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