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

   1 #include "mjpeg_encoder.h"
   2
   3 #include <assert.h>
   4 #include <jpeglib.h>
   5 #include <unistd.h>
   6 #if __SSE2__
   7 #include <immintrin.h>
   8 #endif
   9 #include <list>
  10
  11 extern "C" {
  12 #include <libavformat/avformat.h>
  13 }
  14
  15 #include "defs.h"
  16 #include "shared/ffmpeg_raii.h"
  17 #include "flags.h"
  18 #include "shared/httpd.h"
  19 #include "shared/memcpy_interleaved.h"
  20 #include "shared/metrics.h"
  21 #include "pbo_frame_allocator.h"
  22 #include "shared/timebase.h"
  23 #include "va_display_with_cleanup.h"
  24
  25 #include <movit/colorspace_conversion_effect.h>
  26
  27 #include <va/va.h>
  28 #include <va/va_drm.h>
  29 #include <va/va_x11.h>
  30
  31 using namespace Eigen;
  32 using namespace bmusb;
  33 using namespace movit;
  34 using namespace std;
  35
  36 static VAImageFormat uyvy_format, nv12_format;
  37
  38 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
  39
  40 // The inverse of memcpy_interleaved(), with (slow) support for pitch.
  41 void interleave_with_pitch(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, size_t src_width, size_t dst_pitch, size_t height)
  42 {
  43 #if __SSE2__
  44         if (dst_pitch == src_width * 2 && (src_width * height) % 16 == 0) {
  45                 __m128i *dptr = reinterpret_cast<__m128i *>(dst);
  46                 const __m128i *sptr1 = reinterpret_cast<const __m128i *>(src1);
  47                 const __m128i *sptr2 = reinterpret_cast<const __m128i *>(src2);
  48                 for (size_t i = 0; i < src_width * height / 16; ++i) {
  49                         __m128i data1 = _mm_loadu_si128(sptr1++);
  50                         __m128i data2 = _mm_loadu_si128(sptr2++);
  51                         _mm_storeu_si128(dptr++, _mm_unpacklo_epi8(data1, data2));
  52                         _mm_storeu_si128(dptr++, _mm_unpackhi_epi8(data1, data2));
  53                 }
  54                 return;
  55         }
  56 #endif
  57
  58         for (size_t y = 0; y < height; ++y) {
  59                 uint8_t *dptr = dst + y * dst_pitch;
  60                 const uint8_t *sptr1 = src1 + y * src_width;
  61                 const uint8_t *sptr2 = src2 + y * src_width;
  62                 for (size_t x = 0; x < src_width; ++x) {
  63                         *dptr++ = *sptr1++;
  64                         *dptr++ = *sptr2++;
  65                 }
  66         }
  67 }
  68
  69 // From libjpeg (although it's of course identical between implementations).
  70 static const int jpeg_natural_order[DCTSIZE2] = {
  71          0,  1,  8, 16,  9,  2,  3, 10,
  72         17, 24, 32, 25, 18, 11,  4,  5,
  73         12, 19, 26, 33, 40, 48, 41, 34,
  74         27, 20, 13,  6,  7, 14, 21, 28,
  75         35, 42, 49, 56, 57, 50, 43, 36,
  76         29, 22, 15, 23, 30, 37, 44, 51,
  77         58, 59, 52, 45, 38, 31, 39, 46,
  78         53, 60, 61, 54, 47, 55, 62, 63,
  79 };
  80
  81 struct VectorDestinationManager {
  82         jpeg_destination_mgr pub;
  83         std::vector<uint8_t> dest;
  84
  85         VectorDestinationManager()
  86         {
  87                 pub.init_destination = init_destination_thunk;
  88                 pub.empty_output_buffer = empty_output_buffer_thunk;
  89                 pub.term_destination = term_destination_thunk;
  90         }
  91
  92         static void init_destination_thunk(j_compress_ptr ptr)
  93         {
  94                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
  95         }
  96
  97         inline void init_destination()
  98         {
  99                 make_room(0);
 100         }
 101
 102         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
 103         {
 104                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
 105         }
 106
 107         inline bool empty_output_buffer()
 108         {
 109                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
 110                 return true;
 111         }
 112
 113         inline void make_room(size_t bytes_used)
 114         {
 115                 dest.resize(bytes_used + 4096);
 116                 dest.resize(dest.capacity());
 117                 pub.next_output_byte = dest.data() + bytes_used;
 118                 pub.free_in_buffer = dest.size() - bytes_used;
 119         }
 120
 121         static void term_destination_thunk(j_compress_ptr ptr)
 122         {
 123                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
 124         }
 125
 126         inline void term_destination()
 127         {
 128                 dest.resize(dest.size() - pub.free_in_buffer);
 129         }
 130 };
 131 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 132
 133 int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 134 {
 135         WritePacket2Context *ctx = (WritePacket2Context *)opaque;
 136         return ctx->mjpeg_encoder->write_packet2(ctx->stream_id, buf, buf_size, type, time);
 137 }
 138
 139 int MJPEGEncoder::write_packet2(HTTPD::StreamID stream_id, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 140 {
 141         string *mux_header = &streams[stream_id].mux_header;
 142         if (type == AVIO_DATA_MARKER_HEADER) {
 143                 mux_header->append((char *)buf, buf_size);
 144                 httpd->set_header(stream_id, *mux_header);
 145         } else {
 146                 httpd->add_data(stream_id, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
 147         }
 148         return buf_size;
 149 }
 150
 151 namespace {
 152
 153 void add_video_stream(AVFormatContext *avctx)
 154 {
 155         AVStream *stream = avformat_new_stream(avctx, nullptr);
 156         if (stream == nullptr) {
 157                 fprintf(stderr, "avformat_new_stream() failed\n");
 158                 abort();
 159         }
 160
 161         // FFmpeg is very picky about having audio at 1/48000 timebase,
 162         // no matter what we write. Even though we'd prefer our usual 1/120000,
 163         // put the video on the same one, so that we can have locked audio.
