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

   1 #include "mjpeg_encoder.h"
   2
   3 #include <jpeglib.h>
   4 #include <unistd.h>
   5 #if __SSE2__
   6 #include <immintrin.h>
   7 #endif
   8 #include <list>
   9
  10 extern "C" {
  11 #include <libavformat/avformat.h>
  12 }
  13
  14 #include "defs.h"
  15 #include "shared/ffmpeg_raii.h"
  16 #include "flags.h"
  17 #include "shared/httpd.h"
  18 #include "shared/memcpy_interleaved.h"
  19 #include "pbo_frame_allocator.h"
  20 #include "shared/timebase.h"
  21 #include "va_display_with_cleanup.h"
  22
  23 #include <va/va.h>
  24 #include <va/va_drm.h>
  25 #include <va/va_x11.h>
  26
  27 using namespace bmusb;
  28 using namespace std;
  29
  30 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
  31
  32 #define CHECK_VASTATUS(va_status, func)                                 \
  33     if (va_status != VA_STATUS_SUCCESS) {                               \
  34         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  35         exit(1);                                                        \
  36     }
  37
  38 // From libjpeg (although it's of course identical between implementations).
  39 static const int jpeg_natural_order[DCTSIZE2] = {
  40          0,  1,  8, 16,  9,  2,  3, 10,
  41         17, 24, 32, 25, 18, 11,  4,  5,
  42         12, 19, 26, 33, 40, 48, 41, 34,
  43         27, 20, 13,  6,  7, 14, 21, 28,
  44         35, 42, 49, 56, 57, 50, 43, 36,
  45         29, 22, 15, 23, 30, 37, 44, 51,
  46         58, 59, 52, 45, 38, 31, 39, 46,
  47         53, 60, 61, 54, 47, 55, 62, 63,
  48 };
  49
  50 struct VectorDestinationManager {
  51         jpeg_destination_mgr pub;
  52         std::vector<uint8_t> dest;
  53
  54         VectorDestinationManager()
  55         {
  56                 pub.init_destination = init_destination_thunk;
  57                 pub.empty_output_buffer = empty_output_buffer_thunk;
  58                 pub.term_destination = term_destination_thunk;
  59         }
  60
  61         static void init_destination_thunk(j_compress_ptr ptr)
  62         {
  63                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  64         }
  65
  66         inline void init_destination()
  67         {
  68                 make_room(0);
  69         }
  70
  71         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  72         {
  73                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  74         }
  75
  76         inline bool empty_output_buffer()
  77         {
  78                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  79                 return true;
  80         }
  81
  82         inline void make_room(size_t bytes_used)
  83         {
  84                 dest.resize(bytes_used + 4096);
  85                 dest.resize(dest.capacity());
  86                 pub.next_output_byte = dest.data() + bytes_used;
  87                 pub.free_in_buffer = dest.size() - bytes_used;
  88         }
  89
  90         static void term_destination_thunk(j_compress_ptr ptr)
  91         {
  92                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  93         }
  94
  95         inline void term_destination()
  96         {
  97                 dest.resize(dest.size() - pub.free_in_buffer);
  98         }
  99 };
 100 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 101
 102 int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 103 {
 104         MJPEGEncoder *engine = (MJPEGEncoder *)opaque;
 105         return engine->write_packet2(buf, buf_size, type, time);
 106 }
 107
 108 int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 109 {
 110         if (type == AVIO_DATA_MARKER_HEADER) {
 111                 mux_header.append((char *)buf, buf_size);
 112                 httpd->set_header(HTTPD::MULTICAM_STREAM, mux_header);
 113         } else {
 114                 httpd->add_data(HTTPD::MULTICAM_STREAM, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
 115         }
 116         return buf_size;
 117 }
 118
 119 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
 120         : httpd(httpd)
 121 {
 122         // Set up the mux. We don't use the Mux wrapper, because it's geared towards
 123         // a situation with only one video stream (and possibly one audio stream)
 124         // with known width/height, and we don't need the extra functionality it provides.
 125         avctx.reset(avformat_alloc_context());
 126         avctx->oformat = av_guess_format("mp4", nullptr, nullptr);
 127
 128         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 129         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 130         avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
 131         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 132
 133         for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
 134                 AVStream *stream = avformat_new_stream(avctx.get(), nullptr);
 135                 if (stream == nullptr) {
 136                         fprintf(stderr, "avformat_new_stream() failed\n");
 137                         exit(1);
 138                 }
 139                 stream->time_base = AVRational{ 1, TIMEBASE };
 140                 stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
 141                 stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
 142
 143                 // Used for aspect ratio only. Can change without notice (the mux won't care).
