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 static VAImageFormat uyvy_format;
  32
  33 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
  34
  35 // From libjpeg (although it's of course identical between implementations).
  36 static const int jpeg_natural_order[DCTSIZE2] = {
  37          0,  1,  8, 16,  9,  2,  3, 10,
  38         17, 24, 32, 25, 18, 11,  4,  5,
  39         12, 19, 26, 33, 40, 48, 41, 34,
  40         27, 20, 13,  6,  7, 14, 21, 28,
  41         35, 42, 49, 56, 57, 50, 43, 36,
  42         29, 22, 15, 23, 30, 37, 44, 51,
  43         58, 59, 52, 45, 38, 31, 39, 46,
  44         53, 60, 61, 54, 47, 55, 62, 63,
  45 };
  46
  47 struct VectorDestinationManager {
  48         jpeg_destination_mgr pub;
  49         std::vector<uint8_t> dest;
  50
  51         VectorDestinationManager()
  52         {
  53                 pub.init_destination = init_destination_thunk;
  54                 pub.empty_output_buffer = empty_output_buffer_thunk;
  55                 pub.term_destination = term_destination_thunk;
  56         }
  57
  58         static void init_destination_thunk(j_compress_ptr ptr)
  59         {
  60                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  61         }
  62
  63         inline void init_destination()
  64         {
  65                 make_room(0);
  66         }
  67
  68         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
  69         {
  70                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
  71         }
  72
  73         inline bool empty_output_buffer()
  74         {
  75                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
  76                 return true;
  77         }
  78
  79         inline void make_room(size_t bytes_used)
  80         {
  81                 dest.resize(bytes_used + 4096);
  82                 dest.resize(dest.capacity());
  83                 pub.next_output_byte = dest.data() + bytes_used;
  84                 pub.free_in_buffer = dest.size() - bytes_used;
  85         }
  86
  87         static void term_destination_thunk(j_compress_ptr ptr)
  88         {
  89                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
  90         }
  91
  92         inline void term_destination()
  93         {
  94                 dest.resize(dest.size() - pub.free_in_buffer);
  95         }
  96 };
  97 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
  98
  99 int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 100 {
 101         MJPEGEncoder *engine = (MJPEGEncoder *)opaque;
 102         return engine->write_packet2(buf, buf_size, type, time);
 103 }
 104
 105 int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 106 {
 107         if (type == AVIO_DATA_MARKER_HEADER) {
 108                 mux_header.append((char *)buf, buf_size);
 109                 httpd->set_header(HTTPD::MULTICAM_STREAM, mux_header);
 110         } else {
 111                 httpd->add_data(HTTPD::MULTICAM_STREAM, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
 112         }
 113         return buf_size;
 114 }
 115
 116 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
 117         : httpd(httpd)
 118 {
 119         // Set up the mux. We don't use the Mux wrapper, because it's geared towards
 120         // a situation with only one video stream (and possibly one audio stream)
 121         // with known width/height, and we don't need the extra functionality it provides.
 122         avctx.reset(avformat_alloc_context());
 123         avctx->oformat = av_guess_format("mp4", nullptr, nullptr);
 124
 125         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 126         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 127         avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
 128         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 129
 130         for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
 131                 AVStream *stream = avformat_new_stream(avctx.get(), nullptr);
 132                 if (stream == nullptr) {
 133                         fprintf(stderr, "avformat_new_stream() failed\n");
 134                         exit(1);
 135                 }
 136                 stream->time_base = AVRational{ 1, TIMEBASE };
 137                 stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
 138                 stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
 139
 140                 // Used for aspect ratio only. Can change without notice (the mux won't care).
 141                 stream->codecpar->width = global_flags.width;
 142                 stream->codecpar->height = global_flags.height;
 143
 144                 // TODO: We could perhaps use the interpretation for each card here
 145                 // (or at least the command-line flags) instead of the defaults,
 146                 // but what would we do when they change?