 164         stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
 165         stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
 166         stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
 167
 168         // Used for aspect ratio only. Can change without notice (the mux won't care).
 169         stream->codecpar->width = global_flags.width;
 170         stream->codecpar->height = global_flags.height;
 171
 172         // TODO: We could perhaps use the interpretation for each card here
 173         // (or at least the command-line flags) instead of the defaults,
 174         // but what would we do when they change?
 175         stream->codecpar->color_primaries = AVCOL_PRI_BT709;
 176         stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
 177         stream->codecpar->color_space = AVCOL_SPC_BT709;
 178         stream->codecpar->color_range = AVCOL_RANGE_MPEG;
 179         stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
 180         stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
 181 }
 182
 183 void add_audio_stream(AVFormatContext *avctx)
 184 {
 185         AVStream *stream = avformat_new_stream(avctx, nullptr);
 186         if (stream == nullptr) {
 187                 fprintf(stderr, "avformat_new_stream() failed\n");
 188                 abort();
 189         }
 190         stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
 191         stream->codecpar->codec_type = AVMEDIA_TYPE_AUDIO;
 192         stream->codecpar->codec_id = AV_CODEC_ID_PCM_S32LE;
 193         stream->codecpar->channel_layout = AV_CH_LAYOUT_STEREO;
 194         stream->codecpar->channels = 2;
 195         stream->codecpar->sample_rate = OUTPUT_FREQUENCY;
 196 }
 197
 198 void finalize_mux(AVFormatContext *avctx)
 199 {
 200         AVDictionary *options = NULL;
 201         vector<pair<string, string>> opts = MUX_OPTS;
 202         for (pair<string, string> opt : opts) {
 203                 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
 204         }
 205         if (avformat_write_header(avctx, &options) < 0) {
 206                 fprintf(stderr, "avformat_write_header() failed\n");
 207                 abort();
 208         }
 209 }
 210
 211 }  // namespace
 212
 213 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
 214         : httpd(httpd)
 215 {
 216         create_ffmpeg_context(HTTPD::StreamID{ HTTPD::MULTICAM_STREAM, 0 });
 217         for (unsigned stream_idx = 0; stream_idx < MAX_VIDEO_CARDS; ++stream_idx) {
 218                 create_ffmpeg_context(HTTPD::StreamID{ HTTPD::SIPHON_STREAM, stream_idx });
 219         }
 220
 221         add_stream(HTTPD::StreamID{ HTTPD::MULTICAM_STREAM, 0 });
 222
 223         // Initialize VA-API.
 224         string error;
 225         va_dpy = try_open_va(va_display, &error, &config_id_422, &config_id_420);
 226         if (va_dpy == nullptr) {
 227                 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
 228         }
 229
 230         encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
 231         if (va_dpy != nullptr) {
 232                 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
 233         }
 234
 235         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
 236         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
 237         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
 238         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
 239         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
 240         global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
 241
 242         running = true;
 243 }
 244
 245 MJPEGEncoder::~MJPEGEncoder()
 246 {
 247         for (auto &id_and_stream : streams) {
 248                 av_free(id_and_stream.second.avctx->pb->buffer);
 249         }
 250
 251         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
 252         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
 253         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
 254         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
 255         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
 256         global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
 257 }
 258
 259 void MJPEGEncoder::stop()
 260 {
 261         if (!running) {
 262                 return;
 263         }
 264         running = false;
 265         should_quit = true;
 266         any_frames_to_be_encoded.notify_all();
 267         any_frames_encoding.notify_all();
 268         encoder_thread.join();
 269         if (va_dpy != nullptr) {
 270                 va_receiver_thread.join();
 271         }
 272 }
 273
 274 unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id_422, VAConfigID *config_id_420)
 275 {
 276         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
 277         if (va_dpy == nullptr) {
 278                 if (error) *error = "Opening VA display failed";
 279                 return nullptr;
 280         }
 281         int major_ver, minor_ver;
 282         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
 283         if (va_status != VA_STATUS_SUCCESS) {
 284                 char buf[256];
 285                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
 286                 if (error != nullptr) *error = buf;
 287                 return nullptr;
 288         }
 289
 290         {
 291                 VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
 292                 va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
 293                         &attr, 1, config_id_422);
 294                 if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
 295                         if (error != nullptr) *error = "No 4:2:2 hardware support";
 296                         return nullptr;
 297                 } else if (va_status != VA_STATUS_SUCCESS) {
 298                         char buf[256];
 299                         snprintf(buf, sizeof(buf), "vaCreateConfig() for 4:2:2 failed with status %d\n", va_status);
 300                         if (error != nullptr) *error = buf;
 301                         return nullptr;
 302                 }
 303         }
 304         {
 305                 VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV420 };
 306                 va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
 307                         &attr, 1, config_id_420);
 308                 if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
 309                         if (error != nullptr) *error = "No 4:2:0 hardware support";
 310                         return nullptr;
 311                 } else if (va_status != VA_STATUS_SUCCESS) {
 312                         char buf[256];
 313                         snprintf(buf, sizeof(buf), "vaCreateConfig() for 4:2:0 failed with status %d\n", va_status);
 314                         if (error != nullptr) *error = buf;
 315                         return nullptr;
 316                 }
 317         }
 318
 319         // TODO: Unify with the code in Futatabi.
