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
 270 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)
 271 {
 272         assert(!should_quit.should_quit());
 273
 274         if ((current_mode_flags & bmdDisplayModeColorspaceRec601) && output_ycbcr_coefficients == YCBCR_REC_709) {
 275                 if (!last_frame_had_mode_mismatch) {
 276                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 277                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 278                 }
 279                 last_frame_had_mode_mismatch = true;
 280                 ++metric_decklink_output_color_mismatch_frames;
 281         } else if ((current_mode_flags & bmdDisplayModeColorspaceRec709) && output_ycbcr_coefficients == YCBCR_REC_601) {
 282                 if (!last_frame_had_mode_mismatch) {
 283                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 284                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 285                 }
 286                 last_frame_had_mode_mismatch = true;
 287                 ++metric_decklink_output_color_mismatch_frames;
 288         } else {
 289                 last_frame_had_mode_mismatch = false;
 290         }
 291
 292         unique_ptr<Frame> frame = get_frame();
 293         if (global_flags.ten_bit_output) {
 294                 chroma_subsampler->create_v210(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 295         } else {
 296                 chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 297         }
 298
 299         // Download the UYVY texture to the PBO.
 300         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 301         check_error();
 302
 303         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 304         check_error();
 305
 306         if (global_flags.ten_bit_output) {
 307                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 308                 check_error();
 309                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, BUFFER_OFFSET(0));
 310                 check_error();
 311         } else {
 312                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 313                 check_error();
 314                 glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 315                 check_error();
 316         }
 317
 318         glBindTexture(GL_TEXTURE_2D, 0);
 319         check_error();
 320         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 321         check_error();
 322
 323         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 324         check_error();
 325
 326         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 327         check_error();
 328         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 329         check_error();
 330
 331         frame->input_frames = input_frames;
 332         frame->received_ts = find_received_timestamp(input_frames);
 333         frame->pts = pts;
 334         frame->duration = duration;
 335
 336         {
 337                 lock_guard<mutex> lock(frame_queue_mutex);
 338                 pending_video_frames.push(move(frame));
 339         }
 340         frame_queues_changed.notify_all();
 341 }
 342
 343 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 344 {
 345         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 346         for (size_t i = 0; i < samples.size(); ++i) {
 347                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 348         }
 349
 350         uint32_t frames_written;
 351         HRESULT result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 352                 pts, TIMEBASE, &frames_written);
 353         if (result != S_OK) {
 354                 fprintf(stderr, "ScheduleAudioSamples(pts=%" PRId64 ") failed (result=0x%08x)\n", pts, result);
 355         } else {
 356                 if (frames_written != samples.size() / 2) {
 357                         fprintf(stderr, "ScheduleAudioSamples() returned short write (%u/%zu)\n", frames_written, samples.size() / 2);
 358                 }
 359         }
 360         metric_decklink_output_scheduled_samples += samples.size() / 2;
 361 }
 362
 363 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
 364 {
 365         assert(!should_quit.should_quit());
 366
 367         *dropped_frames = 0;
 368         *frame_duration = this->frame_duration;
 369
 370         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 371         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 372         BMDTimeValue target_time = pts - buffer;
 373
 374         // While prerolling, we send out frames as quickly as we can.
 375         if (target_time < base_pts) {
 376                 *is_preroll = true;
 377                 ++metric_decklink_output_scheduled_frames_preroll;
 378                 return;
 379         }
 380
 381         *is_preroll = !playback_started;
 382
 383         if (!playback_started) {
 384                 if (output->EndAudioPreroll() != S_OK) {
 385                         fprintf(stderr, "Could not end audio preroll\n");
 386                         abort();  // TODO
 387                 }
 388                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 389                         fprintf(stderr, "Could not start playback\n");
 390                         abort();  // TODO
 391                 }
 392                 playback_started = true;
 393         }
 394
 395         BMDTimeValue stream_frame_time;
 396         double playback_speed;
 397         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 398
 399         *frame_timestamp = steady_clock::now() +
 400                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 401
 402         metric_decklink_output_margin_seconds.count_event(
 403                 (target_time - stream_frame_time) / double(TIMEBASE));
 404
 405         // If we're ahead of time, wait for the frame to (approximately) start.
 406         if (stream_frame_time < target_time) {
 407                 should_quit.sleep_until(*frame_timestamp);
 408                 ++metric_decklink_output_scheduled_frames_normal;
 409                 return;
 410         }
 411
 412         // If we overshot the previous frame by just a little,
 413         // fire off one immediately.
