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                 if (has_dequeue_callbacks) {
 323                         dequeue_init_callback();
 324                 }
 325                 done_init = true;
 326         }
 327
 328         steady_clock::time_point now = steady_clock::now();
 329
 330         FrameAllocator::Frame current_video_frame, current_audio_frame;
 331         VideoFormat video_format;
 332         AudioFormat audio_format;
 333
 334         video_format.frame_rate_nom = time_scale;
 335         video_format.frame_rate_den = frame_duration;
 336         // TODO: Respect the TFF/BFF flag.
 337         video_format.interlaced = (field_dominance == bmdLowerFieldFirst || field_dominance == bmdUpperFieldFirst);
 338         video_format.second_field_start = 1;
 339
 340         if (video_frame) {
 341                 video_format.has_signal = !(video_frame->GetFlags() & bmdFrameHasNoInputSource);
 342
 343                 const int width = video_frame->GetWidth();
 344                 const int height = video_frame->GetHeight();
 345                 const int stride = video_frame->GetRowBytes();
 346                 const BMDPixelFormat format = video_frame->GetPixelFormat();
 347                 assert(format == pixel_format_to_bmd(current_pixel_format));
 348                 if (global_flags.ten_bit_input) {
 349                         assert(stride == int(v210Converter::get_v210_stride(width)));
 350                 } else {
 351                         assert(stride == width * 2);
 352                 }
 353
 354                 current_video_frame = video_frame_allocator->alloc_frame();
 355                 if (current_video_frame.data != nullptr) {
 356                         const uint8_t *frame_bytes;
 357                         video_frame->GetBytes((void **)&frame_bytes);
 358                         size_t num_bytes = stride * height;
 359
 360                         if (current_video_frame.interleaved) {
 361                                 uint8_t *data = current_video_frame.data;
 362                                 uint8_t *data2 = current_video_frame.data2;
 363 #ifdef __SSE2__
 364                                 size_t consumed = memcpy_interleaved_fastpath(data, data2, frame_bytes, num_bytes);
 365                                 frame_bytes += consumed;
 366                                 data += consumed / 2;
 367                                 data2 += consumed / 2;
 368                                 if (num_bytes % 2) {
 369                                         swap(data, data2);
 370                                 }
 371                                 current_video_frame.len += consumed;
 372                                 num_bytes -= consumed;
 373 #endif
 374
 375                                 if (num_bytes > 0) {
 376                                         memcpy_interleaved(data, data2, frame_bytes, num_bytes);
 377                                 }
 378                         } else {
 379                                 memcpy(current_video_frame.data, frame_bytes, num_bytes);
 380                         }
 381                         current_video_frame.len += num_bytes;
 382
 383                         video_format.width = width;
 384                         video_format.height = height;
 385                         video_format.stride = stride;
 386                 }
 387         }
 388
 389         if (audio_frame) {
 390                 int num_samples = audio_frame->GetSampleFrameCount();
 391
 392                 current_audio_frame = audio_frame_allocator->alloc_frame();
 393                 if (current_audio_frame.data != nullptr) {
 394                         const uint8_t *frame_bytes;
 395                         audio_frame->GetBytes((void **)&frame_bytes);
 396                         current_audio_frame.len = sizeof(int32_t) * 2 * num_samples;
 397
 398                         memcpy(current_audio_frame.data, frame_bytes, current_audio_frame.len);
 399
 400                         audio_format.bits_per_sample = 32;
 401                         audio_format.num_channels = 2;
 402                 }
 403         }
 404
 405         current_video_frame.received_timestamp = now;
 406         current_audio_frame.received_timestamp = now;
 407
 408         if (current_video_frame.data != nullptr || current_audio_frame.data != nullptr) {
 409                 // TODO: Put into a queue and put into a dequeue thread, if the
 410                 // BlackMagic drivers don't already do that for us?
