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         assert(!playback_initiated);
  67
  68         should_quit = false;
  69         playback_initiated = true;
  70         playback_started = false;
  71         this->base_pts = base_pts;
  72
  73         IDeckLinkConfiguration *config = nullptr;
  74         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
  75                 fprintf(stderr, "Failed to get configuration interface for output card\n");
  76                 exit(1);
  77         }
  78         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
  79                 fprintf(stderr, "Failed to set low latency output\n");
  80                 exit(1);
  81         }
  82         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
  83                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
  84                 exit(1);
  85         }
  86         if (config->SetFlag(bmdDeckLinkConfigUse1080pNotPsF, true) != S_OK) {
  87                 fprintf(stderr, "Failed to set PsF flag for card\n");
  88                 exit(1);
  89         }
  90
  91         BMDDisplayModeSupport support;
  92         IDeckLinkDisplayMode *display_mode;
  93         if (output->DoesSupportVideoMode(mode, bmdFormat8BitYUV, bmdVideoOutputFlagDefault,
  94                                          &support, &display_mode) != S_OK) {
  95                 fprintf(stderr, "Couldn't ask for format support\n");
  96                 exit(1);
  97         }
  98
  99         if (support == bmdDisplayModeNotSupported) {
 100                 fprintf(stderr, "Requested display mode not supported\n");
 101                 exit(1);
 102         }
 103
 104         BMDDisplayModeFlags flags = display_mode->GetFlags();
 105         if ((flags & bmdDisplayModeColorspaceRec601) && global_flags.ycbcr_rec709_coefficients) {
 106                 fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
 107                 fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
 108         } else if ((flags & bmdDisplayModeColorspaceRec709) && !global_flags.ycbcr_rec709_coefficients) {
 109                 fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
 110                 fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
 111         }
 112
 113         BMDTimeValue time_value;
 114         BMDTimeScale time_scale;
 115         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
 116                 fprintf(stderr, "Couldn't get frame rate\n");
 117                 exit(1);
 118         }
 119
 120         frame_duration = time_value * TIMEBASE / time_scale;
 121
 122         display_mode->Release();
 123
 124         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
 125         if (result != S_OK) {
 126                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
 127                 exit(1);
 128         }
 129         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
 130                 fprintf(stderr, "Couldn't set callback\n");
 131                 exit(1);
 132         }
 133         assert(OUTPUT_FREQUENCY == 48000);
 134         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
 135                 fprintf(stderr, "Couldn't enable audio output\n");
 136                 exit(1);
 137         }
 138         if (output->BeginAudioPreroll() != S_OK) {
 139                 fprintf(stderr, "Couldn't begin audio preroll\n");
 140                 exit(1);
 141         }
 142
 143         present_thread = thread([this]{
 144                 QOpenGLContext *context = create_context(this->surface);
 145                 eglBindAPI(EGL_OPENGL_API);
 146                 if (!make_current(context, this->surface)) {
 147                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
 148                                 eglGetError());
 149                         exit(1);
 150                 }
 151                 present_thread_func();
 152                 delete_context(context);
 153         });
 154 }
 155
 156 void DeckLinkOutput::end_output()
 157 {
 158         if (!playback_initiated) {
 159                 return;
 160         }
 161
 162         should_quit = true;
 163         frame_queues_changed.notify_all();
 164         present_thread.join();
 165         playback_initiated = false;
 166
 167         output->StopScheduledPlayback(0, nullptr, 0);
 168         output->DisableVideoOutput();
 169         output->DisableAudioOutput();
 170
 171         // Wait until all frames are accounted for, and free them.
 172         {
 173                 unique_lock<mutex> lock(frame_queue_mutex);
 174                 while (!(frame_freelist.empty() && num_frames_in_flight == 0)) {
 175                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
 176                         frame_freelist.pop();
 177                 }
 178         }
 179 }
 180
 181 void DeckLinkOutput::send_frame(GLuint y_tex, GLuint cbcr_tex, const vector<RefCountedFrame> &input_frames, int64_t pts, int64_t duration)
 182 {
 183         unique_ptr<Frame> frame = move(get_frame());
 184         chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
 185
 186         // Download the UYVY texture to the PBO.
