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 <movit/colorspace_conversion_effect.h>
  25
  26 #include <va/va.h>
  27 #include <va/va_drm.h>
  28 #include <va/va_x11.h>
  29
  30 using namespace Eigen;
  31 using namespace bmusb;
  32 using namespace movit;
  33 using namespace std;
  34
  35 static VAImageFormat uyvy_format;
  36
  37 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
  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 namespace {
 121
 122 void add_video_stream(AVFormatContext *avctx)
 123 {
 124         AVStream *stream = avformat_new_stream(avctx, nullptr);
 125         if (stream == nullptr) {
 126                 fprintf(stderr, "avformat_new_stream() failed\n");
 127                 abort();
 128         }
 129
 130         // FFmpeg is very picky about having audio at 1/48000 timebase,
 131         // no matter what we write. Even though we'd prefer our usual 1/120000,
 132         // put the video on the same one, so that we can have locked audio.
 133         stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
 134         stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
 135         stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
 136
 137         // Used for aspect ratio only. Can change without notice (the mux won't care).
 138         stream->codecpar->width = global_flags.width;
 139         stream->codecpar->height = global_flags.height;
 140
 141         // TODO: We could perhaps use the interpretation for each card here
 142         // (or at least the command-line flags) instead of the defaults,
 143         // but what would we do when they change?
 144         stream->codecpar->color_primaries = AVCOL_PRI_BT709;
 145         stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
 146         stream->codecpar->color_space = AVCOL_SPC_BT709;
 147         stream->codecpar->color_range = AVCOL_RANGE_MPEG;
 148         stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
 149         stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
 150 }
 151
 152 void add_audio_stream(AVFormatContext *avctx)
 153 {
 154         AVStream *stream = avformat_new_stream(avctx, nullptr);
 155         if (stream == nullptr) {
 156                 fprintf(stderr, "avformat_new_stream() failed\n");
 157                 abort();
 158         }
 159         stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
 160         stream->codecpar->codec_type = AVMEDIA_TYPE_AUDIO;
 161         stream->codecpar->codec_id = AV_CODEC_ID_PCM_S32LE;
 162         stream->codecpar->channel_layout = AV_CH_LAYOUT_STEREO;
 163         stream->codecpar->channels = 2;
 164         stream->codecpar->sample_rate = OUTPUT_FREQUENCY;
 165 }
 166
 167 void finalize_mux(AVFormatContext *avctx)
 168 {
 169         AVDictionary *options = NULL;
 170         vector<pair<string, string>> opts = MUX_OPTS;
 171         for (pair<string, string> opt : opts) {
 172                 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
 173         }
 174         if (avformat_write_header(avctx, &options) < 0) {
 175                 fprintf(stderr, "avformat_write_header() failed\n");
 176                 abort();
 177         }
 178 }
 179
 180 }  // namespace
 181
 182 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
 183         : httpd(httpd)
 184 {
 185         // Set up the mux. We don't use the Mux wrapper, because it's geared towards
 186         // a situation with only one video stream (and possibly one audio stream)
 187         // with known width/height, and we don't need the extra functionality it provides.
 188         avctx.reset(avformat_alloc_context());
 189         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 190
 191         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 192         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
 193         avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
 194         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 195
 196         for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
 197                 add_video_stream(avctx.get());
 198         }
 199         for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
 200                 add_audio_stream(avctx.get());
 201         }
 202         finalize_mux(avctx.get());
 203
 204         // Initialize VA-API.
 205         string error;
 206         va_dpy = try_open_va(va_display, &error, &config_id);
 207         if (va_dpy == nullptr) {
 208                 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
 209         }
 210
 211         encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
 212         if (va_dpy != nullptr) {
 213                 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
 214         }
 215
 216         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
 217         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
 218         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
 219         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
 220         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
 221         global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
 222
 223         running = true;
 224 }
 225
 226 MJPEGEncoder::~MJPEGEncoder()
 227 {
 228         av_free(avctx->pb->buffer);
 229
 230         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
 231         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
 232         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
 233         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
 234         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
 235         global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
 236 }
 237
 238 void MJPEGEncoder::stop()
 239 {
 240         if (!running) {
 241                 return;
 242         }
 243         running = false;
 244         should_quit = true;
 245         any_frames_to_be_encoded.notify_all();
 246         any_frames_encoding.notify_all();
 247         encoder_thread.join();
 248         if (va_dpy != nullptr) {
 249                 va_receiver_thread.join();
 250         }
 251 }
 252
 253 unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
 254 {
 255         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
 256         if (va_dpy == nullptr) {
 257                 if (error) *error = "Opening VA display failed";
 258                 return nullptr;
 259         }
 260         int major_ver, minor_ver;
 261         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
 262         if (va_status != VA_STATUS_SUCCESS) {
 263                 char buf[256];
 264                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
 265                 if (error != nullptr) *error = buf;
 266                 return nullptr;
 267         }
 268
 269         VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
 270         va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
 271                 &attr, 1, config_id);
 272         if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
 273                 if (error != nullptr) *error = "No hardware support";
 274                 return nullptr;
 275         } else if (va_status != VA_STATUS_SUCCESS) {
 276                 char buf[256];
 277                 snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
 278                 if (error != nullptr) *error = buf;
 279                 return nullptr;
 280         }
 281
 282         // TODO: Unify with the code in Futatabi.
