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 bool DeckLinkOutput::set_device(IDeckLink *decklink)
  86 {
  87         if (decklink->QueryInterface(IID_IDeckLinkOutput, (void**)&output) != S_OK) {
  88                 fprintf(stderr, "Warning: Card %u has no outputs\n", card_index);
  89                 return false;
  90         }
  91
  92         IDeckLinkDisplayModeIterator *mode_it;
  93         if (output->GetDisplayModeIterator(&mode_it) != S_OK) {
  94                 fprintf(stderr, "Warning: Failed to enumerate output display modes for card %u\n", card_index);
  95                 return false;
  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         return true;
 122 }
 123
 124 void DeckLinkOutput::start_output(uint32_t mode, int64_t base_pts)
 125 {
 126         assert(output);
 127         assert(!playback_initiated);
 128
 129         if (video_modes.empty()) {
 130                 fprintf(stderr, "ERROR: No matching output modes for %dx%d found\n", width, height);
 131                 exit(1);
 132         }
 133
 134         should_quit.unquit();
 135         playback_initiated = true;
 136         playback_started = false;
 137         this->base_pts = base_pts;
 138
 139         IDeckLinkConfiguration *config = nullptr;
 140         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
 141                 fprintf(stderr, "Failed to get configuration interface for output card\n");
 142                 exit(1);
 143         }
 144         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
 145                 fprintf(stderr, "Failed to set low latency output\n");
 146                 exit(1);
 147         }
 148         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
 149                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
 150                 exit(1);
 151         }
 152         if (config->SetFlag(bmdDeckLinkConfigUse1080pNotPsF, true) != S_OK) {
 153                 fprintf(stderr, "Failed to set PsF flag for card\n");
 154                 exit(1);
 155         }
 156         if (config->SetFlag(bmdDeckLinkConfigSMPTELevelAOutput, true) != S_OK) {
 157                 // This affects at least some no-name SDI->HDMI converters.
 158                 // Warn, but don't die.
 159                 fprintf(stderr, "WARNING: Failed to enable SMTPE Level A; resolutions like 1080p60 might have issues.\n");
 160         }
 161
 162         BMDDisplayModeSupport support;
 163         IDeckLinkDisplayMode *display_mode;
 164         BMDPixelFormat pixel_format = global_flags.ten_bit_output ? bmdFormat10BitYUV : bmdFormat8BitYUV;
 165         if (output->DoesSupportVideoMode(mode, pixel_format, bmdVideoOutputFlagDefault,
 166                                          &support, &display_mode) != S_OK) {
 167                 fprintf(stderr, "Couldn't ask for format support\n");
 168                 exit(1);
 169         }
 170
 171         if (support == bmdDisplayModeNotSupported) {
 172                 fprintf(stderr, "Requested display mode not supported\n");
 173                 exit(1);
 174         }
 175
 176         current_mode_flags = display_mode->GetFlags();
 177
 178         BMDTimeValue time_value;
 179         BMDTimeScale time_scale;
 180         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
 181                 fprintf(stderr, "Couldn't get frame rate\n");
 182                 exit(1);
 183         }
 184
 185         metric_decklink_output_width_pixels = width;
 186         metric_decklink_output_height_pixels = height;
 187         metric_decklink_output_frame_rate_nom = time_value;
 188         metric_decklink_output_frame_rate_den = time_scale;
 189
 190         frame_duration = time_value * TIMEBASE / time_scale;
 191
 192         display_mode->Release();
 193
 194         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
 195         if (result != S_OK) {
 196                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
 197                 exit(1);
 198         }
 199         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
 200                 fprintf(stderr, "Couldn't set callback\n");
 201                 exit(1);
 202         }
 203         assert(OUTPUT_FREQUENCY == 48000);
 204         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
 205                 fprintf(stderr, "Couldn't enable audio output\n");
 206                 exit(1);
 207         }
 208         if (output->BeginAudioPreroll() != S_OK) {
 209                 fprintf(stderr, "Couldn't begin audio preroll\n");
 210                 exit(1);
 211         }
 212
 213         present_thread = thread([this]{
 214                 QOpenGLContext *context = create_context(this->surface);
 215                 eglBindAPI(EGL_OPENGL_API);
 216                 if (!make_current(context, this->surface)) {
 217                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
 218                                 eglGetError());
 219                         exit(1);
 220                 }
 221                 present_thread_func();
 222                 delete_context(context);
 223         });
 224 }
 225
 226 void DeckLinkOutput::end_output()
 227 {
 228         if (!playback_initiated) {
 229                 return;
 230         }
 231
 232         should_quit.quit();
 233         frame_queues_changed.notify_all();
 234         present_thread.join();
 235         playback_initiated = false;
 236
 237         output->StopScheduledPlayback(0, nullptr, 0);
 238         output->DisableVideoOutput();
 239         output->DisableAudioOutput();
 240
 241         // Wait until all frames are accounted for, and free them.