 187         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 188         check_error();
 189
 190         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 191         check_error();
 192
 193         glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
 194         check_error();
 195         glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
 196         check_error();
 197
 198         glBindTexture(GL_TEXTURE_2D, 0);
 199         check_error();
 200         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 201         check_error();
 202
 203         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
 204         check_error();
 205
 206         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 207         check_error();
 208         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
 209         check_error();
 210
 211         frame->input_frames = input_frames;
 212         frame->received_ts = find_received_timestamp(input_frames);
 213         frame->pts = pts;
 214         frame->duration = duration;
 215
 216         {
 217                 unique_lock<mutex> lock(frame_queue_mutex);
 218                 pending_video_frames.push(move(frame));
 219         }
 220         frame_queues_changed.notify_all();
 221 }
 222
 223 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
 224 {
 225         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
 226         for (size_t i = 0; i < samples.size(); ++i) {
 227                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
 228         }
 229
 230         uint32_t frames_written;
 231         HRESULT result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
 232                 pts, TIMEBASE, &frames_written);
 233         if (result != S_OK) {
 234                 fprintf(stderr, "ScheduleAudioSamples(pts=%ld) failed (result=0x%08x)\n", pts, result);
 235         } else {
 236                 if (frames_written != samples.size() / 2) {
 237                         fprintf(stderr, "ScheduleAudioSamples() returned short write (%u/%ld)\n", frames_written, samples.size() / 2);
 238                 }
 239         }
 240 }
 241
 242 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration)
 243 {
 244         *dropped_frames = 0;
 245         *frame_duration = this->frame_duration;
 246
 247         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
 248         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
 249         BMDTimeValue target_time = pts - buffer;
 250
 251         // While prerolling, we send out frames as quickly as we can.
 252         if (target_time < base_pts) {
 253                 return;
 254         }
 255
 256         if (!playback_started) {
 257                 if (output->EndAudioPreroll() != S_OK) {
 258                         fprintf(stderr, "Could not end audio preroll\n");
 259                         exit(1);  // TODO
 260                 }
 261                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
 262                         fprintf(stderr, "Could not start playback\n");
 263                         exit(1);  // TODO
 264                 }
 265                 playback_started = true;
 266         }
 267
 268         BMDTimeValue stream_frame_time;
 269         double playback_speed;
 270         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
 271
 272         // If we're ahead of time, wait for the frame to (approximately) start.
 273         if (stream_frame_time < target_time) {
 274                 steady_clock::time_point t = steady_clock::now() +
 275                         nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
 276                 this_thread::sleep_until(t);
 277                 return;
 278         }
 279
 280         // If we overshot the previous frame by just a little,
 281         // fire off one immediately.
 282         if (stream_frame_time < target_time + max_overshoot) {
 283                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
 284                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
 285                 return;
 286         }
 287
 288         // Oops, we missed by more than one frame. Return immediately,
 289         // but drop so that we catch up.
 290         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
 291         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
 292 }
 293
 294 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
 295 {
 296         Frame *frame = static_cast<Frame *>(completedFrame);
 297         switch (result) {
 298         case bmdOutputFrameCompleted:
 299                 break;
 300         case bmdOutputFrameDisplayedLate:
 301                 fprintf(stderr, "Output frame displayed late (pts=%ld)\n", frame->pts);
 302                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 303                 break;
 304         case bmdOutputFrameDropped:
 305                 fprintf(stderr, "Output frame was dropped (pts=%ld)\n", frame->pts);
 306                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
 307                 break;
 308         case bmdOutputFrameFlushed:
 309                 fprintf(stderr, "Output frame was flushed (pts=%ld)\n", frame->pts);
 310                 break;
 311         default:
 312                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
 313                 break;
 314         }
 315
 316         static int hei = 0;
 317         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &hei);
 318
 319         {
 320                 lock_guard<mutex> lock(frame_queue_mutex);
 321                 frame_freelist.push(unique_ptr<Frame>(frame));
