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

   1 #include "DeckLinkAPI.h"
   2 #include "DeckLinkAPIConfiguration.h"
   3 #include "DeckLinkAPIModes.h"
   4 #include "DeckLinkAPITypes.h"
   5 #include "LinuxCOM.h"
   6 #include <atomic>
   7 #include <assert.h>
   8 #include <chrono>
   9 #include <inttypes.h>
  10 #include <math.h>
  11 #include <stddef.h>
  12 #include <stdint.h>
  13 #include <stdio.h>
  14 #include <string.h>
  15 #include <memory>
  16 #include <movit/image_format.h>
  17 #include <movit/util.h>
  18 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
  19 #include <pthread.h>
  20 #include <thread>
  21
  22 #include <mutex>
  23
  24 #include <epoxy/egl.h>
  25 #include <utility>
  26 #include <vector>
  27
  28 #include "chroma_subsampler.h"
  29 #include "decklink_output.h"
  30 #include "decklink_util.h"
  31 #include "flags.h"
  32 #include "ref_counted_frame.h"
  33 #include "shared/metrics.h"
  34 #include "print_latency.h"
  35 #include "shared/ref_counted_gl_sync.h"
  36 #include "shared/shared_defs.h"
  37 #include "shared/timebase.h"
  38 #include "v210_converter.h"
  39
  40 using namespace movit;
  41 using namespace std;
  42 using namespace std::chrono;
  43
  44 namespace {
  45
  46 // This class can be deleted during regular use, so make all the metrics static.
  47 once_flag decklink_metrics_inited;
  48 LatencyHistogram latency_histogram;
  49 atomic<int64_t> metric_decklink_output_width_pixels{-1};
  50 atomic<int64_t> metric_decklink_output_height_pixels{-1};
  51 atomic<int64_t> metric_decklink_output_frame_rate_den{-1};
  52 atomic<int64_t> metric_decklink_output_frame_rate_nom{-1};
  53 atomic<int64_t> metric_decklink_output_inflight_frames{0};
  54 atomic<int64_t> metric_decklink_output_color_mismatch_frames{0};
  55
  56 atomic<int64_t> metric_decklink_output_scheduled_frames_dropped{0};
  57 atomic<int64_t> metric_decklink_output_scheduled_frames_late{0};
  58 atomic<int64_t> metric_decklink_output_scheduled_frames_normal{0};
  59 atomic<int64_t> metric_decklink_output_scheduled_frames_preroll{0};
  60
  61 atomic<int64_t> metric_decklink_output_completed_frames_completed{0};
  62 atomic<int64_t> metric_decklink_output_completed_frames_dropped{0};
  63 atomic<int64_t> metric_decklink_output_completed_frames_flushed{0};
  64 atomic<int64_t> metric_decklink_output_completed_frames_late{0};
  65 atomic<int64_t> metric_decklink_output_completed_frames_unknown{0};
  66
  67 atomic<int64_t> metric_decklink_output_scheduled_samples{0};
  68
  69 Summary metric_decklink_output_margin_seconds;
  70
  71 }  // namespace
  72
  73 DeckLinkOutput::DeckLinkOutput(ResourcePool *resource_pool, QSurface *surface, unsigned width, unsigned height, unsigned card_index)
  74         : resource_pool(resource_pool), surface(surface), width(width), height(height), card_index(card_index)
  75 {
  76         chroma_subsampler.reset(new ChromaSubsampler(resource_pool));
  77
  78         call_once(decklink_metrics_inited, [](){
  79                 latency_histogram.init("decklink_output");
  80                 global_metrics.add("decklink_output_width_pixels", &metric_decklink_output_width_pixels, Metrics::TYPE_GAUGE);
  81                 global_metrics.add("decklink_output_height_pixels", &metric_decklink_output_height_pixels, Metrics::TYPE_GAUGE);
  82                 global_metrics.add("decklink_output_frame_rate_den", &metric_decklink_output_frame_rate_den, Metrics::TYPE_GAUGE);
  83                 global_metrics.add("decklink_output_frame_rate_nom", &metric_decklink_output_frame_rate_nom, Metrics::TYPE_GAUGE);
  84                 global_metrics.add("decklink_output_inflight_frames", &metric_decklink_output_inflight_frames, Metrics::TYPE_GAUGE);
  85                 global_metrics.add("decklink_output_color_mismatch_frames", &metric_decklink_output_color_mismatch_frames);
  86
  87                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "dropped" }}, &metric_decklink_output_scheduled_frames_dropped);
  88                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "late" }}, &metric_decklink_output_scheduled_frames_late);
  89                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "normal" }}, &metric_decklink_output_scheduled_frames_normal);
  90                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "preroll" }}, &metric_decklink_output_scheduled_frames_preroll);
  91
  92                 global_metrics.add("decklink_output_completed_frames", {{ "status", "completed" }}, &metric_decklink_output_completed_frames_completed);
  93                 global_metrics.add("decklink_output_completed_frames", {{ "status", "dropped" }}, &metric_decklink_output_completed_frames_dropped);
