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

   1 #include "decklink_capture.h"
   2
   3 #include <assert.h>
   4 #include <stdint.h>
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <cstddef>
   9 #ifdef __SSE2__
  10 #include <immintrin.h>
  11 #endif
  12
  13 #include <DeckLinkAPI.h>
  14 #include <DeckLinkAPIConfiguration.h>
  15 #include <DeckLinkAPIDiscovery.h>
  16 #include <DeckLinkAPIModes.h>
  17 #include "bmusb/bmusb.h"
  18
  19 #define FRAME_SIZE (8 << 20)  // 8 MB.
  20
  21 using namespace std;
  22 using namespace std::placeholders;
  23
  24 namespace {
  25
  26 // TODO: Support stride.
  27 void memcpy_interleaved(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size_t n)
  28 {
  29         assert(n % 2 == 0);
  30         uint8_t *dptr1 = dest1;
  31         uint8_t *dptr2 = dest2;
  32
  33         for (size_t i = 0; i < n; i += 2) {
  34                 *dptr1++ = *src++;
  35                 *dptr2++ = *src++;
  36         }
  37 }
  38
  39 #ifdef __SSE2__
  40
  41 // Returns the number of bytes consumed.
  42 size_t memcpy_interleaved_fastpath(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size_t n)
  43 {
  44         const uint8_t *limit = src + n;
  45         size_t consumed = 0;
  46
  47         // Align end to 32 bytes.
  48         limit = (const uint8_t *)(intptr_t(limit) & ~31);
  49
  50         if (src >= limit) {
  51                 return 0;
  52         }
  53
  54         // Process [0,31] bytes, such that start gets aligned to 32 bytes.
  55         const uint8_t *aligned_src = (const uint8_t *)(intptr_t(src + 31) & ~31);
  56         if (aligned_src != src) {
  57                 size_t n2 = aligned_src - src;
  58                 memcpy_interleaved(dest1, dest2, src, n2);
  59                 dest1 += n2 / 2;
  60                 dest2 += n2 / 2;
  61                 if (n2 % 1) {
  62                         swap(dest1, dest2);
  63                 }
  64                 src = aligned_src;
  65                 consumed += n2;
  66         }
  67
  68         // Make the length a multiple of 64.
  69         if (((limit - src) % 64) != 0) {
  70                 limit -= 32;
  71         }
  72         assert(((limit - src) % 64) == 0);
  73
  74 #if __AVX2__
  75         const __restrict __m256i *in = (const __m256i *)src;
  76         __restrict __m256i *out1 = (__m256i *)dest1;
  77         __restrict __m256i *out2 = (__m256i *)dest2;
  78
  79         __m256i shuffle_cw = _mm256_set_epi8(
  80                 15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0,
  81                 15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0);
  82         while (in < (const __m256i *)limit) {
  83                 // Note: For brevity, comments show lanes as if they were 2x64-bit (they're actually 2x128).
  84                 __m256i data1 = _mm256_stream_load_si256(in);         // AaBbCcDd EeFfGgHh
  85                 __m256i data2 = _mm256_stream_load_si256(in + 1);     // IiJjKkLl MmNnOoPp
  86
  87                 data1 = _mm256_shuffle_epi8(data1, shuffle_cw);       // ABCDabcd EFGHefgh
  88                 data2 = _mm256_shuffle_epi8(data2, shuffle_cw);       // IJKLijkl MNOPmnop
  89
  90                 data1 = _mm256_permute4x64_epi64(data1, 0b11011000);  // ABCDEFGH abcdefgh
  91                 data2 = _mm256_permute4x64_epi64(data2, 0b11011000);  // IJKLMNOP ijklmnop
  92
  93                 __m256i lo = _mm256_permute2x128_si256(data1, data2, 0b00100000);
  94                 __m256i hi = _mm256_permute2x128_si256(data1, data2, 0b00110001);
  95
  96                 _mm256_storeu_si256(out1, lo);
  97                 _mm256_storeu_si256(out2, hi);
  98
  99                 in += 2;
 100                 ++out1;
 101                 ++out2;
 102                 consumed += 64;
 103         }
 104 #else
 105         const __restrict __m128i *in = (const __m128i *)src;
 106         __restrict __m128i *out1 = (__m128i *)dest1;
 107         __restrict __m128i *out2 = (__m128i *)dest2;
 108
 109         __m128i mask_lower_byte = _mm_set1_epi16(0x00ff);
 110         while (in < (const __m128i *)limit) {
 111                 __m128i data1 = _mm_load_si128(in);
 112                 __m128i data2 = _mm_load_si128(in + 1);
 113                 __m128i data1_lo = _mm_and_si128(data1, mask_lower_byte);
 114                 __m128i data2_lo = _mm_and_si128(data2, mask_lower_byte);
 115                 __m128i data1_hi = _mm_srli_epi16(data1, 8);
 116                 __m128i data2_hi = _mm_srli_epi16(data2, 8);
 117                 __m128i lo = _mm_packus_epi16(data1_lo, data2_lo);