 322                 --num_frames_in_flight;
 323         }
 324
 325         return S_OK;
 326 }
 327
 328 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
 329 {
 330         printf("playback stopped!\n");
 331         return S_OK;
 332 }
 333
 334 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
 335 {
 336         lock_guard<mutex> lock(frame_queue_mutex);
 337
 338         if (!frame_freelist.empty()) {
 339                 unique_ptr<Frame> frame = move(frame_freelist.front());
 340                 frame_freelist.pop();
 341                 return frame;
 342         }
 343
 344         unique_ptr<Frame> frame(new Frame);
 345
 346         frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
 347
 348         glGenBuffers(1, &frame->pbo);
 349         check_error();
 350         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
 351         check_error();
 352         glBufferStorage(GL_PIXEL_PACK_BUFFER, width * height * 2, NULL, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 353         check_error();
 354         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, width * height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 355         check_error();
 356         frame->uyvy_ptr_local.reset(new uint8_t[width * height * 2]);
 357         frame->resource_pool = resource_pool;
 358
 359         return frame;
 360 }
 361
 362 void DeckLinkOutput::present_thread_func()
 363 {
 364         for ( ;; ) {
 365                 unique_ptr<Frame> frame;
 366                 {
 367                         unique_lock<mutex> lock(frame_queue_mutex);
 368                         frame_queues_changed.wait(lock, [this]{
 369                                 return should_quit || !pending_video_frames.empty();
 370                         });
 371                         if (should_quit) {
 372                                 return;
 373                         }
 374                         frame = move(pending_video_frames.front());
 375                         pending_video_frames.pop();
 376                         ++num_frames_in_flight;
 377                 }
 378
 379                 glWaitSync(frame->fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 380                 check_error();
 381                 frame->fence.reset();
 382
 383                 memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
 384
 385                 // Release any input frames we needed to render this frame.
 386                 frame->input_frames.clear();
 387
 388                 BMDTimeValue pts = frame->pts;
 389                 BMDTimeValue duration = frame->duration;
 390                 HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
 391                 if (res == S_OK) {
 392                         frame.release();  // Owned by the driver now.
 393                 } else {
 394                         fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
 395
 396                         lock_guard<mutex> lock(frame_queue_mutex);
 397                         frame_freelist.push(move(frame));
 398                         --num_frames_in_flight;
 399                 }
 400         }
 401 }
 402
 403 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
 404 {
 405         return E_NOINTERFACE;
 406 }
 407
 408 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
 409 {
 410         return refcount.fetch_add(1) + 1;
 411 }
 412
 413 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
 414 {
 415         int new_ref = refcount.fetch_sub(1) - 1;
 416         if (new_ref == 0)
 417                 delete this;
 418         return new_ref;
 419 }
 420
 421 DeckLinkOutput::Frame::~Frame()
 422 {
 423         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
 424         check_error();
 425         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 426         check_error();
 427         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 428         check_error();
 429         glDeleteBuffers(1, &pbo);
 430         check_error();
 431         resource_pool->release_2d_texture(uyvy_tex);
 432         check_error();
 433 }
 434
 435 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
 436 {
 437         return E_NOINTERFACE;
 438 }
 439
 440 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
 441 {
 442         return refcount.fetch_add(1) + 1;
 443 }
 444
 445 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
 446 {
 447         int new_ref = refcount.fetch_sub(1) - 1;
 448         if (new_ref == 0)
 449                 delete this;
 450         return new_ref;
 451 }
 452
 453 long DeckLinkOutput::Frame::GetWidth()
 454 {
 455         return global_flags.width;
 456 }
 457
 458 long DeckLinkOutput::Frame::GetHeight()
 459 {
 460         return global_flags.height;
 461 }
 462
 463 long DeckLinkOutput::Frame::GetRowBytes()
 464 {
 465         return global_flags.width * 2;
 466 }
 467
 468 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
 469 {
 470         return bmdFormat8BitYUV;
 471 }
 472
 473 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
 474 {
 475         return bmdFrameFlagDefault;
 476 }
 477
 478 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
 479 {
 480         *buffer = uyvy_ptr_local.get();
 481         return S_OK;
 482 }
 483
 484 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
 485 {
 486         fprintf(stderr, "STUB: GetTimecode()\n");
 487         return E_NOTIMPL;
 488 }
 489
 490 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
 491 {
 492         fprintf(stderr, "STUB: GetAncillaryData()\n");
 493         return E_NOTIMPL;
 494 }