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