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 "shared/metrics.h"
  20 #include "pbo_frame_allocator.h"
  21 #include "shared/timebase.h"
  22 #include "va_display_with_cleanup.h"
  23
  24 #include <va/va.h>
  25 #include <va/va_drm.h>
  26 #include <va/va_x11.h>
  27
  28 using namespace bmusb;
  29 using namespace std;
  30
  31 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
  32
  33 #define CHECK_VASTATUS(va_status, func)                                 \
  34     if (va_status != VA_STATUS_SUCCESS) {                               \
  35         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  36         exit(1);                                                        \
  37     }
  38
  39 // From libjpeg (although it's of course identical between implementations).
  40 static const int jpeg_natural_order[DCTSIZE2] = {
  41          0,  1,  8, 16,  9,  2,  3, 10,
  42         17, 24, 32, 25, 18, 11,  4,  5,
  43         12, 19, 26, 33, 40, 48, 41, 34,
  44         27, 20, 13,  6,  7, 14, 21, 28,
  45         35, 42, 49, 56, 57, 50, 43, 36,
  46         29, 22, 15, 23, 30, 37, 44, 51,
  47         58, 59, 52, 45, 38, 31, 39, 46,
  48         53, 60, 61, 54, 47, 55, 62, 63,
  49 };
  50
  51 struct VectorDestinationManager {
  52         jpeg_destination_mgr pub;
  53         std::vector<uint8_t> dest;
  54
  55         VectorDestinationManager()
  56         {
  57                 pub.init_destination = init_destination_thunk;
  58                 pub.empty_output_buffer = empty_output_buffer_thunk;
  59                 pub.term_destination = term_destination_thunk;
  60         }
  61
  62         static void init_destination_thunk(j_compress_ptr ptr)
  63         {
  64                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  65         }
  66
  67         inline void init_destination()
  68         {
  69                 make_room(0);
  70         }
  71
  72         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  73         {
  74                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  75         }
  76
  77         inline bool empty_output_buffer()
  78         {
  79                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  80                 return true;
  81         }
  82
  83         inline void make_room(size_t bytes_used)
  84         {
  85                 dest.resize(bytes_used + 4096);
  86                 dest.resize(dest.capacity());
  87                 pub.next_output_byte = dest.data() + bytes_used;
  88                 pub.free_in_buffer = dest.size() - bytes_used;
  89         }
  90
  91         static void term_destination_thunk(j_compress_ptr ptr)
  92         {
  93                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  94         }
  95
  96         inline void term_destination()
  97         {
  98                 dest.resize(dest.size() - pub.free_in_buffer);
  99         }
 100 };
 101 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 102
 103 int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 104 {
 105         MJPEGEncoder *engine = (MJPEGEncoder *)opaque;
 106         return engine->write_packet2(buf, buf_size, type, time);
 107 }
 108
 109 int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 110 {
 111         if (type == AVIO_DATA_MARKER_HEADER) {
 112                 mux_header.append((char *)buf, buf_size);
 113                 httpd->set_header(HTTPD::MULTICAM_STREAM, mux_header);
 114         } else {
 115                 httpd->add_data(HTTPD::MULTICAM_STREAM, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
 116         }
 117         return buf_size;
 118 }
 119
 120 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
 121         : httpd(httpd)
 122 {
 123         // Set up the mux. We don't use the Mux wrapper, because it's geared towards
 124         // a situation with only one video stream (and possibly one audio stream)
 125         // with known width/height, and we don't need the extra functionality it provides.
 126         avctx.reset(avformat_alloc_context());
 127         avctx->oformat = av_guess_format("mp4", nullptr, nullptr);
 128
 129         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 130         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 131         avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
 132         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 133
 134         for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
 135                 AVStream *stream = avformat_new_stream(avctx.get(), nullptr);
 136                 if (stream == nullptr) {
 137                         fprintf(stderr, "avformat_new_stream() failed\n");
 138                         exit(1);
 139                 }
 140                 stream->time_base = AVRational{ 1, TIMEBASE };
 141                 stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
 142                 stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
 143
 144                 // Used for aspect ratio only. Can change without notice (the mux won't care).
 145                 stream->codecpar->width = global_flags.width;
 146                 stream->codecpar->height = global_flags.height;
 147
 148                 // TODO: We could perhaps use the interpretation for each card here
 149                 // (or at least the command-line flags) instead of the defaults,
 150                 // but what would we do when they change?