 147                 stream->codecpar->color_primaries = AVCOL_PRI_BT709;
 148                 stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
 149                 stream->codecpar->color_space = AVCOL_SPC_BT709;
 150                 stream->codecpar->color_range = AVCOL_RANGE_MPEG;
 151                 stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
 152                 stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
 153         }
 154
 155         AVDictionary *options = NULL;
 156         vector<pair<string, string>> opts = MUX_OPTS;
 157         for (pair<string, string> opt : opts) {
 158                 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
 159         }
 160         if (avformat_write_header(avctx.get(), &options) < 0) {
 161                 fprintf(stderr, "avformat_write_header() failed\n");
 162                 exit(1);
 163         }
 164
 165         // Initialize VA-API.
 166         string error;
 167         va_dpy = try_open_va(va_display, &error, &config_id);
 168         if (va_dpy == nullptr) {
 169                 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
 170         }
 171
 172         encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
 173         if (va_dpy != nullptr) {
 174                 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
 175         }
 176
 177         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
 178         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
 179         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
 180         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
 181         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
 182         global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
 183
 184         running = true;
 185 }
 186
 187 MJPEGEncoder::~MJPEGEncoder()
 188 {
 189         av_free(avctx->pb->buffer);
 190
 191         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
 192         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
 193         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
 194         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
 195         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
 196         global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
 197 }
 198
 199 void MJPEGEncoder::stop()
 200 {
 201         if (!running) {
 202                 return;
 203         }
 204         running = false;
 205         should_quit = true;
 206         any_frames_to_be_encoded.notify_all();
 207         any_frames_encoding.notify_all();
 208         encoder_thread.join();
 209         if (va_dpy != nullptr) {
 210                 va_receiver_thread.join();
 211         }
 212 }
 213
 214 unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
 215 {
 216         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
 217         if (va_dpy == nullptr) {
 218                 if (error) *error = "Opening VA display failed";
 219                 return nullptr;
 220         }
 221         int major_ver, minor_ver;
 222         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
 223         if (va_status != VA_STATUS_SUCCESS) {
 224                 char buf[256];
 225                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
 226                 if (error != nullptr) *error = buf;
 227                 return nullptr;
 228         }
 229
 230         VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
 231         va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
 232                 &attr, 1, config_id);
 233         if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
 234                 if (error != nullptr) *error = "No hardware support";
 235                 return nullptr;
 236         } else if (va_status != VA_STATUS_SUCCESS) {
 237                 char buf[256];
 238                 snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
 239                 if (error != nullptr) *error = buf;
 240                 return nullptr;
 241         }
 242
 243         // TODO: Unify with the code in Futatabi.
 244         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
 245         assert(num_formats > 0);
 246
 247         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
 248         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
 249         if (va_status != VA_STATUS_SUCCESS) {
 250                 char buf[256];
 251                 snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
 252                 if (error != nullptr) *error = buf;
 253                 return nullptr;
 254         }
 255
 256         bool found = false;
 257         for (int i = 0; i < num_formats; ++i) {
 258                 if (formats[i].fourcc == VA_FOURCC_UYVY) {
 259                         memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
 260                         found = true;
 261                         break;
 262                 }
 263         }
 264         if (!found) {
 265                 if (error != nullptr) *error = "UYVY format not found";
 266                 return nullptr;
 267         }
 268
 269         return va_dpy;
 270 }
 271
 272 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)
 273 {
 274         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 275         if (video_format.width == 0 || video_format.height == 0) {
 276                 ++metric_mjpeg_frames_zero_size_dropped;
 277                 return;
 278         }
 279         if (video_format.interlaced) {
 280                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
 281                 ++metric_mjpeg_frames_interlaced_dropped;
 282                 return;
 283         }
 284         if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
 285             !frame->interleaved) {
 286                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
 287                 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
 288                 return;
 289         }
 290         if (video_format.width > 4096 || video_format.height > 4096) {
 291                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
 292                 ++metric_mjpeg_frames_oversized_dropped;
 293                 return;
 294         }
 295
 296         lock_guard<mutex> lock(mu);
 297         if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
 298                 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
 299                 ++metric_mjpeg_overrun_dropped;
 300                 return;
 301         }
 302         ++metric_mjpeg_overrun_submitted;
 303         frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset });
 304         any_frames_to_be_encoded.notify_all();
 305 }
 306
 307 void MJPEGEncoder::finish_frame(RefCountedFrame frame)
 308 {
 309         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 310
 311         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 312                 VAResources resources __attribute__((unused)) = move(userdata->va_resources);
 313                 ReleaseVAResources release = move(userdata->va_resources_release);
 314
 315                 VAStatus va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 316                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 317         }
 318 }
 319
 320 int MJPEGEncoder::get_mjpeg_stream_for_card(unsigned card_index)
 321 {
 322         // Only bother doing MJPEG encoding if there are any connected clients
 323         // that want the stream.
