git.sesse.net Git - bmusb/blob - bmusb.cpp

   1 // Intensity Shuttle USB3 prototype capture driver, v0.3
   2 // Can download 8-bit and 10-bit UYVY/v210 frames from HDMI, quite stable
   3 // (can do captures for hours at a time with no drops), except during startup
   4 // 576p60/720p60/1080i60 works, 1080p60 does not work (firmware limitation)
   5 // Audio comes out as 8-channel 24-bit raw audio.
   6
   7 #include <assert.h>
   8 #include <errno.h>
   9 #include <libusb.h>
  10 #include <netinet/in.h>
  11 #include <sched.h>
  12 #include <stdint.h>
  13 #include <stdio.h>
  14 #include <stdlib.h>
  15 #include <string.h>
  16 #ifdef __SSE4_1__
  17 #include <immintrin.h>
  18 #endif
  19 #include "bmusb.h"
  20
  21 #include <algorithm>
  22 #include <atomic>
  23 #include <condition_variable>
  24 #include <cstddef>
  25 #include <cstdint>
  26 #include <deque>
  27 #include <functional>
  28 #include <memory>
  29 #include <mutex>
  30 #include <stack>
  31 #include <thread>
  32
  33 using namespace std;
  34 using namespace std::placeholders;
  35
  36 #define WIDTH 1280
  37 #define HEIGHT 750  /* 30 lines ancillary data? */
  38 //#define WIDTH 1920
  39 //#define HEIGHT 1125  /* ??? lines ancillary data? */
  40 #define HEADER_SIZE 44
  41 //#define HEADER_SIZE 0
  42 #define AUDIO_HEADER_SIZE 4
  43
  44 //#define FRAME_SIZE (WIDTH * HEIGHT * 2 + HEADER_SIZE)  // UYVY
  45 //#define FRAME_SIZE (WIDTH * HEIGHT * 2 * 4 / 3 + HEADER_SIZE)  // v210
  46 #define FRAME_SIZE (8 << 20)
  47
  48 FILE *audiofp;
  49
  50 thread usb_thread;
  51 atomic<bool> should_quit;
  52
  53 FrameAllocator::~FrameAllocator() {}
  54
  55 // Audio is more important than video, and also much cheaper.
  56 // By having many more audio frames available, hopefully if something
  57 // starts to drop, we'll have CPU load go down (from not having to
  58 // process as much video) before we have to drop audio.
  59 #define NUM_QUEUED_VIDEO_FRAMES 16
  60 #define NUM_QUEUED_AUDIO_FRAMES 64
  61
  62 class MallocFrameAllocator : public FrameAllocator {
  63 public:
  64         MallocFrameAllocator(size_t frame_size, size_t num_queued_frames);
  65         Frame alloc_frame() override;
  66         void release_frame(Frame frame) override;
  67
  68 private:
  69         size_t frame_size;
  70
  71         mutex freelist_mutex;
  72         stack<unique_ptr<uint8_t[]>> freelist;  // All of size <frame_size>.
  73 };
  74
  75 MallocFrameAllocator::MallocFrameAllocator(size_t frame_size, size_t num_queued_frames)
  76         : frame_size(frame_size)
  77 {
  78         for (size_t i = 0; i < num_queued_frames; ++i) {
  79                 freelist.push(unique_ptr<uint8_t[]>(new uint8_t[frame_size]));
  80         }
  81 }
  82
  83 FrameAllocator::Frame MallocFrameAllocator::alloc_frame()
  84 {
  85         Frame vf;
  86         vf.owner = this;
  87
  88         unique_lock<mutex> lock(freelist_mutex);  // Meh.
  89         if (freelist.empty()) {
  90                 printf("Frame overrun (no more spare frames of size %ld), dropping frame!\n",
  91                         frame_size);
  92         } else {
  93                 vf.data = freelist.top().release();
  94                 vf.size = frame_size;
  95                 freelist.pop();  // Meh.
  96         }
  97         return vf;
  98 }
  99
 100 void MallocFrameAllocator::release_frame(Frame frame)
 101 {
 102         if (frame.overflow > 0) {
 103                 printf("%d bytes overflow after last (malloc) frame\n", int(frame.overflow));
 104         }
 105         unique_lock<mutex> lock(freelist_mutex);
 106         freelist.push(unique_ptr<uint8_t[]>(frame.data));
 107 }
 108
 109 bool uint16_less_than_with_wraparound(uint16_t a, uint16_t b)
 110 {
 111         if (a == b) {
 112                 return false;
 113         } else if (a < b) {
 114                 return (b - a < 0x8000);
 115         } else {
 116                 int wrap_b = 0x10000 + int(b);
 117                 return (wrap_b - a < 0x8000);
 118         }
 119 }
 120
 121 void BMUSBCapture::queue_frame(uint16_t format, uint16_t timecode, FrameAllocator::Frame frame, deque<QueuedFrame> *q)
 122 {
 123         if (!q->empty() && !uint16_less_than_with_wraparound(q->back().timecode, timecode)) {
 124                 printf("Blocks going backwards: prev=0x%04x, cur=0x%04x (dropped)\n",
 125                         q->back().timecode, timecode);
 126                 frame.owner->release_frame(frame);
 127                 return;
 128         }
 129
 130         QueuedFrame qf;
 131         qf.format = format;
 132         qf.timecode = timecode;
 133         qf.frame = frame;
 134
 135         {
 136                 unique_lock<mutex> lock(queue_lock);
 137                 q->push_back(move(qf));
 138         }
 139         queues_not_empty.notify_one();  // might be spurious
 140 }
 141
 142 void dump_frame(const char *filename, uint8_t *frame_start, size_t frame_len)
 143 {
 144         FILE *fp = fopen(filename, "wb");
 145         if (fwrite(frame_start + HEADER_SIZE, frame_len - HEADER_SIZE, 1, fp) != 1) {
 146                 printf("short write!\n");
 147         }
 148         fclose(fp);
 149 }
 150
 151 void dump_audio_block(uint8_t *audio_start, size_t audio_len)
 152 {
 153         fwrite(audio_start + AUDIO_HEADER_SIZE, 1, audio_len - AUDIO_HEADER_SIZE, audiofp);
 154 }
 155
 156 void BMUSBCapture::dequeue_thread_func()
 157 {
 158         if (has_dequeue_callbacks) {
 159                 dequeue_init_callback();
 160         }
 161         while (!dequeue_thread_should_quit) {
