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

   1 // Intensity Shuttle USB3 capture driver, v0.4
   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 #if (defined(__i386__) || defined(__x86_64__)) && defined(__GNUC__)
   8 #define HAS_MULTIVERSIONING 1
   9 #endif
  10
  11 #include <assert.h>
  12 #include <errno.h>
  13 #include <libusb.h>
  14 #include <unistd.h>
  15 #include <netinet/in.h>
  16 #include <sched.h>
  17 #include <stdint.h>
  18 #include <stdio.h>
  19 #include <stdlib.h>
  20 #include <string.h>
  21 #if HAS_MULTIVERSIONING
  22 #include <immintrin.h>
  23 #endif
  24 #include "bmusb/bmusb.h"
  25
  26 #include <algorithm>
  27 #include <atomic>
  28 #include <condition_variable>
  29 #include <cstddef>
  30 #include <cstdint>
  31 #include <deque>
  32 #include <functional>
  33 #include <memory>
  34 #include <mutex>
  35 #include <stack>
  36 #include <string>
  37 #include <thread>
  38
  39 using namespace std;
  40 using namespace std::placeholders;
  41
  42 #define USB_VENDOR_BLACKMAGIC 0x1edb
  43 #define MIN_WIDTH 640
  44 #define HEADER_SIZE 44
  45 //#define HEADER_SIZE 0
  46 #define AUDIO_HEADER_SIZE 4
  47
  48 #define FRAME_SIZE (8 << 20)  // 8 MB.
  49 #define USB_VIDEO_TRANSFER_SIZE (128 << 10)  // 128 kB.
  50
  51 namespace bmusb {
  52
  53 card_connected_callback_t BMUSBCapture::card_connected_callback = nullptr;
  54 bool BMUSBCapture::hotplug_existing_devices = false;
  55
  56 namespace {
  57
  58 FILE *audiofp;
  59
  60 thread usb_thread;
  61 atomic<bool> should_quit;
  62
  63 int find_xfer_size_for_width(int width)
  64 {
  65         // Video seems to require isochronous packets scaled with the width;
  66         // seemingly six lines is about right, rounded up to the required 1kB
  67         // multiple.
  68         int size = width * 2 * 6;
  69         // Note that for 10-bit input, you'll need to increase size accordingly.
  70         //size = size * 4 / 3;
  71         if (size % 1024 != 0) {
  72                 size &= ~1023;
  73                 size += 1024;
  74         }
  75         return size;
  76 }
  77
  78 void change_xfer_size_for_width(int width, libusb_transfer *xfr)
  79 {
  80         assert(width >= MIN_WIDTH);
  81         size_t size = find_xfer_size_for_width(width);
  82         int num_iso_pack = xfr->length / size;
  83         if (num_iso_pack != xfr->num_iso_packets ||
  84             size != xfr->iso_packet_desc[0].length) {
  85                 xfr->num_iso_packets = num_iso_pack;
  86                 libusb_set_iso_packet_lengths(xfr, size);
  87         }
  88 }
  89
  90 struct VideoFormatEntry {
  91         uint16_t normalized_video_format;
  92         unsigned width, height, second_field_start;
  93         unsigned extra_lines_top, extra_lines_bottom;
  94         unsigned frame_rate_nom, frame_rate_den;
  95         bool interlaced;
  96 };
  97
  98 // Get details for the given video format; returns false if detection was incomplete.
  99 bool decode_video_format(uint16_t video_format, VideoFormat *decoded_video_format)
 100 {
 101         decoded_video_format->id = video_format;
 102         decoded_video_format->interlaced = false;
 103
 104         // TODO: Add these for all formats as we find them.
 105         decoded_video_format->extra_lines_top = decoded_video_format->extra_lines_bottom = decoded_video_format->second_field_start = 0;
 106
 107         if (video_format == 0x0800) {
 108                 // No video signal. These green pseudo-frames seem to come at about 30.13 Hz.
 109                 // It's a strange thing, but what can you do.
 110                 decoded_video_format->width = 720;
 111                 decoded_video_format->height = 525;
 112                 decoded_video_format->extra_lines_top = 0;
 113                 decoded_video_format->extra_lines_bottom = 0;
 114                 decoded_video_format->frame_rate_nom = 3013;
 115                 decoded_video_format->frame_rate_den = 100;
 116                 decoded_video_format->has_signal = false;
 117                 return true;
 118         }
 119         if ((video_format & 0xe800) != 0xe800) {
 120                 printf("Video format 0x%04x does not appear to be a video format. Assuming 60 Hz.\n",
 121                         video_format);
 122                 decoded_video_format->width = 0;
 123                 decoded_video_format->height = 0;
 124                 decoded_video_format->extra_lines_top = 0;
 125                 decoded_video_format->extra_lines_bottom = 0;
 126                 decoded_video_format->frame_rate_nom = 60;
 127                 decoded_video_format->frame_rate_den = 1;
 128                 decoded_video_format->has_signal = false;
 129                 return false;
 130         }
 131
 132         decoded_video_format->has_signal = true;
 133
 134         // NTSC (480i59.94, I suppose). A special case, see below.
 135         if (video_format == 0xe901 || video_format == 0xe9c1 || video_format == 0xe801) {
 136                 decoded_video_format->width = 720;
 137                 decoded_video_format->height = 480;
 138                 decoded_video_format->extra_lines_top = 17;
 139                 decoded_video_format->extra_lines_bottom = 28;
 140                 decoded_video_format->frame_rate_nom = 30000;
 141                 decoded_video_format->frame_rate_den = 1001;
 142                 decoded_video_format->second_field_start = 280;
 143                 decoded_video_format->interlaced = true;
 144                 return true;
 145         }
 146
 147         // PAL (576i50, I suppose). A special case, see below.
 148         if (video_format == 0xe909 || video_format == 0xe9c9 || video_format == 0xe809 || video_format == 0xebe9 || video_format == 0xebe1) {
 149                 decoded_video_format->width = 720;
 150                 decoded_video_format->height = 576;
 151                 decoded_video_format->extra_lines_top = 22;
 152                 decoded_video_format->extra_lines_bottom = 27;
 153                 decoded_video_format->frame_rate_nom = 25;
 154                 decoded_video_format->frame_rate_den = 1;
 155                 decoded_video_format->second_field_start = 335;
 156                 decoded_video_format->interlaced = true;
 157                 return true;
 158         }
 159
 160         // 0x8 seems to be a flag about availability of deep color on the input,
 161         // except when it's not (e.g. it's the only difference between NTSC
 162         // and PAL). Rather confusing. But we clear it here nevertheless, because
 163         // usually it doesn't mean anything.
 164         //
 165         // 0x4 is a flag I've only seen from the D4. I don't know what it is.
 166         uint16_t normalized_video_format = video_format & ~0xe80c;
 167         constexpr VideoFormatEntry entries[] = {
 168                 { 0x01f1,  720,  480,   0, 40,  5, 60000, 1001, false },  // 480p59.94 (believed).
 169                 { 0x0131,  720,  576,   0, 44,  5,    50,    1, false },  // 576p50.
 170                 { 0x0011,  720,  576,   0, 44,  5,    50,    1, false },  // 576p50 (5:4).
 171                 { 0x0143, 1280,  720,   0, 25,  5,    50,    1, false },  // 720p50.
 172                 { 0x0103, 1280,  720,   0, 25,  5,    60,    1, false },  // 720p60.
 173                 { 0x0125, 1280,  720,   0, 25,  5,    60,    1, false },  // 720p60.
 174                 { 0x0121, 1280,  720,   0, 25,  5, 60000, 1001, false },  // 720p59.94.
 175                 { 0x01c3, 1920, 1080,   0,  0,  0,    30,    1, false },  // 1080p30.
 176                 { 0x0003, 1920, 1080, 583, 20, 25,    30,    1,  true },  // 1080i60.
 177                 { 0x01e1, 1920, 1080,   0,  0,  0, 30000, 1001, false },  // 1080p29.97.
 178                 { 0x0021, 1920, 1080, 583, 20, 25, 30000, 1001,  true },  // 1080i59.94.
 179                 { 0x0063, 1920, 1080,   0,  0,  0,    25,    1, false },  // 1080p25.
 180                 { 0x0043, 1920, 1080,   0,  0,  0,    25,    1,  true },  // 1080p50.