  94                 global_metrics.add("decklink_output_completed_frames", {{ "status", "flushed" }}, &metric_decklink_output_completed_frames_flushed);
  95                 global_metrics.add("decklink_output_completed_frames", {{ "status", "late" }}, &metric_decklink_output_completed_frames_late);
  96                 global_metrics.add("decklink_output_completed_frames", {{ "status", "unknown" }}, &metric_decklink_output_completed_frames_unknown);
  97
  98                 global_metrics.add("decklink_output_scheduled_samples", &metric_decklink_output_scheduled_samples);
  99                 vector<double> quantiles{0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99};
 100                 metric_decklink_output_margin_seconds.init(quantiles, 60.0);
 101                 global_metrics.add("decklink_output_margin_seconds", &metric_decklink_output_margin_seconds);
 102         });
 103 }
 104
 105 DeckLinkOutput::~DeckLinkOutput()
 106 {
 107         if (output != nullptr) {
 108                 output->Release();
 109         }
 110 }
 111
 112 bool DeckLinkOutput::set_device(IDeckLink *decklink, IDeckLinkInput *input_arg)
 113 {
 114         input = input_arg;
 115         if (decklink->QueryInterface(IID_IDeckLinkOutput, (void**)&output) != S_OK) {
 116                 fprintf(stderr, "Warning: Card %u has no outputs\n", card_index);
 117                 return false;
 118         }
 119
 120         IDeckLinkDisplayModeIterator *mode_it;
 121         if (output->GetDisplayModeIterator(&mode_it) != S_OK) {
 122                 fprintf(stderr, "Warning: Failed to enumerate output display modes for card %u\n", card_index);
 123                 return false;
 124         }
 125
 126         video_modes.clear();
 127
 128         for (const auto &it : summarize_video_modes(mode_it, card_index)) {
 129                 if (it.second.width != width || it.second.height != height) {
 130                         continue;
 131                 }
 132
 133                 // We could support interlaced modes, but let's stay out of it for now,
 134                 // since we don't have interlaced stream output.
 135                 if (it.second.interlaced) {
 136                         continue;
 137                 }
 138
 139                 video_modes.insert(it);
 140         }
 141
 142         mode_it->Release();
 143
 144         // HDMI or SDI generally mean “both HDMI and SDI at the same time” on DeckLink cards
 145         // that support both; pick_default_video_connection() will generally pick one of those
 146         // if they exist. (--prefer-hdmi-input would also affect the selection despite the name
 147         // of the option, but since either generally means both, it's inconsequential.)
 148         // We're not very likely to need analog outputs, so we don't need a way to change
 149         // beyond that.
 150         video_connection = pick_default_video_connection(decklink, BMDDeckLinkVideoOutputConnections, card_index);
 151         return true;
 152 }
 153
 154 void DeckLinkOutput::start_output(uint32_t mode, int64_t base_pts, bool is_master_card_arg)
 155 {
 156         assert(output);
 157         assert(!playback_initiated);
 158         this->is_master_card = is_master_card_arg;
 159
 160         if (video_modes.empty()) {
 161                 fprintf(stderr, "ERROR: No matching output modes for %dx%d found\n", width, height);
 162                 abort();
 163         }
 164
 165         should_quit.unquit();
 166         playback_initiated = true;
 167         playback_started = false;
 168         this->base_pts = base_pts;
 169
 170         IDeckLinkConfiguration *config = nullptr;
 171         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
 172                 fprintf(stderr, "Failed to get configuration interface for output card\n");
 173                 abort();
 174         }
 175         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
 176                 fprintf(stderr, "Failed to set low latency output\n");
 177                 abort();
 178         }
 179         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
 180                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
 181                 abort();
 182         }
 183         if (config->SetFlag(bmdDeckLinkConfigOutput1080pAsPsF, true) != S_OK) {
 184                 fprintf(stderr, "Failed to set PsF flag for card\n");
 185                 abort();
 186         }
 187         if (config->SetFlag(bmdDeckLinkConfigSMPTELevelAOutput, true) != S_OK) {
 188                 // This affects at least some no-name SDI->HDMI converters.