 162                 unique_lock<mutex> lock(queue_lock);
 163                 queues_not_empty.wait(lock, [this]{ return dequeue_thread_should_quit || (!pending_video_frames.empty() && !pending_audio_frames.empty()); });
 164
 165                 uint16_t video_timecode = pending_video_frames.front().timecode;
 166                 uint16_t audio_timecode = pending_audio_frames.front().timecode;
 167                 if (uint16_less_than_with_wraparound(video_timecode, audio_timecode)) {
 168                         printf("Video block 0x%04x without corresponding audio block, dropping.\n",
 169                                 video_timecode);
 170                         video_frame_allocator->release_frame(pending_video_frames.front().frame);
 171                         pending_video_frames.pop_front();
 172                 } else if (uint16_less_than_with_wraparound(audio_timecode, video_timecode)) {
 173                         printf("Audio block 0x%04x without corresponding video block, sending blank frame.\n",
 174                                 audio_timecode);
 175                         QueuedFrame audio_frame = pending_audio_frames.front();
 176                         pending_audio_frames.pop_front();
 177                         lock.unlock();
 178                         frame_callback(audio_timecode,
 179                                        FrameAllocator::Frame(), 0, 0x0000,
 180                                        audio_frame.frame, AUDIO_HEADER_SIZE, audio_frame.format);
 181                 } else {
 182                         QueuedFrame video_frame = pending_video_frames.front();
 183                         QueuedFrame audio_frame = pending_audio_frames.front();
 184                         pending_audio_frames.pop_front();
 185                         pending_video_frames.pop_front();
 186                         lock.unlock();
 187
 188 #if 0
 189                         char filename[255];
 190                         snprintf(filename, sizeof(filename), "%04x%04x.uyvy", video_frame.format, video_timecode);
 191                         dump_frame(filename, video_frame.frame.data, video_frame.data_len);
 192                         dump_audio_block(audio_frame.frame.data, audio_frame.data_len);
 193 #endif
 194
 195                         frame_callback(video_timecode,
 196                                        video_frame.frame, HEADER_SIZE, video_frame.format,
 197                                        audio_frame.frame, AUDIO_HEADER_SIZE, audio_frame.format);
 198                 }
 199         }
 200         if (has_dequeue_callbacks) {
 201                 dequeue_cleanup_callback();
 202         }
 203 }
 204
 205 void BMUSBCapture::start_new_frame(const uint8_t *start)
 206 {
 207         uint16_t format = (start[3] << 8) | start[2];
 208         uint16_t timecode = (start[1] << 8) | start[0];
 209
 210         if (current_video_frame.len > 0) {
 211                 // If format is 0x0800 (no signal), add a fake (empty) audio
 212                 // frame to get it out of the queue.
 213                 // TODO: Figure out if there are other formats that come with
 214                 // no audio, and treat them the same.
 215                 if (format == 0x0800) {
 216                         FrameAllocator::Frame fake_audio_frame = audio_frame_allocator->alloc_frame();
 217                         if (fake_audio_frame.data == nullptr) {
 218                                 // Oh well, it's just a no-signal frame anyway.
 219                                 printf("Couldn't allocate fake audio frame, also dropping no-signal video frame.\n");
 220                                 current_video_frame.owner->release_frame(current_video_frame);
 221                                 current_video_frame = video_frame_allocator->alloc_frame();
 222                                 return;
 223                         }
 224                         queue_frame(format, timecode, fake_audio_frame, &pending_audio_frames);
 225                 }
 226                 //dump_frame();
 227                 queue_frame(format, timecode, current_video_frame, &pending_video_frames);
 228         }
 229         //printf("Found frame start, format 0x%04x timecode 0x%04x, previous frame length was %d/%d\n",
 230         //      format, timecode,
 231         //      //start[7], start[6], start[5], start[4],
 232         //      read_current_frame, FRAME_SIZE);
 233
 234         current_video_frame = video_frame_allocator->alloc_frame();
 235         //if (current_video_frame.data == nullptr) {
 236         //      read_current_frame = -1;
 237         //} else {
 238         //      read_current_frame = 0;
 239         //}
 240 }
 241
 242 void BMUSBCapture::start_new_audio_block(const uint8_t *start)
 243 {
 244         uint16_t format = (start[3] << 8) | start[2];
 245         uint16_t timecode = (start[1] << 8) | start[0];
 246         if (current_audio_frame.len > 0) {
 247                 //dump_audio_block();
 248                 queue_frame(format, timecode, current_audio_frame, &pending_audio_frames);
 249         }
 250         //printf("Found audio block start, format 0x%04x timecode 0x%04x, previous block length was %d\n",
 251         //      format, timecode, read_current_audio_block);
 252         current_audio_frame = audio_frame_allocator->alloc_frame();
 253 }
 254
 255 #if 0
 256 static void dump_pack(const libusb_transfer *xfr, int offset, const libusb_iso_packet_descriptor *pack)
 257 {
 258         //      printf("ISO pack%u length:%u, actual_length:%u, offset:%u\n", i, pack->length, pack->actual_length, offset);