 242         {
 243                 unique_lock<mutex> lock(frame_queue_mutex);
 244                 while (!(frame_freelist.empty() && num_frames_in_flight == 0)) {
 245                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
 246                         frame_freelist.pop();
 247                 }
 248         }
 249 }
 250
 251 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)
 252 {
 253         assert(!should_quit.should_quit());
 254
 255         if ((current_mode_flags & bmdDisplayModeColorspaceRec601) && output_ycbcr_coefficients == YCBCR_REC_709) {
 256                 if (!last_frame_had_mode_mismatch) {
 257                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 258                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 259                 }
 260                 last_frame_had_mode_mismatch = true;
 261                 ++metric_decklink_output_color_mismatch_frames;
 262         } else if ((current_mode_flags & bmdDisplayModeColorspaceRec709) && output_ycbcr_coefficients == YCBCR_REC_601) {
 263                 if (!last_frame_had_mode_mismatch) {
 264                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 265                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 266                 }
 267                 last_frame_had_mode_mismatch = true;
 268                 ++metric_decklink_output_color_mismatch_frames;
 269         } else {
 270                 last_frame_had_mode_mismatch = false;
 271         }
 272
 273         unique_ptr<Frame> frame = get_frame();
 274         if (global_flags.ten_bit_output) {
 275                 chroma_subsampler->create_v210(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 276         } else {
 277                 chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 278         }
 279
 280         // Download the UYVY texture to the PBO.
 281         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 282         check_error();
 283
 284         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 285         check_error();
 286
 287         if (global_flags.ten_bit_output) {
 288                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 289                 check_error();
 290                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, BUFFER_OFFSET(0));
 291                 check_error();
 292         } else {
 293                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 294                 check_error();
 295                 glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 296                 check_error();
 297         }
 298
 299         glBindTexture(GL_TEXTURE_2D, 0);
 300         check_error();
 301         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 302         check_error();
 303
 304         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 305         check_error();
 306
 307         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 308         check_error();
 309         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 310         check_error();
 311
 312         frame->input_frames = input_frames;
 313         frame->received_ts = find_received_timestamp(input_frames);
 314         frame->pts = pts;
 315         frame->duration = duration;
 316
 317         {
 318                 unique_lock<mutex> lock(frame_queue_mutex);
 319                 pending_video_frames.push(move(frame));
 320         }
 321         frame_queues_changed.notify_all();
 322 }
 323
 324 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 325 {
 326         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 327         for (size_t i = 0; i < samples.size(); ++i) {
 328                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 329         }
 330
 331         uint32_t frames_written;
 332         HRESULT result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 333                 pts, TIMEBASE, &frames_written);
 334         if (result != S_OK) {
 335                 fprintf(stderr, "ScheduleAudioSamples(pts=%ld) failed (result=0x%08x)\n", pts, result);
 336         } else {
 337                 if (frames_written != samples.size() / 2) {
 338                         fprintf(stderr, "ScheduleAudioSamples() returned short write (%u/%ld)\n", frames_written, samples.size() / 2);
 339                 }
 340         }
 341         metric_decklink_output_scheduled_samples += samples.size() / 2;
 342 }
 343
 344 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
 345 {
 346         assert(!should_quit.should_quit());
 347
 348         *dropped_frames = 0;
 349         *frame_duration = this->frame_duration;
 350
 351         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 352         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 353         BMDTimeValue target_time = pts - buffer;
 354
 355         // While prerolling, we send out frames as quickly as we can.
 356         if (target_time < base_pts) {
 357                 *is_preroll = true;
 358                 ++metric_decklink_output_scheduled_frames_preroll;
 359                 return;
 360         }
 361
 362         *is_preroll = !playback_started;
 363
 364         if (!playback_started) {
 365                 if (output->EndAudioPreroll() != S_OK) {
 366                         fprintf(stderr, "Could not end audio preroll\n");
 367                         exit(1);  // TODO
 368                 }
 369                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 370                         fprintf(stderr, "Could not start playback\n");
 371                         exit(1);  // TODO
 372                 }
 373                 playback_started = true;
 374         }
 375
 376         BMDTimeValue stream_frame_time;
 377         double playback_speed;
 378         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 379
 380         *frame_timestamp = steady_clock::now() +
 381                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 382
 383         metric_decklink_output_margin_seconds.count_event(
 384                 (target_time - stream_frame_time) / double(TIMEBASE));
 385
 386         // If we're ahead of time, wait for the frame to (approximately) start.