 320         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
 321         assert(num_formats > 0);
 322
 323         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
 324         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
 325         if (va_status != VA_STATUS_SUCCESS) {
 326                 char buf[256];
 327                 snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
 328                 if (error != nullptr) *error = buf;
 329                 return nullptr;
 330         }
 331
 332         bool uyvy_found = false, nv12_found = false;
 333         for (int i = 0; i < num_formats; ++i) {
 334                 if (formats[i].fourcc == VA_FOURCC_UYVY) {
 335                         memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
 336                         uyvy_found = true;
 337                 }
 338                 if (formats[i].fourcc == VA_FOURCC_NV12) {
 339                         memcpy(&nv12_format, &formats[i], sizeof(VAImageFormat));
 340                         nv12_found = true;
 341                 }
 342         }
 343         if (!uyvy_found) {
 344                 if (error != nullptr) *error = "UYVY format not found";
 345                 return nullptr;
 346         }
 347         if (!nv12_found) {
 348                 if (error != nullptr) *error = "NV12 format not found";
 349                 return nullptr;
 350         }
 351
 352         return va_dpy;
 353 }
 354
 355 namespace {
 356
 357 bool is_uyvy(RefCountedFrame frame)
 358 {
 359         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 360         return userdata->pixel_format == PixelFormat_8BitYCbCr && frame->interleaved;
 361 }
 362
 363 bool is_i420(RefCountedFrame frame)
 364 {
 365         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 366         return userdata->pixel_format == PixelFormat_8BitYCbCrPlanar &&
 367                 userdata->ycbcr_format.chroma_subsampling_x == 2 &&
 368                 userdata->ycbcr_format.chroma_subsampling_y == 2;
 369 }
 370
 371 }  // namespace
 372
 373 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)
 374 {
 375         if (video_format.width == 0 || video_format.height == 0) {
 376                 ++metric_mjpeg_frames_zero_size_dropped;
 377                 return;
 378         }
 379         if (video_format.interlaced) {
 380                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
 381                 ++metric_mjpeg_frames_interlaced_dropped;
 382                 return;
 383         }
 384         if (!is_uyvy(frame) && !is_i420(frame)) {
 385                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
 386                 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
 387                 return;
 388         }
 389         if (video_format.width > 4096 || video_format.height > 4096) {
 390                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
 391                 ++metric_mjpeg_frames_oversized_dropped;
 392                 return;
 393         }
 394
 395         lock_guard<mutex> lock(mu);
 396         if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
 397                 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
 398                 ++metric_mjpeg_overrun_dropped;
 399                 return;
 400         }
 401         ++metric_mjpeg_overrun_submitted;
 402         frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio), white_balance });
 403         any_frames_to_be_encoded.notify_all();
 404 }
 405
 406 bool MJPEGEncoder::should_encode_mjpeg_for_card(unsigned card_index)
 407 {
 408         // Only bother doing MJPEG encoding if there are any connected clients
 409         // that want the stream.
 410         if (httpd->get_num_connected_multicam_clients() == 0 &&
 411             httpd->get_num_connected_siphon_clients(card_index) == 0) {
 412                 return false;
 413         }
 414
 415         auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
 416         return (it != global_flags.card_to_mjpeg_stream_export.end());
 417 }
 418
 419 void MJPEGEncoder::encoder_thread_func()
 420 {
 421         pthread_setname_np(pthread_self(), "MJPEG_Encode");
 422         posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
 423         posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
 424         posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
 425         posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
 426
 427         for (;;) {
 428                 QueuedFrame qf;
 429                 {
 430                         unique_lock<mutex> lock(mu);
 431                         any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
 432                         if (should_quit) break;
 433                         qf = move(frames_to_be_encoded.front());
 434                         frames_to_be_encoded.pop();
 435                 }
 436
 437                 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index));  // Or should_encode_mjpeg_for_card() would have returned false.
 438                 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
 439
 440                 if (va_dpy != nullptr) {
 441                         // Will call back in the receiver thread.
 442                         encode_jpeg_va(move(qf));
 443                 } else {
 444                         update_siphon_streams();
 445
 446                         HTTPD::StreamID multicam_id{ HTTPD::MULTICAM_STREAM, 0 };
 447                         HTTPD::StreamID siphon_id{ HTTPD::SIPHON_STREAM, qf.card_index };
 448                         assert(streams.count(multicam_id));
 449
 450                         // Write audio before video, since Futatabi expects it.
 451                         if (qf.audio.size() > 0) {
 452                                 write_audio_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index + global_flags.card_to_mjpeg_stream_export.size(), qf.audio);
 453                                 if (streams.count(siphon_id)) {
 454                                         write_audio_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/1, qf.audio);
 455                                 }
 456                         }
 457
 458                         // Encode synchronously, in the same thread.
