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

   1 #include "decklink_capture.h"
   2
   3 #include <DeckLinkAPI.h>
   4 #include <DeckLinkAPIConfiguration.h>
   5 #include <DeckLinkAPIDiscovery.h>
   6 #include <DeckLinkAPIModes.h>
   7 #include <assert.h>
   8 #ifdef __SSE2__
   9 #include <immintrin.h>
  10 #endif
  11 #include <pthread.h>
  12 #include <stdint.h>
  13 #include <stdio.h>
  14 #include <stdlib.h>
  15 #include <string.h>
  16 #include <chrono>
  17 #include <cstdint>
  18 #include <utility>
  19 #include <vector>
  20
  21 #include "bmusb/bmusb.h"
  22 #include "decklink_util.h"
  23 #include "flags.h"
  24 #include "v210_converter.h"
  25
  26 #define FRAME_SIZE (8 << 20)  // 8 MB.
  27
  28 using namespace std;
  29 using namespace std::chrono;
  30 using namespace std::placeholders;
  31 using namespace bmusb;
  32
  33 namespace {
  34
  35 // TODO: Support stride.
  36 void memcpy_interleaved(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size_t n)
  37 {
  38         assert(n % 2 == 0);
  39         uint8_t *dptr1 = dest1;
  40         uint8_t *dptr2 = dest2;
  41
  42         for (size_t i = 0; i < n; i += 2) {
  43                 *dptr1++ = *src++;
  44                 *dptr2++ = *src++;
  45         }
  46 }
  47
  48 #ifdef __SSE2__
  49
  50 // Returns the number of bytes consumed.
  51 size_t memcpy_interleaved_fastpath(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size_t n)
  52 {
  53         const uint8_t *limit = src + n;
  54         size_t consumed = 0;
  55
  56         // Align end to 32 bytes.
  57         limit = (const uint8_t *)(intptr_t(limit) & ~31);
  58
  59         if (src >= limit) {
  60                 return 0;
  61         }
  62
  63         // Process [0,31] bytes, such that start gets aligned to 32 bytes.
  64         const uint8_t *aligned_src = (const uint8_t *)(intptr_t(src + 31) & ~31);
  65         if (aligned_src != src) {
  66                 size_t n2 = aligned_src - src;
  67                 memcpy_interleaved(dest1, dest2, src, n2);
  68                 dest1 += n2 / 2;
  69                 dest2 += n2 / 2;
  70                 if (n2 % 2) {
  71                         swap(dest1, dest2);
  72                 }
  73                 src = aligned_src;
  74                 consumed += n2;
  75         }
  76
  77         // Make the length a multiple of 64.
  78         if (((limit - src) % 64) != 0) {
  79                 limit -= 32;
  80         }
  81         assert(((limit - src) % 64) == 0);
  82
  83 #if __AVX2__
  84         const __m256i * __restrict in = (const __m256i *)src;
  85         __m256i * __restrict out1 = (__m256i *)dest1;
  86         __m256i * __restrict out2 = (__m256i *)dest2;
  87
  88         __m256i shuffle_cw = _mm256_set_epi8(
  89                 15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0,
  90                 15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0);
  91         while (in < (const __m256i *)limit) {
  92                 // Note: For brevity, comments show lanes as if they were 2x64-bit (they're actually 2x128).