 118                 _mm_storeu_si128(out1, lo);
 119                 __m128i hi = _mm_packus_epi16(data1_hi, data2_hi);
 120                 _mm_storeu_si128(out2, hi);
 121
 122                 in += 2;
 123                 ++out1;
 124                 ++out2;
 125                 consumed += 32;
 126         }
 127 #endif
 128
 129         return consumed;
 130 }
 131
 132 #endif  // __SSE2__
 133
 134 }  // namespace
 135
 136 DeckLinkCapture::DeckLinkCapture(IDeckLink *card, int card_index)
 137 {
 138         {
 139                 const char *model_name;
 140                 char buf[256];
 141                 if (card->GetModelName(&model_name) == S_OK) {
 142                         snprintf(buf, sizeof(buf), "Card %d: %s", card_index, model_name);
 143                 } else {
 144                         snprintf(buf, sizeof(buf), "Card %d: Unknown DeckLink card", card_index);
 145                 }
 146                 description = buf;
 147         }
 148
 149         if (card->QueryInterface(IID_IDeckLinkInput, (void**)&input) != S_OK) {
 150                 fprintf(stderr, "Card %d has no inputs\n", card_index);
 151                 exit(1);
 152         }
 153
 154         /* Set up the video and audio sources. */
 155         IDeckLinkConfiguration *config;
 156         if (card->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
 157                 fprintf(stderr, "Failed to get configuration interface for card %d\n", card_index);
 158                 exit(1);
 159         }
 160
 161         if (config->SetInt(bmdDeckLinkConfigVideoInputConnection, bmdVideoConnectionHDMI) != S_OK) {
 162                 fprintf(stderr, "Failed to set video input connection for card %d\n", card_index);
 163                 exit(1);
 164         }
 165
 166         if (config->SetInt(bmdDeckLinkConfigAudioInputConnection, bmdAudioConnectionEmbedded) != S_OK) {
 167                 fprintf(stderr, "Failed to set video input connection for card %d\n", card_index);
 168                 exit(1);
 169         }
 170
 171         // TODO: Make the user mode selectable.
 172         BMDDisplayModeSupport support;
 173         IDeckLinkDisplayMode *display_mode;
 174         if (input->DoesSupportVideoMode(bmdModeHD720p5994, bmdFormat8BitYUV, /*flags=*/0, &support, &display_mode)) {
 175                 fprintf(stderr, "Failed to query display mode for card %d\n", card_index);
 176                 exit(1);
 177         }
 178
 179         if (support == bmdDisplayModeNotSupported) {
 180                 fprintf(stderr, "Card %d does not support display mode\n", card_index);
 181                 exit(1);
 182         }
 183
 184         if (display_mode->GetFrameRate(&frame_duration, &time_scale) != S_OK) {
 185                 fprintf(stderr, "Could not get frame rate for card %d\n", card_index);
 186                 exit(1);
 187         }
 188
 189         if (input->EnableVideoInput(bmdModeHD720p5994, bmdFormat8BitYUV, 0) != S_OK) {
 190                 fprintf(stderr, "Failed to set 720p59.94 connection for card %d\n", card_index);
 191                 exit(1);
 192         }
 193
 194         if (input->EnableAudioInput(48000, bmdAudioSampleType32bitInteger, 2) != S_OK) {
 195                 fprintf(stderr, "Failed to enable audio input for card %d\n", card_index);
 196                 exit(1);
 197         }
 198
 199         input->SetCallback(this);
 200 }
 201
 202 DeckLinkCapture::~DeckLinkCapture()
 203 {
 204         if (has_dequeue_callbacks) {
 205                 dequeue_cleanup_callback();
 206         }
 207 }
 208
 209 HRESULT STDMETHODCALLTYPE DeckLinkCapture::QueryInterface(REFIID, LPVOID *)
 210 {
 211         return E_NOINTERFACE;
 212 }
 213
 214 ULONG STDMETHODCALLTYPE DeckLinkCapture::AddRef(void)
 215 {
 216         return refcount.fetch_add(1) + 1;
 217 }
 218
 219 ULONG STDMETHODCALLTYPE DeckLinkCapture::Release(void)
 220 {
 221         int new_ref = refcount.fetch_sub(1) - 1;
 222         if (new_ref == 0)
 223                 delete this;
 224         return new_ref;
 225 }
 226
 227 HRESULT STDMETHODCALLTYPE DeckLinkCapture::VideoInputFormatChanged(
 228         BMDVideoInputFormatChangedEvents,
 229         IDeckLinkDisplayMode* display_mode,
 230         BMDDetectedVideoInputFormatFlags)
 231 {
 232         if (display_mode->GetFrameRate(&frame_duration, &time_scale) != S_OK) {
 233                 fprintf(stderr, "Could not get new frame rate\n");
 234                 exit(1);
 235         }
 236         return S_OK;
 237 }
 238
 239 HRESULT STDMETHODCALLTYPE DeckLinkCapture::VideoInputFrameArrived(
 240         IDeckLinkVideoInputFrame *video_frame,