 144                 stream->codecpar->width = global_flags.width;
 145                 stream->codecpar->height = global_flags.height;
 146
 147                 // TODO: We could perhaps use the interpretation for each card here
 148                 // (or at least the command-line flags) instead of the defaults,
 149                 // but what would we do when they change?
 150                 stream->codecpar->color_primaries = AVCOL_PRI_BT709;
 151                 stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
 152                 stream->codecpar->color_space = AVCOL_SPC_BT709;
 153                 stream->codecpar->color_range = AVCOL_RANGE_MPEG;
 154                 stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
 155                 stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
 156         }
 157
 158         AVDictionary *options = NULL;
 159         vector<pair<string, string>> opts = MUX_OPTS;
 160         for (pair<string, string> opt : opts) {
 161                 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
 162         }
 163         if (avformat_write_header(avctx.get(), &options) < 0) {
 164                 fprintf(stderr, "avformat_write_header() failed\n");
 165                 exit(1);
 166         }
 167
 168         // Initialize VA-API.
 169         string error;
 170         va_dpy = try_open_va(va_display, &error, &config_id);
 171         if (va_dpy == nullptr) {
 172                 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
 173         }
 174
 175         encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
 176         if (va_dpy != nullptr) {
 177                 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
 178         }
 179
 180         running = true;
 181 }
 182
 183 MJPEGEncoder::~MJPEGEncoder()
 184 {
 185         av_free(avctx->pb->buffer);
 186 }
 187
 188 void MJPEGEncoder::stop()
 189 {
 190         if (!running) {
 191                 return;
 192         }
 193         running = false;
 194         should_quit = true;
 195         any_frames_to_be_encoded.notify_all();
 196         encoder_thread.join();
 197         if (va_dpy != nullptr) {
 198                 va_receiver_thread.join();
 199         }
 200 }
 201
 202 unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
 203 {
 204         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
 205         if (va_dpy == nullptr) {
 206                 if (error) *error = "Opening VA display failed";
 207                 return nullptr;
 208         }
 209         int major_ver, minor_ver;
 210         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
 211         if (va_status != VA_STATUS_SUCCESS) {
 212                 char buf[256];
 213                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
 214                 if (error != nullptr) *error = buf;
 215                 return nullptr;
 216         }
 217
 218         VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
 219         va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
 220                 &attr, 1, config_id);
 221         if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
 222                 if (error != nullptr) *error = "No hardware support";
 223                 return nullptr;
 224         } else if (va_status != VA_STATUS_SUCCESS) {
 225                 char buf[256];
 226                 snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
 227                 if (error != nullptr) *error = buf;
 228                 return nullptr;
 229         }
 230
 231         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
 232         assert(num_formats > 0);
 233
 234         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
 235         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
 236         if (va_status != VA_STATUS_SUCCESS) {
 237                 char buf[256];
 238                 snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
 239                 if (error != nullptr) *error = buf;
 240                 return nullptr;
 241         }
 242
 243         return va_dpy;
 244 }
 245
 246 void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFrame frame, const bmusb::VideoFormat &video_format, size_t y_offset, size_t cbcr_offset)
 247 {
 248         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 249         if (video_format.width == 0 || video_format.height == 0) {
 250                 return;
 251         }
 252         if (video_format.interlaced) {
 253                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
 254                 return;
 255         }
 256         if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
 257             !frame->interleaved) {
 258                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
 259                 return;
 260         }
 261         if (video_format.width > 4096 || video_format.height > 4096) {
 262                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
 263                 return;
 264         }
 265
 266         lock_guard<mutex> lock(mu);
 267         if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
 268                 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
 269                 return;
 270         }
 271         frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset });
 272         any_frames_to_be_encoded.notify_all();
 273 }
 274
 275 void MJPEGEncoder::encoder_thread_func()
 276 {
 277         pthread_setname_np(pthread_self(), "MJPEG_Encode");
 278         posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
 279         posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
 280         posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
 281         posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
 282
 283         for (;;) {
 284                 QueuedFrame qf;
 285                 {
 286                         unique_lock<mutex> lock(mu);
 287                         any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
 288                         if (should_quit) break;
 289                         qf = move(frames_to_be_encoded.front());
 290                         frames_to_be_encoded.pop();
 291                 }
 292
 293                 if (va_dpy != nullptr) {
 294                         // Will call back in the receiver thread.
 295                         encode_jpeg_va(move(qf));
 296                 } else {
 297                         // Encode synchronously, in the same thread.