  93                 __m256i data1 = _mm256_stream_load_si256(in);         // AaBbCcDd EeFfGgHh
  94                 __m256i data2 = _mm256_stream_load_si256(in + 1);     // IiJjKkLl MmNnOoPp
  95
  96                 data1 = _mm256_shuffle_epi8(data1, shuffle_cw);       // ABCDabcd EFGHefgh
  97                 data2 = _mm256_shuffle_epi8(data2, shuffle_cw);       // IJKLijkl MNOPmnop
  98
  99                 data1 = _mm256_permute4x64_epi64(data1, 0b11011000);  // ABCDEFGH abcdefgh
 100                 data2 = _mm256_permute4x64_epi64(data2, 0b11011000);  // IJKLMNOP ijklmnop
 101
 102                 __m256i lo = _mm256_permute2x128_si256(data1, data2, 0b00100000);
 103                 __m256i hi = _mm256_permute2x128_si256(data1, data2, 0b00110001);
 104
 105                 _mm256_storeu_si256(out1, lo);
 106                 _mm256_storeu_si256(out2, hi);
 107
 108                 in += 2;
 109                 ++out1;
 110                 ++out2;
 111                 consumed += 64;
 112         }
 113 #else
 114         const __m128i * __restrict in = (const __m128i *)src;
 115         __m128i * __restrict out1 = (__m128i *)dest1;
 116         __m128i * __restrict out2 = (__m128i *)dest2;
 117
 118         __m128i mask_lower_byte = _mm_set1_epi16(0x00ff);
 119         while (in < (const __m128i *)limit) {
 120                 __m128i data1 = _mm_load_si128(in);
 121                 __m128i data2 = _mm_load_si128(in + 1);
 122                 __m128i data1_lo = _mm_and_si128(data1, mask_lower_byte);
 123                 __m128i data2_lo = _mm_and_si128(data2, mask_lower_byte);
 124                 __m128i data1_hi = _mm_srli_epi16(data1, 8);
 125                 __m128i data2_hi = _mm_srli_epi16(data2, 8);
 126                 __m128i lo = _mm_packus_epi16(data1_lo, data2_lo);
 127                 _mm_storeu_si128(out1, lo);
 128                 __m128i hi = _mm_packus_epi16(data1_hi, data2_hi);
 129                 _mm_storeu_si128(out2, hi);
 130
 131                 in += 2;
 132                 ++out1;
 133                 ++out2;
 134                 consumed += 32;
 135         }
 136 #endif
 137
 138         return consumed;
 139 }
 140
 141 #endif  // __SSE2__
 142
 143 BMDPixelFormat pixel_format_to_bmd(PixelFormat pixel_format)
 144 {
 145         switch (pixel_format) {
 146         case PixelFormat_8BitYCbCr:
 147                 return bmdFormat8BitYUV;
 148         case PixelFormat_10BitYCbCr:
 149                 return bmdFormat10BitYUV;
 150         default:
 151                 assert(false);
 152         }
 153 }
 154
 155 }  // namespace
 156
 157 DeckLinkCapture::DeckLinkCapture(IDeckLink *card, int card_index)
 158         : card_index(card_index), card(card)
 159 {
 160         {
 161                 const char *model_name;
 162                 char buf[256];
 163                 if (card->GetModelName(&model_name) == S_OK) {
 164                         snprintf(buf, sizeof(buf), "PCI card %d: %s", card_index, model_name);
 165                 } else {
 166                         snprintf(buf, sizeof(buf), "PCI card %d: Unknown DeckLink card", card_index);
 167                 }
 168                 description = buf;
 169         }
 170
 171         if (card->QueryInterface(IID_IDeckLinkInput, (void**)&input) != S_OK) {
 172                 fprintf(stderr, "Card %d has no inputs\n", card_index);
 173                 exit(1);
 174         }
 175
 176         IDeckLinkAttributes *attr;
 177         if (card->QueryInterface(IID_IDeckLinkAttributes, (void**)&attr) != S_OK) {
 178                 fprintf(stderr, "Card %d has no attributes\n", card_index);
 179                 exit(1);
 180         }
 181
 182         // Get the list of available video inputs.
 183         int64_t video_input_mask;
 184         if (attr->GetInt(BMDDeckLinkVideoInputConnections, &video_input_mask) != S_OK) {
 185                 fprintf(stderr, "Failed to enumerate video inputs for card %d\n", card_index);
 186                 exit(1);
 187         }
 188         const vector<pair<BMDVideoConnection, string>> video_input_types = {
 189                 { bmdVideoConnectionSDI, "SDI" },
 190                 { bmdVideoConnectionHDMI, "HDMI" },
 191                 { bmdVideoConnectionOpticalSDI, "Optical SDI" },
 192                 { bmdVideoConnectionComponent, "Component" },
 193                 { bmdVideoConnectionComposite, "Composite" },
 194                 { bmdVideoConnectionSVideo, "S-Video" }
 195         };
 196         for (const auto &video_input : video_input_types) {
 197                 if (video_input_mask & video_input.first) {
 198                         video_inputs.emplace(video_input.first, video_input.second);
 199                 }
 200         }
 201
 202         // And then the available audio inputs.
