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

   1 #include <movit/effect_util.h>
   2 #include <movit/util.h>
   3 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
   4 #include <pthread.h>
   5 #include <unistd.h>
   6
   7 #include <mutex>
   8
   9 #include <epoxy/egl.h>
  10
  11 #include "chroma_subsampler.h"
  12 #include "decklink_output.h"
  13 #include "decklink_util.h"
  14 #include "flags.h"
  15 #include "shared/metrics.h"
  16 #include "print_latency.h"
  17 #include "shared/timebase.h"
  18 #include "v210_converter.h"
  19
  20 using namespace movit;
  21 using namespace std;
  22 using namespace std::chrono;
  23
  24 namespace {
  25
  26 // This class can be deleted during regular use, so make all the metrics static.
  27 once_flag decklink_metrics_inited;
  28 LatencyHistogram latency_histogram;
  29 atomic<int64_t> metric_decklink_output_width_pixels{-1};
  30 atomic<int64_t> metric_decklink_output_height_pixels{-1};
  31 atomic<int64_t> metric_decklink_output_frame_rate_den{-1};
  32 atomic<int64_t> metric_decklink_output_frame_rate_nom{-1};
  33 atomic<int64_t> metric_decklink_output_inflight_frames{0};
  34 atomic<int64_t> metric_decklink_output_color_mismatch_frames{0};
  35
  36 atomic<int64_t> metric_decklink_output_scheduled_frames_dropped{0};
  37 atomic<int64_t> metric_decklink_output_scheduled_frames_late{0};
  38 atomic<int64_t> metric_decklink_output_scheduled_frames_normal{0};
  39 atomic<int64_t> metric_decklink_output_scheduled_frames_preroll{0};
  40
  41 atomic<int64_t> metric_decklink_output_completed_frames_completed{0};
  42 atomic<int64_t> metric_decklink_output_completed_frames_dropped{0};
  43 atomic<int64_t> metric_decklink_output_completed_frames_flushed{0};
  44 atomic<int64_t> metric_decklink_output_completed_frames_late{0};
  45 atomic<int64_t> metric_decklink_output_completed_frames_unknown{0};
  46
  47 atomic<int64_t> metric_decklink_output_scheduled_samples{0};
  48
  49 Summary metric_decklink_output_margin_seconds;
  50
  51 }  // namespace
  52
  53 DeckLinkOutput::DeckLinkOutput(ResourcePool *resource_pool, QSurface *surface, unsigned width, unsigned height, unsigned card_index)
  54         : resource_pool(resource_pool), surface(surface), width(width), height(height), card_index(card_index)
  55 {
  56         chroma_subsampler.reset(new ChromaSubsampler(resource_pool));
  57
  58         call_once(decklink_metrics_inited, [](){
  59                 latency_histogram.init("decklink_output");
  60                 global_metrics.add("decklink_output_width_pixels", &metric_decklink_output_width_pixels, Metrics::TYPE_GAUGE);
  61                 global_metrics.add("decklink_output_height_pixels", &metric_decklink_output_height_pixels, Metrics::TYPE_GAUGE);
  62                 global_metrics.add("decklink_output_frame_rate_den", &metric_decklink_output_frame_rate_den, Metrics::TYPE_GAUGE);
  63                 global_metrics.add("decklink_output_frame_rate_nom", &metric_decklink_output_frame_rate_nom, Metrics::TYPE_GAUGE);
  64                 global_metrics.add("decklink_output_inflight_frames", &metric_decklink_output_inflight_frames, Metrics::TYPE_GAUGE);
  65                 global_metrics.add("decklink_output_color_mismatch_frames", &metric_decklink_output_color_mismatch_frames);
  66
  67                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "dropped" }}, &metric_decklink_output_scheduled_frames_dropped);
  68                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "late" }}, &metric_decklink_output_scheduled_frames_late);
  69                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "normal" }}, &metric_decklink_output_scheduled_frames_normal);
  70                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "preroll" }}, &metric_decklink_output_scheduled_frames_preroll);
  71
  72                 global_metrics.add("decklink_output_completed_frames", {{ "status", "completed" }}, &metric_decklink_output_completed_frames_completed);
  73                 global_metrics.add("decklink_output_completed_frames", {{ "status", "dropped" }}, &metric_decklink_output_completed_frames_dropped);
  74                 global_metrics.add("decklink_output_completed_frames", {{ "status", "flushed" }}, &metric_decklink_output_completed_frames_flushed);
