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

   1 #include <movit/effect_util.h>
   2 #include <movit/util.h>
   3 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
   4 #include <pthread.h>
   5
   6 #include <epoxy/egl.h>
   7
   8 #include "chroma_subsampler.h"
   9 #include "decklink_output.h"
  10 #include "decklink_util.h"
  11 #include "flags.h"
  12 #include "print_latency.h"
  13 #include "resource_pool.h"
  14 #include "timebase.h"
  15
  16 using namespace movit;
  17 using namespace std;
  18 using namespace std::chrono;
  19
  20 DeckLinkOutput::DeckLinkOutput(ResourcePool *resource_pool, QSurface *surface, unsigned width, unsigned height, unsigned card_index)
  21         : resource_pool(resource_pool), surface(surface), width(width), height(height), card_index(card_index)
  22 {
  23         chroma_subsampler.reset(new ChromaSubsampler(resource_pool));
  24 }
  25
  26 void DeckLinkOutput::set_device(IDeckLink *decklink)
  27 {
  28         if (decklink->QueryInterface(IID_IDeckLinkOutput, (void**)&output) != S_OK) {
  29                 fprintf(stderr, "Card %u has no outputs\n", card_index);
  30                 exit(1);
  31         }
  32
  33         IDeckLinkDisplayModeIterator *mode_it;
  34         if (output->GetDisplayModeIterator(&mode_it) != S_OK) {
  35                 fprintf(stderr, "Failed to enumerate output display modes for card %u\n", card_index);
  36                 exit(1);
  37         }
  38
  39         video_modes.clear();
  40
  41         for (const auto &it : summarize_video_modes(mode_it, card_index)) {
  42                 if (it.second.width != width || it.second.height != height) {
  43                         continue;
  44                 }
  45
  46                 // We could support interlaced modes, but let's stay out of it for now,
  47                 // since we don't have interlaced stream output.
  48                 if (it.second.interlaced) {
  49                         continue;
  50                 }
  51
  52                 video_modes.insert(it);
  53         }
  54
  55         mode_it->Release();
  56
  57         // HDMI or SDI generally mean “both HDMI and SDI at the same time” on DeckLink cards
  58         // that support both; pick_default_video_connection() will generally pick one of those
  59         // if they exist. We're not very likely to need analog outputs, so we don't need a way
  60         // to change beyond that.
  61         video_connection = pick_default_video_connection(decklink, BMDDeckLinkVideoOutputConnections, card_index);
  62 }
  63
  64 void DeckLinkOutput::start_output(uint32_t mode, int64_t base_pts)
  65 {
  66         assert(output);
  67         assert(!playback_initiated);
  68
  69         should_quit = false;
  70         playback_initiated = true;
  71         playback_started = false;
  72         this->base_pts = base_pts;
  73
  74         IDeckLinkConfiguration *config = nullptr;
  75         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
  76                 fprintf(stderr, "Failed to get configuration interface for output card\n");
  77                 exit(1);
  78         }
  79         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
  80                 fprintf(stderr, "Failed to set low latency output\n");
  81                 exit(1);
  82         }
  83         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
  84                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
  85                 exit(1);
  86         }
  87         if (config->SetFlag(bmdDeckLinkConfigUse1080pNotPsF, true) != S_OK) {
  88                 fprintf(stderr, "Failed to set PsF flag for card\n");
  89                 exit(1);
  90         }
  91         if (config->SetFlag(bmdDeckLinkConfigSMPTELevelAOutput, true) != S_OK) {
  92                 // This affects at least some no-name SDI->HDMI converters.
  93                 // Warn, but don't die.