 181                 { 0x008e, 1920, 1080,   0,  0,  0,    24,    1, false },  // 1080p24.
 182                 { 0x00a1, 1920, 1080,   0,  0,  0, 24000, 1001, false },  // 1080p23.98.
 183         };
 184         for (const VideoFormatEntry &entry : entries) {
 185                 if (normalized_video_format == entry.normalized_video_format) {
 186                         decoded_video_format->width = entry.width;
 187                         decoded_video_format->height = entry.height;
 188                         decoded_video_format->second_field_start = entry.second_field_start;
 189                         decoded_video_format->extra_lines_top = entry.extra_lines_top;
 190                         decoded_video_format->extra_lines_bottom = entry.extra_lines_bottom;
 191                         decoded_video_format->frame_rate_nom = entry.frame_rate_nom;
 192                         decoded_video_format->frame_rate_den = entry.frame_rate_den;
 193                         decoded_video_format->interlaced = entry.interlaced;
 194                         return true;
 195                 }
 196         }
 197
 198         printf("Unknown video format 0x%04x (normalized 0x%04x). Assuming 720p60.\n", video_format, normalized_video_format);
 199         decoded_video_format->width = 1280;
 200         decoded_video_format->height = 720;
 201         decoded_video_format->frame_rate_nom = 60;
 202         decoded_video_format->frame_rate_den = 1;
 203         return false;
 204 }
 205
 206 }  // namespace
 207
 208 FrameAllocator::~FrameAllocator() {}
 209
 210 MallocFrameAllocator::MallocFrameAllocator(size_t frame_size, size_t num_queued_frames)
 211         : frame_size(frame_size)
 212 {
 213         for (size_t i = 0; i < num_queued_frames; ++i) {
 214                 freelist.push(unique_ptr<uint8_t[]>(new uint8_t[frame_size]));
 215         }
 216 }
 217
 218 FrameAllocator::Frame MallocFrameAllocator::alloc_frame()
 219 {
 220         Frame vf;
 221         vf.owner = this;
 222
 223         unique_lock<mutex> lock(freelist_mutex);  // Meh.
 224         if (freelist.empty()) {
 225                 printf("Frame overrun (no more spare frames of size %ld), dropping frame!\n",
 226                         frame_size);
 227         } else {
 228                 vf.data = freelist.top().release();
 229                 vf.size = frame_size;
 230                 freelist.pop();  // Meh.
 231         }
 232         return vf;
 233 }
 234
 235 void MallocFrameAllocator::release_frame(Frame frame)
 236 {
 237         if (frame.overflow > 0) {
 238                 printf("%d bytes overflow after last (malloc) frame\n", int(frame.overflow));
 239         }
 240         unique_lock<mutex> lock(freelist_mutex);
 241         freelist.push(unique_ptr<uint8_t[]>(frame.data));
 242 }
 243
 244 bool uint16_less_than_with_wraparound(uint16_t a, uint16_t b)
 245 {
 246         if (a == b) {
 247                 return false;
 248         } else if (a < b) {
 249                 return (b - a < 0x8000);
 250         } else {
 251                 int wrap_b = 0x10000 + int(b);
 252                 return (wrap_b - a < 0x8000);
 253         }
 254 }
 255
 256 void BMUSBCapture::queue_frame(uint16_t format, uint16_t timecode, FrameAllocator::Frame frame, deque<QueuedFrame> *q)
 257 {
 258         unique_lock<mutex> lock(queue_lock);
 259         if (!q->empty() && !uint16_less_than_with_wraparound(q->back().timecode, timecode)) {
 260                 printf("Blocks going backwards: prev=0x%04x, cur=0x%04x (dropped)\n",
 261                         q->back().timecode, timecode);
 262                 frame.owner->release_frame(frame);
 263                 return;
 264         }
 265
 266         QueuedFrame qf;
 267         qf.format = format;
 268         qf.timecode = timecode;
 269         qf.frame = frame;
 270         q->push_back(move(qf));
 271         queues_not_empty.notify_one();  // might be spurious
 272 }
 273
 274 void dump_frame(const char *filename, uint8_t *frame_start, size_t frame_len)
 275 {
 276         FILE *fp = fopen(filename, "wb");
 277         if (fwrite(frame_start + HEADER_SIZE, frame_len - HEADER_SIZE, 1, fp) != 1) {
 278                 printf("short write!\n");
 279         }
 280         fclose(fp);
 281 }
 282
 283 void dump_audio_block(uint8_t *audio_start, size_t audio_len)
 284 {
 285         fwrite(audio_start + AUDIO_HEADER_SIZE, 1, audio_len - AUDIO_HEADER_SIZE, audiofp);
 286 }
 287
 288 void BMUSBCapture::dequeue_thread_func()
 289 {
 290         if (has_dequeue_callbacks) {
 291                 dequeue_init_callback();
 292         }
 293         while (!dequeue_thread_should_quit) {
 294                 unique_lock<mutex> lock(queue_lock);
 295                 queues_not_empty.wait(lock, [this]{ return dequeue_thread_should_quit || (!pending_video_frames.empty() && !pending_audio_frames.empty()); });
 296
 297                 if (dequeue_thread_should_quit) break;
 298
 299                 uint16_t video_timecode = pending_video_frames.front().timecode;
 300                 uint16_t audio_timecode = pending_audio_frames.front().timecode;
 301                 AudioFormat audio_format;
 302                 audio_format.bits_per_sample = 24;
 303                 audio_format.num_channels = 8;
 304                 if (uint16_less_than_with_wraparound(video_timecode, audio_timecode)) {
 305                         printf("Video block 0x%04x without corresponding audio block, dropping.\n",
 306                                 video_timecode);
 307                         QueuedFrame video_frame = pending_video_frames.front();
 308                         pending_video_frames.pop_front();
 309                         lock.unlock();
 310                         video_frame_allocator->release_frame(video_frame.frame);
 311                 } else if (uint16_less_than_with_wraparound(audio_timecode, video_timecode)) {
 312                         printf("Audio block 0x%04x without corresponding video block, sending blank frame.\n",
 313                                 audio_timecode);
 314                         QueuedFrame audio_frame = pending_audio_frames.front();
 315                         pending_audio_frames.pop_front();
 316                         lock.unlock();
 317                         audio_format.id = audio_frame.format;
 318
 319                         // Use the video format of the pending frame.
 320                         QueuedFrame video_frame = pending_video_frames.front();
 321                         VideoFormat video_format;
 322                         decode_video_format(video_frame.format, &video_format);
 323
 324                         frame_callback(audio_timecode,
 325                                        FrameAllocator::Frame(), 0, video_format,
 326                                        audio_frame.frame, AUDIO_HEADER_SIZE, audio_format);
 327                 } else {
 328                         QueuedFrame video_frame = pending_video_frames.front();
 329                         QueuedFrame audio_frame = pending_audio_frames.front();
 330                         pending_audio_frames.pop_front();
 331                         pending_video_frames.pop_front();
 332                         lock.unlock();
 333
 334 #if 0
 335                         char filename[255];
 336                         snprintf(filename, sizeof(filename), "%04x%04x.uyvy", video_frame.format, video_timecode);
 337                         dump_frame(filename, video_frame.frame.data, video_frame.data_len);
 338                         dump_audio_block(audio_frame.frame.data, audio_frame.data_len);
 339 #endif
 340
 341                         VideoFormat video_format;
 342                         audio_format.id = audio_frame.format;
 343                         if (decode_video_format(video_frame.format, &video_format)) {
 344                                 frame_callback(video_timecode,
 345                                                video_frame.frame, HEADER_SIZE, video_format,
 346                                                audio_frame.frame, AUDIO_HEADER_SIZE, audio_format);
 347                         } else {
 348                                 frame_callback(video_timecode,
 349                                                FrameAllocator::Frame(), 0, video_format,
 350                                                audio_frame.frame, AUDIO_HEADER_SIZE, audio_format);
 351                         }
 352                 }
 353         }
 354         if (has_dequeue_callbacks) {
 355                 dequeue_cleanup_callback();
 356         }
 357 }
 358
 359 void BMUSBCapture::start_new_frame(const uint8_t *start)
 360 {
 361         uint16_t format = (start[3] << 8) | start[2];
 362         uint16_t timecode = (start[1] << 8) | start[0];
 363
 364         if (current_video_frame.len > 0) {
 365                 // If format is 0x0800 (no signal), add a fake (empty) audio
 366                 // frame to get it out of the queue.