 283         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
 284         assert(num_formats > 0);
 285
 286         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
 287         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
 288         if (va_status != VA_STATUS_SUCCESS) {
 289                 char buf[256];
 290                 snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
 291                 if (error != nullptr) *error = buf;
 292                 return nullptr;
 293         }
 294
 295         bool found = false;
 296         for (int i = 0; i < num_formats; ++i) {
 297                 if (formats[i].fourcc == VA_FOURCC_UYVY) {
 298                         memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
 299                         found = true;
 300                         break;
 301                 }
 302         }
 303         if (!found) {
 304                 if (error != nullptr) *error = "UYVY format not found";
 305                 return nullptr;
 306         }
 307
 308         return va_dpy;
 309 }
 310
 311 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, vector<int32_t> audio, const RGBTriplet &white_balance)
 312 {
 313         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 314         if (video_format.width == 0 || video_format.height == 0) {
 315                 ++metric_mjpeg_frames_zero_size_dropped;
 316                 return;
 317         }
 318         if (video_format.interlaced) {
 319                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
 320                 ++metric_mjpeg_frames_interlaced_dropped;
 321                 return;
 322         }
 323         if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
 324             !frame->interleaved) {
 325                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
 326                 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
 327                 return;
 328         }
 329         if (video_format.width > 4096 || video_format.height > 4096) {
 330                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
 331                 ++metric_mjpeg_frames_oversized_dropped;
 332                 return;
 333         }
 334
 335         lock_guard<mutex> lock(mu);
 336         if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
 337                 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
 338                 ++metric_mjpeg_overrun_dropped;
 339                 return;
 340         }
 341         ++metric_mjpeg_overrun_submitted;
 342         frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio), white_balance });
 343         any_frames_to_be_encoded.notify_all();
 344 }
 345
 346 bool MJPEGEncoder::should_encode_mjpeg_for_card(unsigned card_index)
 347 {
 348         // Only bother doing MJPEG encoding if there are any connected clients
 349         // that want the stream.
 350         if (httpd->get_num_connected_multicam_clients() == 0) {
 351                 return false;
 352         }
 353
 354         auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
 355         return (it != global_flags.card_to_mjpeg_stream_export.end());
 356 }
 357
 358 void MJPEGEncoder::encoder_thread_func()
 359 {
 360         pthread_setname_np(pthread_self(), "MJPEG_Encode");
 361         posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
 362         posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
 363         posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
 364         posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
 365
 366         for (;;) {
 367                 QueuedFrame qf;
 368                 {
 369                         unique_lock<mutex> lock(mu);
 370                         any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
 371                         if (should_quit) break;
 372                         qf = move(frames_to_be_encoded.front());
 373                         frames_to_be_encoded.pop();
 374                 }
 375
 376                 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index));  // Or should_encode_mjpeg_for_card() would have returned false.
 377                 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
 378
 379                 if (va_dpy != nullptr) {
 380                         // Will call back in the receiver thread.
 381                         encode_jpeg_va(move(qf));
 382                 } else {
 383                         // Write audio before video, since Futatabi expects it.
 384                         if (qf.audio.size() > 0) {
 385                                 write_audio_packet(qf.pts, stream_index, qf.audio);
 386                         }
 387
 388                         // Encode synchronously, in the same thread.