  75                 global_metrics.add("decklink_output_completed_frames", {{ "status", "late" }}, &metric_decklink_output_completed_frames_late);
  76                 global_metrics.add("decklink_output_completed_frames", {{ "status", "unknown" }}, &metric_decklink_output_completed_frames_unknown);
  77
  78                 global_metrics.add("decklink_output_scheduled_samples", &metric_decklink_output_scheduled_samples);
  79                 vector<double> quantiles{0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99};
  80                 metric_decklink_output_margin_seconds.init(quantiles, 60.0);
  81                 global_metrics.add("decklink_output_margin_seconds", &metric_decklink_output_margin_seconds);
  82         });
  83 }
  84
  85 bool DeckLinkOutput::set_device(IDeckLink *decklink, IDeckLinkInput *input_arg)
  86 {
  87         input = input_arg;
  88         if (decklink->QueryInterface(IID_IDeckLinkOutput, (void**)&output) != S_OK) {
  89                 fprintf(stderr, "Warning: Card %u has no outputs\n", card_index);
  90                 return false;
  91         }
  92
  93         IDeckLinkDisplayModeIterator *mode_it;
  94         if (output->GetDisplayModeIterator(&mode_it) != S_OK) {
  95                 fprintf(stderr, "Warning: Failed to enumerate output display modes for card %u\n", card_index);
  96                 return false;
  97         }
  98
  99         video_modes.clear();
 100
 101         for (const auto &it : summarize_video_modes(mode_it, card_index)) {
 102                 if (it.second.width != width || it.second.height != height) {
 103                         continue;
 104                 }
 105
 106                 // We could support interlaced modes, but let's stay out of it for now,
 107                 // since we don't have interlaced stream output.
 108                 if (it.second.interlaced) {
 109                         continue;
 110                 }
 111
 112                 video_modes.insert(it);
 113         }
 114
 115         mode_it->Release();
 116
 117         // HDMI or SDI generally mean “both HDMI and SDI at the same time” on DeckLink cards
 118         // that support both; pick_default_video_connection() will generally pick one of those
 119         // if they exist. (--prefer-hdmi-input would also affect the selection despite the name
 120         // of the option, but since either generally means both, it's inconsequential.)
 121         // We're not very likely to need analog outputs, so we don't need a way to change
 122         // beyond that.
 123         video_connection = pick_default_video_connection(decklink, BMDDeckLinkVideoOutputConnections, card_index);
 124         return true;
 125 }
 126
 127 void DeckLinkOutput::start_output(uint32_t mode, int64_t base_pts)
 128 {
 129         assert(output);
 130         assert(!playback_initiated);
 131
 132         if (video_modes.empty()) {
 133                 fprintf(stderr, "ERROR: No matching output modes for %dx%d found\n", width, height);
 134                 abort();
 135         }
 136
 137         should_quit.unquit();
 138         playback_initiated = true;
 139         playback_started = false;
 140         this->base_pts = base_pts;
 141
 142         IDeckLinkConfiguration *config = nullptr;
 143         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
 144                 fprintf(stderr, "Failed to get configuration interface for output card\n");
 145                 abort();
 146         }
 147         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
 148                 fprintf(stderr, "Failed to set low latency output\n");
 149                 abort();
 150         }
 151         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
 152                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
 153                 abort();
 154         }
 155         if (config->SetFlag(bmdDeckLinkConfigOutput1080pAsPsF, true) != S_OK) {
 156                 fprintf(stderr, "Failed to set PsF flag for card\n");
 157                 abort();
 158         }
 159         if (config->SetFlag(bmdDeckLinkConfigSMPTELevelAOutput, true) != S_OK) {
 160                 // This affects at least some no-name SDI->HDMI converters.
 161                 // Warn, but don't die.