 259         for (unsigned j = 0; j < pack->actual_length; j++) {
 260         //for (int j = 0; j < min(pack->actual_length, 16u); j++) {
 261                 printf("%02x", xfr->buffer[j + offset]);
 262                 if ((j % 16) == 15)
 263                         printf("\n");
 264                 else if ((j % 8) == 7)
 265                         printf("  ");
 266                 else
 267                         printf(" ");
 268         }
 269 }
 270 #endif
 271
 272 void memcpy_interleaved(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size_t n)
 273 {
 274         assert(n % 2 == 0);
 275         uint8_t *dptr1 = dest1;
 276         uint8_t *dptr2 = dest2;
 277
 278         for (size_t i = 0; i < n; i += 2) {
 279                 *dptr1++ = *src++;
 280                 *dptr2++ = *src++;
 281         }
 282 }
 283
 284 void add_to_frame(FrameAllocator::Frame *current_frame, const char *frame_type_name, const uint8_t *start, const uint8_t *end)
 285 {
 286         if (current_frame->data == nullptr ||
 287             current_frame->len > current_frame->size ||
 288             start == end) {
 289                 return;
 290         }
 291
 292         int bytes = end - start;
 293         if (current_frame->len + bytes > current_frame->size) {
 294                 current_frame->overflow = current_frame->len + bytes - current_frame->size;
 295                 current_frame->len = current_frame->size;
 296                 if (current_frame->overflow > 1048576) {
 297                         printf("%d bytes overflow after last %s frame\n",
 298                                 int(current_frame->overflow), frame_type_name);
 299                         current_frame->overflow = 0;
 300                 }
 301                 //dump_frame();
 302         } else {
 303                 if (current_frame->interleaved) {
 304                         uint8_t *data = current_frame->data + current_frame->len / 2;
 305                         uint8_t *data2 = current_frame->data2 + current_frame->len / 2;
 306                         if (current_frame->len % 2 == 1) {
 307                                 ++data;
 308                                 swap(data, data2);
 309                         }
 310                         if (bytes % 2 == 1) {
 311                                 *data++ = *start++;
 312                                 swap(data, data2);
 313                                 ++current_frame->len;
 314                                 --bytes;
 315                         }
 316                         memcpy_interleaved(data, data2, start, bytes);
 317                         current_frame->len += bytes;
 318                 } else {
 319                         memcpy(current_frame->data + current_frame->len, start, bytes);
 320                         current_frame->len += bytes;
 321                 }
 322         }
 323 }
 324
 325 #ifdef __SSE4_1__
 326
 327 #if 0
 328 void avx2_dump(const char *name, __m256i n)
 329 {
 330         printf("%-10s:", name);
 331         printf(" %02x", _mm256_extract_epi8(n, 0));
 332         printf(" %02x", _mm256_extract_epi8(n, 1));
 333         printf(" %02x", _mm256_extract_epi8(n, 2));
 334         printf(" %02x", _mm256_extract_epi8(n, 3));
 335         printf(" %02x", _mm256_extract_epi8(n, 4));
 336         printf(" %02x", _mm256_extract_epi8(n, 5));
 337         printf(" %02x", _mm256_extract_epi8(n, 6));
 338         printf(" %02x", _mm256_extract_epi8(n, 7));
 339         printf(" ");
 340         printf(" %02x", _mm256_extract_epi8(n, 8));
 341         printf(" %02x", _mm256_extract_epi8(n, 9));
 342         printf(" %02x", _mm256_extract_epi8(n, 10));
 343         printf(" %02x", _mm256_extract_epi8(n, 11));
 344         printf(" %02x", _mm256_extract_epi8(n, 12));
 345         printf(" %02x", _mm256_extract_epi8(n, 13));
 346         printf(" %02x", _mm256_extract_epi8(n, 14));
 347         printf(" %02x", _mm256_extract_epi8(n, 15));
 348         printf(" ");
 349         printf(" %02x", _mm256_extract_epi8(n, 16));
 350         printf(" %02x", _mm256_extract_epi8(n, 17));
 351         printf(" %02x", _mm256_extract_epi8(n, 18));
 352         printf(" %02x", _mm256_extract_epi8(n, 19));
 353         printf(" %02x", _mm256_extract_epi8(n, 20));
 354         printf(" %02x", _mm256_extract_epi8(n, 21));
 355         printf(" %02x", _mm256_extract_epi8(n, 22));
 356         printf(" %02x", _mm256_extract_epi8(n, 23));
 357         printf(" ");
 358         printf(" %02x", _mm256_extract_epi8(n, 24));
 359         printf(" %02x", _mm256_extract_epi8(n, 25));
 360         printf(" %02x", _mm256_extract_epi8(n, 26));
 361         printf(" %02x", _mm256_extract_epi8(n, 27));
 362         printf(" %02x", _mm256_extract_epi8(n, 28));
 363         printf(" %02x", _mm256_extract_epi8(n, 29));
 364         printf(" %02x", _mm256_extract_epi8(n, 30));
 365         printf(" %02x", _mm256_extract_epi8(n, 31));
 366         printf("\n");
 367 }
 368 #endif
 369
 370 // Does a memcpy and memchr in one to reduce processing time.
 371 // Note that the benefit is somewhat limited if your L3 cache is small,
 372 // as you'll (unfortunately) spend most of the time loading the data
 373 // from main memory.
 374 //
 375 // Complicated cases are left to the slow path; it basically stops copying
 376 // up until the first instance of "sync_char" (usually a bit before, actually).
 377 // This is fine, since 0x00 bytes shouldn't really show up in normal picture
 378 // data, and what we really need this for is the 00 00 ff ff marker in video data.