 414         if (stream_frame_time < target_time + max_overshoot) {
 415                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 416                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 417                 ++metric_decklink_output_scheduled_frames_late;
 418                 return;
 419         }
 420
 421         // Oops, we missed by more than one frame. Return immediately,
 422         // but drop so that we catch up.
 423         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 424         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
 425         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
 426         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 427         metric_decklink_output_scheduled_frames_dropped += *dropped_frames;
 428         ++metric_decklink_output_scheduled_frames_normal;
 429 }
 430
 431 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
 432 {
 433         if (video_modes.count(mode)) {
 434                 return mode;
 435         }
 436
 437         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
 438         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 1 }, { 30000, 1001 } }) {
 439                 for (const auto &it : video_modes) {
 440                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
 441                                 return it.first;
 442                         }
 443                 }
 444         }
 445
 446         uint32_t best_mode = 0;
 447         double best_fps = 0.0;
 448         for (const auto &it : video_modes) {
 449                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
 450                 if (fps > best_fps) {
 451                         best_mode = it.first;
 452                         best_fps = fps;
 453                 }
 454         }
 455         return best_mode;
 456 }
 457
 458 YCbCrLumaCoefficients DeckLinkOutput::preferred_ycbcr_coefficients() const
 459 {
 460         if (current_mode_flags & bmdDisplayModeColorspaceRec601) {
 461                 return YCBCR_REC_601;
 462         } else {
 463                 // Don't bother checking bmdDisplayModeColorspaceRec709;
 464                 // if none is set, 709 is a good default anyway.
 465                 return YCBCR_REC_709;
 466         }
 467 }
 468
 469 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 470 {
 471         Frame *frame = static_cast<Frame *>(completedFrame);
 472         switch (result) {
 473         case bmdOutputFrameCompleted:
 474                 ++metric_decklink_output_completed_frames_completed;
 475                 break;
 476         case bmdOutputFrameDisplayedLate:
 477                 fprintf(stderr, "Output frame displayed late (pts=%" PRId64 ")\n", frame->pts);
 478                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 479                 ++metric_decklink_output_completed_frames_late;
 480                 break;
 481         case bmdOutputFrameDropped:
 482                 fprintf(stderr, "Output frame was dropped (pts=%" PRId64 ")\n", frame->pts);
 483                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 484                 ++metric_decklink_output_completed_frames_dropped;
 485                 break;
 486         case bmdOutputFrameFlushed:
 487                 fprintf(stderr, "Output frame was flushed (pts=%" PRId64 ")\n", frame->pts);
 488                 ++metric_decklink_output_completed_frames_flushed;
 489                 break;
 490         default:
 491                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 492                 ++metric_decklink_output_completed_frames_unknown;
 493                 break;
 494         }
 495
 496         static int frameno = 0;
 497         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &frameno, &latency_histogram);
 498
 499         {
 500                 lock_guard<mutex> lock(frame_queue_mutex);
 501                 frame_freelist.push(unique_ptr<Frame>(frame));
 502                 assert(scheduled_frames.count(frame));
 503                 scheduled_frames.erase(frame);
 504                 --metric_decklink_output_inflight_frames;
 505         }
 506
 507         return S_OK;
 508 }
 509
 510 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 511 {
 512         printf("playback stopped!\n");
 513         return S_OK;
 514 }
 515
 516 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 517 {
 518         lock_guard<mutex> lock(frame_queue_mutex);
 519
 520         if (!frame_freelist.empty()) {
 521                 unique_ptr<Frame> frame = move(frame_freelist.front());
 522                 frame_freelist.pop();
 523                 return frame;
 524         }
 525
 526         unique_ptr<Frame> frame(new Frame);
 527
 528         size_t stride;
 529         if (global_flags.ten_bit_output) {
 530                 stride = v210Converter::get_v210_stride(width);
 531                 GLint v210_width = stride / sizeof(uint32_t);
 532                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGB10_A2, v210_width, height);
 533
 534                 // We need valid texture state, or NVIDIA won't allow us to write to the texture.