 389                         vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
 390                         write_mjpeg_packet(qf.pts, stream_index, jpeg.data(), jpeg.size());
 391                 }
 392         }
 393
 394         free(tmp_y);
 395         free(tmp_cbcr);
 396         free(tmp_cb);
 397         free(tmp_cr);
 398 }
 399
 400 void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const uint8_t *jpeg, size_t jpeg_size)
 401 {
 402         AVPacket pkt;
 403         memset(&pkt, 0, sizeof(pkt));
 404         pkt.buf = nullptr;
 405         pkt.data = const_cast<uint8_t *>(jpeg);
 406         pkt.size = jpeg_size;
 407         pkt.stream_index = card_index;
 408         pkt.flags = AV_PKT_FLAG_KEY;
 409         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
 410         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
 411         pkt.duration = 0;
 412
 413         if (av_write_frame(avctx.get(), &pkt) < 0) {
 414                 fprintf(stderr, "av_write_frame() failed\n");
 415                 abort();
 416         }
 417 }
 418
 419 void MJPEGEncoder::write_audio_packet(int64_t pts, unsigned card_index, const vector<int32_t> &audio)
 420 {
 421         AVPacket pkt;
 422         memset(&pkt, 0, sizeof(pkt));
 423         pkt.buf = nullptr;
 424         pkt.data = reinterpret_cast<uint8_t *>(const_cast<int32_t *>(&audio[0]));
 425         pkt.size = audio.size() * sizeof(audio[0]);
 426         pkt.stream_index = card_index + global_flags.card_to_mjpeg_stream_export.size();
 427         pkt.flags = AV_PKT_FLAG_KEY;
 428         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
 429         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
 430         size_t num_stereo_samples = audio.size() / 2;
 431         pkt.duration = av_rescale_q(num_stereo_samples, AVRational{ 1, OUTPUT_FREQUENCY }, time_base);
 432
 433         if (av_write_frame(avctx.get(), &pkt) < 0) {
 434                 fprintf(stderr, "av_write_frame() failed\n");
 435                 abort();
 436         }
 437 }
 438
 439 class VABufferDestroyer {
 440 public:
 441         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 442                 : dpy(dpy), buf(buf) {}
 443
 444         ~VABufferDestroyer() {
 445                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 446                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 447         }
 448
 449 private:
 450         VADisplay dpy;
 451         VABufferID buf;
 452 };
 453
 454 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
 455 {
 456         {
 457                 lock_guard<mutex> lock(va_resources_mutex);
 458                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
 459                         if (it->width == width && it->height == height) {
 460                                 VAResources ret = *it;
 461                                 va_resources_freelist.erase(it);
 462                                 return ret;
 463                         }
 464                 }
 465         }
 466
 467         VAResources ret;
 468
 469         ret.width = width;
 470         ret.height = height;
 471
 472         VASurfaceAttrib attrib;
 473         attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
 474         attrib.type = VASurfaceAttribPixelFormat;
 475         attrib.value.type = VAGenericValueTypeInteger;
 476         attrib.value.value.i = VA_FOURCC_UYVY;
 477
 478         VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
 479                 width, height,
 480                 &ret.surface, 1, &attrib, 1);
 481         CHECK_VASTATUS(va_status, "vaCreateSurfaces");
 482
 483         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
 484         CHECK_VASTATUS(va_status, "vaCreateContext");
 485
 486         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
 487         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 488
 489         va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
 490         CHECK_VASTATUS(va_status, "vaCreateImage");
 491
 492         return ret;
 493 }
 494
 495 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
 496 {
 497         lock_guard<mutex> lock(va_resources_mutex);
 498         if (va_resources_freelist.size() > 50) {
 499                 auto it = va_resources_freelist.end();
 500                 --it;
 501
 502                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
 503                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 504
 505                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
 506                 CHECK_VASTATUS(va_status, "vaDestroyContext");
 507
 508                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
 509                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
 510
 511                 va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
 512                 CHECK_VASTATUS(va_status, "vaDestroyImage");
 513
 514                 va_resources_freelist.erase(it);
 515         }
 516