 459                         vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
 460                         write_mjpeg_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index, jpeg.data(), jpeg.size());
 461                         if (streams.count(siphon_id)) {
 462                                 write_mjpeg_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/0, jpeg.data(), jpeg.size());
 463                         }
 464                 }
 465         }
 466
 467         free(tmp_y);
 468         free(tmp_cbcr);
 469         free(tmp_cb);
 470         free(tmp_cr);
 471 }
 472
 473 void MJPEGEncoder::write_mjpeg_packet(AVFormatContext *avctx, int64_t pts, unsigned stream_index, const uint8_t *jpeg, size_t jpeg_size)
 474 {
 475         AVPacket pkt;
 476         memset(&pkt, 0, sizeof(pkt));
 477         pkt.buf = nullptr;
 478         pkt.data = const_cast<uint8_t *>(jpeg);
 479         pkt.size = jpeg_size;
 480         pkt.stream_index = stream_index;
 481         pkt.flags = AV_PKT_FLAG_KEY;
 482         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
 483         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
 484         pkt.duration = 0;
 485
 486         if (av_write_frame(avctx, &pkt) < 0) {
 487                 fprintf(stderr, "av_write_frame() failed\n");
 488                 abort();
 489         }
 490 }
 491
 492 void MJPEGEncoder::write_audio_packet(AVFormatContext *avctx, int64_t pts, unsigned stream_index, const vector<int32_t> &audio)
 493 {
 494         AVPacket pkt;
 495         memset(&pkt, 0, sizeof(pkt));
 496         pkt.buf = nullptr;
 497         pkt.data = reinterpret_cast<uint8_t *>(const_cast<int32_t *>(&audio[0]));
 498         pkt.size = audio.size() * sizeof(audio[0]);
 499         pkt.stream_index = stream_index;
 500         pkt.flags = AV_PKT_FLAG_KEY;
 501         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
 502         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
 503         size_t num_stereo_samples = audio.size() / 2;
 504         pkt.duration = av_rescale_q(num_stereo_samples, AVRational{ 1, OUTPUT_FREQUENCY }, time_base);
 505
 506         if (av_write_frame(avctx, &pkt) < 0) {
 507                 fprintf(stderr, "av_write_frame() failed\n");
 508                 abort();
 509         }
 510 }
 511
 512 class VABufferDestroyer {
 513 public:
 514         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 515                 : dpy(dpy), buf(buf) {}
 516
 517         ~VABufferDestroyer() {
 518                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 519                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 520         }
 521
 522 private:
 523         VADisplay dpy;
 524         VABufferID buf;
 525 };
 526
 527 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height, uint32_t fourcc)
 528 {
 529         {
 530                 lock_guard<mutex> lock(va_resources_mutex);
 531                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
 532                         if (it->width == width && it->height == height && it->fourcc == fourcc) {
 533                                 VAResources ret = *it;
 534                                 va_resources_freelist.erase(it);
 535                                 return ret;
 536                         }
 537                 }
 538         }
 539
 540         VAResources ret;
 541
 542         ret.width = width;
 543         ret.height = height;
 544         ret.fourcc = fourcc;
 545
 546         VASurfaceAttrib attrib;
 547         attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
 548         attrib.type = VASurfaceAttribPixelFormat;
 549         attrib.value.type = VAGenericValueTypeInteger;
 550         attrib.value.value.i = fourcc;
 551
 552         VAStatus va_status;
 553         VAConfigID config_id;
 554         if (fourcc == VA_FOURCC_UYVY) {
 555                 va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422, width, height, &ret.surface, 1, &attrib, 1);
 556                 config_id = config_id_422;
 557         } else {
 558                 assert(fourcc == VA_FOURCC_NV12);
 559                 va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV420, width, height, &ret.surface, 1, &attrib, 1);
 560                 config_id = config_id_420;
 561         }
 562
 563         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
 564         CHECK_VASTATUS(va_status, "vaCreateContext");
 565
 566         va_status = vaCreateBuffer(va_dpy->va_dpy, ret.context, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
 567         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 568
 569         if (fourcc == VA_FOURCC_UYVY) {
 570                 va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
 571                 CHECK_VASTATUS(va_status, "vaCreateImage");
 572         } else {
 573                 assert(fourcc == VA_FOURCC_NV12);
 574                 va_status = vaCreateImage(va_dpy->va_dpy, &nv12_format, width, height, &ret.image);
 575                 CHECK_VASTATUS(va_status, "vaCreateImage");
 576         }
 577
 578         return ret;
 579 }
 580
 581 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
 582 {
 583         lock_guard<mutex> lock(va_resources_mutex);
 584         if (va_resources_freelist.size() > 50) {
 585                 auto it = va_resources_freelist.end();
 586                 --it;
 587
 588                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
 589                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 590
 591                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
 592                 CHECK_VASTATUS(va_status, "vaDestroyContext");
 593
 594                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
