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