 203         int64_t audio_input_mask;
 204         if (attr->GetInt(BMDDeckLinkAudioInputConnections, &audio_input_mask) != S_OK) {
 205                 fprintf(stderr, "Failed to enumerate audio inputs for card %d\n", card_index);
 206                 exit(1);
 207         }
 208         const vector<pair<BMDAudioConnection, string>> audio_input_types = {
 209                 { bmdAudioConnectionEmbedded, "Embedded" },
 210                 { bmdAudioConnectionAESEBU, "AES/EBU" },
 211                 { bmdAudioConnectionAnalog, "Analog" },
 212                 { bmdAudioConnectionAnalogXLR, "Analog XLR" },
 213                 { bmdAudioConnectionAnalogRCA, "Analog RCA" },
 214                 { bmdAudioConnectionMicrophone, "Microphone" },
 215                 { bmdAudioConnectionHeadphones, "Headphones" }
 216         };
 217         for (const auto &audio_input : audio_input_types) {
 218                 if (audio_input_mask & audio_input.first) {
 219                         audio_inputs.emplace(audio_input.first, audio_input.second);
 220                 }
 221         }
 222
 223         // Check if we the card supports input autodetection.
 224         if (attr->GetFlag(BMDDeckLinkSupportsInputFormatDetection, &supports_autodetect) != S_OK) {
 225                 fprintf(stderr, "Warning: Failed to ask card %d whether it supports input format autodetection\n", card_index);
 226                 supports_autodetect = false;
 227         }
 228
 229         // If there's more than one subdevice on this card, label them.
 230         int64_t num_subdevices, subdevice_idx;
 231         if (attr->GetInt(BMDDeckLinkNumberOfSubDevices, &num_subdevices) == S_OK && num_subdevices > 1) {
 232                 if (attr->GetInt(BMDDeckLinkSubDeviceIndex, &subdevice_idx) == S_OK) {
 233                         char buf[256];
 234                         snprintf(buf, sizeof(buf), " (subdevice %d)", int(subdevice_idx));
 235                         description += buf;
 236                 }
 237         }
 238
 239         attr->Release();
 240
 241         /* Set up the video and audio sources. */
 242         if (card->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
 243                 fprintf(stderr, "Failed to get configuration interface for card %d\n", card_index);
 244                 exit(1);
 245         }
 246
 247         BMDVideoConnection connection = pick_default_video_connection(card, BMDDeckLinkVideoInputConnections, card_index);
 248
 249         set_video_input(connection);
 250         set_audio_input(bmdAudioConnectionEmbedded);
 251
 252         IDeckLinkDisplayModeIterator *mode_it;
 253         if (input->GetDisplayModeIterator(&mode_it) != S_OK) {
 254                 fprintf(stderr, "Failed to enumerate display modes for card %d\n", card_index);
 255                 exit(1);
 256         }
 257
 258         video_modes = summarize_video_modes(mode_it, card_index);
 259         mode_it->Release();
 260
 261         set_video_mode_no_restart(bmdModeHD720p5994);
 262
 263         input->SetCallback(this);
 264 }
 265