 535                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 536                 check_error();
 537                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 538                 check_error();
 539         } else {
 540                 stride = width * 2;
 541                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 542         }
 543
 544         glGenBuffers(1, &frame->pbo);
 545         check_error();
 546         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 547         check_error();
 548         glBufferStorage(GL_PIXEL_PACK_BUFFER, stride * height, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 549         check_error();
 550         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, stride * height, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 551         check_error();
 552         frame->uyvy_ptr_local.reset(new uint8_t[stride * height]);
 553         frame->resource_pool = resource_pool;
 554
 555         return frame;
 556 }
 557
 558 void DeckLinkOutput::present_thread_func()
 559 {
 560         pthread_setname_np(pthread_self(), "DeckLinkOutput");
 561         for ( ;; ) {
 562                 unique_ptr<Frame> frame;
 563                 {
 564                         unique_lock<mutex> lock(frame_queue_mutex);
 565                         frame_queues_changed.wait(lock, [this]{
 566                                 return should_quit.should_quit() || !pending_video_frames.empty();
 567                         });
 568                         if (should_quit.should_quit()) {
 569                                 return;
 570                         }
 571                         frame = move(pending_video_frames.front());
 572                         pending_video_frames.pop();
 573                 }
 574
 575                 for ( ;; ) {
 576                         int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
 577                         if (err == GL_TIMEOUT_EXPIRED) {
 578                                 // NVIDIA likes to busy-wait; yield instead.
 579                                 this_thread::sleep_for(milliseconds(1));
 580                         } else {
 581                                 break;
 582                         }
 583                 }
 584                 check_error();
 585                 frame->fence.reset();
 586
 587                 if (global_flags.ten_bit_output) {
 588                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, v210Converter::get_v210_stride(width) * height);
 589                 } else {
 590                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 591                 }
 592
 593                 // Release any input frames we needed to render this frame.
 594                 frame->input_frames.clear();
 595
 596                 BMDTimeValue pts = frame->pts;
 597                 BMDTimeValue duration = frame->duration;
 598                 HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 599                 lock_guard<mutex> lock(frame_queue_mutex);
 600                 if (res == S_OK) {
 601                         scheduled_frames.insert(frame.release());  // Owned by the driver now.
 602                         ++metric_decklink_output_inflight_frames;
 603                 } else {
 604                         fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 605
 606                         frame_freelist.push(move(frame));
 607                 }
 608         }
 609 }
 610
 611 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 612 {
 613         return E_NOINTERFACE;
 614 }
 615
 616 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 617 {
 618         return refcount.fetch_add(1) + 1;
 619 }
 620
 621 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 622 {
 623         int new_ref = refcount.fetch_sub(1) - 1;
 624         if (new_ref == 0)
 625                 delete this;
 626         return new_ref;
 627 }
 628
 629 DeckLinkOutput::Frame::~Frame()
 630 {
 631         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 632         check_error();
 633         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 634         check_error();
 635         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 636         check_error();
 637         glDeleteBuffers(1, &pbo);
 638         check_error();
 639         resource_pool->release_2d_texture(uyvy_tex);
 640         check_error();
 641 }
 642
 643 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 644 {
 645         return E_NOINTERFACE;
 646 }
 647
 648 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 649 {
 650         return refcount.fetch_add(1) + 1;
 651 }
 652
 653 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 654 {
 655         int new_ref = refcount.fetch_sub(1) - 1;
 656         if (new_ref == 0)
 657                 delete this;
 658         return new_ref;
 659 }
 660
 661 long DeckLinkOutput::Frame::GetWidth()
 662 {
 663         return global_flags.width;
 664 }
 665
 666 long DeckLinkOutput::Frame::GetHeight()
 667 {
 668         return global_flags.height;
 669 }
 670
 671 long DeckLinkOutput::Frame::GetRowBytes()
 672 {
 673         if (global_flags.ten_bit_output) {
 674                 return v210Converter::get_v210_stride(global_flags.width);
 675         } else {
 676                 return global_flags.width * 2;
 677         }
 678 }
 679
 680 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 681 {
 682         if (global_flags.ten_bit_output) {
 683                 return bmdFormat10BitYUV;
 684         } else {
 685                 return bmdFormat8BitYUV;
 686         }
 687 }
 688
 689 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 690 {
 691         return bmdFrameFlagDefault;
 692 }
 693
 694 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 695 {
 696         *buffer = uyvy_ptr_local.get();
 697         return S_OK;
 698 }
 699
 700 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 701 {
 702         fprintf(stderr, "STUB: GetTimecode()\n");
 703         return E_NOTIMPL;
 704 }
 705
 706 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 707 {
 708         fprintf(stderr, "STUB: GetAncillaryData()\n");
 709         return E_NOTIMPL;
 710 }