 632
 633 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 634 {
 635         return E_NOINTERFACE;
 636 }
 637
 638 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 639 {
 640         return refcount.fetch_add(1) + 1;
 641 }
 642
 643 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 644 {
 645         int new_ref = refcount.fetch_sub(1) - 1;
 646         if (new_ref == 0)
 647                 delete this;
 648         return new_ref;
 649 }
 650
 651 DeckLinkOutput::Frame::~Frame()
 652 {
 653         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 654         check_error();
 655         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 656         check_error();
 657         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 658         check_error();
 659         glDeleteBuffers(1, &pbo);
 660         check_error();
 661         resource_pool->release_2d_texture(uyvy_tex);
 662         check_error();
 663 }
 664
 665 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 666 {
 667         return E_NOINTERFACE;
 668 }
 669
 670 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 671 {
 672         return refcount.fetch_add(1) + 1;
 673 }
 674
 675 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 676 {
 677         int new_ref = refcount.fetch_sub(1) - 1;
 678         if (new_ref == 0)
 679                 delete this;
 680         return new_ref;
 681 }
 682
 683 long DeckLinkOutput::Frame::GetWidth()
 684 {
 685         return global_flags.width;
 686 }
 687
 688 long DeckLinkOutput::Frame::GetHeight()
 689 {
 690         return global_flags.height;
 691 }
 692
 693 long DeckLinkOutput::Frame::GetRowBytes()
 694 {
 695         if (global_flags.bit_depth > 8) {
 696                 return v210Converter::get_v210_stride(global_flags.width);
 697         } else {
 698                 return global_flags.width * 2;
 699         }
 700 }
 701
 702 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 703 {
 704         if (global_flags.bit_depth > 8) {
 705                 return bmdFormat10BitYUV;
 706         } else {
 707                 return bmdFormat8BitYUV;
 708         }
 709 }
 710
 711 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 712 {
 713         return bmdFrameFlagDefault;
 714 }
 715
 716 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 717 {
 718         *buffer = uyvy_ptr_local.get();
 719         return S_OK;
 720 }
 721
 722 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 723 {
 724         fprintf(stderr, "STUB: GetTimecode()\n");
 725         return E_NOTIMPL;
 726 }
 727
 728 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 729 {
 730         fprintf(stderr, "STUB: GetAncillaryData()\n");
 731         return E_NOTIMPL;
 732 }