 324         if (httpd->get_num_connected_multicam_clients() == 0) {
 325                 return -1;
 326         }
 327
 328         auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
 329         if (it == global_flags.card_to_mjpeg_stream_export.end()) {
 330                 return -1;
 331         }
 332         return it->second;
 333 }
 334
 335 void MJPEGEncoder::encoder_thread_func()
 336 {
 337         pthread_setname_np(pthread_self(), "MJPEG_Encode");
 338         posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
 339         posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
 340         posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
 341         posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
 342
 343         for (;;) {
 344                 QueuedFrame qf;
 345                 {
 346                         unique_lock<mutex> lock(mu);
 347                         any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
 348                         if (should_quit) break;
 349                         qf = move(frames_to_be_encoded.front());
 350                         frames_to_be_encoded.pop();
 351                 }
 352
 353                 if (va_dpy != nullptr) {
 354                         // Will call back in the receiver thread.
 355                         encode_jpeg_va(move(qf));
 356                 } else {
 357                         // Encode synchronously, in the same thread.
 358                         vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
 359                         write_mjpeg_packet(qf.pts, qf.card_index, jpeg.data(), jpeg.size());
 360                 }
 361         }
 362
 363         free(tmp_y);
 364         free(tmp_cbcr);
 365         free(tmp_cb);
 366         free(tmp_cr);
 367 }
 368
 369 void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const uint8_t *jpeg, size_t jpeg_size)
 370 {
 371         AVPacket pkt;
 372         memset(&pkt, 0, sizeof(pkt));
 373         pkt.buf = nullptr;
 374         pkt.data = const_cast<uint8_t *>(jpeg);
 375         pkt.size = jpeg_size;
 376         pkt.stream_index = card_index;
 377         pkt.flags = AV_PKT_FLAG_KEY;
 378         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
 379         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
 380
 381         if (av_write_frame(avctx.get(), &pkt) < 0) {
 382                 fprintf(stderr, "av_write_frame() failed\n");
 383                 exit(1);
 384         }
 385 }
 386
 387 class VABufferDestroyer {
 388 public:
 389         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 390                 : dpy(dpy), buf(buf) {}
 391
 392         ~VABufferDestroyer() {
 393                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 394                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 395         }
 396
 397 private:
 398         VADisplay dpy;
 399         VABufferID buf;
 400 };
 401
 402 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
 403 {
 404         {
 405                 lock_guard<mutex> lock(va_resources_mutex);
 406                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
 407                         if (it->width == width && it->height == height) {
 408                                 VAResources ret = *it;
 409                                 va_resources_freelist.erase(it);
 410                                 return ret;
 411                         }
 412                 }
 413         }
 414
 415         VAResources ret;
 416
 417         ret.width = width;
 418         ret.height = height;
 419
 420         VASurfaceAttrib attrib;
 421         attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
 422         attrib.type = VASurfaceAttribPixelFormat;
 423         attrib.value.type = VAGenericValueTypeInteger;
 424         attrib.value.value.i = VA_FOURCC_UYVY;
 425
 426         VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
 427                 width, height,
 428                 &ret.surface, 1, &attrib, 1);
 429         CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 430
 431         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
 432         CHECK_VASTATUS(va_status, "vaCreateContext");
 433
 434         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