 151                 stream->codecpar->color_primaries = AVCOL_PRI_BT709;
 152                 stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
 153                 stream->codecpar->color_space = AVCOL_SPC_BT709;
 154                 stream->codecpar->color_range = AVCOL_RANGE_MPEG;
 155                 stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
 156                 stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
 157         }
 158
 159         AVDictionary *options = NULL;
 160         vector<pair<string, string>> opts = MUX_OPTS;
 161         for (pair<string, string> opt : opts) {
 162                 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
 163         }
 164         if (avformat_write_header(avctx.get(), &options) < 0) {
 165                 fprintf(stderr, "avformat_write_header() failed\n");
 166                 exit(1);
 167         }
 168
 169         // Initialize VA-API.
 170         string error;
 171         va_dpy = try_open_va(va_display, &error, &config_id);
 172         if (va_dpy == nullptr) {
 173                 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
 174         }
 175
 176         encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
 177         if (va_dpy != nullptr) {
 178                 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
 179         }
 180
 181         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
 182         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
 183         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
 184         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
 185         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
 186         global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
 187
 188         running = true;
 189 }
 190
 191 MJPEGEncoder::~MJPEGEncoder()
 192 {
 193         av_free(avctx->pb->buffer);
 194
 195         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
 196         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
 197         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
 198         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
 199         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
 200         global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
 201 }
 202
 203 void MJPEGEncoder::stop()
 204 {
 205         if (!running) {
 206                 return;
 207         }
 208         running = false;
 209         should_quit = true;
 210         any_frames_to_be_encoded.notify_all();
 211         encoder_thread.join();
 212         if (va_dpy != nullptr) {
 213                 va_receiver_thread.join();
 214         }
 215 }
 216
 217 unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
 218 {
 219         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
 220         if (va_dpy == nullptr) {
 221                 if (error) *error = "Opening VA display failed";
 222                 return nullptr;
 223         }
 224         int major_ver, minor_ver;
 225         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
 226         if (va_status != VA_STATUS_SUCCESS) {
 227                 char buf[256];
 228                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
 229                 if (error != nullptr) *error = buf;
 230                 return nullptr;
 231         }
 232
 233         VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
 234         va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
 235                 &attr, 1, config_id);
 236         if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
 237                 if (error != nullptr) *error = "No hardware support";
 238                 return nullptr;
 239         } else if (va_status != VA_STATUS_SUCCESS) {
 240                 char buf[256];
 241                 snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
 242                 if (error != nullptr) *error = buf;
 243                 return nullptr;
 244         }
 245
 246         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
 247         assert(num_formats > 0);
 248
 249         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
 250         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
 251         if (va_status != VA_STATUS_SUCCESS) {
 252                 char buf[256];
 253                 snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
 254                 if (error != nullptr) *error = buf;
 255                 return nullptr;
 256         }
 257
 258         return va_dpy;
 259 }
 260
 261 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)
 262 {
 263         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 264         if (video_format.width == 0 || video_format.height == 0) {
 265                 ++metric_mjpeg_frames_zero_size_dropped;
 266                 return;
 267         }
 268         if (video_format.interlaced) {
 269                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
 270                 ++metric_mjpeg_frames_interlaced_dropped;
 271                 return;
 272         }
 273         if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
 274             !frame->interleaved) {
 275                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
 276                 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
 277                 return;
 278         }
 279         if (video_format.width > 4096 || video_format.height > 4096) {
 280                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
 281                 ++metric_mjpeg_frames_oversized_dropped;
 282                 return;
 283         }
 284
 285         lock_guard<mutex> lock(mu);
 286         if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
 287                 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
 288                 ++metric_mjpeg_overrun_dropped;
 289                 return;
 290         }
 291         ++metric_mjpeg_overrun_submitted;
 292         frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset });
 293         any_frames_to_be_encoded.notify_all();
 294 }
 295
 296 void MJPEGEncoder::encoder_thread_func()
 297 {
 298         pthread_setname_np(pthread_self(), "MJPEG_Encode");
 299         posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
 300         posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
 301         posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
 302         posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
 303
 304         for (;;) {
 305                 QueuedFrame qf;
 306                 {
 307                         unique_lock<mutex> lock(mu);
 308                         any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
 309                         if (should_quit) break;
 310                         qf = move(frames_to_be_encoded.front());
 311                         frames_to_be_encoded.pop();
 312                 }
 313
 314                 if (va_dpy != nullptr) {
 315                         // Will call back in the receiver thread.
 316                         encode_jpeg_va(move(qf));
 317                 } else {
 318                         // Encode synchronously, in the same thread.