 189                 // Warn, but don't die.
 190                 fprintf(stderr, "WARNING: Failed to enable SMTPE Level A; resolutions like 1080p60 might have issues.\n");
 191         }
 192
 193         BMDDisplayModeSupport support;
 194         IDeckLinkDisplayMode *display_mode;
 195         BMDPixelFormat pixel_format = global_flags.bit_depth > 8 ? bmdFormat10BitYUV : bmdFormat8BitYUV;
 196         if (output->DoesSupportVideoMode(mode, pixel_format, bmdVideoOutputFlagDefault,
 197                                          &support, &display_mode) != S_OK) {
 198                 fprintf(stderr, "Couldn't ask for format support\n");
 199                 abort();
 200         }
 201
 202         if (support == bmdDisplayModeNotSupported) {
 203                 fprintf(stderr, "Requested display mode not supported\n");
 204                 abort();
 205         }
 206
 207         current_mode_flags = display_mode->GetFlags();
 208
 209         BMDTimeValue time_value;
 210         BMDTimeScale time_scale;
 211         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
 212                 fprintf(stderr, "Couldn't get frame rate\n");
 213                 abort();
 214         }
 215
 216         metric_decklink_output_width_pixels = width;
 217         metric_decklink_output_height_pixels = height;
 218         metric_decklink_output_frame_rate_nom = time_value;
 219         metric_decklink_output_frame_rate_den = time_scale;
 220
 221         frame_duration = time_value * TIMEBASE / time_scale;
 222
 223         display_mode->Release();
 224
 225         if (input != nullptr) {
 226                 if (input->DisableVideoInput() != S_OK) {
 227                         fprintf(stderr, "Warning: Failed to disable video input for card %d\n", card_index);
 228                 }
 229                 if (input->DisableAudioInput() != S_OK) {
 230                         fprintf(stderr, "Warning: Failed to disable audio input for card %d\n", card_index);
 231                 }
 232         }
 233
 234         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
 235         if (result != S_OK) {
 236                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
 237                 abort();
 238         }
 239         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
 240                 fprintf(stderr, "Couldn't set callback\n");
 241                 abort();
 242         }
 243         assert(OUTPUT_FREQUENCY == 48000);
 244         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
 245                 fprintf(stderr, "Couldn't enable audio output\n");
 246                 abort();
 247         }
 248         if (is_master_card) {
 249                 if (output->BeginAudioPreroll() != S_OK) {
 250                         fprintf(stderr, "Couldn't begin audio preroll\n");
 251                         abort();
 252                 }
 253         } else {
 254                 playback_started = true;
 255         }
 256
 257         present_thread = thread([this]{
 258                 QOpenGLContext *context = create_context(this->surface);
 259                 eglBindAPI(EGL_OPENGL_API);
 260                 if (!make_current(context, this->surface)) {
 261                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
 262                                 eglGetError());
 263                         abort();
 264                 }
 265                 present_thread_func();
 266                 delete_context(context);
 267         });
 268 }
 269
 270 void DeckLinkOutput::end_output()
 271 {
 272         if (!playback_initiated) {
 273                 return;
 274         }
 275
 276         should_quit.quit();
 277         frame_queues_changed.notify_all();
 278         present_thread.join();
 279         playback_initiated = false;
 280
 281         if (is_master_card) {
 282                 output->StopScheduledPlayback(0, nullptr, 0);
 283         }
 284         output->DisableVideoOutput();
 285         output->DisableAudioOutput();
 286
 287         // Wait until all frames are accounted for, and free them.
 288         {
 289                 unique_lock<mutex> lock(frame_queue_mutex);
 290                 while (!(frame_freelist.empty() && scheduled_frames.empty())) {
 291                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
 292                         frame_freelist.pop();
 293                 }
 294         }
 295 }
 296
 297 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)
 298 {
 299         assert(!should_quit.should_quit());
 300
 301         if ((current_mode_flags & bmdDisplayModeColorspaceRec601) && output_ycbcr_coefficients == YCBCR_REC_709) {
 302                 if (!last_frame_had_mode_mismatch) {
 303                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 304                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 305                 }
 306                 last_frame_had_mode_mismatch = true;
 307                 ++metric_decklink_output_color_mismatch_frames;
 308         } else if ((current_mode_flags & bmdDisplayModeColorspaceRec709) && output_ycbcr_coefficients == YCBCR_REC_601) {
 309                 if (!last_frame_had_mode_mismatch) {
 310                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 311                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 312                 }
 313                 last_frame_had_mode_mismatch = true;
 314                 ++metric_decklink_output_color_mismatch_frames;
 315         } else {
 316                 last_frame_had_mode_mismatch = false;
 317         }
 318
 319         unique_ptr<Frame> frame = get_frame();
 320         if (global_flags.bit_depth > 8) {
 321                 chroma_subsampler->create_v210(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 322         } else {
 323                 chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 324         }
 325
 326         // Download the UYVY texture to the PBO.