 162                 fprintf(stderr, "WARNING: Failed to enable SMTPE Level A; resolutions like 1080p60 might have issues.\n");
 163         }
 164
 165         BMDDisplayModeSupport support;
 166         IDeckLinkDisplayMode *display_mode;
 167         BMDPixelFormat pixel_format = global_flags.ten_bit_output ? bmdFormat10BitYUV : bmdFormat8BitYUV;
 168         if (output->DoesSupportVideoMode(mode, pixel_format, bmdVideoOutputFlagDefault,
 169                                          &support, &display_mode) != S_OK) {
 170                 fprintf(stderr, "Couldn't ask for format support\n");
 171                 abort();
 172         }
 173
 174         if (support == bmdDisplayModeNotSupported) {
 175                 fprintf(stderr, "Requested display mode not supported\n");
 176                 abort();
 177         }
 178
 179         current_mode_flags = display_mode->GetFlags();
 180
 181         BMDTimeValue time_value;
 182         BMDTimeScale time_scale;
 183         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
 184                 fprintf(stderr, "Couldn't get frame rate\n");
 185                 abort();
 186         }
 187
 188         metric_decklink_output_width_pixels = width;
 189         metric_decklink_output_height_pixels = height;
 190         metric_decklink_output_frame_rate_nom = time_value;
 191         metric_decklink_output_frame_rate_den = time_scale;
 192
 193         frame_duration = time_value * TIMEBASE / time_scale;
 194
 195         display_mode->Release();
 196
 197         if (input != nullptr) {
 198                 if (input->DisableVideoInput() != S_OK) {
 199                         fprintf(stderr, "Warning: Failed to disable video input for card %d\n", card_index);
 200                 }
 201                 if (input->DisableAudioInput() != S_OK) {
 202                         fprintf(stderr, "Warning: Failed to disable audio input for card %d\n", card_index);
 203                 }
 204         }
 205
 206         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
 207         if (result != S_OK) {
 208                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
 209                 abort();
 210         }
 211         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
 212                 fprintf(stderr, "Couldn't set callback\n");
 213                 abort();
 214         }
 215         assert(OUTPUT_FREQUENCY == 48000);
 216         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
 217                 fprintf(stderr, "Couldn't enable audio output\n");
 218                 abort();
 219         }
 220         if (output->BeginAudioPreroll() != S_OK) {
 221                 fprintf(stderr, "Couldn't begin audio preroll\n");
 222                 abort();
 223         }
 224
 225         present_thread = thread([this]{
 226                 QOpenGLContext *context = create_context(this->surface);
 227                 eglBindAPI(EGL_OPENGL_API);
 228                 if (!make_current(context, this->surface)) {
 229                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
 230                                 eglGetError());
 231                         abort();
 232                 }
 233                 present_thread_func();
 234                 delete_context(context);
 235         });
 236 }
 237
 238 void DeckLinkOutput::end_output()
 239 {
 240         if (!playback_initiated) {
 241                 return;
 242         }
 243
 244         should_quit.quit();
 245         frame_queues_changed.notify_all();
 246         present_thread.join();
 247         playback_initiated = false;
 248
 249         output->StopScheduledPlayback(0, nullptr, 0);
 250         output->DisableVideoOutput();
 251         output->DisableAudioOutput();
 252
 253         // Wait until all frames are accounted for, and free them.
 254         {
 255                 unique_lock<mutex> lock(frame_queue_mutex);
 256                 while (!(frame_freelist.empty() && scheduled_frames.empty())) {
 257                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
 258                         frame_freelist.pop();
 259                 }
 260         }
 261
 262         if (input != nullptr) {
 263                 input->Release();
 264                 input = nullptr;
 265         }
 266         if (output != nullptr) {
 267                 output->Release();
 268                 output = nullptr;
 269         }
 270 }
 271
 272 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)
 273 {
 274         assert(!should_quit.should_quit());
 275
 276         if ((current_mode_flags & bmdDisplayModeColorspaceRec601) && output_ycbcr_coefficients == YCBCR_REC_709) {
 277                 if (!last_frame_had_mode_mismatch) {
 278                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 279                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 280                 }
 281                 last_frame_had_mode_mismatch = true;
 282                 ++metric_decklink_output_color_mismatch_frames;
 283         } else if ((current_mode_flags & bmdDisplayModeColorspaceRec709) && output_ycbcr_coefficients == YCBCR_REC_601) {
 284                 if (!last_frame_had_mode_mismatch) {
 285                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 286                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 287                 }
 288                 last_frame_had_mode_mismatch = true;
 289                 ++metric_decklink_output_color_mismatch_frames;
 290         } else {
 291                 last_frame_had_mode_mismatch = false;
 292         }
 293
 294         unique_ptr<Frame> frame = get_frame();
 295         if (global_flags.ten_bit_output) {
 296                 chroma_subsampler->create_v210(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 297         } else {
 298                 chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 299         }
 300
 301         // Download the UYVY texture to the PBO.