  94                 fprintf(stderr, "WARNING: Failed to enable SMTPE Level A; resolutions like 1080p60 might have issues.\n");
  95         }
  96
  97         BMDDisplayModeSupport support;
  98         IDeckLinkDisplayMode *display_mode;
  99         if (output->DoesSupportVideoMode(mode, bmdFormat8BitYUV, bmdVideoOutputFlagDefault,
 100                                          &support, &display_mode) != S_OK) {
 101                 fprintf(stderr, "Couldn't ask for format support\n");
 102                 exit(1);
 103         }
 104
 105         if (support == bmdDisplayModeNotSupported) {
 106                 fprintf(stderr, "Requested display mode not supported\n");
 107                 exit(1);
 108         }
 109
 110         BMDDisplayModeFlags flags = display_mode->GetFlags();
 111         if ((flags & bmdDisplayModeColorspaceRec601) && global_flags.ycbcr_rec709_coefficients) {
 112                 fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 113                 fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 114         } else if ((flags & bmdDisplayModeColorspaceRec709) && !global_flags.ycbcr_rec709_coefficients) {
 115                 fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 116                 fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 117         }
 118
 119         BMDTimeValue time_value;
 120         BMDTimeScale time_scale;
 121         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
 122                 fprintf(stderr, "Couldn't get frame rate\n");
 123                 exit(1);
 124         }
 125
 126         frame_duration = time_value * TIMEBASE / time_scale;
 127
 128         display_mode->Release();
 129
 130         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
 131         if (result != S_OK) {
 132                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
 133                 exit(1);
 134         }
 135         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
 136                 fprintf(stderr, "Couldn't set callback\n");
 137                 exit(1);
 138         }
 139         assert(OUTPUT_FREQUENCY == 48000);
 140         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
 141                 fprintf(stderr, "Couldn't enable audio output\n");
 142                 exit(1);
 143         }
 144         if (output->BeginAudioPreroll() != S_OK) {
 145                 fprintf(stderr, "Couldn't begin audio preroll\n");
 146                 exit(1);
 147         }
 148
 149         present_thread = thread([this]{
 150                 QOpenGLContext *context = create_context(this->surface);
 151                 eglBindAPI(EGL_OPENGL_API);
 152                 if (!make_current(context, this->surface)) {
 153                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
 154                                 eglGetError());
 155                         exit(1);
 156                 }
 157                 present_thread_func();
 158                 delete_context(context);
 159         });
 160 }
 161
 162 void DeckLinkOutput::end_output()
 163 {
 164         if (!playback_initiated) {
 165                 return;
 166         }
 167
 168         should_quit = true;
 169         frame_queues_changed.notify_all();
 170         present_thread.join();
 171         playback_initiated = false;
 172
 173         output->StopScheduledPlayback(0, nullptr, 0);
 174         output->DisableVideoOutput();
 175         output->DisableAudioOutput();
 176
 177         // Wait until all frames are accounted for, and free them.
 178         {
 179                 unique_lock<mutex> lock(frame_queue_mutex);
 180                 while (!(frame_freelist.empty() && num_frames_in_flight == 0)) {
 181                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
 182                         frame_freelist.pop();
 183                 }
 184         }
 185 }
 186
 187 void DeckLinkOutput::send_frame(GLuint y_tex, GLuint cbcr_tex, const vector<RefCountedFrame> &input_frames, int64_t pts, int64_t duration)
 188 {
 189         assert(!should_quit);
 190
 191         unique_ptr<Frame> frame = move(get_frame());
 192         chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 193
 194         // Download the UYVY texture to the PBO.
 195         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 196         check_error();
 197
 198         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 199         check_error();
 200
 201         glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 202         check_error();
 203         glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 204         check_error();
 205
 206         glBindTexture(GL_TEXTURE_2D, 0);
 207         check_error();
 208         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 209         check_error();
 210
 211         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 212         check_error();
 213
 214         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 215         check_error();
 216         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 217         check_error();
 218
 219         frame->input_frames = input_frames;
 220         frame->received_ts = find_received_timestamp(input_frames);
 221         frame->pts = pts;
 222         frame->duration = duration;
 223
 224         {
 225                 unique_lock<mutex> lock(frame_queue_mutex);
 226                 pending_video_frames.push(move(frame));
 227         }
 228         frame_queues_changed.notify_all();
 229 }
 230
 231 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 232 {
 233         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 234         for (size_t i = 0; i < samples.size(); ++i) {
 235                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 236         }
 237
 238         uint32_t frames_written;
 239         HRESULT result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 240                 pts, TIMEBASE, &frames_written);
 241         if (result != S_OK) {
 242                 fprintf(stderr, "ScheduleAudioSamples(pts=%ld) failed (result=0x%08x)\n", pts, result);
 243         } else {
 244                 if (frames_written != samples.size() / 2) {
 245                         fprintf(stderr, "ScheduleAudioSamples() returned short write (%u/%ld)\n", frames_written, samples.size() / 2);
 246                 }
 247         }
 248 }
 249
 250 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
 251 {
 252         assert(!should_quit);
 253
 254         *dropped_frames = 0;
 255         *frame_duration = this->frame_duration;
 256
 257         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 258         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 259         BMDTimeValue target_time = pts - buffer;
 260
 261         // While prerolling, we send out frames as quickly as we can.