 327         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 328         check_error();
 329
 330         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 331         check_error();
 332
 333         if (global_flags.bit_depth > 8) {
 334                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 335                 check_error();
 336                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, BUFFER_OFFSET(0));
 337                 check_error();
 338         } else {
 339                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 340                 check_error();
 341                 glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 342                 check_error();
 343         }
 344
 345         glBindTexture(GL_TEXTURE_2D, 0);
 346         check_error();
 347         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 348         check_error();
 349
 350         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 351         check_error();
 352
 353         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 354         check_error();
 355         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 356         check_error();
 357
 358         frame->input_frames = input_frames;
 359         frame->received_ts = find_received_timestamp(input_frames);
 360         frame->pts = pts;
 361         frame->duration = duration;
 362
 363         {
 364                 lock_guard<mutex> lock(frame_queue_mutex);
 365                 pending_video_frames.push(move(frame));
 366         }
 367         frame_queues_changed.notify_all();
 368 }
 369
 370 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 371 {
 372         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 373         for (size_t i = 0; i < samples.size(); ++i) {
 374                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 375         }
 376
 377         uint32_t frames_written;
 378         HRESULT result;
 379         if (is_master_card) {
 380                 result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 381                         pts, TIMEBASE, &frames_written);
 382         } else {
 383                 result = output->WriteAudioSamplesSync(int_samples.get(), samples.size() / 2,
 384                         &frames_written);
 385         }
 386         if (result != S_OK) {
 387                 fprintf(stderr, "write audio to DeckLink (pts=%" PRId64 ") failed (result=0x%08x)\n", pts, result);
 388         } else {
 389                 // Non-master card is not really synchronized on audio at all, so we don't warn on it.
 390                 if (frames_written != samples.size() / 2 && is_master_card) {
 391                         fprintf(stderr, "write audio to DeckLink returned short write (%u/%zu)\n", frames_written, samples.size() / 2);
 392                 }
 393         }
 394         metric_decklink_output_scheduled_samples += samples.size() / 2;
 395 }
 396
 397 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
 398 {
 399         assert(!should_quit.should_quit());
 400
 401         *dropped_frames = 0;
 402         *frame_duration = this->frame_duration;
 403
 404         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 405         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 406         BMDTimeValue target_time = pts - buffer;
 407
 408         // While prerolling, we send out frames as quickly as we can.
 409         if (target_time < base_pts) {
 410                 *is_preroll = true;
 411                 ++metric_decklink_output_scheduled_frames_preroll;
 412                 return;
 413         }
 414
 415         *is_preroll = !playback_started;
 416
 417         if (!playback_started) {
 418                 if (output->EndAudioPreroll() != S_OK) {
 419                         fprintf(stderr, "Could not end audio preroll\n");
 420                         abort();  // TODO
 421                 }
 422                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 423                         fprintf(stderr, "Could not start playback\n");
 424                         abort();  // TODO
 425                 }
 426                 playback_started = true;
 427         }
 428
 429         BMDTimeValue stream_frame_time;
 430         double playback_speed;
 431         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 432
 433         *frame_timestamp = steady_clock::now() +
 434                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 435
 436         metric_decklink_output_margin_seconds.count_event(
 437                 (target_time - stream_frame_time) / double(TIMEBASE));
 438
 439         // If we're ahead of time, wait for the frame to (approximately) start.
 440         if (stream_frame_time < target_time) {
 441                 should_quit.sleep_until(*frame_timestamp);
 442                 ++metric_decklink_output_scheduled_frames_normal;
 443                 return;
 444         }
 445
 446         // If we overshot the previous frame by just a little,
 447         // fire off one immediately.