 319                         vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
 320                         write_mjpeg_packet(qf.pts, qf.card_index, jpeg);
 321                 }
 322         }
 323
 324         free(tmp_y);
 325         free(tmp_cbcr);
 326         free(tmp_cb);
 327         free(tmp_cr);
 328 }
 329
 330 void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const vector<uint8_t> &jpeg)
 331 {
 332         AVPacket pkt;
 333         memset(&pkt, 0, sizeof(pkt));
 334         pkt.buf = nullptr;
 335         pkt.data = const_cast<uint8_t *>(&jpeg[0]);
 336         pkt.size = jpeg.size();
 337         pkt.stream_index = card_index;
 338         pkt.flags = AV_PKT_FLAG_KEY;
 339         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
 340         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
 341
 342         if (av_write_frame(avctx.get(), &pkt) < 0) {
 343                 fprintf(stderr, "av_write_frame() failed\n");
 344                 exit(1);
 345         }
 346 }
 347
 348 class VABufferDestroyer {
 349 public:
 350         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 351                 : dpy(dpy), buf(buf) {}
 352
 353         ~VABufferDestroyer() {
 354                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 355                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 356         }
 357
 358 private:
 359         VADisplay dpy;
 360         VABufferID buf;
 361 };
 362
 363 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
 364 {
 365         {
 366                 lock_guard<mutex> lock(va_resources_mutex);
 367                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
 368                         if (it->width == width && it->height == height) {
 369                                 VAResources ret = *it;
 370                                 va_resources_freelist.erase(it);
 371                                 return ret;
 372                         }
 373                 }
 374         }
 375
 376         VAResources ret;
 377
 378         ret.width = width;
 379         ret.height = height;
 380
 381         VASurfaceAttrib attrib;
 382         attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
 383         attrib.type = VASurfaceAttribPixelFormat;
 384         attrib.value.type = VAGenericValueTypeInteger;
 385         attrib.value.value.i = VA_FOURCC_UYVY;
 386
 387         VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
 388                 width, height,
 389                 &ret.surface, 1, &attrib, 1);
 390         CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 391
 392         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
 393         CHECK_VASTATUS(va_status, "vaCreateContext");
 394
 395         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
 396         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 397
 398         return ret;
 399 }
 400
 401 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
 402 {
 403         lock_guard<mutex> lock(va_resources_mutex);
 404         if (va_resources_freelist.size() > 10) {
 405                 auto it = va_resources_freelist.end();
 406                 --it;
 407
 408                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
 409                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 410
 411                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
 412                 CHECK_VASTATUS(va_status, "vaDestroyContext");
 413
 414                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
 415                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
 416
 417                 va_resources_freelist.erase(it);
 418         }
 419
 420         va_resources_freelist.push_front(resources);
 421 }
 422
 423 void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
 424 {
 425         jpeg_error_mgr jerr;
 426         cinfo->err = jpeg_std_error(&jerr);
 427         jpeg_create_compress(cinfo);
 428
 429         cinfo->dest = (jpeg_destination_mgr *)dest;
 430
 431         cinfo->input_components = 3;
 432         jpeg_set_defaults(cinfo);
 433         jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
 434
 435         cinfo->image_width = width;
 436         cinfo->image_height = height;
 437         cinfo->raw_data_in = true;
 438         jpeg_set_colorspace(cinfo, JCS_YCbCr);
 439         cinfo->comp_info[0].h_samp_factor = 2;
 440         cinfo->comp_info[0].v_samp_factor = 1;
 441         cinfo->comp_info[1].h_samp_factor = 1;
 442         cinfo->comp_info[1].v_samp_factor = 1;
 443         cinfo->comp_info[2].h_samp_factor = 1;
 444         cinfo->comp_info[2].v_samp_factor = 1;
 445         cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 446         jpeg_start_compress(cinfo, true);
 447
 448         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 449         // (and nothing else).
 450         jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 451 }
 452
 453 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
 454 {
 455         VectorDestinationManager dest;
 456         init_jpeg_422(width, height, &dest, cinfo);
 457
 458         // Make a dummy black image; there's seemingly no other easy way of
 459         // making libjpeg outputting all of its headers.
 460         JSAMPROW yptr[8], cbptr[8], crptr[8];
 461         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 462         memset(tmp_y, 0, 4096);
 463         memset(tmp_cb, 0, 4096);
 464         memset(tmp_cr, 0, 4096);
 465         for (unsigned yy = 0; yy < 8; ++yy) {
 466                 yptr[yy] = tmp_y;
 467                 cbptr[yy] = tmp_cb;
 468                 crptr[yy] = tmp_cr;
 469         }
 470         for (unsigned y = 0; y < height; y += 8) {
 471                 jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
 472         }
 473         jpeg_finish_compress(cinfo);
 474
 475         // We're only interested in the header, not the data after it.