 379 const uint8_t *add_to_frame_fastpath(FrameAllocator::Frame *current_frame, const uint8_t *start, const uint8_t *limit, const char sync_char)
 380 {
 381         if (current_frame->data == nullptr ||
 382             current_frame->len > current_frame->size ||
 383             start == limit) {
 384                 return start;
 385         }
 386         size_t orig_bytes = limit - start;
 387         if (orig_bytes < 128) {
 388                 // Don't bother.
 389                 return start;
 390         }
 391
 392         // Don't read more bytes than we can write.
 393         limit = min(limit, start + (current_frame->size - current_frame->len));
 394
 395         // Align end to 32 bytes.
 396         limit = (const uint8_t *)(intptr_t(limit) & ~31);
 397
 398         if (start >= limit) {
 399                 return start;
 400         }
 401
 402         // Process [0,31] bytes, such that start gets aligned to 32 bytes.
 403         const uint8_t *aligned_start = (const uint8_t *)(intptr_t(start + 31) & ~31);
 404         if (aligned_start != start) {
 405                 const uint8_t *sync_start = (const uint8_t *)memchr(start, sync_char, aligned_start - start);
 406                 if (sync_start == nullptr) {
 407                         add_to_frame(current_frame, "", start, aligned_start);
 408                 } else {
 409                         add_to_frame(current_frame, "", start, sync_start);
 410                         return sync_start;
 411                 }
 412         }
 413
 414         // Make the length a multiple of 64.
 415         if (current_frame->interleaved) {
 416                 if (((limit - aligned_start) % 64) != 0) {
 417                         limit -= 32;
 418                 }
 419                 assert(((limit - aligned_start) % 64) == 0);
 420         }
 421
 422 #if __AVX2__
 423         const __m256i needle = _mm256_set1_epi8(sync_char);
 424
 425         const __restrict __m256i *in = (const __m256i *)aligned_start;
 426         if (current_frame->interleaved) {
 427                 __restrict __m256i *out1 = (__m256i *)(current_frame->data + (current_frame->len + 1) / 2);
 428                 __restrict __m256i *out2 = (__m256i *)(current_frame->data2 + current_frame->len / 2);
 429                 if (current_frame->len % 2 == 1) {
 430                         swap(out1, out2);
 431                 }
 432
 433                 __m256i shuffle_cw = _mm256_set_epi8(
 434                         15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0,
 435                         15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0);
 436                 while (in < (const __m256i *)limit) {
 437                         // Note: For brevity, comments show lanes as if they were 2x64-bit (they're actually 2x128).
 438                         __m256i data1 = _mm256_stream_load_si256(in);         // AaBbCcDd EeFfGgHh
 439                         __m256i data2 = _mm256_stream_load_si256(in + 1);     // IiJjKkLl MmNnOoPp
 440
 441                         __m256i found1 = _mm256_cmpeq_epi8(data1, needle);
 442                         __m256i found2 = _mm256_cmpeq_epi8(data2, needle);
 443                         __m256i found = _mm256_or_si256(found1, found2);
 444
 445                         data1 = _mm256_shuffle_epi8(data1, shuffle_cw);       // ABCDabcd EFGHefgh
 446                         data2 = _mm256_shuffle_epi8(data2, shuffle_cw);       // IJKLijkl MNOPmnop
 447
 448                         data1 = _mm256_permute4x64_epi64(data1, 0b11011000);  // ABCDEFGH abcdefgh
 449                         data2 = _mm256_permute4x64_epi64(data2, 0b11011000);  // IJKLMNOP ijklmnop
 450
 451                         __m256i lo = _mm256_permute2x128_si256(data1, data2, 0b00100000);
 452                         __m256i hi = _mm256_permute2x128_si256(data1, data2, 0b00110001);
 453
 454                         _mm256_storeu_si256(out1, lo);  // Store as early as possible, even if the data isn't used.
 455                         _mm256_storeu_si256(out2, hi);
 456
 457                         if (!_mm256_testz_si256(found, found)) {
 458                                 break;
 459                         }
 460
 461                         in += 2;
 462                         ++out1;
 463                         ++out2;
 464                 }
 465                 current_frame->len += (uint8_t *)in - aligned_start;
 466         } else {
 467                 __m256i *out = (__m256i *)(current_frame->data + current_frame->len);
 468                 while (in < (const __m256i *)limit) {
 469                         __m256i data = _mm256_load_si256(in);
 470                         _mm256_storeu_si256(out, data);  // Store as early as possible, even if the data isn't used.
 471                         __m256i found = _mm256_cmpeq_epi8(data, needle);
 472                         if (!_mm256_testz_si256(found, found)) {
 473                                 break;
 474                         }
 475
 476                         ++in;
 477                         ++out;
 478                 }
 479                 current_frame->len = (uint8_t *)out - current_frame->data;
 480         }
 481 #else
 482         const __m128i needle = _mm_set1_epi8(sync_char);
 483
 484         const __m128i *in = (const __m128i *)aligned_start;
 485         if (current_frame->interleaved) {
 486                 __m128i *out1 = (__m128i *)(current_frame->data + (current_frame->len + 1) / 2);
 487                 __m128i *out2 = (__m128i *)(current_frame->data2 + current_frame->len / 2);
 488                 if (current_frame->len % 2 == 1) {
 489                         swap(out1, out2);
 490                 }
 491
 492                 __m128i mask_lower_byte = _mm_set1_epi16(0x00ff);
 493                 while (in < (const __m128i *)limit) {
 494                         __m128i data1 = _mm_load_si128(in);
 495                         __m128i data2 = _mm_load_si128(in + 1);
 496                         __m128i data1_lo = _mm_and_si128(data1, mask_lower_byte);
 497                         __m128i data2_lo = _mm_and_si128(data2, mask_lower_byte);
 498                         __m128i data1_hi = _mm_srli_epi16(data1, 8);
 499                         __m128i data2_hi = _mm_srli_epi16(data2, 8);
 500                         __m128i lo = _mm_packus_epi16(data1_lo, data2_lo);
 501                         _mm_storeu_si128(out1, lo);  // Store as early as possible, even if the data isn't used.