 595                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
 596
 597                 va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
 598                 CHECK_VASTATUS(va_status, "vaDestroyImage");
 599
 600                 va_resources_freelist.erase(it);
 601         }
 602
 603         va_resources_freelist.push_front(resources);
 604 }
 605
 606 namespace {
 607
 608 void push16(uint16_t val, string *str)
 609 {
 610         str->push_back(val >> 8);
 611         str->push_back(val & 0xff);
 612 }
 613
 614 void push32(uint32_t val, string *str)
 615 {
 616         str->push_back(val >> 24);
 617         str->push_back((val >> 16) & 0xff);
 618         str->push_back((val >> 8) & 0xff);
 619         str->push_back(val & 0xff);
 620 }
 621
 622 }  // namespace
 623
 624 void MJPEGEncoder::init_jpeg(unsigned width, unsigned height, const RGBTriplet &white_balance, VectorDestinationManager *dest, jpeg_compress_struct *cinfo, int y_h_samp_factor, int y_v_samp_factor)
 625 {
 626         jpeg_error_mgr jerr;
 627         cinfo->err = jpeg_std_error(&jerr);
 628         jpeg_create_compress(cinfo);
 629
 630         cinfo->dest = (jpeg_destination_mgr *)dest;
 631
 632         cinfo->input_components = 3;
 633         jpeg_set_defaults(cinfo);
 634         jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
 635
 636         cinfo->image_width = width;
 637         cinfo->image_height = height;
 638         cinfo->raw_data_in = true;
 639         jpeg_set_colorspace(cinfo, JCS_YCbCr);
 640         cinfo->comp_info[0].h_samp_factor = y_h_samp_factor;
 641         cinfo->comp_info[0].v_samp_factor = y_v_samp_factor;
 642         cinfo->comp_info[1].h_samp_factor = 1;
 643         cinfo->comp_info[1].v_samp_factor = 1;
 644         cinfo->comp_info[2].h_samp_factor = 1;
 645         cinfo->comp_info[2].v_samp_factor = 1;
 646         cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 647         jpeg_start_compress(cinfo, true);
 648
 649         if (fabs(white_balance.r - 1.0f) > 1e-3 ||
 650             fabs(white_balance.g - 1.0f) > 1e-3 ||
 651             fabs(white_balance.b - 1.0f) > 1e-3) {
 652                 // Convert from (linear) RGB to XYZ.
 653                 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
 654                 Vector3d xyz = rgb_to_xyz_matrix * Vector3d(white_balance.r, white_balance.g, white_balance.b);
 655
 656                 // Convert from XYZ to xyz by normalizing.
 657                 xyz /= (xyz[0] + xyz[1] + xyz[2]);
 658
 659                 // Create a very rudimentary EXIF header to hold our white point.
 660                 string exif;
 661
 662                 // Exif header, followed by some padding.
 663                 exif = "Exif";
 664                 push16(0, &exif);
 665
 666                 // TIFF header first:
 667                 exif += "MM";  // Big endian.
 668
 669                 // Magic number.
 670                 push16(42, &exif);
 671
 672                 // Offset of first IFD (relative to the MM, immediately after the header).
 673                 push32(exif.size() - 6 + 4, &exif);
 674
 675                 // Now the actual IFD.
 676
 677                 // One entry.
 678                 push16(1, &exif);
 679
 680                 // WhitePoint tag ID.
 681                 push16(0x13e, &exif);
 682
 683                 // Rational type.
 684                 push16(5, &exif);
 685
 686                 // Two values (x and y; z is implicit due to normalization).
 687                 push32(2, &exif);
 688
 689                 // Offset (relative to the MM, immediately after the last IFD).
 690                 push32(exif.size() - 6 + 8, &exif);
 691
 692                 // No more IFDs.
 693                 push32(0, &exif);
 694
 695                 // The actual values.
 696                 push32(lrintf(xyz[0] * 10000.0f), &exif);
 697                 push32(10000, &exif);
 698                 push32(lrintf(xyz[1] * 10000.0f), &exif);
 699                 push32(10000, &exif);
 700
 701                 jpeg_write_marker(cinfo, JPEG_APP0 + 1, (const JOCTET *)exif.data(), exif.size());
 702         }
 703
 704         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 705         // (and nothing else).
 706         jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 707 }
 708
 709 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, const RGBTriplet &white_balance, int y_h_samp_factor, int y_v_samp_factor, jpeg_compress_struct *cinfo)
 710 {
 711         VectorDestinationManager dest;
 712         init_jpeg(width, height, white_balance, &dest, cinfo, y_h_samp_factor, y_v_samp_factor);
 713
 714         // Make a dummy black image; there's seemingly no other easy way of
 715         // making libjpeg outputting all of its headers.
 716         assert(y_v_samp_factor <= 2);  // Or we'd need larger JSAMPROW arrays below.
 717         size_t block_height_y = 8 * y_v_samp_factor;
 718         size_t block_height_cbcr = 8;
 719
 720         JSAMPROW yptr[16], cbptr[16], crptr[16];
 721         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 722         memset(tmp_y, 0, 4096);
 723         memset(tmp_cb, 0, 4096);
 724         memset(tmp_cr, 0, 4096);
 725         for (unsigned yy = 0; yy < block_height_y; ++yy) {
 726                 yptr[yy] = tmp_y;
 727         }
 728         for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
 729                 cbptr[yy] = tmp_cb;
 730                 crptr[yy] = tmp_cr;
 731         }
 732         for (unsigned y = 0; y < height; y += block_height_y) {
 733                 jpeg_write_raw_data(cinfo, data, block_height_y);
 734         }
 735         jpeg_finish_compress(cinfo);
 736
 737         // We're only interested in the header, not the data after it.