 298                         vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
 299                         write_mjpeg_packet(qf.pts, qf.card_index, jpeg);
 300                 }
 301         }
 302
 303         free(tmp_y);
 304         free(tmp_cbcr);
 305         free(tmp_cb);
 306         free(tmp_cr);
 307 }
 308
 309 void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const vector<uint8_t> &jpeg)
 310 {
 311         AVPacket pkt;
 312         memset(&pkt, 0, sizeof(pkt));
 313         pkt.buf = nullptr;
 314         pkt.data = const_cast<uint8_t *>(&jpeg[0]);
 315         pkt.size = jpeg.size();
 316         pkt.stream_index = card_index;
 317         pkt.flags = AV_PKT_FLAG_KEY;
 318         pkt.pts = pkt.dts = pts;
 319
 320         if (av_write_frame(avctx.get(), &pkt) < 0) {
 321                 fprintf(stderr, "av_write_frame() failed\n");
 322                 exit(1);
 323         }
 324 }
 325
 326 class VABufferDestroyer {
 327 public:
 328         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 329                 : dpy(dpy), buf(buf) {}
 330
 331         ~VABufferDestroyer() {
 332                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 333                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 334         }
 335
 336 private:
 337         VADisplay dpy;
 338         VABufferID buf;
 339 };
 340
 341 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
 342 {
 343         {
 344                 lock_guard<mutex> lock(va_resources_mutex);
 345                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
 346                         if (it->width == width && it->height == height) {
 347                                 VAResources ret = *it;
 348                                 va_resources_freelist.erase(it);
 349                                 return ret;
 350                         }
 351                 }
 352         }
 353
 354         VAResources ret;
 355
 356         ret.width = width;
 357         ret.height = height;
 358
 359         VASurfaceAttrib attrib;
 360         attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
 361         attrib.type = VASurfaceAttribPixelFormat;
 362         attrib.value.type = VAGenericValueTypeInteger;
 363         attrib.value.value.i = VA_FOURCC_UYVY;
 364
 365         VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
 366                 width, height,
 367                 &ret.surface, 1, &attrib, 1);
 368         CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 369
 370         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
 371         CHECK_VASTATUS(va_status, "vaCreateContext");
 372
 373         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
 374         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 375
 376         return ret;
 377 }
 378
 379 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
 380 {
 381         lock_guard<mutex> lock(va_resources_mutex);
 382         if (va_resources_freelist.size() > 10) {
 383                 auto it = va_resources_freelist.end();
 384                 --it;
 385
 386                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
 387                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 388
 389                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
 390                 CHECK_VASTATUS(va_status, "vaDestroyContext");
 391
 392                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
 393                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
 394
 395                 va_resources_freelist.erase(it);
 396         }
 397
 398         va_resources_freelist.push_front(resources);
 399 }
 400
 401 void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
 402 {
 403         jpeg_error_mgr jerr;
 404         cinfo->err = jpeg_std_error(&jerr);
 405         jpeg_create_compress(cinfo);
 406
 407         cinfo->dest = (jpeg_destination_mgr *)dest;
 408
 409         cinfo->input_components = 3;
 410         jpeg_set_defaults(cinfo);
 411         jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
 412
 413         cinfo->image_width = width;
 414         cinfo->image_height = height;
 415         cinfo->raw_data_in = true;
 416         jpeg_set_colorspace(cinfo, JCS_YCbCr);
 417         cinfo->comp_info[0].h_samp_factor = 2;
 418         cinfo->comp_info[0].v_samp_factor = 1;
 419         cinfo->comp_info[1].h_samp_factor = 1;
 420         cinfo->comp_info[1].v_samp_factor = 1;
 421         cinfo->comp_info[2].h_samp_factor = 1;
 422         cinfo->comp_info[2].v_samp_factor = 1;
 423         cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 424         jpeg_start_compress(cinfo, true);
 425
 426         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 427         // (and nothing else).
 428         jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 429 }
 430
 431 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
 432 {
 433         VectorDestinationManager dest;
 434         init_jpeg_422(width, height, &dest, cinfo);
 435
 436         // Make a dummy black image; there's seemingly no other easy way of
 437         // making libjpeg outputting all of its headers.
 438         JSAMPROW yptr[8], cbptr[8], crptr[8];
 439         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 440         memset(tmp_y, 0, 4096);
 441         memset(tmp_cb, 0, 4096);
 442         memset(tmp_cr, 0, 4096);
 443         for (unsigned yy = 0; yy < 8; ++yy) {
 444                 yptr[yy] = tmp_y;
 445                 cbptr[yy] = tmp_cb;
 446                 crptr[yy] = tmp_cr;
 447         }
 448         for (unsigned y = 0; y < height; y += 8) {
 449                 jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
 450         }
 451         jpeg_finish_compress(cinfo);
 452
 453         // We're only interested in the header, not the data after it.
