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