 738         dest.term_destination();
 739         for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
 740                 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
 741                         unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
 742                         dest.dest.resize(i + len + 2);
 743                         break;
 744                 }
 745         }
 746
 747         return dest.dest;
 748 }
 749
 750 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_parameters(unsigned width, unsigned height, unsigned y_h_samp_factor, unsigned y_v_samp_factor, const RGBTriplet &white_balance)
 751 {
 752         VAKey key{width, height, y_h_samp_factor, y_v_samp_factor, white_balance};
 753         if (va_data_for_parameters.count(key)) {
 754                 return va_data_for_parameters[key];
 755         }
 756
 757         // Use libjpeg to generate a header and set sane defaults for e.g.
 758         // quantization tables. Then do the actual encode with VA-API.
 759         jpeg_compress_struct cinfo;
 760         vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, y_h_samp_factor, y_v_samp_factor, &cinfo);
 761
 762         // Picture parameters.
 763         VAEncPictureParameterBufferJPEG pic_param;
 764         memset(&pic_param, 0, sizeof(pic_param));
 765         pic_param.reconstructed_picture = VA_INVALID_ID;
 766         pic_param.picture_width = cinfo.image_width;
 767         pic_param.picture_height = cinfo.image_height;
 768         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 769                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 770                 pic_param.component_id[component_idx] = comp->component_id;
 771                 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
 772         }
 773         pic_param.num_components = cinfo.num_components;
 774         pic_param.num_scan = 1;
 775         pic_param.sample_bit_depth = 8;
 776         pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
 777         pic_param.pic_flags.bits.huffman = 1;
 778         pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
 779
 780         // Quantization matrices.
 781         VAQMatrixBufferJPEG q;
 782         memset(&q, 0, sizeof(q));
 783
 784         q.load_lum_quantiser_matrix = true;
 785         q.load_chroma_quantiser_matrix = true;
 786         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 787                 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
 788                 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
 789                 if (qtbl == nullptr) continue;
 790
 791                 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
 792                 for (int i = 0; i < 64; ++i) {
 793                         if (qtbl->quantval[i] > 255) {
 794                                 fprintf(stderr, "Baseline JPEG only!\n");
 795                                 abort();
 796                         }
 797                         qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
 798                 }
 799         }
 800
 801         // Huffman tables (arithmetic is not supported).
 802         VAHuffmanTableBufferJPEGBaseline huff;
 803         memset(&huff, 0, sizeof(huff));
 804
 805         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 806                 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 807                 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 808                 if (ac_hufftbl == nullptr) {
 809                         assert(dc_hufftbl == nullptr);
 810                         huff.load_huffman_table[huff_tbl_idx] = 0;
 811                 } else {
 812                         assert(dc_hufftbl != nullptr);
 813                         huff.load_huffman_table[huff_tbl_idx] = 1;
 814
 815                         for (int i = 0; i < 16; ++i) {
 816                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 817                         }
 818                         for (int i = 0; i < 12; ++i) {
 819                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 820                         }
 821                         for (int i = 0; i < 16; ++i) {
 822                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 823                         }
 824                         for (int i = 0; i < 162; ++i) {
 825                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 826                         }
 827                 }
 828         }
 829
 830         // Slice parameters (metadata about the slice).
 831         VAEncSliceParameterBufferJPEG parms;
 832         memset(&parms, 0, sizeof(parms));
 833         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 834                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 835                 parms.components[component_idx].component_selector = comp->component_id;
 836                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 837                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 838                 if (parms.components[component_idx].dc_table_selector > 1 ||
 839                     parms.components[component_idx].ac_table_selector > 1) {
 840                         fprintf(stderr, "Uses too many Huffman tables\n");
 841                         abort();
 842                 }
 843         }
 844         parms.num_components = cinfo.num_components;
 845         parms.restart_interval = cinfo.restart_interval;
 846
 847         jpeg_destroy_compress(&cinfo);
 848
 849         VAData ret;
 850         ret.jpeg_header = move(jpeg_header);
 851         ret.pic_param = pic_param;
 852         ret.q = q;
 853         ret.huff = huff;
 854         ret.parms = parms;
 855         va_data_for_parameters[key] = ret;
 856         return ret;
 857 }
 858
 859 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
 860 {
 861         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
 862         unsigned width = qf.video_format.width;
 863         unsigned height = qf.video_format.height;
 864
 865         VAResources resources;
 866         ReleaseVAResources release;
 867         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 868                 assert(is_uyvy(qf.frame));
 869                 resources = move(userdata->va_resources);
 870                 release = move(userdata->va_resources_release);
 871         } else {
 872                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 873                 if (is_uyvy(qf.frame)) {
 874                         resources = get_va_resources(width, height, VA_FOURCC_UYVY);
 875                 } else {
 876                         assert(is_i420(qf.frame));
 877                         // We'd prefer VA_FOURCC_I420, but it's not supported by Intel's driver.
 878                         resources = get_va_resources(width, height, VA_FOURCC_NV12);
 879                 }
 880                 release = ReleaseVAResources(this, resources);
 881         }
 882
 883         int y_h_samp_factor, y_v_samp_factor;
 884         if (is_uyvy(qf.frame)) {
 885                 // 4:2:2 (sample Y' twice as often vertically as Cb or Cr).