 387         if (stream_frame_time < target_time) {
 388                 should_quit.sleep_until(*frame_timestamp);
 389                 ++metric_decklink_output_scheduled_frames_normal;
 390                 return;
 391         }
 392
 393         // If we overshot the previous frame by just a little,
 394         // fire off one immediately.
 395         if (stream_frame_time < target_time + max_overshoot) {
 396                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 397                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 398                 ++metric_decklink_output_scheduled_frames_late;
 399                 return;
 400         }
 401
 402         // Oops, we missed by more than one frame. Return immediately,
 403         // but drop so that we catch up.
 404         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 405         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
 406         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
 407         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 408         metric_decklink_output_scheduled_frames_dropped += *dropped_frames;
 409         ++metric_decklink_output_scheduled_frames_normal;
 410 }
 411
 412 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
 413 {
 414         if (video_modes.count(mode)) {
 415                 return mode;
 416         }
 417
 418         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
 419         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 0 }, { 30000, 1001 } }) {
 420                 for (const auto &it : video_modes) {
 421                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
 422                                 return it.first;
 423                         }
 424                 }
 425         }
 426
 427         uint32_t best_mode = 0;
 428         double best_fps = 0.0;
 429         for (const auto &it : video_modes) {
 430                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
 431                 if (fps > best_fps) {
 432                         best_mode = it.first;
 433                         best_fps = fps;
 434                 }
 435         }
 436         return best_mode;
 437 }
 438
 439 YCbCrLumaCoefficients DeckLinkOutput::preferred_ycbcr_coefficients() const
 440 {
 441         if (current_mode_flags & bmdDisplayModeColorspaceRec601) {
 442                 return YCBCR_REC_601;
 443         } else {
 444                 // Don't bother checking bmdDisplayModeColorspaceRec709;
 445                 // if none is set, 709 is a good default anyway.
 446                 return YCBCR_REC_709;
 447         }
 448 }
 449
 450 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 451 {
 452         Frame *frame = static_cast<Frame *>(completedFrame);
 453         switch (result) {
 454         case bmdOutputFrameCompleted:
 455                 ++metric_decklink_output_completed_frames_completed;
 456                 break;
 457         case bmdOutputFrameDisplayedLate:
 458                 fprintf(stderr, "Output frame displayed late (pts=%ld)\n", frame->pts);
 459                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 460                 ++metric_decklink_output_completed_frames_late;
 461                 break;
 462         case bmdOutputFrameDropped:
 463                 fprintf(stderr, "Output frame was dropped (pts=%ld)\n", frame->pts);
 464                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 465                 ++metric_decklink_output_completed_frames_dropped;
 466                 break;
 467         case bmdOutputFrameFlushed:
 468                 fprintf(stderr, "Output frame was flushed (pts=%ld)\n", frame->pts);
 469                 ++metric_decklink_output_completed_frames_flushed;
 470                 break;
 471         default:
 472                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 473                 ++metric_decklink_output_completed_frames_unknown;
 474                 break;
 475         }
 476
 477         static int frameno = 0;
 478         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &frameno, &latency_histogram);
 479
 480         {
 481                 lock_guard<mutex> lock(frame_queue_mutex);
 482                 frame_freelist.push(unique_ptr<Frame>(frame));
 483                 --num_frames_in_flight;
 484                 --metric_decklink_output_inflight_frames;
 485         }
 486
 487         return S_OK;
 488 }
 489
 490 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 491 {
 492         printf("playback stopped!\n");
 493         return S_OK;
 494 }
 495
 496 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 497 {
 498         lock_guard<mutex> lock(frame_queue_mutex);
 499
 500         if (!frame_freelist.empty()) {
 501                 unique_ptr<Frame> frame = move(frame_freelist.front());
 502                 frame_freelist.pop();
 503                 return frame;
 504         }
 505
 506         unique_ptr<Frame> frame(new Frame);
 507
 508         size_t stride;
 509         if (global_flags.ten_bit_output) {
 510                 stride = v210Converter::get_v210_stride(width);
 511                 GLint v210_width = stride / sizeof(uint32_t);
 512                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGB10_A2, v210_width, height);
 513
 514                 // We need valid texture state, or NVIDIA won't allow us to write to the texture.