 266 DeckLinkCapture::~DeckLinkCapture()
 267 {
 268         if (has_dequeue_callbacks) {
 269                 dequeue_cleanup_callback();
 270         }
 271         input->Release();
 272         config->Release();
 273         card->Release();
 274 }
 275
 276 HRESULT STDMETHODCALLTYPE DeckLinkCapture::QueryInterface(REFIID, LPVOID *)
 277 {
 278         return E_NOINTERFACE;
 279 }
 280
 281 ULONG STDMETHODCALLTYPE DeckLinkCapture::AddRef(void)
 282 {
 283         return refcount.fetch_add(1) + 1;
 284 }
 285
 286 ULONG STDMETHODCALLTYPE DeckLinkCapture::Release(void)
 287 {
 288         int new_ref = refcount.fetch_sub(1) - 1;
 289         if (new_ref == 0)
 290                 delete this;
 291         return new_ref;
 292 }
 293
 294 HRESULT STDMETHODCALLTYPE DeckLinkCapture::VideoInputFormatChanged(
 295         BMDVideoInputFormatChangedEvents,
 296         IDeckLinkDisplayMode* display_mode,
 297         BMDDetectedVideoInputFormatFlags format_flags)
 298 {
 299         if (format_flags & bmdDetectedVideoInputRGB444) {
 300                 fprintf(stderr, "WARNING: Input detected as 4:4:4 RGB, but Nageru can't consume that yet.\n");
 301                 fprintf(stderr, "Doing hardware conversion to 4:2:2 Y'CbCr.\n");
 302         }
 303         if (supports_autodetect && display_mode->GetDisplayMode() != current_video_mode) {
 304                 set_video_mode(display_mode->GetDisplayMode());
 305         }
 306         if (display_mode->GetFrameRate(&frame_duration, &time_scale) != S_OK) {
 307                 fprintf(stderr, "Could not get new frame rate\n");
 308                 exit(1);
 309         }
 310         field_dominance = display_mode->GetFieldDominance();
 311         return S_OK;
 312 }
 313
 314 HRESULT STDMETHODCALLTYPE DeckLinkCapture::VideoInputFrameArrived(
 315         IDeckLinkVideoInputFrame *video_frame,
 316         IDeckLinkAudioInputPacket *audio_frame)
 317 {
 318         if (!done_init) {
 319                 char thread_name[16];
 320                 snprintf(thread_name, sizeof(thread_name), "DeckLink_C_%d", card_index);
 321                 pthread_setname_np(pthread_self(), thread_name);
 322
 323                 sched_param param;
 324                 memset(&param, 0, sizeof(param));
 325                 param.sched_priority = 1;
 326                 if (sched_setscheduler(0, SCHED_RR, &param) == -1) {
 327                         printf("couldn't set realtime priority for DeckLink thread: %s\n", strerror(errno));
 328                 }
 329
 330                 if (has_dequeue_callbacks) {
 331                         dequeue_init_callback();
 332                 }
 333                 done_init = true;
 334         }
 335
 336         steady_clock::time_point now = steady_clock::now();
 337
 338         FrameAllocator::Frame current_video_frame, current_audio_frame;
 339         VideoFormat video_format;
 340         AudioFormat audio_format;
 341
 342         video_format.frame_rate_nom = time_scale;
 343         video_format.frame_rate_den = frame_duration;
 344         // TODO: Respect the TFF/BFF flag.