 302         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 303         check_error();
 304
 305         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 306         check_error();
 307
 308         if (global_flags.ten_bit_output) {
 309                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 310                 check_error();
 311                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, BUFFER_OFFSET(0));
 312                 check_error();
 313         } else {
 314                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 315                 check_error();
 316                 glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 317                 check_error();
 318         }
 319
 320         glBindTexture(GL_TEXTURE_2D, 0);
 321         check_error();
 322         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 323         check_error();
 324
 325         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 326         check_error();
 327
 328         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 329         check_error();
 330         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 331         check_error();
 332
 333         frame->input_frames = input_frames;
 334         frame->received_ts = find_received_timestamp(input_frames);
 335         frame->pts = pts;
 336         frame->duration = duration;
 337
 338         {
 339                 lock_guard<mutex> lock(frame_queue_mutex);
 340                 pending_video_frames.push(move(frame));
 341         }
 342         frame_queues_changed.notify_all();
 343 }
 344
 345 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 346 {
 347         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 348         for (size_t i = 0; i < samples.size(); ++i) {
 349                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 350         }
 351
 352         uint32_t frames_written;
 353         HRESULT result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 354                 pts, TIMEBASE, &frames_written);
 355         if (result != S_OK) {
 356                 fprintf(stderr, "ScheduleAudioSamples(pts=%" PRId64 ") failed (result=0x%08x)\n", pts, result);
 357         } else {
 358                 if (frames_written != samples.size() / 2) {
 359                         fprintf(stderr, "ScheduleAudioSamples() returned short write (%u/%zu)\n", frames_written, samples.size() / 2);
 360                 }
 361         }
 362         metric_decklink_output_scheduled_samples += samples.size() / 2;
 363 }
 364
 365 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
 366 {
 367         assert(!should_quit.should_quit());
 368
 369         *dropped_frames = 0;
 370         *frame_duration = this->frame_duration;
 371
 372         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 373         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 374         BMDTimeValue target_time = pts - buffer;
 375
 376         // While prerolling, we send out frames as quickly as we can.
 377         if (target_time < base_pts) {
 378                 *is_preroll = true;
 379                 ++metric_decklink_output_scheduled_frames_preroll;
 380                 return;
 381         }
 382
 383         *is_preroll = !playback_started;
 384
 385         if (!playback_started) {
 386                 if (output->EndAudioPreroll() != S_OK) {
 387                         fprintf(stderr, "Could not end audio preroll\n");
 388                         abort();  // TODO
 389                 }
 390                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 391                         fprintf(stderr, "Could not start playback\n");
 392                         abort();  // TODO
 393                 }
 394                 playback_started = true;
 395         }
 396
 397         BMDTimeValue stream_frame_time;
 398         double playback_speed;
 399         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 400
 401         *frame_timestamp = steady_clock::now() +
 402                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 403
 404         metric_decklink_output_margin_seconds.count_event(
 405                 (target_time - stream_frame_time) / double(TIMEBASE));
 406
 407         // If we're ahead of time, wait for the frame to (approximately) start.
 408         if (stream_frame_time < target_time) {
 409                 should_quit.sleep_until(*frame_timestamp);
 410                 ++metric_decklink_output_scheduled_frames_normal;
 411                 return;
 412         }
 413
 414         // If we overshot the previous frame by just a little,
 415         // fire off one immediately.
 416         if (stream_frame_time < target_time + max_overshoot) {
 417                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 418                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 419                 ++metric_decklink_output_scheduled_frames_late;
 420                 return;
 421         }
 422
 423         // Oops, we missed by more than one frame. Return immediately,
 424         // but drop so that we catch up.