 448         if (stream_frame_time < target_time + max_overshoot) {
 449                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 450                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 451                 ++metric_decklink_output_scheduled_frames_late;
 452                 return;
 453         }
 454
 455         // Oops, we missed by more than one frame. Return immediately,
 456         // but drop so that we catch up.
 457         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 458         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
 459         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
 460         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 461         metric_decklink_output_scheduled_frames_dropped += *dropped_frames;
 462         ++metric_decklink_output_scheduled_frames_normal;
 463 }
 464
 465 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
 466 {
 467         if (video_modes.count(mode)) {
 468                 return mode;
 469         }
 470
 471         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
 472         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 1 }, { 30000, 1001 } }) {
 473                 for (const auto &it : video_modes) {
 474                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
 475                                 return it.first;
 476                         }
 477                 }
 478         }
 479
 480         uint32_t best_mode = 0;
 481         double best_fps = 0.0;
 482         for (const auto &it : video_modes) {
 483                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
 484                 if (fps > best_fps) {
 485                         best_mode = it.first;
 486                         best_fps = fps;
 487                 }
 488         }
 489         return best_mode;
 490 }
 491
 492 YCbCrLumaCoefficients DeckLinkOutput::preferred_ycbcr_coefficients() const
 493 {
 494         if (current_mode_flags & bmdDisplayModeColorspaceRec601) {
 495                 return YCBCR_REC_601;
 496         } else {
 497                 // Don't bother checking bmdDisplayModeColorspaceRec709;
 498                 // if none is set, 709 is a good default anyway.
 499                 return YCBCR_REC_709;
 500         }
 501 }
 502
 503 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 504 {
 505         Frame *frame = static_cast<Frame *>(completedFrame);
 506         switch (result) {
 507         case bmdOutputFrameCompleted:
 508                 ++metric_decklink_output_completed_frames_completed;
 509                 break;
 510         case bmdOutputFrameDisplayedLate:
 511                 fprintf(stderr, "Output frame displayed late (pts=%" PRId64 ")\n", frame->pts);
 512                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 513                 ++metric_decklink_output_completed_frames_late;
 514                 break;
 515         case bmdOutputFrameDropped:
 516                 fprintf(stderr, "Output frame was dropped (pts=%" PRId64 ")\n", frame->pts);
 517                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 518                 ++metric_decklink_output_completed_frames_dropped;
 519                 break;
 520         case bmdOutputFrameFlushed:
 521                 fprintf(stderr, "Output frame was flushed (pts=%" PRId64 ")\n", frame->pts);
 522                 ++metric_decklink_output_completed_frames_flushed;
 523                 break;
 524         default:
 525                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 526                 ++metric_decklink_output_completed_frames_unknown;
 527                 break;
 528         }
 529
 530         static int frameno = 0;
 531         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &frameno, &latency_histogram);
 532
 533         {
 534                 lock_guard<mutex> lock(frame_queue_mutex);
 535                 frame_freelist.push(unique_ptr<Frame>(frame));
 536                 assert(scheduled_frames.count(frame));
 537                 scheduled_frames.erase(frame);
 538                 --metric_decklink_output_inflight_frames;
 539         }
 540
 541         return S_OK;
 542 }
 543
 544 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 545 {
 546         printf("playback stopped!\n");
 547         return S_OK;
 548 }
 549
 550 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 551 {
 552         lock_guard<mutex> lock(frame_queue_mutex);
 553
 554         if (!frame_freelist.empty()) {
 555                 unique_ptr<Frame> frame = move(frame_freelist.front());
 556                 frame_freelist.pop();
 557                 return frame;
 558         }
 559
 560         unique_ptr<Frame> frame(new Frame);
 561
 562         size_t stride;
 563         if (global_flags.bit_depth > 8) {
 564                 stride = v210Converter::get_v210_stride(width);
 565                 GLint v210_width = stride / sizeof(uint32_t);
 566                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGB10_A2, v210_width, height);
 567
 568                 // We need valid texture state, or NVIDIA won't allow us to write to the texture.