 345         video_format.interlaced = (field_dominance == bmdLowerFieldFirst || field_dominance == bmdUpperFieldFirst);
 346         video_format.second_field_start = 1;
 347
 348         if (video_frame) {
 349                 video_format.has_signal = !(video_frame->GetFlags() & bmdFrameHasNoInputSource);
 350
 351                 const int width = video_frame->GetWidth();
 352                 const int height = video_frame->GetHeight();
 353                 const int stride = video_frame->GetRowBytes();
 354                 const BMDPixelFormat format = video_frame->GetPixelFormat();
 355                 assert(format == pixel_format_to_bmd(current_pixel_format));
 356                 if (global_flags.ten_bit_input) {
 357                         assert(stride == int(v210Converter::get_v210_stride(width)));
 358                 } else {
 359                         assert(stride == width * 2);
 360                 }
 361
 362                 current_video_frame = video_frame_allocator->alloc_frame();
 363                 if (current_video_frame.data != nullptr) {
 364                         const uint8_t *frame_bytes;
 365                         video_frame->GetBytes((void **)&frame_bytes);
 366                         size_t num_bytes = stride * height;
 367
 368                         if (current_video_frame.interleaved) {
 369                                 uint8_t *data = current_video_frame.data;
 370                                 uint8_t *data2 = current_video_frame.data2;
 371 #ifdef __SSE2__
 372                                 size_t consumed = memcpy_interleaved_fastpath(data, data2, frame_bytes, num_bytes);
 373                                 frame_bytes += consumed;
 374                                 data += consumed / 2;
 375                                 data2 += consumed / 2;
 376                                 if (num_bytes % 2) {
 377                                         swap(data, data2);
 378                                 }
 379                                 current_video_frame.len += consumed;
 380                                 num_bytes -= consumed;
 381 #endif
 382
 383                                 if (num_bytes > 0) {
 384                                         memcpy_interleaved(data, data2, frame_bytes, num_bytes);
 385                                 }
 386                         } else {
 387                                 memcpy(current_video_frame.data, frame_bytes, num_bytes);
 388                         }
 389                         current_video_frame.len += num_bytes;
 390
 391                         video_format.width = width;
 392                         video_format.height = height;
 393                         video_format.stride = stride;
 394                 }
 395         }
 396
 397         if (audio_frame) {
 398                 int num_samples = audio_frame->GetSampleFrameCount();
 399
 400                 current_audio_frame = audio_frame_allocator->alloc_frame();
 401                 if (current_audio_frame.data != nullptr) {
 402                         const uint8_t *frame_bytes;
 403                         audio_frame->GetBytes((void **)&frame_bytes);
 404                         current_audio_frame.len = sizeof(int32_t) * 2 * num_samples;
 405
 406                         memcpy(current_audio_frame.data, frame_bytes, current_audio_frame.len);
 407
 408                         audio_format.bits_per_sample = 32;
 409                         audio_format.num_channels = 2;
 410                 }
 411         }
 412
 413         current_video_frame.received_timestamp = now;
 414         current_audio_frame.received_timestamp = now;
 415
 416         if (current_video_frame.data != nullptr || current_audio_frame.data != nullptr) {
 417                 // TODO: Put into a queue and put into a dequeue thread, if the
 418                 // BlackMagic drivers don't already do that for us?
 419                 frame_callback(timecode,
 420                         current_video_frame, /*video_offset=*/0, video_format,
 421                         current_audio_frame, /*audio_offset=*/0, audio_format);
 422         }
 423
 424         timecode++;
 425         return S_OK;
 426 }
 427
 428 void DeckLinkCapture::configure_card()
 429 {
 430         if (video_frame_allocator == nullptr) {
 431                 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
 432                 set_video_frame_allocator(owned_video_frame_allocator.get());
 433         }
 434         if (audio_frame_allocator == nullptr) {
 435                 owned_audio_frame_allocator.reset(new MallocFrameAllocator(65536, NUM_QUEUED_AUDIO_FRAMES));
 436                 set_audio_frame_allocator(owned_audio_frame_allocator.get());
 437         }
 438 }
 439
 440 void DeckLinkCapture::start_bm_capture()
 441 {