 262         if (target_time < base_pts) {
 263                 *is_preroll = true;
 264                 return;
 265         }
 266
 267         *is_preroll = !playback_started;
 268
 269         if (!playback_started) {
 270                 if (output->EndAudioPreroll() != S_OK) {
 271                         fprintf(stderr, "Could not end audio preroll\n");
 272                         exit(1);  // TODO
 273                 }
 274                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 275                         fprintf(stderr, "Could not start playback\n");
 276                         exit(1);  // TODO
 277                 }
 278                 playback_started = true;
 279         }
 280
 281         BMDTimeValue stream_frame_time;
 282         double playback_speed;
 283         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 284
 285         *frame_timestamp = steady_clock::now() +
 286                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 287
 288         // If we're ahead of time, wait for the frame to (approximately) start.
 289         if (stream_frame_time < target_time) {
 290                 this_thread::sleep_until(*frame_timestamp);
 291                 return;
 292         }
 293
 294         // If we overshot the previous frame by just a little,
 295         // fire off one immediately.
 296         if (stream_frame_time < target_time + max_overshoot) {
 297                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 298                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 299                 return;
 300         }
 301
 302         // Oops, we missed by more than one frame. Return immediately,
 303         // but drop so that we catch up.
 304         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 305         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
 306         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
 307         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 308 }
 309
 310 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
 311 {
 312         if (video_modes.count(mode)) {
 313                 return mode;
 314         }
 315
 316         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
 317         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 0 }, { 30000, 1001 } }) {
 318                 for (const auto &it : video_modes) {
 319                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
 320                                 return it.first;
 321                         }
 322                 }
 323         }
 324
 325         uint32_t best_mode = 0;
 326         double best_fps = 0.0;
 327         for (const auto &it : video_modes) {
 328                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
 329                 if (fps > best_fps) {
 330                         best_mode = it.first;
 331                         best_fps = fps;
 332                 }
 333         }
 334         return best_mode;
 335 }
 336
 337 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 338 {
 339         Frame *frame = static_cast<Frame *>(completedFrame);
 340         switch (result) {
 341         case bmdOutputFrameCompleted:
 342                 break;
 343         case bmdOutputFrameDisplayedLate:
 344                 fprintf(stderr, "Output frame displayed late (pts=%ld)\n", frame->pts);
 345                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 346                 break;
 347         case bmdOutputFrameDropped:
 348                 fprintf(stderr, "Output frame was dropped (pts=%ld)\n", frame->pts);
 349                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 350                 break;
 351         case bmdOutputFrameFlushed:
 352                 fprintf(stderr, "Output frame was flushed (pts=%ld)\n", frame->pts);
 353                 break;
 354         default:
 355                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 356                 break;
 357         }
 358
 359         static int hei = 0;
 360         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &hei);
 361
 362         {
 363                 lock_guard<mutex> lock(frame_queue_mutex);
 364                 frame_freelist.push(unique_ptr<Frame>(frame));
 365                 --num_frames_in_flight;
 366         }
 367
 368         return S_OK;
 369 }
 370
 371 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 372 {
 373         printf("playback stopped!\n");
 374         return S_OK;
 375 }
 376
 377 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 378 {
 379         lock_guard<mutex> lock(frame_queue_mutex);
 380
 381         if (!frame_freelist.empty()) {
 382                 unique_ptr<Frame> frame = move(frame_freelist.front());
 383                 frame_freelist.pop();
 384                 return frame;
 385         }
 386
 387         unique_ptr<Frame> frame(new Frame);
 388
 389         frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 390