 367                 // TODO: Figure out if there are other formats that come with
 368                 // no audio, and treat them the same.
 369                 if (format == 0x0800) {
 370                         FrameAllocator::Frame fake_audio_frame = audio_frame_allocator->alloc_frame();
 371                         if (fake_audio_frame.data == nullptr) {
 372                                 // Oh well, it's just a no-signal frame anyway.
 373                                 printf("Couldn't allocate fake audio frame, also dropping no-signal video frame.\n");
 374                                 current_video_frame.owner->release_frame(current_video_frame);
 375                                 current_video_frame = video_frame_allocator->alloc_frame();
 376                                 return;
 377                         }
 378                         queue_frame(format, timecode, fake_audio_frame, &pending_audio_frames);
 379                 }
 380                 //dump_frame();
 381                 queue_frame(format, timecode, current_video_frame, &pending_video_frames);
 382
 383                 // Update the assumed frame width. We might be one frame too late on format changes,
 384                 // but it's much better than asking the user to choose manually.
 385                 VideoFormat video_format;
 386                 if (decode_video_format(format, &video_format)) {
 387                         assumed_frame_width = video_format.width;
 388                 }
 389         }
 390         //printf("Found frame start, format 0x%04x timecode 0x%04x, previous frame length was %d/%d\n",
 391         //      format, timecode,
 392         //      //start[7], start[6], start[5], start[4],
 393         //      read_current_frame, FRAME_SIZE);
 394
 395         current_video_frame = video_frame_allocator->alloc_frame();
 396         //if (current_video_frame.data == nullptr) {
 397         //      read_current_frame = -1;
 398         //} else {
 399         //      read_current_frame = 0;
 400         //}
 401 }
 402
 403 void BMUSBCapture::start_new_audio_block(const uint8_t *start)
 404 {
 405         uint16_t format = (start[3] << 8) | start[2];
 406         uint16_t timecode = (start[1] << 8) | start[0];
 407         if (current_audio_frame.len > 0) {
 408                 //dump_audio_block();
 409                 queue_frame(format, timecode, current_audio_frame, &pending_audio_frames);
 410         }
 411         //printf("Found audio block start, format 0x%04x timecode 0x%04x, previous block length was %d\n",
 412         //      format, timecode, read_current_audio_block);
 413         current_audio_frame = audio_frame_allocator->alloc_frame();
 414 }
 415
 416 #if 0
 417 static void dump_pack(const libusb_transfer *xfr, int offset, const libusb_iso_packet_descriptor *pack)
 418 {
 419         //      printf("ISO pack%u length:%u, actual_length:%u, offset:%u\n", i, pack->length, pack->actual_length, offset);
 420         for (unsigned j = 0; j < pack->actual_length; j++) {
 421         //for (int j = 0; j < min(pack->actual_length, 16u); j++) {
 422                 printf("%02x", xfr->buffer[j + offset]);
 423                 if ((j % 16) == 15)
 424                         printf("\n");
 425                 else if ((j % 8) == 7)
 426                         printf("  ");
 427                 else
 428                         printf(" ");
 429         }
 430 }
 431 #endif
 432
 433 void memcpy_interleaved(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size_t n)
 434 {
 435         assert(n % 2 == 0);
 436         uint8_t *dptr1 = dest1;
 437         uint8_t *dptr2 = dest2;
 438
 439         for (size_t i = 0; i < n; i += 2) {
 440                 *dptr1++ = *src++;
 441                 *dptr2++ = *src++;
 442         }
 443 }
 444
 445 void add_to_frame(FrameAllocator::Frame *current_frame, const char *frame_type_name, const uint8_t *start, const uint8_t *end)
 446 {
 447         if (current_frame->data == nullptr ||
 448             current_frame->len > current_frame->size ||
 449             start == end) {
 450                 return;
 451         }
 452
 453         int bytes = end - start;
 454         if (current_frame->len + bytes > current_frame->size) {
 455                 current_frame->overflow = current_frame->len + bytes - current_frame->size;
 456                 current_frame->len = current_frame->size;
 457                 if (current_frame->overflow > 1048576) {
 458                         printf("%d bytes overflow after last %s frame\n",
 459                                 int(current_frame->overflow), frame_type_name);
 460                         current_frame->overflow = 0;
 461                 }
 462                 //dump_frame();
 463         } else {
 464                 if (current_frame->interleaved) {
 465                         uint8_t *data = current_frame->data + current_frame->len / 2;
 466                         uint8_t *data2 = current_frame->data2 + current_frame->len / 2;
 467                         if (current_frame->len % 2 == 1) {
 468                                 ++data;
 469                                 swap(data, data2);
 470                         }
 471                         if (bytes % 2 == 1) {
 472                                 *data++ = *start++;
 473                                 swap(data, data2);
 474                                 ++current_frame->len;
 475                                 --bytes;
 476                         }
 477                         memcpy_interleaved(data, data2, start, bytes);
 478                         current_frame->len += bytes;
 479                 } else {
 480                         memcpy(current_frame->data + current_frame->len, start, bytes);
 481                         current_frame->len += bytes;
 482                 }
 483         }
 484 }
 485
 486 #if 0
 487 void avx2_dump(const char *name, __m256i n)
 488 {
 489         printf("%-10s:", name);
 490         printf(" %02x", _mm256_extract_epi8(n, 0));
 491         printf(" %02x", _mm256_extract_epi8(n, 1));
 492         printf(" %02x", _mm256_extract_epi8(n, 2));
 493         printf(" %02x", _mm256_extract_epi8(n, 3));
 494         printf(" %02x", _mm256_extract_epi8(n, 4));
 495         printf(" %02x", _mm256_extract_epi8(n, 5));
 496         printf(" %02x", _mm256_extract_epi8(n, 6));
 497         printf(" %02x", _mm256_extract_epi8(n, 7));
 498         printf(" ");
 499         printf(" %02x", _mm256_extract_epi8(n, 8));
 500         printf(" %02x", _mm256_extract_epi8(n, 9));
 501         printf(" %02x", _mm256_extract_epi8(n, 10));
 502         printf(" %02x", _mm256_extract_epi8(n, 11));
 503         printf(" %02x", _mm256_extract_epi8(n, 12));
 504         printf(" %02x", _mm256_extract_epi8(n, 13));
 505         printf(" %02x", _mm256_extract_epi8(n, 14));
 506         printf(" %02x", _mm256_extract_epi8(n, 15));
 507         printf(" ");
 508         printf(" %02x", _mm256_extract_epi8(n, 16));
 509         printf(" %02x", _mm256_extract_epi8(n, 17));
 510         printf(" %02x", _mm256_extract_epi8(n, 18));
 511         printf(" %02x", _mm256_extract_epi8(n, 19));
 512         printf(" %02x", _mm256_extract_epi8(n, 20));
 513         printf(" %02x", _mm256_extract_epi8(n, 21));
 514         printf(" %02x", _mm256_extract_epi8(n, 22));
 515         printf(" %02x", _mm256_extract_epi8(n, 23));
 516         printf(" ");
 517         printf(" %02x", _mm256_extract_epi8(n, 24));
 518         printf(" %02x", _mm256_extract_epi8(n, 25));
 519         printf(" %02x", _mm256_extract_epi8(n, 26));
 520         printf(" %02x", _mm256_extract_epi8(n, 27));
 521         printf(" %02x", _mm256_extract_epi8(n, 28));
 522         printf(" %02x", _mm256_extract_epi8(n, 29));
 523         printf(" %02x", _mm256_extract_epi8(n, 30));
 524         printf(" %02x", _mm256_extract_epi8(n, 31));
 525         printf("\n");
 526 }
 527 #endif
 528
 529 #ifndef HAS_MULTIVERSIONING
 530
 531 const uint8_t *add_to_frame_fastpath(FrameAllocator::Frame *current_frame, const uint8_t *start, const uint8_t *limit, const char sync_char)
 532 {
 533         // No fast path possible unless we have multiversioning.
 534         return start;
 535 }
 536
 537 #else  // defined(HAS_MULTIVERSIONING)
 538
 539 const uint8_t *add_to_frame_fastpath_core(FrameAllocator::Frame *current_frame, const uint8_t *aligned_start, const uint8_t *limit, const char sync_char);
 540
 541 // Does a memcpy and memchr in one to reduce processing time.