 442         if (running) {
 443                 return;
 444         }
 445         if (input->EnableVideoInput(current_video_mode, pixel_format_to_bmd(current_pixel_format), supports_autodetect ? bmdVideoInputEnableFormatDetection : 0) != S_OK) {
 446                 fprintf(stderr, "Failed to set video mode 0x%04x for card %d\n", current_video_mode, card_index);
 447                 exit(1);
 448         }
 449         if (input->EnableAudioInput(48000, bmdAudioSampleType32bitInteger, 2) != S_OK) {
 450                 fprintf(stderr, "Failed to enable audio input for card %d\n", card_index);
 451                 exit(1);
 452         }
 453
 454         if (input->StartStreams() != S_OK) {
 455                 fprintf(stderr, "StartStreams failed\n");
 456                 exit(1);
 457         }
 458         running = true;
 459 }
 460
 461 void DeckLinkCapture::stop_dequeue_thread()
 462 {
 463         if (!running) {
 464                 return;
 465         }
 466         HRESULT result = input->StopStreams();
 467         if (result != S_OK) {
 468                 fprintf(stderr, "StopStreams failed with error 0x%x\n", result);
 469                 exit(1);
 470         }
 471         if (input->DisableVideoInput() != S_OK) {
 472                 fprintf(stderr, "Failed to disable video input for card %d\n", card_index);
 473                 exit(1);
 474         }
 475         if (input->DisableAudioInput() != S_OK) {
 476                 fprintf(stderr, "Failed to disable audio input for card %d\n", card_index);
 477                 exit(1);
 478         }
 479         running = false;
 480 }
 481
 482 void DeckLinkCapture::set_video_mode(uint32_t video_mode_id)
 483 {
 484         if (running) {
 485                 if (input->PauseStreams() != S_OK) {
 486                         fprintf(stderr, "PauseStreams failed\n");
 487                         exit(1);
 488                 }
 489                 if (input->FlushStreams() != S_OK) {
 490                         fprintf(stderr, "FlushStreams failed\n");
 491                         exit(1);
 492                 }
 493         }
 494
 495         set_video_mode_no_restart(video_mode_id);
 496
 497         if (running) {
 498                 if (input->StartStreams() != S_OK) {
 499                         fprintf(stderr, "StartStreams failed\n");
 500                         exit(1);
 501                 }
 502         }
 503 }
 504
 505 void DeckLinkCapture::set_pixel_format(PixelFormat pixel_format)
 506 {
 507         current_pixel_format = pixel_format;
 508         set_video_mode(current_video_mode);
 509 }
 510
 511 void DeckLinkCapture::set_video_mode_no_restart(uint32_t video_mode_id)
 512 {
 513         BMDDisplayModeSupport support;
 514         IDeckLinkDisplayMode *display_mode;
 515         if (input->DoesSupportVideoMode(video_mode_id, pixel_format_to_bmd(current_pixel_format), /*flags=*/0, &support, &display_mode)) {
 516                 fprintf(stderr, "Failed to query display mode for card %d\n", card_index);
 517                 exit(1);
 518         }
 519
 520         if (support == bmdDisplayModeNotSupported) {
 521                 fprintf(stderr, "Card %d does not support display mode\n", card_index);
 522                 exit(1);
 523         }
 524
 525         if (display_mode->GetFrameRate(&frame_duration, &time_scale) != S_OK) {
 526                 fprintf(stderr, "Could not get frame rate for card %d\n", card_index);
 527                 exit(1);
 528         }
 529
 530         field_dominance = display_mode->GetFieldDominance();
 531
 532         if (running) {
 533                 if (input->EnableVideoInput(video_mode_id, pixel_format_to_bmd(current_pixel_format), supports_autodetect ? bmdVideoInputEnableFormatDetection : 0) != S_OK) {
 534                         fprintf(stderr, "Failed to set video mode 0x%04x for card %d\n", video_mode_id, card_index);
 535                         exit(1);
 536                 }
 537         }
 538
 539         current_video_mode = video_mode_id;
 540 }
 541
 542 void DeckLinkCapture::set_video_input(uint32_t video_input_id)
 543 {
 544         if (config->SetInt(bmdDeckLinkConfigVideoInputConnection, video_input_id) != S_OK) {
 545                 fprintf(stderr, "Failed to set video input connection for card %d\n", card_index);
 546                 exit(1);
 547         }
 548
 549         current_video_input = video_input_id;
 550 }
 551
 552 void DeckLinkCapture::set_audio_input(uint32_t audio_input_id)
 553 {
 554         if (config->SetInt(bmdDeckLinkConfigAudioInputConnection, audio_input_id) != S_OK) {
 555                 fprintf(stderr, "Failed to set audio input connection for card %d\n", card_index);
 556                 exit(1);
 557         }
 558
 559         current_audio_input = audio_input_id;
 560 }