 517         va_resources_freelist.push_front(resources);
 518 }
 519
 520 namespace {
 521
 522 void push16(uint16_t val, string *str)
 523 {
 524         str->push_back(val >> 8);
 525         str->push_back(val & 0xff);
 526 }
 527
 528 void push32(uint32_t val, string *str)
 529 {
 530         str->push_back(val >> 24);
 531         str->push_back((val >> 16) & 0xff);
 532         str->push_back((val >> 8) & 0xff);
 533         str->push_back(val & 0xff);
 534 }
 535
 536 }  // namespace
 537
 538 void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, const RGBTriplet &white_balance, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
 539 {
 540         jpeg_error_mgr jerr;
 541         cinfo->err = jpeg_std_error(&jerr);
 542         jpeg_create_compress(cinfo);
 543
 544         cinfo->dest = (jpeg_destination_mgr *)dest;
 545
 546         cinfo->input_components = 3;
 547         jpeg_set_defaults(cinfo);
 548         jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
 549
 550         cinfo->image_width = width;
 551         cinfo->image_height = height;
 552         cinfo->raw_data_in = true;
 553         jpeg_set_colorspace(cinfo, JCS_YCbCr);
 554         cinfo->comp_info[0].h_samp_factor = 2;
 555         cinfo->comp_info[0].v_samp_factor = 1;
 556         cinfo->comp_info[1].h_samp_factor = 1;
 557         cinfo->comp_info[1].v_samp_factor = 1;
 558         cinfo->comp_info[2].h_samp_factor = 1;
 559         cinfo->comp_info[2].v_samp_factor = 1;
 560         cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 561         jpeg_start_compress(cinfo, true);
 562
 563         if (fabs(white_balance.r - 1.0f) > 1e-3 ||
 564             fabs(white_balance.g - 1.0f) > 1e-3 ||
 565             fabs(white_balance.b - 1.0f) > 1e-3) {
 566                 // Convert from (linear) RGB to XYZ.
 567                 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
 568                 Vector3d xyz = rgb_to_xyz_matrix * Vector3d(white_balance.r, white_balance.g, white_balance.b);
 569
 570                 // Convert from XYZ to xyz by normalizing.
 571                 xyz /= (xyz[0] + xyz[1] + xyz[2]);
 572
 573                 // Create a very rudimentary EXIF header to hold our white point.
 574                 string exif;
 575
 576                 // Exif header, followed by some padding.
 577                 exif = "Exif";
 578                 push16(0, &exif);
 579
 580                 // TIFF header first:
 581                 exif += "MM";  // Big endian.
 582
 583                 // Magic number.
 584                 push16(42, &exif);
 585
 586                 // Offset of first IFD (relative to the MM, immediately after the header).
 587                 push32(exif.size() - 6 + 4, &exif);
 588
 589                 // Now the actual IFD.
 590
 591                 // One entry.
 592                 push16(1, &exif);
 593
 594                 // WhitePoint tag ID.
 595                 push16(0x13e, &exif);
 596
 597                 // Rational type.
 598                 push16(5, &exif);
 599
 600                 // Two values (x and y; z is implicit due to normalization).
 601                 push32(2, &exif);
 602
 603                 // Offset (relative to the MM, immediately after the last IFD).
 604                 push32(exif.size() - 6 + 8, &exif);
 605
 606                 // No more IFDs.
 607                 push32(0, &exif);
 608
 609                 // The actual values.
 610                 push32(lrintf(xyz[0] * 10000.0f), &exif);
 611                 push32(10000, &exif);
 612                 push32(lrintf(xyz[1] * 10000.0f), &exif);
 613                 push32(10000, &exif);
 614
 615                 jpeg_write_marker(cinfo, JPEG_APP0 + 1, (const JOCTET *)exif.data(), exif.size());
 616         }
 617
 618         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 619         // (and nothing else).
 620         jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 621 }
 622
 623 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, const RGBTriplet &white_balance, jpeg_compress_struct *cinfo)
 624 {
 625         VectorDestinationManager dest;
 626         init_jpeg_422(width, height, white_balance, &dest, cinfo);
 627
 628         // Make a dummy black image; there's seemingly no other easy way of
 629         // making libjpeg outputting all of its headers.
 630         JSAMPROW yptr[8], cbptr[8], crptr[8];
 631         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 632         memset(tmp_y, 0, 4096);
 633         memset(tmp_cb, 0, 4096);
 634         memset(tmp_cr, 0, 4096);
 635         for (unsigned yy = 0; yy < 8; ++yy) {
 636                 yptr[yy] = tmp_y;
 637                 cbptr[yy] = tmp_cb;
 638                 crptr[yy] = tmp_cr;
 639         }
 640         for (unsigned y = 0; y < height; y += 8) {
 641                 jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
 642         }
 643         jpeg_finish_compress(cinfo);
 644
 645         // We're only interested in the header, not the data after it.