 542 // Note that the benefit is somewhat limited if your L3 cache is small,
 543 // as you'll (unfortunately) spend most of the time loading the data
 544 // from main memory.
 545 //
 546 // Complicated cases are left to the slow path; it basically stops copying
 547 // up until the first instance of "sync_char" (usually a bit before, actually).
 548 // This is fine, since 0x00 bytes shouldn't really show up in normal picture
 549 // data, and what we really need this for is the 00 00 ff ff marker in video data.
 550 __attribute__((target("default")))
 551 const uint8_t *add_to_frame_fastpath(FrameAllocator::Frame *current_frame, const uint8_t *start, const uint8_t *limit, const char sync_char)
 552 {
 553         // No fast path possible unless we have SSE 4.1 or higher.
 554         return start;
 555 }
 556
 557 __attribute__((target("sse4.1", "avx2")))
 558 const uint8_t *add_to_frame_fastpath(FrameAllocator::Frame *current_frame, const uint8_t *start, const uint8_t *limit, const char sync_char)
 559 {
 560         if (current_frame->data == nullptr ||
 561             current_frame->len > current_frame->size ||
 562             start == limit) {
 563                 return start;
 564         }
 565         size_t orig_bytes = limit - start;
 566         if (orig_bytes < 128) {
 567                 // Don't bother.
 568                 return start;
 569         }
 570
 571         // Don't read more bytes than we can write.
 572         limit = min(limit, start + (current_frame->size - current_frame->len));
 573
 574         // Align end to 32 bytes.
 575         limit = (const uint8_t *)(intptr_t(limit) & ~31);
 576
 577         if (start >= limit) {
 578                 return start;
 579         }
 580
 581         // Process [0,31] bytes, such that start gets aligned to 32 bytes.
 582         const uint8_t *aligned_start = (const uint8_t *)(intptr_t(start + 31) & ~31);
 583         if (aligned_start != start) {
 584                 const uint8_t *sync_start = (const uint8_t *)memchr(start, sync_char, aligned_start - start);
 585                 if (sync_start == nullptr) {
 586                         add_to_frame(current_frame, "", start, aligned_start);
 587                 } else {
 588                         add_to_frame(current_frame, "", start, sync_start);
 589                         return sync_start;
 590                 }
 591         }
 592
 593         // Make the length a multiple of 64.
 594         if (current_frame->interleaved) {
 595                 if (((limit - aligned_start) % 64) != 0) {
 596                         limit -= 32;
 597                 }
 598                 assert(((limit - aligned_start) % 64) == 0);
 599         }
 600
 601         return add_to_frame_fastpath_core(current_frame, aligned_start, limit, sync_char);
 602 }
 603
 604 __attribute__((target("avx2")))
 605 const uint8_t *add_to_frame_fastpath_core(FrameAllocator::Frame *current_frame, const uint8_t *aligned_start, const uint8_t *limit, const char sync_char)
 606 {
 607         const __m256i needle = _mm256_set1_epi8(sync_char);
 608
 609         const __restrict __m256i *in = (const __m256i *)aligned_start;
 610         if (current_frame->interleaved) {
 611                 __restrict __m256i *out1 = (__m256i *)(current_frame->data + (current_frame->len + 1) / 2);
 612                 __restrict __m256i *out2 = (__m256i *)(current_frame->data2 + current_frame->len / 2);
 613                 if (current_frame->len % 2 == 1) {
 614                         swap(out1, out2);
 615                 }
 616
 617                 __m256i shuffle_cw = _mm256_set_epi8(
 618                         15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0,
 619                         15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0);
 620                 while (in < (const __m256i *)limit) {
 621                         // Note: For brevity, comments show lanes as if they were 2x64-bit (they're actually 2x128).
 622                         __m256i data1 = _mm256_stream_load_si256(in);         // AaBbCcDd EeFfGgHh
 623                         __m256i data2 = _mm256_stream_load_si256(in + 1);     // IiJjKkLl MmNnOoPp
 624
 625                         __m256i found1 = _mm256_cmpeq_epi8(data1, needle);
 626                         __m256i found2 = _mm256_cmpeq_epi8(data2, needle);
 627                         __m256i found = _mm256_or_si256(found1, found2);
 628
 629                         data1 = _mm256_shuffle_epi8(data1, shuffle_cw);       // ABCDabcd EFGHefgh
 630                         data2 = _mm256_shuffle_epi8(data2, shuffle_cw);       // IJKLijkl MNOPmnop
 631
 632                         data1 = _mm256_permute4x64_epi64(data1, 0b11011000);  // ABCDEFGH abcdefgh
 633                         data2 = _mm256_permute4x64_epi64(data2, 0b11011000);  // IJKLMNOP ijklmnop
 634
 635                         __m256i lo = _mm256_permute2x128_si256(data1, data2, 0b00100000);
 636                         __m256i hi = _mm256_permute2x128_si256(data1, data2, 0b00110001);
 637
 638                         _mm256_storeu_si256(out1, lo);  // Store as early as possible, even if the data isn't used.
 639                         _mm256_storeu_si256(out2, hi);
 640
 641                         if (!_mm256_testz_si256(found, found)) {
 642                                 break;
 643                         }
 644
 645                         in += 2;
 646                         ++out1;
 647                         ++out2;
 648                 }
 649                 current_frame->len += (uint8_t *)in - aligned_start;
 650         } else {
 651                 __m256i *out = (__m256i *)(current_frame->data + current_frame->len);
 652                 while (in < (const __m256i *)limit) {
 653                         __m256i data = _mm256_load_si256(in);
 654                         _mm256_storeu_si256(out, data);  // Store as early as possible, even if the data isn't used.
 655                         __m256i found = _mm256_cmpeq_epi8(data, needle);
 656                         if (!_mm256_testz_si256(found, found)) {
 657                                 break;
 658                         }
 659
 660                         ++in;
 661                         ++out;
 662                 }
 663                 current_frame->len = (uint8_t *)out - current_frame->data;
 664         }
 665
 666         //printf("managed to fastpath %ld/%ld bytes\n", (const uint8_t *)in - (const uint8_t *)aligned_start, orig_bytes);
 667         return (const uint8_t *)in;
 668 }
 669
 670 __attribute__((target("sse4.1")))
 671 const uint8_t *add_to_frame_fastpath_core(FrameAllocator::Frame *current_frame, const uint8_t *aligned_start, const uint8_t *limit, const char sync_char)
 672 {
 673         const __m128i needle = _mm_set1_epi8(sync_char);
 674
 675         const __m128i *in = (const __m128i *)aligned_start;
 676         if (current_frame->interleaved) {
 677                 __m128i *out1 = (__m128i *)(current_frame->data + (current_frame->len + 1) / 2);
 678                 __m128i *out2 = (__m128i *)(current_frame->data2 + current_frame->len / 2);
 679                 if (current_frame->len % 2 == 1) {
 680                         swap(out1, out2);
 681                 }
 682
 683                 __m128i mask_lower_byte = _mm_set1_epi16(0x00ff);
 684                 while (in < (const __m128i *)limit) {
 685                         __m128i data1 = _mm_load_si128(in);
 686                         __m128i data2 = _mm_load_si128(in + 1);
 687                         __m128i data1_lo = _mm_and_si128(data1, mask_lower_byte);
 688                         __m128i data2_lo = _mm_and_si128(data2, mask_lower_byte);
 689                         __m128i data1_hi = _mm_srli_epi16(data1, 8);
 690                         __m128i data2_hi = _mm_srli_epi16(data2, 8);
 691                         __m128i lo = _mm_packus_epi16(data1_lo, data2_lo);
 692                         _mm_storeu_si128(out1, lo);  // Store as early as possible, even if the data isn't used.
 693                         __m128i hi = _mm_packus_epi16(data1_hi, data2_hi);
 694                         _mm_storeu_si128(out2, hi);
 695                         __m128i found1 = _mm_cmpeq_epi8(data1, needle);
 696                         __m128i found2 = _mm_cmpeq_epi8(data2, needle);
 697                         if (!_mm_testz_si128(found1, found1) ||
 698                             !_mm_testz_si128(found2, found2)) {
 699                                 break;
 700                         }
 701
 702                         in += 2;
 703                         ++out1;
 704                         ++out2;
 705                 }
 706                 current_frame->len += (uint8_t *)in - aligned_start;
 707         } else {
 708                 __m128i *out = (__m128i *)(current_frame->data + current_frame->len);
 709                 while (in < (const __m128i *)limit) {
 710                         __m128i data = _mm_load_si128(in);
 711                         _mm_storeu_si128(out, data);  // Store as early as possible, even if the data isn't used.