 502                         __m128i hi = _mm_packus_epi16(data1_hi, data2_hi);
 503                         _mm_storeu_si128(out2, hi);
 504                         __m128i found1 = _mm_cmpeq_epi8(data1, needle);
 505                         __m128i found2 = _mm_cmpeq_epi8(data2, needle);
 506                         if (!_mm_testz_si128(found1, found1) ||
 507                             !_mm_testz_si128(found2, found2)) {
 508                                 break;
 509                         }
 510
 511                         in += 2;
 512                         ++out1;
 513                         ++out2;
 514                 }
 515                 current_frame->len += (uint8_t *)in - aligned_start;
 516         } else {
 517                 __m128i *out = (__m128i *)(current_frame->data + current_frame->len);
 518                 while (in < (const __m128i *)limit) {
 519                         __m128i data = _mm_load_si128(in);
 520                         _mm_storeu_si128(out, data);  // Store as early as possible, even if the data isn't used.
 521                         __m128i found = _mm_cmpeq_epi8(data, needle);
 522                         if (!_mm_testz_si128(found, found)) {
 523                                 break;
 524                         }
 525
 526                         ++in;
 527                         ++out;
 528                 }
 529                 current_frame->len = (uint8_t *)out - current_frame->data;
 530         }
 531 #endif
 532
 533         //printf("managed to fastpath %ld/%ld bytes\n", (const uint8_t *)in - (const uint8_t *)aligned_start, orig_bytes);
 534
 535         return (const uint8_t *)in;
 536 }
 537 #endif
 538
 539 void decode_packs(const libusb_transfer *xfr,
 540                   const char *sync_pattern,
 541                   int sync_length,
 542                   FrameAllocator::Frame *current_frame,
 543                   const char *frame_type_name,
 544                   function<void(const uint8_t *start)> start_callback)
 545 {
 546         int offset = 0;
 547         for (int i = 0; i < xfr->num_iso_packets; i++) {
 548                 const libusb_iso_packet_descriptor *pack = &xfr->iso_packet_desc[i];
 549
 550                 if (pack->status != LIBUSB_TRANSFER_COMPLETED) {
 551                         fprintf(stderr, "Error: pack %u/%u status %d\n", i, xfr->num_iso_packets, pack->status);
 552                         continue;
 553 //exit(5);
 554                 }
 555
 556                 const uint8_t *start = xfr->buffer + offset;
 557                 const uint8_t *limit = start + pack->actual_length;
 558                 while (start < limit) {  // Usually runs only one iteration.
 559 #ifdef __SSE4_1__
 560                         start = add_to_frame_fastpath(current_frame, start, limit, sync_pattern[0]);
 561                         if (start == limit) break;
 562                         assert(start < limit);
 563 #endif
 564
 565                         const unsigned char* start_next_frame = (const unsigned char *)memmem(start, limit - start, sync_pattern, sync_length);
 566                         if (start_next_frame == nullptr) {
 567                                 // add the rest of the buffer
 568                                 add_to_frame(current_frame, frame_type_name, start, limit);
 569                                 break;
 570                         } else {
 571                                 add_to_frame(current_frame, frame_type_name, start, start_next_frame);
 572                                 start = start_next_frame + sync_length;  // skip sync
 573                                 start_callback(start);
 574                         }
 575                 }
 576 #if 0
 577                 dump_pack(xfr, offset, pack);
 578 #endif
 579                 offset += pack->length;
 580         }
 581 }
 582
 583 void BMUSBCapture::cb_xfr(struct libusb_transfer *xfr)
 584 {
 585         if (xfr->status != LIBUSB_TRANSFER_COMPLETED) {
 586                 fprintf(stderr, "transfer status %d\n", xfr->status);
 587                 libusb_free_transfer(xfr);
 588                 exit(3);
 589         }
 590
 591         assert(xfr->user_data != nullptr);
 592         BMUSBCapture *usb = static_cast<BMUSBCapture *>(xfr->user_data);
 593
 594         if (xfr->type == LIBUSB_TRANSFER_TYPE_ISOCHRONOUS) {
 595                 if (xfr->endpoint == 0x84) {
 596                         decode_packs(xfr, "DeckLinkAudioResyncT", 20, &usb->current_audio_frame, "audio", bind(&BMUSBCapture::start_new_audio_block, usb, _1));
 597                 } else {
 598                         decode_packs(xfr, "\x00\x00\xff\xff", 4, &usb->current_video_frame, "video", bind(&BMUSBCapture::start_new_frame, usb, _1));
 599                 }
 600         }
 601         if (xfr->type == LIBUSB_TRANSFER_TYPE_CONTROL) {
 602                 //const libusb_control_setup *setup = libusb_control_transfer_get_setup(xfr);
 603                 uint8_t *buf = libusb_control_transfer_get_data(xfr);
 604 #if 0
 605                 if (setup->wIndex == 44) {
 606                         printf("read timer register: 0x%02x%02x%02x%02x\n", buf[0], buf[1], buf[2], buf[3]);
 607                 } else {
 608                         printf("read register %2d:                      0x%02x%02x%02x%02x\n",
 609                                 setup->wIndex, buf[0], buf[1], buf[2], buf[3]);
 610                 }
 611 #else
 612                 memcpy(usb->register_file + usb->current_register, buf, 4);
 613                 usb->current_register = (usb->current_register + 4) % NUM_BMUSB_REGISTERS;