 425         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 426         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
 427         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
 428         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 429         metric_decklink_output_scheduled_frames_dropped += *dropped_frames;
 430         ++metric_decklink_output_scheduled_frames_normal;
 431 }
 432
 433 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
 434 {
 435         if (video_modes.count(mode)) {
 436                 return mode;
 437         }
 438
 439         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
 440         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 1 }, { 30000, 1001 } }) {
 441                 for (const auto &it : video_modes) {
 442                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
 443                                 return it.first;
 444                         }
 445                 }
 446         }
 447
 448         uint32_t best_mode = 0;
 449         double best_fps = 0.0;
 450         for (const auto &it : video_modes) {
 451                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
 452                 if (fps > best_fps) {
 453                         best_mode = it.first;
 454                         best_fps = fps;
 455                 }
 456         }
 457         return best_mode;
 458 }
 459
 460 YCbCrLumaCoefficients DeckLinkOutput::preferred_ycbcr_coefficients() const
 461 {
 462         if (current_mode_flags & bmdDisplayModeColorspaceRec601) {
 463                 return YCBCR_REC_601;
 464         } else {
 465                 // Don't bother checking bmdDisplayModeColorspaceRec709;
 466                 // if none is set, 709 is a good default anyway.
 467                 return YCBCR_REC_709;
 468         }
 469 }
 470
 471 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 472 {
 473         Frame *frame = static_cast<Frame *>(completedFrame);
 474         switch (result) {
 475         case bmdOutputFrameCompleted:
 476                 ++metric_decklink_output_completed_frames_completed;
 477                 break;
 478         case bmdOutputFrameDisplayedLate:
 479                 fprintf(stderr, "Output frame displayed late (pts=%" PRId64 ")\n", frame->pts);
 480                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 481                 ++metric_decklink_output_completed_frames_late;
 482                 break;
 483         case bmdOutputFrameDropped:
 484                 fprintf(stderr, "Output frame was dropped (pts=%" PRId64 ")\n", frame->pts);
 485                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 486                 ++metric_decklink_output_completed_frames_dropped;
 487                 break;
 488         case bmdOutputFrameFlushed:
 489                 fprintf(stderr, "Output frame was flushed (pts=%" PRId64 ")\n", frame->pts);
 490                 ++metric_decklink_output_completed_frames_flushed;
 491                 break;
 492         default:
 493                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 494                 ++metric_decklink_output_completed_frames_unknown;
 495                 break;
 496         }
 497
 498         static int frameno = 0;
 499         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &frameno, &latency_histogram);
 500
 501         {
 502                 lock_guard<mutex> lock(frame_queue_mutex);
 503                 frame_freelist.push(unique_ptr<Frame>(frame));
 504                 assert(scheduled_frames.front() == frame);
 505                 scheduled_frames.pop_front();
 506                 --metric_decklink_output_inflight_frames;
 507         }
 508
 509         return S_OK;
 510 }
 511
 512 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 513 {
 514         printf("playback stopped!\n");
 515         return S_OK;
 516 }
 517
 518 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 519 {
 520         lock_guard<mutex> lock(frame_queue_mutex);
 521
 522         if (!frame_freelist.empty()) {
 523                 unique_ptr<Frame> frame = move(frame_freelist.front());
 524                 frame_freelist.pop();
 525                 return frame;
 526         }
 527
 528         unique_ptr<Frame> frame(new Frame);
 529
 530         size_t stride;
 531         if (global_flags.ten_bit_output) {
 532                 stride = v210Converter::get_v210_stride(width);
 533                 GLint v210_width = stride / sizeof(uint32_t);
 534                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGB10_A2, v210_width, height);
 535
 536                 // We need valid texture state, or NVIDIA won't allow us to write to the texture.