 712                         __m128i found = _mm_cmpeq_epi8(data, needle);
 713                         if (!_mm_testz_si128(found, found)) {
 714                                 break;
 715                         }
 716
 717                         ++in;
 718                         ++out;
 719                 }
 720                 current_frame->len = (uint8_t *)out - current_frame->data;
 721         }
 722
 723         //printf("managed to fastpath %ld/%ld bytes\n", (const uint8_t *)in - (const uint8_t *)aligned_start, orig_bytes);
 724         return (const uint8_t *)in;
 725 }
 726
 727 #endif  // defined(HAS_MULTIVERSIONING)
 728
 729 void decode_packs(const libusb_transfer *xfr,
 730                   const char *sync_pattern,
 731                   int sync_length,
 732                   FrameAllocator::Frame *current_frame,
 733                   const char *frame_type_name,
 734                   function<void(const uint8_t *start)> start_callback)
 735 {
 736         int offset = 0;
 737         for (int i = 0; i < xfr->num_iso_packets; i++) {
 738                 const libusb_iso_packet_descriptor *pack = &xfr->iso_packet_desc[i];
 739
 740                 if (pack->status != LIBUSB_TRANSFER_COMPLETED) {
 741                         fprintf(stderr, "Error: pack %u/%u status %d\n", i, xfr->num_iso_packets, pack->status);
 742                         continue;
 743 //exit(5);
 744                 }
 745
 746                 const uint8_t *start = xfr->buffer + offset;
 747                 const uint8_t *limit = start + pack->actual_length;
 748                 while (start < limit) {  // Usually runs only one iteration.
 749                         start = add_to_frame_fastpath(current_frame, start, limit, sync_pattern[0]);
 750                         if (start == limit) break;
 751                         assert(start < limit);
 752
 753                         const unsigned char* start_next_frame = (const unsigned char *)memmem(start, limit - start, sync_pattern, sync_length);
 754                         if (start_next_frame == nullptr) {
 755                                 // add the rest of the buffer
 756                                 add_to_frame(current_frame, frame_type_name, start, limit);
 757                                 break;
 758                         } else {
 759                                 add_to_frame(current_frame, frame_type_name, start, start_next_frame);
 760                                 start = start_next_frame + sync_length;  // skip sync
 761                                 start_callback(start);
 762                         }
 763                 }
 764 #if 0
 765                 dump_pack(xfr, offset, pack);
 766 #endif
 767                 offset += pack->length;
 768         }
 769 }
 770
 771 void BMUSBCapture::cb_xfr(struct libusb_transfer *xfr)
 772 {
 773         if (xfr->status != LIBUSB_TRANSFER_COMPLETED &&
 774             xfr->status != LIBUSB_TRANSFER_NO_DEVICE) {
 775                 fprintf(stderr, "error: transfer status %d\n", xfr->status);
 776                 libusb_free_transfer(xfr);
 777                 exit(3);
 778         }
 779
 780         assert(xfr->user_data != nullptr);
 781         BMUSBCapture *usb = static_cast<BMUSBCapture *>(xfr->user_data);
 782
 783         if (xfr->status == LIBUSB_TRANSFER_NO_DEVICE) {
 784                 if (!usb->disconnected) {
 785                         fprintf(stderr, "Device went away, stopping transfers.\n");
 786                         usb->disconnected = true;
 787                         if (usb->card_disconnected_callback) {
 788                                 usb->card_disconnected_callback();
 789                         }
 790                 }
 791                 // Don't reschedule the transfer; the loop will stop by itself.
 792                 return;
 793         }
 794
 795         if (xfr->type == LIBUSB_TRANSFER_TYPE_ISOCHRONOUS) {
 796                 if (xfr->endpoint == 0x84) {
 797                         decode_packs(xfr, "DeckLinkAudioResyncT", 20, &usb->current_audio_frame, "audio", bind(&BMUSBCapture::start_new_audio_block, usb, _1));
 798                 } else {
 799                         decode_packs(xfr, "\x00\x00\xff\xff", 4, &usb->current_video_frame, "video", bind(&BMUSBCapture::start_new_frame, usb, _1));
 800
 801                         // Update the transfer with the new assumed width, if we're in the process of changing formats.
 802                         change_xfer_size_for_width(usb->assumed_frame_width, xfr);
 803                 }
 804         }
 805         if (xfr->type == LIBUSB_TRANSFER_TYPE_CONTROL) {
 806                 //const libusb_control_setup *setup = libusb_control_transfer_get_setup(xfr);
 807                 uint8_t *buf = libusb_control_transfer_get_data(xfr);
 808 #if 0
 809                 if (setup->wIndex == 44) {
 810                         printf("read timer register: 0x%02x%02x%02x%02x\n", buf[0], buf[1], buf[2], buf[3]);
 811                 } else {
 812                         printf("read register %2d:                      0x%02x%02x%02x%02x\n",
 813                                 setup->wIndex, buf[0], buf[1], buf[2], buf[3]);
 814                 }
 815 #else
 816                 memcpy(usb->register_file + usb->current_register, buf, 4);
 817                 usb->current_register = (usb->current_register + 4) % NUM_BMUSB_REGISTERS;
 818                 if (usb->current_register == 0) {
 819                         // read through all of them
 820                         printf("register dump:");
 821                         for (int i = 0; i < NUM_BMUSB_REGISTERS; i += 4) {
 822                                 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]);
 823                         }
 824                         printf("\n");
 825                 }
 826                 libusb_fill_control_setup(xfr->buffer,
 827                     LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN, /*request=*/214, /*value=*/0,
 828                         /*index=*/usb->current_register, /*length=*/4);
 829 #endif
 830         }
 831
 832 #if 0
 833         printf("length:%u, actual_length:%u\n", xfr->length, xfr->actual_length);
 834         for (i = 0; i < xfr->actual_length; i++) {
 835                 printf("%02x", xfr->buffer[i]);
 836                 if (i % 16)
 837                         printf("\n");
 838                 else if (i % 8)
 839                         printf("  ");
 840                 else
 841                         printf(" ");
 842         }
 843 #endif
 844
 845         int rc = libusb_submit_transfer(xfr);
 846         if (rc < 0) {
 847                 fprintf(stderr, "error re-submitting URB: %s\n", libusb_error_name(rc));
 848                 exit(1);
 849         }
 850 }
 851
 852 int BMUSBCapture::cb_hotplug(libusb_context *ctx, libusb_device *dev, libusb_hotplug_event event, void *user_data)
 853 {
 854         if (card_connected_callback != nullptr) {
 855                 libusb_device_descriptor desc;
 856                 if (libusb_get_device_descriptor(dev, &desc) < 0) {
 857                         fprintf(stderr, "Error getting device descriptor for hotplugged device %p, killing hotplug\n", dev);
 858                         libusb_unref_device(dev);
 859                         return 1;
 860                 }
 861
 862                 if ((desc.idVendor == USB_VENDOR_BLACKMAGIC && desc.idProduct == 0xbd3b) ||
 863                     (desc.idVendor == USB_VENDOR_BLACKMAGIC && desc.idProduct == 0xbd4f)) {
 864                         card_connected_callback(dev);  // Callback takes ownership.