 614                 if (usb->current_register == 0) {
 615                         // read through all of them
 616                         printf("register dump:");
 617                         for (int i = 0; i < NUM_BMUSB_REGISTERS; i += 4) {
 618                                 printf(" 0x%02x%02x%02x%02x", usb->register_file[i], usb->register_file[i + 1], usb->register_file[i + 2], usb->register_file[i + 3]);
 619                         }
 620                         printf("\n");
 621                 }
 622                 libusb_fill_control_setup(xfr->buffer,
 623                     LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN, /*request=*/214, /*value=*/0,
 624                         /*index=*/usb->current_register, /*length=*/4);
 625 #endif
 626         }
 627
 628 #if 0
 629         printf("length:%u, actual_length:%u\n", xfr->length, xfr->actual_length);
 630         for (i = 0; i < xfr->actual_length; i++) {
 631                 printf("%02x", xfr->buffer[i]);
 632                 if (i % 16)
 633                         printf("\n");
 634                 else if (i % 8)
 635                         printf("  ");
 636                 else
 637                         printf(" ");
 638         }
 639 #endif
 640
 641         if (libusb_submit_transfer(xfr) < 0) {
 642                 fprintf(stderr, "error re-submitting URB\n");
 643                 exit(1);
 644         }
 645 }
 646
 647 void BMUSBCapture::usb_thread_func()
 648 {
 649         sched_param param;
 650         memset(&param, 0, sizeof(param));
 651         param.sched_priority = 1;
 652         if (sched_setscheduler(0, SCHED_RR, &param) == -1) {
 653                 printf("couldn't set realtime priority for USB thread: %s\n", strerror(errno));
 654         }
 655         while (!should_quit) {
 656                 int rc = libusb_handle_events(nullptr);
 657                 if (rc != LIBUSB_SUCCESS)
 658                         break;
 659         }
 660 }
 661
 662 void BMUSBCapture::configure_card()
 663 {
 664         if (video_frame_allocator == nullptr) {
 665                 set_video_frame_allocator(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));  // FIXME: leak.
 666         }
 667         if (audio_frame_allocator == nullptr) {
 668                 set_audio_frame_allocator(new MallocFrameAllocator(65536, NUM_QUEUED_AUDIO_FRAMES));  // FIXME: leak.
 669         }
 670         dequeue_thread_should_quit = false;
 671         dequeue_thread = thread(&BMUSBCapture::dequeue_thread_func, this);
 672
 673         int rc;
 674         struct libusb_transfer *xfr;
 675
 676         rc = libusb_init(nullptr);
 677         if (rc < 0) {
 678                 fprintf(stderr, "Error initializing libusb: %s\n", libusb_error_name(rc));
 679                 exit(1);
 680         }
 681
 682         //struct libusb_device_handle *devh = libusb_open_device_with_vid_pid(nullptr, 0x1edb, 0xbd3b);
 683         //struct libusb_device_handle *devh = libusb_open_device_with_vid_pid(nullptr, 0x1edb, 0xbd4f);
 684         struct libusb_device_handle *devh = libusb_open_device_with_vid_pid(nullptr, vid, pid);
 685         if (!devh) {
 686                 fprintf(stderr, "Error finding USB device\n");
 687                 exit(1);
 688         }
 689
 690         libusb_config_descriptor *config;
 691         rc = libusb_get_config_descriptor(libusb_get_device(devh), /*config_index=*/0, &config);
 692         if (rc < 0) {
 693                 fprintf(stderr, "Error getting configuration: %s\n", libusb_error_name(rc));
 694                 exit(1);
 695         }
 696         printf("%d interface\n", config->bNumInterfaces);
 697         for (int interface_number = 0; interface_number < config->bNumInterfaces; ++interface_number) {
 698                 printf("  interface %d\n", interface_number);
 699                 const libusb_interface *interface = &config->interface[interface_number];
 700                 for (int altsetting = 0; altsetting < interface->num_altsetting; ++altsetting) {
 701                         const libusb_interface_descriptor *interface_desc = &interface->altsetting[altsetting];
 702                         printf("    alternate setting %d\n", interface_desc->bAlternateSetting);
 703                         for (int endpoint_number = 0; endpoint_number < interface_desc->bNumEndpoints; ++endpoint_number) {
 704                                 const libusb_endpoint_descriptor *endpoint = &interface_desc->endpoint[endpoint_number];
 705                                 printf("        endpoint address 0x%02x\n", endpoint->bEndpointAddress);
 706                         }
 707                 }
 708         }
 709
 710         rc = libusb_set_configuration(devh, /*configuration=*/1);
 711         if (rc < 0) {
 712                 fprintf(stderr, "Error setting configuration 1: %s\n", libusb_error_name(rc));
 713                 exit(1);
 714         }
 715
 716         rc = libusb_claim_interface(devh, 0);
 717         if (rc < 0) {
 718                 fprintf(stderr, "Error claiming interface 0: %s\n", libusb_error_name(rc));
 719                 exit(1);
 720         }
 721
 722         // Alternate setting 1 is output, alternate setting 2 is input.
 723         // Card is reset when switching alternates, so the driver uses
 724         // this “double switch” when it wants to reset.
 725         //
 726         // There's also alternate settings 3 and 4, which seem to be
 727         // like 1 and 2 except they advertise less bandwidth needed.