 241         IDeckLinkAudioInputPacket *audio_frame)
 242 {
 243         if (!done_init) {
 244                 if (has_dequeue_callbacks) {
 245                         dequeue_init_callback();
 246                 }
 247                 done_init = true;
 248         }
 249
 250         FrameAllocator::Frame current_video_frame, current_audio_frame;
 251         VideoFormat video_format;
 252         AudioFormat audio_format;
 253
 254         if (video_frame) {
 255                 video_format.has_signal = !(video_frame->GetFlags() & bmdFrameHasNoInputSource);
 256
 257                 int width = video_frame->GetWidth();
 258                 int height = video_frame->GetHeight();
 259                 const int stride = video_frame->GetRowBytes();
 260                 assert(stride == width * 2);
 261
 262                 current_video_frame = video_frame_allocator->alloc_frame();
 263                 if (current_video_frame.data != nullptr) {
 264                         const uint8_t *frame_bytes;
 265                         video_frame->GetBytes((void **)&frame_bytes);
 266
 267                         uint8_t *data = current_video_frame.data;
 268                         uint8_t *data2 = current_video_frame.data2;
 269                         size_t num_bytes = width * height * 2;
 270 #ifdef __SSE2__
 271                         size_t consumed = memcpy_interleaved_fastpath(data, data2, frame_bytes, num_bytes);
 272                         frame_bytes += consumed;
 273                         data += consumed / 2;
 274                         data2 += consumed / 2;
 275                         if (num_bytes % 2) {
 276                                 swap(data, data2);
 277                         }
 278                         current_video_frame.len += consumed;
 279                         num_bytes -= consumed;
 280 #endif
 281
 282                         if (num_bytes > 0) {
 283                                 memcpy_interleaved(data, data2, frame_bytes, num_bytes);
 284                         }
 285                         current_video_frame.len += num_bytes;
 286
 287                         video_format.width = width;
 288                         video_format.height = height;
 289                         video_format.frame_rate_nom = time_scale;
 290                         video_format.frame_rate_den = frame_duration;
 291                 }
 292         }
 293
 294         if (audio_frame) {
 295                 int num_samples = audio_frame->GetSampleFrameCount();
 296
 297                 current_audio_frame = audio_frame_allocator->alloc_frame();
 298                 if (current_audio_frame.data != nullptr) {
 299                         const uint8_t *frame_bytes;
 300                         audio_frame->GetBytes((void **)&frame_bytes);
 301                         current_audio_frame.len = sizeof(int32_t) * 2 * num_samples;
 302
 303                         memcpy(current_audio_frame.data, frame_bytes, current_audio_frame.len);
 304
 305                         audio_format.bits_per_sample = 32;
 306                         audio_format.num_channels = 2;
 307                 }
 308         }
 309
 310         if (current_video_frame.data != nullptr || current_audio_frame.data != nullptr) {
 311                 // TODO: Put into a queue and put into a dequeue thread, if the
 312                 // BlackMagic drivers don't already do that for us?
 313                 frame_callback(timecode,
 314                         current_video_frame, /*video_offset=*/0, video_format,
 315                         current_audio_frame, /*audio_offset=*/0, audio_format);
 316         }
 317
 318         timecode++;
 319         return S_OK;
 320 }
 321
 322 void DeckLinkCapture::configure_card()
 323 {
 324         if (video_frame_allocator == nullptr) {
 325                 set_video_frame_allocator(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));  // FIXME: leak.
 326         }
 327         if (audio_frame_allocator == nullptr) {
 328                 set_audio_frame_allocator(new MallocFrameAllocator(65536, NUM_QUEUED_AUDIO_FRAMES));  // FIXME: leak.
 329         }
 330 }
 331
 332 void DeckLinkCapture::start_bm_capture()
 333 {
 334         if (input->StartStreams() != S_OK) {
 335                 fprintf(stderr, "StartStreams failed\n");
 336                 exit(1);
 337         }
 338 }
 339
 340 void DeckLinkCapture::stop_dequeue_thread()
 341 {
 342         if (input->StopStreams() != S_OK) {
 343                 fprintf(stderr, "StopStreams failed\n");
 344                 exit(1);
 345         }
 346 }
 347