 865                         return 0;
 866                 }
 867         }
 868         libusb_unref_device(dev);
 869         return 0;
 870 }
 871
 872 void BMUSBCapture::usb_thread_func()
 873 {
 874         sched_param param;
 875         memset(&param, 0, sizeof(param));
 876         param.sched_priority = 1;
 877         if (sched_setscheduler(0, SCHED_RR, &param) == -1) {
 878                 printf("couldn't set realtime priority for USB thread: %s\n", strerror(errno));
 879         }
 880         while (!should_quit) {
 881                 int rc = libusb_handle_events(nullptr);
 882                 if (rc != LIBUSB_SUCCESS)
 883                         break;
 884         }
 885 }
 886
 887 namespace {
 888
 889 struct USBCardDevice {
 890         uint16_t product;
 891         uint8_t bus, port;
 892         libusb_device *device;
 893 };
 894
 895 const char *get_product_name(uint16_t product)
 896 {
 897         if (product == 0xbd3b) {
 898                 return "Intensity Shuttle";
 899         } else if (product == 0xbd4f) {
 900                 return "UltraStudio SDI";
 901         } else {
 902                 assert(false);
 903                 return nullptr;
 904         }
 905 }
 906
 907 string get_card_description(int id, uint8_t bus, uint8_t port, uint16_t product)
 908 {
 909         const char *product_name = get_product_name(product);
 910
 911         char buf[256];
 912         snprintf(buf, sizeof(buf), "USB card %d: Bus %03u Device %03u  %s",
 913                 id, bus, port, product_name);
 914         return buf;
 915 }
 916
 917 vector<USBCardDevice> find_all_cards()
 918 {
 919         libusb_device **devices;
 920         ssize_t num_devices = libusb_get_device_list(nullptr, &devices);
 921         if (num_devices == -1) {
 922                 fprintf(stderr, "Error finding USB devices\n");
 923                 exit(1);
 924         }
 925         vector<USBCardDevice> found_cards;
 926         for (ssize_t i = 0; i < num_devices; ++i) {
 927                 libusb_device_descriptor desc;
 928                 if (libusb_get_device_descriptor(devices[i], &desc) < 0) {
 929                         fprintf(stderr, "Error getting device descriptor for device %d\n", int(i));
 930                         exit(1);
 931                 }
 932
 933                 uint8_t bus = libusb_get_bus_number(devices[i]);
 934                 uint8_t port = libusb_get_port_number(devices[i]);
 935
 936                 if (!(desc.idVendor == USB_VENDOR_BLACKMAGIC && desc.idProduct == 0xbd3b) &&
 937                     !(desc.idVendor == USB_VENDOR_BLACKMAGIC && desc.idProduct == 0xbd4f)) {
 938                         libusb_unref_device(devices[i]);
 939                         continue;
 940                 }
 941
 942                 found_cards.push_back({ desc.idProduct, bus, port, devices[i] });
 943         }
 944         libusb_free_device_list(devices, 0);
 945
 946         // Sort the devices to get a consistent ordering.
 947         sort(found_cards.begin(), found_cards.end(), [](const USBCardDevice &a, const USBCardDevice &b) {
 948                 if (a.product != b.product)
 949                         return a.product < b.product;
 950                 if (a.bus != b.bus)
 951                         return a.bus < b.bus;
 952                 return a.port < b.port;
 953         });
 954
 955         return found_cards;
 956 }
 957
 958 libusb_device_handle *open_card(int card_index, string *description)
 959 {
 960         vector<USBCardDevice> found_cards = find_all_cards();
 961
 962         for (size_t i = 0; i < found_cards.size(); ++i) {
 963                 string tmp_description = get_card_description(i, found_cards[i].bus, found_cards[i].port, found_cards[i].product);
 964                 fprintf(stderr, "%s\n", tmp_description.c_str());
 965                 if (i == size_t(card_index)) {
 966                         *description = tmp_description;
 967                 }
 968         }
 969
 970         if (size_t(card_index) >= found_cards.size()) {
 971                 fprintf(stderr, "Could not open card %d (only %d found)\n", card_index, int(found_cards.size()));
 972                 exit(1);
 973         }
 974
 975         libusb_device_handle *devh;
 976         int rc = libusb_open(found_cards[card_index].device, &devh);
 977         if (rc < 0) {
 978                 fprintf(stderr, "Error opening card %d: %s\n", card_index, libusb_error_name(rc));
 979                 exit(1);
 980         }
 981
 982         for (size_t i = 0; i < found_cards.size(); ++i) {
 983                 libusb_unref_device(found_cards[i].device);
 984         }
 985
 986         return devh;
 987 }
 988
 989 libusb_device_handle *open_card(unsigned card_index, libusb_device *dev, string *description)
 990 {
 991         uint8_t bus = libusb_get_bus_number(dev);
 992         uint8_t port = libusb_get_port_number(dev);
 993
 994         libusb_device_descriptor desc;
 995         if (libusb_get_device_descriptor(dev, &desc) < 0) {
 996                 fprintf(stderr, "Error getting device descriptor for device %p\n", dev);
 997                 exit(1);
 998         }
 999
1000         *description = get_card_description(card_index, bus, port, desc.idProduct);
1001
1002         libusb_device_handle *devh;
1003         int rc = libusb_open(dev, &devh);
1004         if (rc < 0) {
1005                 fprintf(stderr, "Error opening card %p: %s\n", dev, libusb_error_name(rc));
1006                 exit(1);
1007         }
1008
1009         return devh;
1010 }
1011
1012 }  // namespace
1013
1014 unsigned BMUSBCapture::num_cards()
1015 {
1016         int rc = libusb_init(nullptr);
1017         if (rc < 0) {
1018                 fprintf(stderr, "Error initializing libusb: %s\n", libusb_error_name(rc));
1019                 exit(1);
1020         }
1021
1022         vector<USBCardDevice> found_cards = find_all_cards();
1023         unsigned ret = found_cards.size();
1024         for (size_t i = 0; i < found_cards.size(); ++i) {
1025                 libusb_unref_device(found_cards[i].device);
1026         }
1027         return ret;
1028 }
1029
1030 void BMUSBCapture::configure_card()
1031 {
1032         if (video_frame_allocator == nullptr) {
1033                 owned_video_frame_allocator.reset(new MallocFrameAllocator(FRAME_SIZE, NUM_QUEUED_VIDEO_FRAMES));
1034                 set_video_frame_allocator(owned_video_frame_allocator.get());
1035         }
1036         if (audio_frame_allocator == nullptr) {
1037                 owned_audio_frame_allocator.reset(new MallocFrameAllocator(65536, NUM_QUEUED_AUDIO_FRAMES));
1038                 set_audio_frame_allocator(owned_audio_frame_allocator.get());
1039         }
1040         dequeue_thread_should_quit = false;
1041         dequeue_thread = thread(&BMUSBCapture::dequeue_thread_func, this);
1042
1043         int rc;
1044         struct libusb_transfer *xfr;
1045
1046         rc = libusb_init(nullptr);
1047         if (rc < 0) {
1048                 fprintf(stderr, "Error initializing libusb: %s\n", libusb_error_name(rc));
1049                 exit(1);
1050         }
1051
1052         if (dev == nullptr) {
1053                 devh = open_card(card_index, &description);
1054         } else {
1055                 devh = open_card(card_index, dev, &description);
1056                 libusb_unref_device(dev);
1057         }
1058         if (!devh) {
1059                 fprintf(stderr, "Error finding USB device\n");
1060                 exit(1);
1061         }
1062
1063         libusb_config_descriptor *config;
1064         rc = libusb_get_config_descriptor(libusb_get_device(devh), /*config_index=*/0, &config);
1065         if (rc < 0) {
1066                 fprintf(stderr, "Error getting configuration: %s\n", libusb_error_name(rc));
1067                 exit(1);
1068         }
1069
1070 #if 0
1071         printf("%d interface\n", config->bNumInterfaces);
1072         for (int interface_number = 0; interface_number < config->bNumInterfaces; ++interface_number) {
1073                 printf("  interface %d\n", interface_number);
1074                 const libusb_interface *interface = &config->interface[interface_number];
1075                 for (int altsetting = 0; altsetting < interface->num_altsetting; ++altsetting) {
1076                         const libusb_interface_descriptor *interface_desc = &interface->altsetting[altsetting];
1077                         printf("    alternate setting %d\n", interface_desc->bAlternateSetting);
1078                         for (int endpoint_number = 0; endpoint_number < interface_desc->bNumEndpoints; ++endpoint_number) {
1079                                 const libusb_endpoint_descriptor *endpoint = &interface_desc->endpoint[endpoint_number];
1080                                 printf("        endpoint address 0x%02x\n", endpoint->bEndpointAddress);
1081                         }
1082                 }
1083         }
1084 #endif
1085
1086         rc = libusb_set_configuration(devh, /*configuration=*/1);
1087         if (rc < 0) {
1088                 fprintf(stderr, "Error setting configuration 1: %s\n", libusb_error_name(rc));
1089                 exit(1);
1090         }
1091
1092         rc = libusb_claim_interface(devh, 0);
1093         if (rc < 0) {
1094                 fprintf(stderr, "Error claiming interface 0: %s\n", libusb_error_name(rc));
1095                 exit(1);
1096         }
1097
1098         // Alternate setting 1 is output, alternate setting 2 is input.