 728         rc = libusb_set_interface_alt_setting(devh, /*interface=*/0, /*alternate_setting=*/1);
 729         if (rc < 0) {
 730                 fprintf(stderr, "Error setting alternate 1: %s\n", libusb_error_name(rc));
 731                 exit(1);
 732         }
 733         rc = libusb_set_interface_alt_setting(devh, /*interface=*/0, /*alternate_setting=*/2);
 734         if (rc < 0) {
 735                 fprintf(stderr, "Error setting alternate 1: %s\n", libusb_error_name(rc));
 736                 exit(1);
 737         }
 738 #if 0
 739         rc = libusb_set_interface_alt_setting(devh, /*interface=*/0, /*alternate_setting=*/1);
 740         if (rc < 0) {
 741                 fprintf(stderr, "Error setting alternate 1: %s\n", libusb_error_name(rc));
 742                 exit(1);
 743         }
 744 #endif
 745
 746 #if 0
 747         rc = libusb_claim_interface(devh, 3);
 748         if (rc < 0) {
 749                 fprintf(stderr, "Error claiming interface 3: %s\n", libusb_error_name(rc));
 750                 exit(1);
 751         }
 752 #endif
 753
 754         // theories:
 755         //   44 is some kind of timer register (first 16 bits count upwards)
 756         //   24 is some sort of watchdog?
 757         //      you can seemingly set it to 0x73c60001 and that bit will eventually disappear
 758         //      (or will go to 0x73c60010?), also seen 0x73c60100
 759         //   12 also changes all the time, unclear why
 760         //   16 seems to be autodetected mode somehow
 761         //      --    this is e00115e0 after reset?
 762         //                    ed0115e0 after mode change [to output?]
 763         //                    2d0015e0 after more mode change [to input]
 764         //                    ed0115e0 after more mode change
 765         //                    2d0015e0 after more mode change
 766         //
 767         //                    390115e0 seems to indicate we have signal
 768         //         changes to 200115e0 when resolution changes/we lose signal, driver resets after a while
 769         //
 770         //                    200015e0 on startup
 771         //         changes to 250115e0 when we sync to the signal
 772         //
 773         //    so only first 16 bits count, and 0x0100 is a mask for ok/stable signal?
 774         //
 775         //    Bottom 16 bits of this register seem to be firmware version number (possibly not all all of them).
 776         //
 777         //    28 and 32 seems to be analog audio input levels (one byte for each of the eight channels).
 778         //    however, if setting 32 with HDMI embedded audio, it is immediately overwritten back (to 0xe137002a).
 779         //
 780         //    4, 8, 20 are unclear. seem to be some sort of bitmask, but we can set them to 0 with no apparent effect.
 781         //    perhaps some of them are related to analog output?
 782         //
 783         //    36 can be set to 0 with no apparent effect (all of this tested on both video and audio),
 784         //    but the driver sets it to 0x8036802a at some point.
 785         //
 786         //    all of this is on request 214/215. other requests (192, 219,
 787         //    222, 223, 224) are used for firmware upgrade. Probably best to
 788         //    stay out of it unless you know what you're doing.
 789         //
 790         //
 791         // register 16:
 792         // first byte is 0x39 for a stable 576p60 signal, 0x2d for a stable 720p60 signal, 0x20 for no signal
 793         //
 794         // theories:
 795         //   0x01 - stable signal
 796         //   0x04 - deep color
 797         //   0x08 - unknown (audio??)
 798         //   0x20 - 720p??
 799         //   0x30 - 576p??
 800
 801         struct ctrl {
 802                 int endpoint;
 803                 int request;
 804                 int index;
 805                 uint32_t data;
 806         };
 807         static const ctrl ctrls[] = {
 808                 { LIBUSB_ENDPOINT_IN,  214, 16, 0 },
 809                 { LIBUSB_ENDPOINT_IN,  214,  0, 0 },
 810
 811                 // seems to capture on HDMI, clearing the 0x20000000 bit seems to activate 10-bit
 812                 // capture (v210).
 813                 // clearing the 0x08000000 bit seems to change the capture format (other source?)
 814                 // 0x10000000 = analog audio instead of embedded audio, it seems
 815                 // 0x3a000000 = component video? (analog audio)
 816                 // 0x3c000000 = composite video? (analog audio)
 817                 // 0x3e000000 = s-video? (analog audio)
 818                 { LIBUSB_ENDPOINT_OUT, 215,  0, 0x29000000 },
 819                 //{ LIBUSB_ENDPOINT_OUT, 215,  0, 0x80000100 },
 820                 //{ LIBUSB_ENDPOINT_OUT, 215,  0, 0x09000000 },
 821                 { LIBUSB_ENDPOINT_OUT, 215, 24, 0x73c60001 },  // latch for frame start?