 454         dest.term_destination();
 455         for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
 456                 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
 457                         unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
 458                         dest.dest.resize(i + len + 2);
 459                         break;
 460                 }
 461         }
 462
 463         return dest.dest;
 464 }
 465
 466 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height)
 467 {
 468         pair<unsigned, unsigned> key(width, height);
 469         if (va_data_for_resolution.count(key)) {
 470                 return va_data_for_resolution[key];
 471         }
 472
 473         // Use libjpeg to generate a header and set sane defaults for e.g.
 474         // quantization tables. Then do the actual encode with VA-API.
 475         jpeg_compress_struct cinfo;
 476         vector<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
 477
 478         // Picture parameters.
 479         VAEncPictureParameterBufferJPEG pic_param;
 480         memset(&pic_param, 0, sizeof(pic_param));
 481         pic_param.reconstructed_picture = VA_INVALID_ID;
 482         pic_param.picture_width = cinfo.image_width;
 483         pic_param.picture_height = cinfo.image_height;
 484         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 485                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 486                 pic_param.component_id[component_idx] = comp->component_id;
 487                 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
 488         }
 489         pic_param.num_components = cinfo.num_components;
 490         pic_param.num_scan = 1;
 491         pic_param.sample_bit_depth = 8;
 492         pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
 493         pic_param.pic_flags.bits.huffman = 1;
 494         pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
 495
 496         // Quantization matrices.
 497         VAQMatrixBufferJPEG q;
 498         memset(&q, 0, sizeof(q));
 499
 500         q.load_lum_quantiser_matrix = true;
 501         q.load_chroma_quantiser_matrix = true;
 502         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 503                 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
 504                 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
 505                 if (qtbl == nullptr) continue;
 506
 507                 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
 508                 for (int i = 0; i < 64; ++i) {
 509                         if (qtbl->quantval[i] > 255) {
 510                                 fprintf(stderr, "Baseline JPEG only!\n");
 511                                 abort();
 512                         }
 513                         qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
 514                 }
 515         }
 516
 517         // Huffman tables (arithmetic is not supported).
 518         VAHuffmanTableBufferJPEGBaseline huff;
 519         memset(&huff, 0, sizeof(huff));
 520
 521         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 522                 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 523                 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 524                 if (ac_hufftbl == nullptr) {
 525                         assert(dc_hufftbl == nullptr);
 526                         huff.load_huffman_table[huff_tbl_idx] = 0;
 527                 } else {
 528                         assert(dc_hufftbl != nullptr);
 529                         huff.load_huffman_table[huff_tbl_idx] = 1;
 530
 531                         for (int i = 0; i < 16; ++i) {
 532                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 533                         }
 534                         for (int i = 0; i < 12; ++i) {
 535                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 536                         }
 537                         for (int i = 0; i < 16; ++i) {
 538                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 539                         }
 540                         for (int i = 0; i < 162; ++i) {
 541                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 542                         }
 543                 }
 544         }
 545
 546         // Slice parameters (metadata about the slice).
 547         VAEncSliceParameterBufferJPEG parms;
 548         memset(&parms, 0, sizeof(parms));
 549         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 550                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 551                 parms.components[component_idx].component_selector = comp->component_id;
 552                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 553                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 554                 if (parms.components[component_idx].dc_table_selector > 1 ||
 555                     parms.components[component_idx].ac_table_selector > 1) {
 556                         fprintf(stderr, "Uses too many Huffman tables\n");
 557                         abort();
 558                 }
 559         }
 560         parms.num_components = cinfo.num_components;
 561         parms.restart_interval = cinfo.restart_interval;
 562
 563         jpeg_destroy_compress(&cinfo);
 564
 565         VAData ret;
 566         ret.jpeg_header = move(jpeg_header);
 567         ret.pic_param = pic_param;
 568         ret.q = q;
 569         ret.huff = huff;
 570         ret.parms = parms;
 571         va_data_for_resolution[key] = ret;
 572         return ret;
 573 }
 574
 575 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
 576 {
 577         unsigned width = qf.video_format.width;
 578         unsigned height = qf.video_format.height;
 579
 580         VAResources resources = get_va_resources(width, height);
 581         ReleaseVAResources release(this, resources);
 582
 583         VAData va_data = get_va_data_for_resolution(width, height);
 584         va_data.pic_param.coded_buf = resources.data_buffer;
 585
 586         VABufferID pic_param_buffer;
 587         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
 588         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 589         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 590
 591         VABufferID q_buffer;
 592         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
 593         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 594         VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
 595
 596         VABufferID huff_buffer;
 597         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
 598         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 599         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 600
 601         VABufferID slice_param_buffer;
 602         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
 603         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 604         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 605
 606         VAImage image;
 607         va_status = vaDeriveImage(va_dpy->va_dpy, resources.surface, &image);
 608         CHECK_VASTATUS(va_status, "vaDeriveImage");
 609
 610         // Upload the pixel data.