 391         glGenBuffers(1, &frame->pbo);
 392         check_error();
 393         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 394         check_error();
 395         glBufferStorage(GL_PIXEL_PACK_BUFFER, width * height * 2, NULL, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 396         check_error();
 397         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, width * height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 398         check_error();
 399         frame->uyvy_ptr_local.reset(new uint8_t[width * height * 2]);
 400         frame->resource_pool = resource_pool;
 401
 402         return frame;
 403 }
 404
 405 void DeckLinkOutput::present_thread_func()
 406 {
 407         pthread_setname_np(pthread_self(), "DeckLinkOutput");
 408         for ( ;; ) {
 409                 unique_ptr<Frame> frame;
 410                 {
 411                         unique_lock<mutex> lock(frame_queue_mutex);
 412                         frame_queues_changed.wait(lock, [this]{
 413                                 return should_quit || !pending_video_frames.empty();
 414                         });
 415                         if (should_quit) {
 416                                 return;
 417                         }
 418                         frame = move(pending_video_frames.front());
 419                         pending_video_frames.pop();
 420                         ++num_frames_in_flight;
 421                 }
 422
 423                 glWaitSync(frame->fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 424                 check_error();
 425                 frame->fence.reset();
 426
 427                 memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 428
 429                 // Release any input frames we needed to render this frame.
 430                 frame->input_frames.clear();
 431
 432                 BMDTimeValue pts = frame->pts;
 433                 BMDTimeValue duration = frame->duration;
 434                 HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 435                 if (res == S_OK) {
 436                         frame.release();  // Owned by the driver now.
 437                 } else {
 438                         fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 439
 440                         lock_guard<mutex> lock(frame_queue_mutex);
 441                         frame_freelist.push(move(frame));
 442                         --num_frames_in_flight;
 443                 }
 444         }
 445 }
 446
 447 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 448 {
 449         return E_NOINTERFACE;
 450 }
 451
 452 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 453 {
 454         return refcount.fetch_add(1) + 1;
 455 }
 456
 457 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 458 {
 459         int new_ref = refcount.fetch_sub(1) - 1;
 460         if (new_ref == 0)
 461                 delete this;
 462         return new_ref;
 463 }
 464
 465 DeckLinkOutput::Frame::~Frame()
 466 {
 467         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 468         check_error();
 469         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 470         check_error();
 471         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 472         check_error();
 473         glDeleteBuffers(1, &pbo);
 474         check_error();
 475         resource_pool->release_2d_texture(uyvy_tex);
 476         check_error();
 477 }
 478
 479 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 480 {
 481         return E_NOINTERFACE;
 482 }
 483
 484 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 485 {
 486         return refcount.fetch_add(1) + 1;
 487 }
 488
 489 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 490 {
 491         int new_ref = refcount.fetch_sub(1) - 1;
 492         if (new_ref == 0)
 493                 delete this;
 494         return new_ref;
 495 }
 496
 497 long DeckLinkOutput::Frame::GetWidth()
 498 {
 499         return global_flags.width;
 500 }
 501
 502 long DeckLinkOutput::Frame::GetHeight()
 503 {
 504         return global_flags.height;
 505 }
 506
 507 long DeckLinkOutput::Frame::GetRowBytes()
 508 {
 509         return global_flags.width * 2;
 510 }
 511
 512 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 513 {
 514         return bmdFormat8BitYUV;
 515 }
 516
 517 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 518 {
 519         return bmdFrameFlagDefault;
 520 }
 521
 522 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 523 {
 524         *buffer = uyvy_ptr_local.get();
 525         return S_OK;
 526 }
 527
 528 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 529 {
 530         fprintf(stderr, "STUB: GetTimecode()\n");
 531         return E_NOTIMPL;
 532 }
 533
 534 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 535 {
 536         fprintf(stderr, "STUB: GetAncillaryData()\n");
 537         return E_NOTIMPL;
 538 }