 435         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 436
 437         va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
 438         CHECK_VASTATUS(va_status, "vaCreateImage");
 439
 440         return ret;
 441 }
 442
 443 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
 444 {
 445         lock_guard<mutex> lock(va_resources_mutex);
 446         if (va_resources_freelist.size() > 50) {
 447                 auto it = va_resources_freelist.end();
 448                 --it;
 449
 450                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
 451                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 452
 453                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
 454                 CHECK_VASTATUS(va_status, "vaDestroyContext");
 455
 456                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
 457                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
 458
 459                 va_resources_freelist.erase(it);
 460         }
 461
 462         va_resources_freelist.push_front(resources);
 463 }
 464
 465 void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
 466 {
 467         jpeg_error_mgr jerr;
 468         cinfo->err = jpeg_std_error(&jerr);
 469         jpeg_create_compress(cinfo);
 470
 471         cinfo->dest = (jpeg_destination_mgr *)dest;
 472
 473         cinfo->input_components = 3;
 474         jpeg_set_defaults(cinfo);
 475         jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
 476
 477         cinfo->image_width = width;
 478         cinfo->image_height = height;
 479         cinfo->raw_data_in = true;
 480         jpeg_set_colorspace(cinfo, JCS_YCbCr);
 481         cinfo->comp_info[0].h_samp_factor = 2;
 482         cinfo->comp_info[0].v_samp_factor = 1;
 483         cinfo->comp_info[1].h_samp_factor = 1;
 484         cinfo->comp_info[1].v_samp_factor = 1;
 485         cinfo->comp_info[2].h_samp_factor = 1;
 486         cinfo->comp_info[2].v_samp_factor = 1;
 487         cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 488         jpeg_start_compress(cinfo, true);
 489
 490         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 491         // (and nothing else).
 492         jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 493 }
 494
 495 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
 496 {
 497         VectorDestinationManager dest;
 498         init_jpeg_422(width, height, &dest, cinfo);
 499
 500         // Make a dummy black image; there's seemingly no other easy way of
 501         // making libjpeg outputting all of its headers.
 502         JSAMPROW yptr[8], cbptr[8], crptr[8];
 503         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 504         memset(tmp_y, 0, 4096);
 505         memset(tmp_cb, 0, 4096);
 506         memset(tmp_cr, 0, 4096);
 507         for (unsigned yy = 0; yy < 8; ++yy) {
 508                 yptr[yy] = tmp_y;
 509                 cbptr[yy] = tmp_cb;
 510                 crptr[yy] = tmp_cr;
 511         }
 512         for (unsigned y = 0; y < height; y += 8) {
 513                 jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
 514         }
 515         jpeg_finish_compress(cinfo);
 516
 517         // We're only interested in the header, not the data after it.
 518         dest.term_destination();
 519         for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
 520                 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
 521                         unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
 522                         dest.dest.resize(i + len + 2);
 523                         break;
 524                 }
 525         }
 526
 527         return dest.dest;
 528 }
 529
 530 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height)
 531 {
 532         pair<unsigned, unsigned> key(width, height);
 533         if (va_data_for_resolution.count(key)) {
 534                 return va_data_for_resolution[key];
 535         }
 536
 537         // Use libjpeg to generate a header and set sane defaults for e.g.
 538         // quantization tables. Then do the actual encode with VA-API.
 539         jpeg_compress_struct cinfo;
 540         vector<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
 541
 542         // Picture parameters.