 886                 y_h_samp_factor = 1;
 887                 y_v_samp_factor = 2;
 888         } else {
 889                 // 4:2:0 (sample Y' twice as often as Cb or Cr, in both directions)
 890                 assert(is_i420(qf.frame));
 891                 y_h_samp_factor = 2;
 892                 y_v_samp_factor = 2;
 893         }
 894
 895         VAData va_data = get_va_data_for_parameters(width, height, y_h_samp_factor, y_v_samp_factor, qf.white_balance);
 896         va_data.pic_param.coded_buf = resources.data_buffer;
 897
 898         VABufferID pic_param_buffer;
 899         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
 900         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 901         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 902
 903         VABufferID q_buffer;
 904         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
 905         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 906         VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
 907
 908         VABufferID huff_buffer;
 909         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
 910         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 911         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 912
 913         VABufferID slice_param_buffer;
 914         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
 915         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 916         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 917
 918         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 919                 // The pixel data is already put into the image by the caller.
 920                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 921                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 922         } else {
 923                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 924
 925                 // Upload the pixel data.
 926                 uint8_t *surface_p = nullptr;
 927                 vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
 928
 929                 if (is_uyvy(qf.frame)) {
 930                         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 931                         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 932
 933                         const uint8_t *src = qf.frame->data_copy + field_start;
 934                         uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
 935                         memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
 936                 } else {
 937                         assert(is_i420(qf.frame));
 938                         assert(!qf.frame->interleaved);  // Makes no sense for I420.
 939
 940                         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 941                         const uint8_t *y_src = qf.frame->data + qf.video_format.width * field_start_line;
 942                         const uint8_t *cb_src = y_src + width * height;
 943                         const uint8_t *cr_src = cb_src + (width / 2) * (height / 2);
 944
 945                         uint8_t *y_dst = (unsigned char *)surface_p + resources.image.offsets[0];
 946                         uint8_t *cbcr_dst = (unsigned char *)surface_p + resources.image.offsets[1];
 947
 948                         memcpy_with_pitch(y_dst, y_src, qf.video_format.width, resources.image.pitches[0], qf.video_format.height);
 949                         interleave_with_pitch(cbcr_dst, cb_src, cr_src, qf.video_format.width / 2, resources.image.pitches[1], qf.video_format.height / 2);
 950                 }
 951
 952                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 953                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 954         }
 955
 956         qf.frame->data_copy = nullptr;
 957
 958         // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
 959         // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
 960         // Exactly why is unclear, but it seems to involve L3 cache usage when there
 961         // are many high-res (1080p+) images in play.
 962         va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
 963         CHECK_VASTATUS(va_status, "vaPutImage");
 964
 965         // Finally, stick in the JPEG header.
 966         VAEncPackedHeaderParameterBuffer header_parm;
 967         header_parm.type = VAEncPackedHeaderRawData;
 968         header_parm.bit_length = 8 * va_data.jpeg_header.size();
 969
 970         VABufferID header_parm_buffer;
 971         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
 972         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 973         VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
 974
 975         VABufferID header_data_buffer;
 976         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
 977         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 978         VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
 979
 980         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 981         CHECK_VASTATUS(va_status, "vaBeginPicture");
 982         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 983         CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
 984         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
 985         CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
 986         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 987         CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
 988         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 989         CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
 990         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
 991         CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
 992         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
 993         CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
 994         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 995         CHECK_VASTATUS(va_status, "vaEndPicture");
 996
 997         qf.resources = move(resources);
 998         qf.resource_releaser = move(release);
 999
1000         lock_guard<mutex> lock(mu);
1001         frames_encoding.push(move(qf));
1002         any_frames_encoding.notify_all();
1003 }
1004
1005 void MJPEGEncoder::va_receiver_thread_func()
1006 {
1007         pthread_setname_np(pthread_self(), "MJPEG_Receive");
1008         for (;;) {
1009                 QueuedFrame qf;
1010                 {
1011                         unique_lock<mutex> lock(mu);
1012                         any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
1013                         if (should_quit) return;
1014                         qf = move(frames_encoding.front());
1015                         frames_encoding.pop();
1016                 }
1017
1018                 update_siphon_streams();
1019
1020                 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index));  // Or should_encode_mjpeg_for_card() would have returned false.
1021                 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
1022
1023                 HTTPD::StreamID multicam_id{ HTTPD::MULTICAM_STREAM, 0 };
1024                 HTTPD::StreamID siphon_id{ HTTPD::SIPHON_STREAM, qf.card_index };
1025                 assert(streams.count(multicam_id));
1026                 assert(streams[multicam_id].avctx != nullptr);
1027
1028                 // Write audio before video, since Futatabi expects it.