 515                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 516                 check_error();
 517                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 518                 check_error();
 519         } else {
 520                 stride = width * 2;
 521                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 522         }
 523
 524         glGenBuffers(1, &frame->pbo);
 525         check_error();
 526         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 527         check_error();
 528         glBufferStorage(GL_PIXEL_PACK_BUFFER, stride * height, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 529         check_error();
 530         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, stride * height, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 531         check_error();
 532         frame->uyvy_ptr_local.reset(new uint8_t[stride * height]);
 533         frame->resource_pool = resource_pool;
 534
 535         return frame;
 536 }
 537
 538 void DeckLinkOutput::present_thread_func()
 539 {
 540         pthread_setname_np(pthread_self(), "DeckLinkOutput");
 541         for ( ;; ) {
 542                 unique_ptr<Frame> frame;
 543                 {
 544                         unique_lock<mutex> lock(frame_queue_mutex);
 545                         frame_queues_changed.wait(lock, [this]{
 546                                 return should_quit.should_quit() || !pending_video_frames.empty();
 547                         });
 548                         if (should_quit.should_quit()) {
 549                                 return;
 550                         }
 551                         frame = move(pending_video_frames.front());
 552                         pending_video_frames.pop();
 553                         ++num_frames_in_flight;
 554                         ++metric_decklink_output_inflight_frames;
 555                 }
 556
 557                 for ( ;; ) {
 558                         int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
 559                         if (err == GL_TIMEOUT_EXPIRED) {
 560                                 // NVIDIA likes to busy-wait; yield instead.
 561                                 this_thread::sleep_for(milliseconds(1));
 562                         } else {
 563                                 break;
 564                         }
 565                 }
 566                 check_error();
 567                 frame->fence.reset();
 568
 569                 if (global_flags.ten_bit_output) {
 570                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, v210Converter::get_v210_stride(width) * height);
 571                 } else {
 572                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 573                 }
 574
 575                 // Release any input frames we needed to render this frame.
 576                 frame->input_frames.clear();
 577
 578                 BMDTimeValue pts = frame->pts;
 579                 BMDTimeValue duration = frame->duration;
 580                 HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 581                 if (res == S_OK) {
 582                         frame.release();  // Owned by the driver now.
 583                 } else {
 584                         fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 585
 586                         lock_guard<mutex> lock(frame_queue_mutex);
 587                         frame_freelist.push(move(frame));
 588                         --num_frames_in_flight;
 589                         --metric_decklink_output_inflight_frames;
 590                 }
 591         }
 592 }
 593
 594 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 595 {
 596         return E_NOINTERFACE;
 597 }
 598
 599 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 600 {
 601         return refcount.fetch_add(1) + 1;
 602 }
 603
 604 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 605 {
 606         int new_ref = refcount.fetch_sub(1) - 1;
 607         if (new_ref == 0)
 608                 delete this;
 609         return new_ref;
 610 }
 611
 612 DeckLinkOutput::Frame::~Frame()
 613 {
 614         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 615         check_error();
 616         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 617         check_error();
 618         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 619         check_error();
 620         glDeleteBuffers(1, &pbo);
 621         check_error();
 622         resource_pool->release_2d_texture(uyvy_tex);
 623         check_error();
 624 }
 625
 626 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 627 {
 628         return E_NOINTERFACE;
 629 }
 630
 631 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 632 {
 633         return refcount.fetch_add(1) + 1;
 634 }
 635
 636 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 637 {
 638         int new_ref = refcount.fetch_sub(1) - 1;
 639         if (new_ref == 0)
 640                 delete this;
 641         return new_ref;
 642 }
 643
 644 long DeckLinkOutput::Frame::GetWidth()
 645 {
 646         return global_flags.width;
 647 }
 648
 649 long DeckLinkOutput::Frame::GetHeight()
 650 {
 651         return global_flags.height;
 652 }
 653
 654 long DeckLinkOutput::Frame::GetRowBytes()
 655 {
 656         if (global_flags.ten_bit_output) {
 657                 return v210Converter::get_v210_stride(global_flags.width);
 658         } else {
 659                 return global_flags.width * 2;
 660         }
 661 }
 662
 663 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 664 {
 665         if (global_flags.ten_bit_output) {
 666                 return bmdFormat10BitYUV;
 667         } else {
 668                 return bmdFormat8BitYUV;
 669         }
 670 }
 671
 672 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 673 {
 674         return bmdFrameFlagDefault;
 675 }
 676
 677 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 678 {
 679         *buffer = uyvy_ptr_local.get();
 680         return S_OK;
 681 }
 682
 683 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 684 {
 685         fprintf(stderr, "STUB: GetTimecode()\n");
 686         return E_NOTIMPL;
 687 }
 688
 689 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 690 {
 691         fprintf(stderr, "STUB: GetAncillaryData()\n");
 692         return E_NOTIMPL;
 693 }