1099         // Card is reset when switching alternates, so the driver uses
1100         // this “double switch” when it wants to reset.
1101         //
1102         // There's also alternate settings 3 and 4, which seem to be
1103         // like 1 and 2 except they advertise less bandwidth needed.
1104         rc = libusb_set_interface_alt_setting(devh, /*interface=*/0, /*alternate_setting=*/1);
1105         if (rc < 0) {
1106                 fprintf(stderr, "Error setting alternate 1: %s\n", libusb_error_name(rc));
1107                 exit(1);
1108         }
1109         rc = libusb_set_interface_alt_setting(devh, /*interface=*/0, /*alternate_setting=*/2);
1110         if (rc < 0) {
1111                 fprintf(stderr, "Error setting alternate 2: %s\n", libusb_error_name(rc));
1112                 exit(1);
1113         }
1114 #if 0
1115         rc = libusb_set_interface_alt_setting(devh, /*interface=*/0, /*alternate_setting=*/1);
1116         if (rc < 0) {
1117                 fprintf(stderr, "Error setting alternate 1: %s\n", libusb_error_name(rc));
1118                 exit(1);
1119         }
1120 #endif
1121
1122 #if 0
1123         rc = libusb_claim_interface(devh, 3);
1124         if (rc < 0) {
1125                 fprintf(stderr, "Error claiming interface 3: %s\n", libusb_error_name(rc));
1126                 exit(1);
1127         }
1128 #endif
1129
1130         // theories:
1131         //   44 is some kind of timer register (first 16 bits count upwards)
1132         //   24 is some sort of watchdog?
1133         //      you can seemingly set it to 0x73c60001 and that bit will eventually disappear
1134         //      (or will go to 0x73c60010?), also seen 0x73c60100
1135         //   12 also changes all the time, unclear why
1136         //   16 seems to be autodetected mode somehow
1137         //      --    this is e00115e0 after reset?
1138         //                    ed0115e0 after mode change [to output?]
1139         //                    2d0015e0 after more mode change [to input]
1140         //                    ed0115e0 after more mode change
1141         //                    2d0015e0 after more mode change
1142         //
1143         //                    390115e0 seems to indicate we have signal
1144         //         changes to 200115e0 when resolution changes/we lose signal, driver resets after a while
1145         //
1146         //                    200015e0 on startup
1147         //         changes to 250115e0 when we sync to the signal
1148         //
1149         //    so only first 16 bits count, and 0x0100 is a mask for ok/stable signal?
1150         //
1151         //    Bottom 16 bits of this register seem to be firmware version number (possibly not all all of them).
1152         //
1153         //    28 and 32 seems to be analog audio input levels (one byte for each of the eight channels).
1154         //    however, if setting 32 with HDMI embedded audio, it is immediately overwritten back (to 0xe137002a).
1155         //
1156         //    4, 8, 20 are unclear. seem to be some sort of bitmask, but we can set them to 0 with no apparent effect.
1157         //    perhaps some of them are related to analog output?
1158         //
1159         //    36 can be set to 0 with no apparent effect (all of this tested on both video and audio),
1160         //    but the driver sets it to 0x8036802a at some point.
1161         //
1162         //    all of this is on request 214/215. other requests (192, 219,
1163         //    222, 223, 224) are used for firmware upgrade. Probably best to
1164         //    stay out of it unless you know what you're doing.
1165         //
1166         //
1167         // register 16:
1168         // first byte is 0x39 for a stable 576p60 signal, 0x2d for a stable 720p60 signal, 0x20 for no signal
1169         //
1170         // theories:
1171         //   0x01 - stable signal
1172         //   0x04 - deep color
1173         //   0x08 - unknown (audio??)
1174         //   0x20 - 720p??
1175         //   0x30 - 576p??
1176
1177         update_capture_mode();
1178
1179         struct ctrl {
1180                 int endpoint;
1181                 int request;
1182                 int index;
1183                 uint32_t data;
1184         };
1185         static const ctrl ctrls[] = {
1186                 { LIBUSB_ENDPOINT_IN,  214, 16, 0 },
1187                 { LIBUSB_ENDPOINT_IN,  214,  0, 0 },
1188
1189                 //{ LIBUSB_ENDPOINT_OUT, 215,  0, 0x80000100 },
1190                 //{ LIBUSB_ENDPOINT_OUT, 215,  0, 0x09000000 },
1191                 { LIBUSB_ENDPOINT_OUT, 215, 24, 0x73c60001 },  // latch for frame start?
1192                 { LIBUSB_ENDPOINT_IN,  214, 24, 0 },  //
1193         };
1194
1195         for (unsigned req = 0; req < sizeof(ctrls) / sizeof(ctrls[0]); ++req) {
1196                 uint32_t flipped = htonl(ctrls[req].data);
1197                 static uint8_t value[4];
1198                 memcpy(value, &flipped, sizeof(flipped));
1199                 int size = sizeof(value);
1200                 //if (ctrls[req].request == 215) size = 0;
1201                 rc = libusb_control_transfer(devh, LIBUSB_REQUEST_TYPE_VENDOR | ctrls[req].endpoint,
1202                         /*request=*/ctrls[req].request, /*value=*/0, /*index=*/ctrls[req].index, value, size, /*timeout=*/0);
1203                 if (rc < 0) {
1204                         fprintf(stderr, "Error on control %d: %s\n", ctrls[req].index, libusb_error_name(rc));
1205                         exit(1);
1206                 }
1207
1208                 if (ctrls[req].index == 16 && rc == 4) {
1209                         printf("Card firmware version: 0x%02x%02x\n", value[2], value[3]);
1210                 }
1211
1212 #if 0
1213                 printf("rc=%d: ep=%d@%d %d -> 0x", rc, ctrls[req].endpoint, ctrls[req].request, ctrls[req].index);
1214                 for (int i = 0; i < rc; ++i) {
1215                         printf("%02x", value[i]);
1216                 }
1217                 printf("\n");
1218 #endif
1219         }
1220
1221 #if 0
1222         // DEBUG
1223         for ( ;; ) {
1224                 static int my_index = 0;
1225                 static uint8_t value[4];
1226                 int size = sizeof(value);
1227                 rc = libusb_control_transfer(devh, LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN,
1228                         /*request=*/214, /*value=*/0, /*index=*/my_index, value, size, /*timeout=*/0);
1229                 if (rc < 0) {
1230                         fprintf(stderr, "Error on control\n");
1231                         exit(1);
1232                 }
1233                 printf("rc=%d index=%d: 0x", rc, my_index);
1234                 for (int i = 0; i < rc; ++i) {
1235                         printf("%02x", value[i]);
1236                 }
1237                 printf("\n");
1238         }
1239 #endif
1240
1241 #if 0
1242         // set up an asynchronous transfer of the timer register
1243         static uint8_t cmdbuf[LIBUSB_CONTROL_SETUP_SIZE + 4];
1244         static int completed = 0;
1245
1246         xfr = libusb_alloc_transfer(0);
1247         libusb_fill_control_setup(cmdbuf,
1248             LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN, /*request=*/214, /*value=*/0,
1249                 /*index=*/44, /*length=*/4);
1250         libusb_fill_control_transfer(xfr, devh, cmdbuf, cb_xfr, &completed, 0);
1251         xfr->user_data = this;
1252         libusb_submit_transfer(xfr);
1253
1254         // set up an asynchronous transfer of register 24
1255         static uint8_t cmdbuf2[LIBUSB_CONTROL_SETUP_SIZE + 4];
1256         static int completed2 = 0;
1257
1258         xfr = libusb_alloc_transfer(0);
1259         libusb_fill_control_setup(cmdbuf2,
1260             LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN, /*request=*/214, /*value=*/0,
1261                 /*index=*/24, /*length=*/4);
1262         libusb_fill_control_transfer(xfr, devh, cmdbuf2, cb_xfr, &completed2, 0);
1263         xfr->user_data = this;
1264         libusb_submit_transfer(xfr);
1265 #endif
1266
1267         // set up an asynchronous transfer of the register dump
1268         static uint8_t cmdbuf3[LIBUSB_CONTROL_SETUP_SIZE + 4];
1269         static int completed3 = 0;
1270
1271         xfr = libusb_alloc_transfer(0);
1272         libusb_fill_control_setup(cmdbuf3,
1273             LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_IN, /*request=*/214, /*value=*/0,
1274                 /*index=*/current_register, /*length=*/4);
1275         libusb_fill_control_transfer(xfr, devh, cmdbuf3, cb_xfr, &completed3, 0);
1276         xfr->user_data = this;
1277         //libusb_submit_transfer(xfr);
1278
1279         //audiofp = fopen("audio.raw", "wb");
1280
1281         // set up isochronous transfers for audio and video
1282         for (int e = 3; e <= 4; ++e) {
1283                 //int num_transfers = (e == 3) ? 6 : 6;
1284                 int num_transfers = 6;
1285                 for (int i = 0; i < num_transfers; ++i) {
1286                         size_t buf_size;
1287                         int num_iso_pack, size;
1288                         if (e == 3) {
1289                                 // Allocate for minimum width (because that will give us the most
1290                                 // number of packets, so we don't need to reallocated, but we'll
1291                                 // default to 720p for the first frame.