 476         dest.term_destination();
 477         for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
 478                 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
 479                         unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
 480                         dest.dest.resize(i + len + 2);
 481                         break;
 482                 }
 483         }
 484
 485         return dest.dest;
 486 }
 487
 488 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height)
 489 {
 490         pair<unsigned, unsigned> key(width, height);
 491         if (va_data_for_resolution.count(key)) {
 492                 return va_data_for_resolution[key];
 493         }
 494
 495         // Use libjpeg to generate a header and set sane defaults for e.g.
 496         // quantization tables. Then do the actual encode with VA-API.
 497         jpeg_compress_struct cinfo;
 498         vector<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
 499
 500         // Picture parameters.
 501         VAEncPictureParameterBufferJPEG pic_param;
 502         memset(&pic_param, 0, sizeof(pic_param));
 503         pic_param.reconstructed_picture = VA_INVALID_ID;
 504         pic_param.picture_width = cinfo.image_width;
 505         pic_param.picture_height = cinfo.image_height;
 506         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 507                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 508                 pic_param.component_id[component_idx] = comp->component_id;
 509                 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
 510         }
 511         pic_param.num_components = cinfo.num_components;
 512         pic_param.num_scan = 1;
 513         pic_param.sample_bit_depth = 8;
 514         pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
 515         pic_param.pic_flags.bits.huffman = 1;
 516         pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
 517
 518         // Quantization matrices.
 519         VAQMatrixBufferJPEG q;
 520         memset(&q, 0, sizeof(q));
 521
 522         q.load_lum_quantiser_matrix = true;
 523         q.load_chroma_quantiser_matrix = true;
 524         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 525                 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
 526                 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
 527                 if (qtbl == nullptr) continue;
 528
 529                 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
 530                 for (int i = 0; i < 64; ++i) {
 531                         if (qtbl->quantval[i] > 255) {
 532                                 fprintf(stderr, "Baseline JPEG only!\n");
 533                                 abort();
 534                         }
 535                         qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
 536                 }
 537         }
 538
 539         // Huffman tables (arithmetic is not supported).
 540         VAHuffmanTableBufferJPEGBaseline huff;
 541         memset(&huff, 0, sizeof(huff));
 542
 543         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 544                 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 545                 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 546                 if (ac_hufftbl == nullptr) {
 547                         assert(dc_hufftbl == nullptr);
 548                         huff.load_huffman_table[huff_tbl_idx] = 0;
 549                 } else {
 550                         assert(dc_hufftbl != nullptr);
 551                         huff.load_huffman_table[huff_tbl_idx] = 1;
 552
 553                         for (int i = 0; i < 16; ++i) {
 554                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 555                         }
 556                         for (int i = 0; i < 12; ++i) {
 557                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 558                         }
 559                         for (int i = 0; i < 16; ++i) {
 560                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 561                         }
 562                         for (int i = 0; i < 162; ++i) {
 563                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 564                         }
 565                 }
 566         }
 567
 568         // Slice parameters (metadata about the slice).
 569         VAEncSliceParameterBufferJPEG parms;
 570         memset(&parms, 0, sizeof(parms));
 571         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 572                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 573                 parms.components[component_idx].component_selector = comp->component_id;
 574                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 575                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 576                 if (parms.components[component_idx].dc_table_selector > 1 ||
 577                     parms.components[component_idx].ac_table_selector > 1) {
 578                         fprintf(stderr, "Uses too many Huffman tables\n");
 579                         abort();
 580                 }
 581         }
 582         parms.num_components = cinfo.num_components;
 583         parms.restart_interval = cinfo.restart_interval;
 584
 585         jpeg_destroy_compress(&cinfo);
 586
 587         VAData ret;
 588         ret.jpeg_header = move(jpeg_header);
 589         ret.pic_param = pic_param;
 590         ret.q = q;
 591         ret.huff = huff;
 592         ret.parms = parms;
 593         va_data_for_resolution[key] = ret;
 594         return ret;
 595 }
 596
 597 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
 598 {
 599         unsigned width = qf.video_format.width;
 600         unsigned height = qf.video_format.height;
 601
 602         VAResources resources = get_va_resources(width, height);
 603         ReleaseVAResources release(this, resources);
 604
 605         VAData va_data = get_va_data_for_resolution(width, height);
 606         va_data.pic_param.coded_buf = resources.data_buffer;
 607
 608         VABufferID pic_param_buffer;
 609         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
 610         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 611         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 612
 613         VABufferID q_buffer;
 614         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
 615         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 616         VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
 617
 618         VABufferID huff_buffer;
 619         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
 620         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 621         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 622
 623         VABufferID slice_param_buffer;
 624         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
 625         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 626         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 627
 628         VAImage image;
 629         va_status = vaDeriveImage(va_dpy->va_dpy, resources.surface, &image);
 630         CHECK_VASTATUS(va_status, "vaDeriveImage");
 631
 632         // Upload the pixel data.