 646         dest.term_destination();
 647         for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
 648                 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
 649                         unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
 650                         dest.dest.resize(i + len + 2);
 651                         break;
 652                 }
 653         }
 654
 655         return dest.dest;
 656 }
 657
 658 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height, const RGBTriplet &white_balance)
 659 {
 660         pair<unsigned, unsigned> key(width, height);
 661         if (va_data_for_resolution.count(key)) {
 662                 return va_data_for_resolution[key];
 663         }
 664
 665         // Use libjpeg to generate a header and set sane defaults for e.g.
 666         // quantization tables. Then do the actual encode with VA-API.
 667         jpeg_compress_struct cinfo;
 668         vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, &cinfo);
 669
 670         // Picture parameters.
 671         VAEncPictureParameterBufferJPEG pic_param;
 672         memset(&pic_param, 0, sizeof(pic_param));
 673         pic_param.reconstructed_picture = VA_INVALID_ID;
 674         pic_param.picture_width = cinfo.image_width;
 675         pic_param.picture_height = cinfo.image_height;
 676         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 677                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 678                 pic_param.component_id[component_idx] = comp->component_id;
 679                 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
 680         }
 681         pic_param.num_components = cinfo.num_components;
 682         pic_param.num_scan = 1;
 683         pic_param.sample_bit_depth = 8;
 684         pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
 685         pic_param.pic_flags.bits.huffman = 1;
 686         pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
 687
 688         // Quantization matrices.
 689         VAQMatrixBufferJPEG q;
 690         memset(&q, 0, sizeof(q));
 691
 692         q.load_lum_quantiser_matrix = true;
 693         q.load_chroma_quantiser_matrix = true;
 694         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 695                 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
 696                 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
 697                 if (qtbl == nullptr) continue;
 698
 699                 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
 700                 for (int i = 0; i < 64; ++i) {
 701                         if (qtbl->quantval[i] > 255) {
 702                                 fprintf(stderr, "Baseline JPEG only!\n");
 703                                 abort();
 704                         }
 705                         qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
 706                 }
 707         }
 708
 709         // Huffman tables (arithmetic is not supported).
 710         VAHuffmanTableBufferJPEGBaseline huff;
 711         memset(&huff, 0, sizeof(huff));
 712
 713         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 714                 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 715                 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 716                 if (ac_hufftbl == nullptr) {
 717                         assert(dc_hufftbl == nullptr);
 718                         huff.load_huffman_table[huff_tbl_idx] = 0;
 719                 } else {
 720                         assert(dc_hufftbl != nullptr);
 721                         huff.load_huffman_table[huff_tbl_idx] = 1;
 722
 723                         for (int i = 0; i < 16; ++i) {
 724                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 725                         }
 726                         for (int i = 0; i < 12; ++i) {
 727                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 728                         }
 729                         for (int i = 0; i < 16; ++i) {
 730                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 731                         }
 732                         for (int i = 0; i < 162; ++i) {
 733                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 734                         }
 735                 }
 736         }
 737
 738         // Slice parameters (metadata about the slice).
 739         VAEncSliceParameterBufferJPEG parms;
 740         memset(&parms, 0, sizeof(parms));
 741         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 742                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 743                 parms.components[component_idx].component_selector = comp->component_id;
 744                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 745                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 746                 if (parms.components[component_idx].dc_table_selector > 1 ||
 747                     parms.components[component_idx].ac_table_selector > 1) {
 748                         fprintf(stderr, "Uses too many Huffman tables\n");
 749                         abort();
 750                 }
 751         }
 752         parms.num_components = cinfo.num_components;
 753         parms.restart_interval = cinfo.restart_interval;
 754
 755         jpeg_destroy_compress(&cinfo);
 756
 757         VAData ret;
 758         ret.jpeg_header = move(jpeg_header);
 759         ret.pic_param = pic_param;
 760         ret.q = q;
 761         ret.huff = huff;
 762         ret.parms = parms;
 763         va_data_for_resolution[key] = ret;
 764         return ret;
 765 }
 766
 767 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
 768 {
 769         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
 770         unsigned width = qf.video_format.width;
 771         unsigned height = qf.video_format.height;
 772
 773         VAResources resources;
 774         ReleaseVAResources release;
 775         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 776                 resources = move(userdata->va_resources);
 777                 release = move(userdata->va_resources_release);
 778         } else {
 779                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 780                 resources = get_va_resources(width, height);
 781                 release = ReleaseVAResources(this, resources);
 782         }
 783
 784         VAData va_data = get_va_data_for_resolution(width, height, qf.white_balance);
 785         va_data.pic_param.coded_buf = resources.data_buffer;
 786
 787         VABufferID pic_param_buffer;
 788         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
 789         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 790         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 791
 792         VABufferID q_buffer;
 793         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
 794         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 795         VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
 796
 797         VABufferID huff_buffer;
 798         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
 799         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 800         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 801
 802         VABufferID slice_param_buffer;
 803         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
 804         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 805         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 806
 807         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 808                 // The pixel data is already put into the image by the caller.