1292                                 size = find_xfer_size_for_width(MIN_WIDTH);
1293                                 num_iso_pack = USB_VIDEO_TRANSFER_SIZE / size;
1294                                 buf_size = USB_VIDEO_TRANSFER_SIZE;
1295                         } else {
1296                                 size = 0xc0;
1297                                 num_iso_pack = 80;
1298                                 buf_size = num_iso_pack * size;
1299                         }
1300                         int num_bytes = num_iso_pack * size;
1301                         assert(size_t(num_bytes) <= buf_size);
1302 #if LIBUSB_API_VERSION >= 0x01000105
1303                         uint8_t *buf = libusb_dev_mem_alloc(devh, num_bytes);
1304 #else
1305                         uint8_t *buf = nullptr;
1306 #endif
1307                         if (buf == nullptr) {
1308                                 fprintf(stderr, "Failed to allocate persistent DMA memory ");
1309 #if LIBUSB_API_VERSION >= 0x01000105
1310                                 fprintf(stderr, "(probably too old kernel; use 4.6.0 or newer).\n");
1311 #else
1312                                 fprintf(stderr, "(compiled against too old libusb-1.0).\n");
1313 #endif
1314                                 fprintf(stderr, "Will go slower, and likely fail due to memory fragmentation after a few hours.\n");
1315                                 buf = new uint8_t[num_bytes];
1316                         }
1317
1318                         xfr = libusb_alloc_transfer(num_iso_pack);
1319                         if (!xfr) {
1320                                 fprintf(stderr, "oom\n");
1321                                 exit(1);
1322                         }
1323
1324                         int ep = LIBUSB_ENDPOINT_IN | e;
1325                         libusb_fill_iso_transfer(xfr, devh, ep, buf, buf_size,
1326                                 num_iso_pack, cb_xfr, nullptr, 0);
1327                         libusb_set_iso_packet_lengths(xfr, size);
1328                         xfr->user_data = this;
1329
1330                         if (e == 3) {
1331                                 change_xfer_size_for_width(assumed_frame_width, xfr);
1332                         }
1333
1334                         iso_xfrs.push_back(xfr);
1335                 }
1336         }
1337 }
1338
1339 void BMUSBCapture::start_bm_capture()
1340 {
1341         int i = 0;
1342         for (libusb_transfer *xfr : iso_xfrs) {
1343                 int rc = libusb_submit_transfer(xfr);
1344                 ++i;
1345                 if (rc < 0) {
1346                         //printf("num_bytes=%d\n", num_bytes);
1347                         fprintf(stderr, "Error submitting iso to endpoint 0x%02x, number %d: %s\n",
1348                                 xfr->endpoint, i, libusb_error_name(rc));
1349                         exit(1);
1350                 }
1351         }
1352
1353
1354 #if 0
1355         libusb_release_interface(devh, 0);
1356 out:
1357         if (devh)
1358                 libusb_close(devh);
1359         libusb_exit(nullptr);
1360         return rc;
1361 #endif
1362 }
1363
1364 void BMUSBCapture::stop_dequeue_thread()
1365 {
1366         dequeue_thread_should_quit = true;
1367         queues_not_empty.notify_all();
1368         dequeue_thread.join();
1369 }
1370
1371 void BMUSBCapture::start_bm_thread()
1372 {
1373         // Devices leaving are discovered by seeing the isochronous packets
1374         // coming back with errors, so only care about devices joining.
1375         if (card_connected_callback != nullptr) {
1376                 if (libusb_hotplug_register_callback(
1377                         nullptr, LIBUSB_HOTPLUG_EVENT_DEVICE_ARRIVED, hotplug_existing_devices ? LIBUSB_HOTPLUG_ENUMERATE : LIBUSB_HOTPLUG_NO_FLAGS,
1378                         USB_VENDOR_BLACKMAGIC, LIBUSB_HOTPLUG_MATCH_ANY, LIBUSB_HOTPLUG_MATCH_ANY,
1379                         &BMUSBCapture::cb_hotplug, nullptr, nullptr) < 0) {
1380                         fprintf(stderr, "libusb_hotplug_register_callback() failed\n");
1381                         exit(1);
1382                 }
1383         }
1384
1385         should_quit = false;
1386         usb_thread = thread(&BMUSBCapture::usb_thread_func);
1387 }
1388
1389 void BMUSBCapture::stop_bm_thread()
1390 {
1391         should_quit = true;
1392         usb_thread.join();
1393 }
1394
1395 map<uint32_t, VideoMode> BMUSBCapture::get_available_video_modes() const
1396 {
1397         // The USB3 cards autodetect, and seem to have no provision for forcing modes.
1398         VideoMode auto_mode;
1399         auto_mode.name = "Autodetect";
1400         auto_mode.autodetect = true;
1401         return {{ 0, auto_mode }};
1402 }
1403
1404 uint32_t BMUSBCapture::get_current_video_mode() const
1405 {
1406         return 0;  // Matches get_available_video_modes().
1407 }
1408
1409 void BMUSBCapture::set_video_mode(uint32_t video_mode_id)
1410 {
1411         assert(video_mode_id == 0);  // Matches get_available_video_modes().
1412 }
1413
1414 std::map<uint32_t, std::string> BMUSBCapture::get_available_video_inputs() const
1415 {
1416         return {
1417                 { 0x00000000, "HDMI/SDI" },
1418                 { 0x02000000, "Component" },
1419                 { 0x04000000, "Composite" },
1420                 { 0x06000000, "S-video" }
1421         };
1422 }
1423
1424 void BMUSBCapture::set_video_input(uint32_t video_input_id)
1425 {
1426         assert((video_input_id & ~0x06000000) == 0);
1427         current_video_input = video_input_id;
1428         update_capture_mode();
1429 }
1430
1431 std::map<uint32_t, std::string> BMUSBCapture::get_available_audio_inputs() const
1432 {
1433         return {
1434                 { 0x00000000, "Embedded" },
1435                 { 0x10000000, "Analog" }
1436         };
1437 }
1438
1439 void BMUSBCapture::set_audio_input(uint32_t audio_input_id)
1440 {
1441         assert((audio_input_id & ~0x10000000) == 0);
1442         current_audio_input = audio_input_id;
1443         update_capture_mode();
1444 }
1445
1446 void BMUSBCapture::update_capture_mode()
1447 {
1448         // clearing the 0x20000000 bit seems to activate 10-bit capture (v210).
1449         // clearing the 0x08000000 bit seems to change the capture format (other source?)
1450         uint32_t mode = htonl(0x29000000 | current_video_input | current_audio_input);
1451
1452         int rc = libusb_control_transfer(devh, LIBUSB_REQUEST_TYPE_VENDOR | LIBUSB_ENDPOINT_OUT,
1453                 /*request=*/215, /*value=*/0, /*index=*/0, (unsigned char *)&mode, sizeof(mode), /*timeout=*/0);
1454         if (rc < 0) {
1455                 fprintf(stderr, "Error on setting mode: %s\n", libusb_error_name(rc));
1456                 exit(1);
1457         }
1458 }
1459
1460 }  // namespace bmusb