 411                 frame_callback(timecode,
 412                         current_video_frame, /*video_offset=*/0, video_format,
 413                         current_audio_frame, /*audio_offset=*/0, audio_format);
 414         }
 415
 416         timecode++;
 417         return S_OK;
 418 }
 419
 420 void DeckLinkCapture::configure_card()
 421 {
 422         if (video_frame_allocator == nullptr) {
 423                 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
 424                 set_video_frame_allocator(owned_video_frame_allocator.get());
 425         }
 426         if (audio_frame_allocator == nullptr) {
 427                 owned_audio_frame_allocator.reset(new MallocFrameAllocator(65536, NUM_QUEUED_AUDIO_FRAMES));
 428                 set_audio_frame_allocator(owned_audio_frame_allocator.get());
 429         }
 430 }
 431
 432 void DeckLinkCapture::start_bm_capture()
 433 {
 434         if (running) {
 435                 return;
 436         }
 437         if (input->EnableVideoInput(current_video_mode, pixel_format_to_bmd(current_pixel_format), supports_autodetect ? bmdVideoInputEnableFormatDetection : 0) != S_OK) {
 438                 fprintf(stderr, "Failed to set video mode 0x%04x for card %d\n", current_video_mode, card_index);
 439                 exit(1);
 440         }
 441         if (input->EnableAudioInput(48000, bmdAudioSampleType32bitInteger, 2) != S_OK) {
 442                 fprintf(stderr, "Failed to enable audio input for card %d\n", card_index);
 443                 exit(1);
 444         }
 445
 446         if (input->StartStreams() != S_OK) {
 447                 fprintf(stderr, "StartStreams failed\n");
 448                 exit(1);
 449         }
 450         running = true;
 451 }
 452
 453 void DeckLinkCapture::stop_dequeue_thread()
 454 {
 455         if (!running) {
 456                 return;
 457         }
 458         HRESULT result = input->StopStreams();
 459         if (result != S_OK) {
 460                 fprintf(stderr, "StopStreams failed with error 0x%x\n", result);
 461                 exit(1);
 462         }
 463         if (input->DisableVideoInput() != S_OK) {
 464                 fprintf(stderr, "Failed to disable video input for card %d\n", card_index);
 465                 exit(1);
 466         }
 467         if (input->DisableAudioInput() != S_OK) {
 468                 fprintf(stderr, "Failed to disable audio input for card %d\n", card_index);
 469                 exit(1);
 470         }
 471         running = false;
 472 }
 473
 474 void DeckLinkCapture::set_video_mode(uint32_t video_mode_id)
 475 {
 476         if (running) {
 477                 if (input->PauseStreams() != S_OK) {
 478                         fprintf(stderr, "PauseStreams failed\n");
 479                         exit(1);
 480                 }
 481                 if (input->FlushStreams() != S_OK) {
 482                         fprintf(stderr, "FlushStreams failed\n");
 483                         exit(1);
 484                 }
 485         }
 486
 487         set_video_mode_no_restart(video_mode_id);
 488
 489         if (running) {
 490                 if (input->StartStreams() != S_OK) {
 491                         fprintf(stderr, "StartStreams failed\n");
 492                         exit(1);
 493                 }
 494         }
 495 }
 496
 497 void DeckLinkCapture::set_pixel_format(PixelFormat pixel_format)
 498 {
 499         current_pixel_format = pixel_format;
 500         set_video_mode(current_video_mode);
 501 }
 502
 503 void DeckLinkCapture::set_video_mode_no_restart(uint32_t video_mode_id)
 504 {
 505         BMDDisplayModeSupport support;
 506         IDeckLinkDisplayMode *display_mode;
 507         if (input->DoesSupportVideoMode(video_mode_id, pixel_format_to_bmd(current_pixel_format), /*flags=*/0, &support, &display_mode)) {
 508                 fprintf(stderr, "Failed to query display mode for card %d\n", card_index);
 509                 exit(1);
 510         }
 511
 512         if (support == bmdDisplayModeNotSupported) {
 513                 fprintf(stderr, "Card %d does not support display mode\n", card_index);
 514                 exit(1);
 515         }
 516
 517         if (display_mode->GetFrameRate(&frame_duration, &time_scale) != S_OK) {
 518                 fprintf(stderr, "Could not get frame rate for card %d\n", card_index);
 519                 exit(1);
 520         }
 521
 522         field_dominance = display_mode->GetFieldDominance();
 523
 524         if (running) {
 525                 if (input->EnableVideoInput(video_mode_id, pixel_format_to_bmd(current_pixel_format), supports_autodetect ? bmdVideoInputEnableFormatDetection : 0) != S_OK) {
 526                         fprintf(stderr, "Failed to set video mode 0x%04x for card %d\n", video_mode_id, card_index);
 527                         exit(1);
 528                 }
 529         }
 530
 531         current_video_mode = video_mode_id;
 532 }
 533
 534 void DeckLinkCapture::set_video_input(uint32_t video_input_id)
 535 {
 536         if (config->SetInt(bmdDeckLinkConfigVideoInputConnection, video_input_id) != S_OK) {
 537                 fprintf(stderr, "Failed to set video input connection for card %d\n", card_index);
 538                 exit(1);
 539         }
 540
 541         current_video_input = video_input_id;
 542 }
 543
 544 void DeckLinkCapture::set_audio_input(uint32_t audio_input_id)
 545 {
 546         if (config->SetInt(bmdDeckLinkConfigAudioInputConnection, audio_input_id) != S_OK) {
 547                 fprintf(stderr, "Failed to set audio input connection for card %d\n", card_index);
 548                 exit(1);
 549         }
 550
 551         current_audio_input = audio_input_id;
 552 }