 569                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 570                 check_error();
 571                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 572                 check_error();
 573         } else {
 574                 stride = width * 2;
 575                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 576         }
 577
 578         glGenBuffers(1, &frame->pbo);
 579         check_error();
 580         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 581         check_error();
 582         glBufferStorage(GL_PIXEL_PACK_BUFFER, stride * height, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 583         check_error();
 584         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, stride * height, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 585         check_error();
 586         frame->uyvy_ptr_local.reset(new uint8_t[stride * height]);
 587         frame->resource_pool = resource_pool;
 588
 589         return frame;
 590 }
 591
 592 void DeckLinkOutput::present_thread_func()
 593 {
 594         pthread_setname_np(pthread_self(), "DeckLinkOutput");
 595         for ( ;; ) {
 596                 unique_ptr<Frame> frame;
 597                 {
 598                         unique_lock<mutex> lock(frame_queue_mutex);
 599                         frame_queues_changed.wait(lock, [this]{
 600                                 return should_quit.should_quit() || !pending_video_frames.empty();
 601                         });
 602                         if (should_quit.should_quit()) {
 603                                 return;
 604                         }
 605                         frame = move(pending_video_frames.front());
 606                         pending_video_frames.pop();
 607                 }
 608
 609                 for ( ;; ) {
 610                         int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
 611                         if (err == GL_TIMEOUT_EXPIRED) {
 612                                 // NVIDIA likes to busy-wait; yield instead.
 613                                 this_thread::sleep_for(milliseconds(1));
 614                         } else {
 615                                 break;
 616                         }
 617                 }
 618                 check_error();
 619                 frame->fence.reset();
 620
 621                 if (global_flags.bit_depth > 8) {
 622                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, v210Converter::get_v210_stride(width) * height);
 623                 } else {
 624                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 625                 }
 626
 627                 // Release any input frames we needed to render this frame.
 628                 frame->input_frames.clear();
 629
 630                 if (is_master_card) {
 631                         BMDTimeValue pts = frame->pts;
 632                         BMDTimeValue duration = frame->duration;
 633                         HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 634                         lock_guard<mutex> lock(frame_queue_mutex);
 635                         if (res == S_OK) {
 636                                 scheduled_frames.insert(frame.release());  // Owned by the driver now.
 637                                 ++metric_decklink_output_inflight_frames;
 638                         } else {
 639                                 fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 640
 641                                 frame_freelist.push(move(frame));
 642                         }
 643                 } else {
 644                         HRESULT res = output->DisplayVideoFrameSync(frame.get());
 645                         if (res != S_OK) {
 646                                 fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 647                         }
 648                         frame_freelist.push(move(frame));
 649                 }
 650         }
 651 }
 652
 653 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 654 {
 655         return E_NOINTERFACE;
 656 }
 657
 658 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 659 {
 660         return refcount.fetch_add(1) + 1;
 661 }
 662
 663 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 664 {
 665         int new_ref = refcount.fetch_sub(1) - 1;
 666         if (new_ref == 0)
 667                 delete this;
 668         return new_ref;
 669 }
 670
 671 DeckLinkOutput::Frame::~Frame()
 672 {
 673         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 674         check_error();
 675         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 676         check_error();
 677         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 678         check_error();
 679         glDeleteBuffers(1, &pbo);
 680         check_error();
 681         resource_pool->release_2d_texture(uyvy_tex);
 682         check_error();
 683 }
 684
 685 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 686 {
 687         return E_NOINTERFACE;
 688 }
 689
 690 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 691 {
 692         return refcount.fetch_add(1) + 1;
 693 }
 694
 695 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 696 {
 697         int new_ref = refcount.fetch_sub(1) - 1;
 698         if (new_ref == 0)
 699                 delete this;
 700         return new_ref;
 701 }
 702
 703 long DeckLinkOutput::Frame::GetWidth()
 704 {
 705         return global_flags.width;
 706 }
 707
 708 long DeckLinkOutput::Frame::GetHeight()
 709 {
 710         return global_flags.height;
 711 }
 712
 713 long DeckLinkOutput::Frame::GetRowBytes()
 714 {
 715         if (global_flags.bit_depth > 8) {
 716                 return v210Converter::get_v210_stride(global_flags.width);
 717         } else {
 718                 return global_flags.width * 2;
 719         }
 720 }
 721
 722 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 723 {
 724         if (global_flags.bit_depth > 8) {
 725                 return bmdFormat10BitYUV;
 726         } else {
 727                 return bmdFormat8BitYUV;
 728         }
 729 }
 730
 731 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 732 {
 733         return bmdFrameFlagDefault;
 734 }
 735
 736 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 737 {
 738         *buffer = uyvy_ptr_local.get();
 739         return S_OK;
 740 }
 741
 742 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 743 {
 744         fprintf(stderr, "STUB: GetTimecode()\n");
 745         return E_NOTIMPL;
 746 }
 747
 748 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 749 {
 750         fprintf(stderr, "STUB: GetAncillaryData()\n");
 751         return E_NOTIMPL;
 752 }