 543         VAEncPictureParameterBufferJPEG pic_param;
 544         memset(&pic_param, 0, sizeof(pic_param));
 545         pic_param.reconstructed_picture = VA_INVALID_ID;
 546         pic_param.picture_width = cinfo.image_width;
 547         pic_param.picture_height = cinfo.image_height;
 548         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 549                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 550                 pic_param.component_id[component_idx] = comp->component_id;
 551                 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
 552         }
 553         pic_param.num_components = cinfo.num_components;
 554         pic_param.num_scan = 1;
 555         pic_param.sample_bit_depth = 8;
 556         pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
 557         pic_param.pic_flags.bits.huffman = 1;
 558         pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
 559
 560         // Quantization matrices.
 561         VAQMatrixBufferJPEG q;
 562         memset(&q, 0, sizeof(q));
 563
 564         q.load_lum_quantiser_matrix = true;
 565         q.load_chroma_quantiser_matrix = true;
 566         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 567                 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
 568                 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
 569                 if (qtbl == nullptr) continue;
 570
 571                 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
 572                 for (int i = 0; i < 64; ++i) {
 573                         if (qtbl->quantval[i] > 255) {
 574                                 fprintf(stderr, "Baseline JPEG only!\n");
 575                                 abort();
 576                         }
 577                         qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
 578                 }
 579         }
 580
 581         // Huffman tables (arithmetic is not supported).
 582         VAHuffmanTableBufferJPEGBaseline huff;
 583         memset(&huff, 0, sizeof(huff));
 584
 585         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 586                 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 587                 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 588                 if (ac_hufftbl == nullptr) {
 589                         assert(dc_hufftbl == nullptr);
 590                         huff.load_huffman_table[huff_tbl_idx] = 0;
 591                 } else {
 592                         assert(dc_hufftbl != nullptr);
 593                         huff.load_huffman_table[huff_tbl_idx] = 1;
 594
 595                         for (int i = 0; i < 16; ++i) {
 596                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 597                         }
 598                         for (int i = 0; i < 12; ++i) {
 599                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 600                         }
 601                         for (int i = 0; i < 16; ++i) {
 602                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 603                         }
 604                         for (int i = 0; i < 162; ++i) {
 605                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 606                         }
 607                 }
 608         }
 609
 610         // Slice parameters (metadata about the slice).
 611         VAEncSliceParameterBufferJPEG parms;
 612         memset(&parms, 0, sizeof(parms));
 613         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 614                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 615                 parms.components[component_idx].component_selector = comp->component_id;
 616                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 617                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 618                 if (parms.components[component_idx].dc_table_selector > 1 ||
 619                     parms.components[component_idx].ac_table_selector > 1) {
 620                         fprintf(stderr, "Uses too many Huffman tables\n");
 621                         abort();
 622                 }
 623         }
 624         parms.num_components = cinfo.num_components;
 625         parms.restart_interval = cinfo.restart_interval;
 626
 627         jpeg_destroy_compress(&cinfo);
 628
 629         VAData ret;
 630         ret.jpeg_header = move(jpeg_header);
 631         ret.pic_param = pic_param;
 632         ret.q = q;
 633         ret.huff = huff;
 634         ret.parms = parms;
 635         va_data_for_resolution[key] = ret;
 636         return ret;
 637 }
 638
 639 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
 640 {
 641         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
 642         unsigned width = qf.video_format.width;
 643         unsigned height = qf.video_format.height;
 644
 645         VAResources resources;
 646         ReleaseVAResources release;
 647         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 648                 resources = move(userdata->va_resources);
 649                 release = move(userdata->va_resources_release);
 650         } else {
 651                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 652                 resources = get_va_resources(width, height);
 653                 release = ReleaseVAResources(this, resources);
 654         }
 655
 656         VAData va_data = get_va_data_for_resolution(width, height);
 657         va_data.pic_param.coded_buf = resources.data_buffer;
 658
 659         VABufferID pic_param_buffer;
 660         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
 661         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 662         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 663
 664         VABufferID q_buffer;
 665         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
 666         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 667         VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
 668
 669         VABufferID huff_buffer;
 670         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
 671         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 672         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 673
 674         VABufferID slice_param_buffer;
 675         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
 676         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 677         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 678
 679         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 680                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 681                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 682                 // The pixel data is already put into the image by the caller.