 611         uint8_t *surface_p = nullptr;
 612         vaMapBuffer(va_dpy->va_dpy, image.buf, (void **)&surface_p);
 613
 614         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 615         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 616
 617         {
 618                 const uint8_t *src = qf.frame->data_copy + field_start;
 619                 uint8_t *dst = (unsigned char *)surface_p + image.offsets[0];
 620                 memcpy_with_pitch(dst, src, qf.video_format.width * 2, image.pitches[0], qf.video_format.height);
 621         }
 622
 623         va_status = vaUnmapBuffer(va_dpy->va_dpy, image.buf);
 624         CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 625         va_status = vaDestroyImage(va_dpy->va_dpy, image.image_id);
 626         CHECK_VASTATUS(va_status, "vaDestroyImage");
 627
 628         // Finally, stick in the JPEG header.
 629         VAEncPackedHeaderParameterBuffer header_parm;
 630         header_parm.type = VAEncPackedHeaderRawData;
 631         header_parm.bit_length = 8 * va_data.jpeg_header.size();
 632
 633         VABufferID header_parm_buffer;
 634         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
 635         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 636         VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
 637
 638         VABufferID header_data_buffer;
 639         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
 640         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 641         VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
 642
 643         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 644         CHECK_VASTATUS(va_status, "vaBeginPicture");
 645         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 646         CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
 647         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
 648         CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
 649         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 650         CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
 651         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 652         CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
 653         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
 654         CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
 655         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
 656         CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
 657         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 658         CHECK_VASTATUS(va_status, "vaEndPicture");
 659
 660         qf.resources = move(resources);
 661         qf.resource_releaser = move(release);
 662
 663         lock_guard<mutex> lock(mu);
 664         frames_encoding.push(move(qf));
 665         any_frames_encoding.notify_all();
 666 }
 667
 668 void MJPEGEncoder::va_receiver_thread_func()
 669 {
 670         pthread_setname_np(pthread_self(), "MJPEG_Receive");
 671         for (;;) {
 672                 QueuedFrame qf;
 673                 {
 674                         unique_lock<mutex> lock(mu);
 675                         any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
 676                         if (should_quit) return;
 677                         qf = move(frames_encoding.front());
 678                         frames_encoding.pop();
 679                 }
 680
 681                 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
 682                 CHECK_VASTATUS(va_status, "vaSyncSurface");
 683
 684                 VACodedBufferSegment *segment;
 685                 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
 686                 CHECK_VASTATUS(va_status, "vaMapBuffer");
 687
 688                 const char *coded_buf = reinterpret_cast<char *>(segment->buf);
 689                 vector<uint8_t> jpeg(coded_buf, coded_buf + segment->size);
 690
 691                 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
 692                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 693
 694                 write_mjpeg_packet(qf.pts, qf.card_index, jpeg);
 695         }
 696 }
 697
 698 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
 699 {
 700         unsigned width = qf.video_format.width;
 701         unsigned height = qf.video_format.height;
 702
 703         VectorDestinationManager dest;
 704         jpeg_compress_struct cinfo;
 705         init_jpeg_422(width, height, &dest, &cinfo);
 706
 707         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 708         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 709
 710         JSAMPROW yptr[8], cbptr[8], crptr[8];
 711         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 712         for (unsigned y = 0; y < qf.video_format.height; y += 8) {
 713                 const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
 714
 715                 memcpy_interleaved(tmp_y, tmp_cbcr, src, qf.video_format.width * 8 * 2);
 716                 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
 717                 for (unsigned yy = 0; yy < 8; ++yy) {
 718                         yptr[yy] = tmp_y + yy * width;
 719                         cbptr[yy] = tmp_cb + yy * width / 2;
 720                         crptr[yy] = tmp_cr + yy * width / 2;
 721                 }
 722                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 723         }
 724         jpeg_finish_compress(&cinfo);
 725
 726         return dest.dest;
 727 }