 537                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 538                 check_error();
 539                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 540                 check_error();
 541         } else {
 542                 stride = width * 2;
 543                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 544         }
 545
 546         glGenBuffers(1, &frame->pbo);
 547         check_error();
 548         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 549         check_error();
 550         glBufferStorage(GL_PIXEL_PACK_BUFFER, stride * height, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 551         check_error();
 552         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, stride * height, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 553         check_error();
 554         frame->uyvy_ptr_local.reset(new uint8_t[stride * height]);
 555         frame->resource_pool = resource_pool;
 556
 557         return frame;
 558 }
 559
 560 void DeckLinkOutput::present_thread_func()
 561 {
 562         pthread_setname_np(pthread_self(), "DeckLinkOutput");
 563         for ( ;; ) {
 564                 unique_ptr<Frame> frame;
 565                 {
 566                         unique_lock<mutex> lock(frame_queue_mutex);
 567                         frame_queues_changed.wait(lock, [this]{
 568                                 return should_quit.should_quit() || !pending_video_frames.empty();
 569                         });
 570                         if (should_quit.should_quit()) {
 571                                 return;
 572                         }
 573                         frame = move(pending_video_frames.front());
 574                         pending_video_frames.pop();
 575                 }
 576
 577                 for ( ;; ) {
 578                         int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
 579                         if (err == GL_TIMEOUT_EXPIRED) {
 580                                 // NVIDIA likes to busy-wait; yield instead.
 581                                 this_thread::sleep_for(milliseconds(1));
 582                         } else {
 583                                 break;
 584                         }
 585                 }
 586                 check_error();
 587                 frame->fence.reset();
 588
 589                 if (global_flags.ten_bit_output) {
 590                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, v210Converter::get_v210_stride(width) * height);
 591                 } else {
 592                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 593                 }
 594
 595                 // Release any input frames we needed to render this frame.
 596                 frame->input_frames.clear();
 597
 598                 BMDTimeValue pts = frame->pts;
 599                 BMDTimeValue duration = frame->duration;
 600                 HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 601                 lock_guard<mutex> lock(frame_queue_mutex);
 602                 if (res == S_OK) {
 603                         scheduled_frames.push_back(frame.release());  // Owned by the driver now.
 604                         ++metric_decklink_output_inflight_frames;
 605                 } else {
 606                         fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 607
 608                         frame_freelist.push(move(frame));
 609                 }
 610         }
 611 }
 612
 613 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 614 {
 615         return E_NOINTERFACE;
 616 }
 617
 618 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 619 {
 620         return refcount.fetch_add(1) + 1;
 621 }
 622
 623 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 624 {
 625         int new_ref = refcount.fetch_sub(1) - 1;
 626         if (new_ref == 0)
 627                 delete this;
 628         return new_ref;
 629 }
 630
 631 DeckLinkOutput::Frame::~Frame()
 632 {
 633         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 634         check_error();
 635         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 636         check_error();
 637         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 638         check_error();
 639         glDeleteBuffers(1, &pbo);
 640         check_error();
 641         resource_pool->release_2d_texture(uyvy_tex);
 642         check_error();
 643 }
 644
 645 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 646 {
 647         return E_NOINTERFACE;
 648 }
 649
 650 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 651 {
 652         return refcount.fetch_add(1) + 1;
 653 }
 654
 655 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 656 {
 657         int new_ref = refcount.fetch_sub(1) - 1;
 658         if (new_ref == 0)
 659                 delete this;
 660         return new_ref;
 661 }
 662
 663 long DeckLinkOutput::Frame::GetWidth()
 664 {
 665         return global_flags.width;
 666 }
 667
 668 long DeckLinkOutput::Frame::GetHeight()
 669 {
 670         return global_flags.height;
 671 }
 672
 673 long DeckLinkOutput::Frame::GetRowBytes()
 674 {
 675         if (global_flags.ten_bit_output) {
 676                 return v210Converter::get_v210_stride(global_flags.width);
 677         } else {
 678                 return global_flags.width * 2;
 679         }
 680 }
 681
 682 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 683 {
 684         if (global_flags.ten_bit_output) {
 685                 return bmdFormat10BitYUV;
 686         } else {
 687                 return bmdFormat8BitYUV;
 688         }
 689 }
 690
 691 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 692 {
 693         return bmdFrameFlagDefault;
 694 }
 695
 696 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 697 {
 698         *buffer = uyvy_ptr_local.get();
 699         return S_OK;
 700 }
 701
 702 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 703 {
 704         fprintf(stderr, "STUB: GetTimecode()\n");
 705         return E_NOTIMPL;
 706 }
 707
 708 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 709 {
 710         fprintf(stderr, "STUB: GetAncillaryData()\n");
 711         return E_NOTIMPL;
 712 }