 809                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 810                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 811         } else {
 812                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 813
 814                 // Upload the pixel data.
 815                 uint8_t *surface_p = nullptr;
 816                 vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
 817
 818                 size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 819                 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 820
 821                 {
 822                         const uint8_t *src = qf.frame->data_copy + field_start;
 823                         uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
 824                         memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
 825                 }
 826
 827                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 828                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 829         }
 830
 831         qf.frame->data_copy = nullptr;
 832
 833         // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
 834         // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
 835         // Exactly why is unclear, but it seems to involve L3 cache usage when there
 836         // are many high-res (1080p+) images in play.
 837         va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
 838         CHECK_VASTATUS(va_status, "vaPutImage");
 839
 840         // Finally, stick in the JPEG header.
 841         VAEncPackedHeaderParameterBuffer header_parm;
 842         header_parm.type = VAEncPackedHeaderRawData;
 843         header_parm.bit_length = 8 * va_data.jpeg_header.size();
 844
 845         VABufferID header_parm_buffer;
 846         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
 847         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 848         VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
 849
 850         VABufferID header_data_buffer;
 851         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
 852         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 853         VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
 854
 855         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 856         CHECK_VASTATUS(va_status, "vaBeginPicture");
 857         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 858         CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
 859         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
 860         CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
 861         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 862         CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
 863         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 864         CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
 865         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
 866         CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
 867         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
 868         CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
 869         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 870         CHECK_VASTATUS(va_status, "vaEndPicture");
 871
 872         qf.resources = move(resources);
 873         qf.resource_releaser = move(release);
 874
 875         lock_guard<mutex> lock(mu);
 876         frames_encoding.push(move(qf));
 877         any_frames_encoding.notify_all();
 878 }
 879
 880 void MJPEGEncoder::va_receiver_thread_func()
 881 {
 882         pthread_setname_np(pthread_self(), "MJPEG_Receive");
 883         for (;;) {
 884                 QueuedFrame qf;
 885                 {
 886                         unique_lock<mutex> lock(mu);
 887                         any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
 888                         if (should_quit) return;
 889                         qf = move(frames_encoding.front());
 890                         frames_encoding.pop();
 891                 }
 892
 893                 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index));  // Or should_encode_mjpeg_for_card() would have returned false.
 894                 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
 895
 896                 // Write audio before video, since Futatabi expects it.
 897                 if (qf.audio.size() > 0) {
 898                         write_audio_packet(qf.pts, stream_index, qf.audio);
 899                 }
 900
 901                 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
 902                 CHECK_VASTATUS(va_status, "vaSyncSurface");
 903
 904                 VACodedBufferSegment *segment;
 905                 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
 906                 CHECK_VASTATUS(va_status, "vaMapBuffer");
 907
 908                 const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
 909                 write_mjpeg_packet(qf.pts, stream_index, coded_buf, segment->size);
 910
 911                 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
 912                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 913         }
 914 }
 915
 916 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
 917 {
 918         unsigned width = qf.video_format.width;
 919         unsigned height = qf.video_format.height;
 920
 921         VectorDestinationManager dest;
 922         jpeg_compress_struct cinfo;
 923         init_jpeg_422(width, height, qf.white_balance, &dest, &cinfo);
 924
 925         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 926         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 927
 928         JSAMPROW yptr[8], cbptr[8], crptr[8];
 929         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 930         for (unsigned y = 0; y < qf.video_format.height; y += 8) {
 931                 const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
 932
 933                 memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
 934                 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
 935                 for (unsigned yy = 0; yy < 8; ++yy) {
 936                         yptr[yy] = tmp_y + yy * width;
 937                         cbptr[yy] = tmp_cb + yy * width / 2;
 938                         crptr[yy] = tmp_cr + yy * width / 2;
 939                 }
 940                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 941         }
 942         jpeg_finish_compress(&cinfo);
 943
 944         return dest.dest;
 945 }