 683         } else {
 684                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 685
 686                 // Upload the pixel data.
 687                 uint8_t *surface_p = nullptr;
 688                 vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
 689
 690                 size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 691                 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 692
 693                 {
 694                         const uint8_t *src = qf.frame->data_copy + field_start;
 695                         uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
 696                         memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
 697                 }
 698
 699                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 700                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 701         }
 702
 703         // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
 704         // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
 705         // Exactly why is unclear, but it seems to involve L3 cache usage when there
 706         // are many high-res (1080p+) images in play.
 707         va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
 708         CHECK_VASTATUS(va_status, "vaPutImage");
 709
 710         // Finally, stick in the JPEG header.
 711         VAEncPackedHeaderParameterBuffer header_parm;
 712         header_parm.type = VAEncPackedHeaderRawData;
 713         header_parm.bit_length = 8 * va_data.jpeg_header.size();
 714
 715         VABufferID header_parm_buffer;
 716         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
 717         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 718         VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
 719
 720         VABufferID header_data_buffer;
 721         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
 722         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 723         VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
 724
 725         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 726         CHECK_VASTATUS(va_status, "vaBeginPicture");
 727         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 728         CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
 729         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
 730         CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
 731         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 732         CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
 733         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 734         CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
 735         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
 736         CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
 737         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
 738         CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
 739         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 740         CHECK_VASTATUS(va_status, "vaEndPicture");
 741
 742         qf.resources = move(resources);
 743         qf.resource_releaser = move(release);
 744
 745         lock_guard<mutex> lock(mu);
 746         frames_encoding.push(move(qf));
 747         any_frames_encoding.notify_all();
 748 }
 749
 750 void MJPEGEncoder::va_receiver_thread_func()
 751 {
 752         pthread_setname_np(pthread_self(), "MJPEG_Receive");
 753         for (;;) {
 754                 QueuedFrame qf;
 755                 {
 756                         unique_lock<mutex> lock(mu);
 757                         any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
 758                         if (should_quit) return;
 759                         qf = move(frames_encoding.front());
 760                         frames_encoding.pop();
 761                 }
 762
 763                 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
 764                 CHECK_VASTATUS(va_status, "vaSyncSurface");
 765
 766                 VACodedBufferSegment *segment;
 767                 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
 768                 CHECK_VASTATUS(va_status, "vaMapBuffer");
 769
 770                 const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
 771                 write_mjpeg_packet(qf.pts, qf.card_index, coded_buf, segment->size);
 772
 773                 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
 774                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 775         }
 776 }
 777
 778 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
 779 {
 780         unsigned width = qf.video_format.width;
 781         unsigned height = qf.video_format.height;
 782
 783         VectorDestinationManager dest;
 784         jpeg_compress_struct cinfo;
 785         init_jpeg_422(width, height, &dest, &cinfo);
 786
 787         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 788         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 789
 790         JSAMPROW yptr[8], cbptr[8], crptr[8];
 791         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 792         for (unsigned y = 0; y < qf.video_format.height; y += 8) {
 793                 const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
 794
 795                 memcpy_interleaved(tmp_y, tmp_cbcr, src, qf.video_format.width * 8 * 2);
 796                 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
 797                 for (unsigned yy = 0; yy < 8; ++yy) {
 798                         yptr[yy] = tmp_y + yy * width;
 799                         cbptr[yy] = tmp_cb + yy * width / 2;
 800                         crptr[yy] = tmp_cr + yy * width / 2;
 801                 }
 802                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 803         }
 804         jpeg_finish_compress(&cinfo);
 805
 806         return dest.dest;
 807 }