 822                 { LIBUSB_ENDPOINT_IN,  214, 24, 0 },  //
 823         };
 824
 825         for (unsigned req = 0; req < sizeof(ctrls) / sizeof(ctrls[0]); ++req) {
 826                 uint32_t flipped = htonl(ctrls[req].data);
 827                 static uint8_t value[4];
 828                 memcpy(value, &flipped, sizeof(flipped));
 829                 int size = sizeof(value);
 830                 //if (ctrls[req].request == 215) size = 0;
 831                 rc = libusb_control_transfer(devh, LIBUSB_REQUEST_TYPE_VENDOR | ctrls[req].endpoint,
 832                         /*request=*/ctrls[req].request, /*value=*/0, /*index=*/ctrls[req].index, value, size, /*timeout=*/0);
 833                 if (rc < 0) {
 834                         fprintf(stderr, "Error on control %d: %s\n", ctrls[req].index, libusb_error_name(rc));
 835                         exit(1);
 836                 }
 837
 838                 printf("rc=%d: ep=%d@%d %d -> 0x", rc, ctrls[req].endpoint, ctrls[req].request, ctrls[req].index);
 839                 for (int i = 0; i < rc; ++i) {
 840                         printf("%02x", value[i]);
 841                 }
 842                 printf("\n");
 843         }
 844
 845 #if 0
 846         // DEBUG
 847         for ( ;; ) {
 848                 static int my_index = 0;
 849                 static uint8_t value[4];
 850                 int size = sizeof(value);
 851                 rc = libusb_control_transfer(devh, LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN,
 852                         /*request=*/214, /*value=*/0, /*index=*/my_index, value, size, /*timeout=*/0);
 853                 if (rc < 0) {
 854                         fprintf(stderr, "Error on control\n");
 855                         exit(1);
 856                 }
 857                 printf("rc=%d index=%d: 0x", rc, my_index);
 858                 for (int i = 0; i < rc; ++i) {
 859                         printf("%02x", value[i]);
 860                 }
 861                 printf("\n");
 862         }
 863 #endif
 864
 865 #if 0
 866         // set up an asynchronous transfer of the timer register
 867         static uint8_t cmdbuf[LIBUSB_CONTROL_SETUP_SIZE + 4];
 868         static int completed = 0;
 869
 870         xfr = libusb_alloc_transfer(0);
 871         libusb_fill_control_setup(cmdbuf,
 872             LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN, /*request=*/214, /*value=*/0,
 873                 /*index=*/44, /*length=*/4);
 874         libusb_fill_control_transfer(xfr, devh, cmdbuf, cb_xfr, &completed, 0);
 875         xfr->user_data = this;
 876         libusb_submit_transfer(xfr);
 877
 878         // set up an asynchronous transfer of register 24
 879         static uint8_t cmdbuf2[LIBUSB_CONTROL_SETUP_SIZE + 4];
 880         static int completed2 = 0;
 881
 882         xfr = libusb_alloc_transfer(0);
 883         libusb_fill_control_setup(cmdbuf2,
 884             LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN, /*request=*/214, /*value=*/0,
 885                 /*index=*/24, /*length=*/4);
 886         libusb_fill_control_transfer(xfr, devh, cmdbuf2, cb_xfr, &completed2, 0);
 887         xfr->user_data = this;
 888         libusb_submit_transfer(xfr);
 889 #endif
 890
 891         // set up an asynchronous transfer of the register dump
 892         static uint8_t cmdbuf3[LIBUSB_CONTROL_SETUP_SIZE + 4];
 893         static int completed3 = 0;
 894
 895         xfr = libusb_alloc_transfer(0);
 896         libusb_fill_control_setup(cmdbuf3,
 897             LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN, /*request=*/214, /*value=*/0,
 898                 /*index=*/current_register, /*length=*/4);
 899         libusb_fill_control_transfer(xfr, devh, cmdbuf3, cb_xfr, &completed3, 0);
 900         xfr->user_data = this;
 901         //libusb_submit_transfer(xfr);
 902
 903         audiofp = fopen("audio.raw", "wb");
 904
 905         // set up isochronous transfers for audio and video
 906         for (int e = 3; e <= 4; ++e) {
 907                 //int num_transfers = (e == 3) ? 6 : 6;
 908                 int num_transfers = 6;
 909                 for (int i = 0; i < num_transfers; ++i) {
 910                         int num_iso_pack, size;
 911                         if (e == 3) {
 912                                 // Video seems to require isochronous packets scaled with the width;
 913                                 // seemingly six lines is about right, rounded up to the required 1kB
 914                                 // multiple.
 915                                 size = WIDTH * 2 * 6;
 916                                 // Note that for 10-bit input, you'll need to increase size accordingly.
 917                                 //size = size * 4 / 3;
 918                                 if (size % 1024 != 0) {
 919                                         size &= ~1023;
 920                                         size += 1024;
 921                                 }
 922                                 num_iso_pack = (2 << 18) / size;  // 512 kB.
 923                                 printf("Picking %d packets of 0x%x bytes each\n", num_iso_pack, size);
 924                         } else {
 925                                 size = 0xc0;
 926                                 num_iso_pack = 80;
 927                         }
 928                         int num_bytes = num_iso_pack * size;
 929                         uint8_t *buf = new uint8_t[num_bytes];
 930
 931                         xfr = libusb_alloc_transfer(num_iso_pack);
 932                         if (!xfr) {
 933                                 fprintf(stderr, "oom\n");
 934                                 exit(1);
 935                         }
 936
 937                         int ep = LIBUSB_ENDPOINT_IN | e;
 938                         libusb_fill_iso_transfer(xfr, devh, ep, buf, num_bytes,
 939                                 num_iso_pack, cb_xfr, nullptr, 0);
 940                         libusb_set_iso_packet_lengths(xfr, size);
 941                         xfr->user_data = this;
 942                         iso_xfrs.push_back(xfr);
 943                 }
 944         }
 945 }
 946
 947 void BMUSBCapture::start_bm_capture()
 948 {
 949         printf("starting capture\n");
 950         int i = 0;
 951         for (libusb_transfer *xfr : iso_xfrs) {
 952                 printf("submitting transfer...\n");
 953                 int rc = libusb_submit_transfer(xfr);
 954                 ++i;
 955                 if (rc < 0) {
 956                         //printf("num_bytes=%d\n", num_bytes);
 957                         fprintf(stderr, "Error submitting iso to endpoint 0x%02x, number %d: %s\n",
 958                                 xfr->endpoint, i, libusb_error_name(rc));
 959                         exit(1);
 960                 }
 961         }
 962
 963
 964 #if 0
 965         libusb_release_interface(devh, 0);
 966 out:
 967         if (devh)
 968                 libusb_close(devh);
 969         libusb_exit(nullptr);
 970         return rc;
 971 #endif
 972 }
 973
 974 void BMUSBCapture::stop_dequeue_thread()
 975 {
 976         dequeue_thread_should_quit = true;
 977         queues_not_empty.notify_all();
 978         dequeue_thread.join();
 979 }
 980
 981 void BMUSBCapture::start_bm_thread()
 982 {
 983         should_quit = false;
 984         usb_thread = thread(&BMUSBCapture::usb_thread_func);
 985 }
 986
 987 void BMUSBCapture::stop_bm_thread()
 988 {
 989         should_quit = true;
 990         usb_thread.join();
 991 }