 633         uint8_t *surface_p = nullptr;
 634         vaMapBuffer(va_dpy->va_dpy, image.buf, (void **)&surface_p);
 635
 636         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 637         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 638
 639         {
 640                 const uint8_t *src = qf.frame->data_copy + field_start;
 641                 uint8_t *dst = (unsigned char *)surface_p + image.offsets[0];
 642                 memcpy_with_pitch(dst, src, qf.video_format.width * 2, image.pitches[0], qf.video_format.height);
 643         }
 644
 645         va_status = vaUnmapBuffer(va_dpy->va_dpy, image.buf);
 646         CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 647         va_status = vaDestroyImage(va_dpy->va_dpy, image.image_id);
 648         CHECK_VASTATUS(va_status, "vaDestroyImage");
 649
 650         // Finally, stick in the JPEG header.
 651         VAEncPackedHeaderParameterBuffer header_parm;
 652         header_parm.type = VAEncPackedHeaderRawData;
 653         header_parm.bit_length = 8 * va_data.jpeg_header.size();
 654
 655         VABufferID header_parm_buffer;
 656         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
 657         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 658         VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
 659
 660         VABufferID header_data_buffer;
 661         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
 662         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 663         VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
 664
 665         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 666         CHECK_VASTATUS(va_status, "vaBeginPicture");
 667         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 668         CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
 669         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
 670         CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
 671         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 672         CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
 673         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 674         CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
 675         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
 676         CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
 677         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
 678         CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
 679         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 680         CHECK_VASTATUS(va_status, "vaEndPicture");
 681
 682         qf.resources = move(resources);
 683         qf.resource_releaser = move(release);
 684
 685         lock_guard<mutex> lock(mu);
 686         frames_encoding.push(move(qf));
 687         any_frames_encoding.notify_all();
 688 }
 689
 690 void MJPEGEncoder::va_receiver_thread_func()
 691 {
 692         pthread_setname_np(pthread_self(), "MJPEG_Receive");
 693         for (;;) {
 694                 QueuedFrame qf;
 695                 {
 696                         unique_lock<mutex> lock(mu);
 697                         any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
 698                         if (should_quit) return;
 699                         qf = move(frames_encoding.front());
 700                         frames_encoding.pop();
 701                 }
 702
 703                 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
 704                 CHECK_VASTATUS(va_status, "vaSyncSurface");
 705
 706                 VACodedBufferSegment *segment;
 707                 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
 708                 CHECK_VASTATUS(va_status, "vaMapBuffer");
 709
 710                 const char *coded_buf = reinterpret_cast<char *>(segment->buf);
 711                 vector<uint8_t> jpeg(coded_buf, coded_buf + segment->size);
 712
 713                 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
 714                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 715
 716                 write_mjpeg_packet(qf.pts, qf.card_index, jpeg);
 717         }
 718 }
 719
 720 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
 721 {
 722         unsigned width = qf.video_format.width;
 723         unsigned height = qf.video_format.height;
 724
 725         VectorDestinationManager dest;
 726         jpeg_compress_struct cinfo;
 727         init_jpeg_422(width, height, &dest, &cinfo);
 728
 729         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 730         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 731
 732         JSAMPROW yptr[8], cbptr[8], crptr[8];
 733         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 734         for (unsigned y = 0; y < qf.video_format.height; y += 8) {
 735                 const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
 736
 737                 memcpy_interleaved(tmp_y, tmp_cbcr, src, qf.video_format.width * 8 * 2);
 738                 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
 739                 for (unsigned yy = 0; yy < 8; ++yy) {
 740                         yptr[yy] = tmp_y + yy * width;
 741                         cbptr[yy] = tmp_cb + yy * width / 2;
 742                         crptr[yy] = tmp_cr + yy * width / 2;
 743                 }
 744                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 745         }
 746         jpeg_finish_compress(&cinfo);
 747
 748         return dest.dest;
 749 }