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