1029                 if (qf.audio.size() > 0) {
1030                         write_audio_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index + global_flags.card_to_mjpeg_stream_export.size(), qf.audio);
1031                         if (streams.count(siphon_id)) {
1032                                 write_audio_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/1, qf.audio);
1033                         }
1034                 }
1035
1036                 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
1037                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1038
1039                 VACodedBufferSegment *segment;
1040                 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
1041                 CHECK_VASTATUS(va_status, "vaMapBuffer");
1042
1043                 const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
1044                 write_mjpeg_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index, coded_buf, segment->size);
1045                 if (streams.count(siphon_id)) {
1046                         write_mjpeg_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/0, coded_buf, segment->size);
1047                 }
1048
1049                 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
1050                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1051         }
1052 }
1053
1054 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
1055 {
1056         unsigned width = qf.video_format.width;
1057         unsigned height = qf.video_format.height;
1058
1059         VectorDestinationManager dest;
1060         jpeg_compress_struct cinfo;
1061
1062         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
1063
1064         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
1065         if (userdata->pixel_format == PixelFormat_8BitYCbCr) {
1066                 init_jpeg(width, height, qf.white_balance, &dest, &cinfo, /*y_h_samp_factor=*/2, /*y_v_samp_factor=*/1);
1067
1068                 assert(qf.frame->interleaved);
1069                 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
1070
1071                 JSAMPROW yptr[8], cbptr[8], crptr[8];
1072                 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
1073                 for (unsigned y = 0; y < qf.video_format.height; y += 8) {
1074                         const uint8_t *src;
1075                         src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
1076
1077                         memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
1078                         memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
1079                         for (unsigned yy = 0; yy < 8; ++yy) {
1080                                 yptr[yy] = tmp_y + yy * width;
1081                                 cbptr[yy] = tmp_cb + yy * width / 2;
1082                                 crptr[yy] = tmp_cr + yy * width / 2;
1083                         }
1084                         jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
1085                 }
1086         } else {
1087                 assert(userdata->pixel_format == PixelFormat_8BitYCbCrPlanar);
1088
1089                 const movit::YCbCrFormat &ycbcr = userdata->ycbcr_format;
1090                 init_jpeg(width, height, qf.white_balance, &dest, &cinfo, ycbcr.chroma_subsampling_x, ycbcr.chroma_subsampling_y);
1091                 assert(ycbcr.chroma_subsampling_y <= 2);  // Or we'd need larger JSAMPROW arrays below.
1092
1093                 size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
1094                 const uint8_t *y_start = qf.frame->data + qf.video_format.width * field_start_line;
1095                 const uint8_t *cb_start = y_start + width * height;
1096                 const uint8_t *cr_start = cb_start + (width / ycbcr.chroma_subsampling_x) * (height / ycbcr.chroma_subsampling_y);
1097
1098                 size_t block_height_y = 8 * ycbcr.chroma_subsampling_y;
1099                 size_t block_height_cbcr = 8;
1100
1101                 JSAMPROW yptr[16], cbptr[16], crptr[16];
1102                 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
1103                 for (unsigned y = 0; y < qf.video_format.height; y += block_height_y) {
1104                         for (unsigned yy = 0; yy < block_height_y; ++yy) {
1105                                 yptr[yy] = const_cast<JSAMPROW>(y_start) + (y + yy) * width;
1106                         }
1107                         unsigned cbcr_y = y / ycbcr.chroma_subsampling_y;
1108                         for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
1109                                 cbptr[yy] = const_cast<JSAMPROW>(cb_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
1110                                 crptr[yy] = const_cast<JSAMPROW>(cr_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
1111                         }
1112                         jpeg_write_raw_data(&cinfo, data, block_height_y);
1113                 }
1114         }
1115         jpeg_finish_compress(&cinfo);
1116
1117         return dest.dest;
1118 }
1119
1120 void MJPEGEncoder::add_stream(HTTPD::StreamID stream_id)
1121 {
1122         AVFormatContextWithCloser avctx;
1123
1124         // Set up the mux. We don't use the Mux wrapper, because it's geared towards
1125         // a situation with only one video stream (and possibly one audio stream)
1126         // with known width/height, and we don't need the extra functionality it provides.
1127         avctx.reset(avformat_alloc_context());
1128         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
1129
1130         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
1131         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, &ffmpeg_contexts[stream_id], nullptr, nullptr, nullptr);
1132         avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
1133         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
1134
1135         if (stream_id.type == HTTPD::MULTICAM_STREAM) {
1136                 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
1137                         add_video_stream(avctx.get());
1138                 }
1139                 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
1140                         add_audio_stream(avctx.get());
1141                 }
1142         } else {
1143                 assert(stream_id.type == HTTPD::SIPHON_STREAM);
1144                 add_video_stream(avctx.get());
1145                 add_audio_stream(avctx.get());
1146         }
1147         finalize_mux(avctx.get());
1148
1149         Stream s;
1150         s.avctx = move(avctx);
1151         streams[stream_id] = move(s);
1152 }
1153
1154 void MJPEGEncoder::update_siphon_streams()
1155 {
1156         // Bring the list of streams into sync with what the clients need.
1157         for (auto it = streams.begin(); it != streams.end(); ) {
1158                 if (it->first.type != HTTPD::SIPHON_STREAM) {
1159                         ++it;
1160                         continue;
1161                 }
1162                 if (httpd->get_num_connected_siphon_clients(it->first.index) == 0) {
1163                         av_free(it->second.avctx->pb->buffer);
1164                         streams.erase(it++);
1165                 } else {
1166                         ++it;
1167                 }
1168         }
1169         for (unsigned stream_idx = 0; stream_idx < MAX_VIDEO_CARDS; ++stream_idx) {
1170                 HTTPD::StreamID stream_id{ HTTPD::SIPHON_STREAM, stream_idx };
1171                 if (streams.count(stream_id) == 0 && httpd->get_num_connected_siphon_clients(stream_idx) > 0) {
1172                         add_stream(stream_id);
1173                 }
1174         }
1175 }
1176
1177 void MJPEGEncoder::create_ffmpeg_context(HTTPD::StreamID stream_id)
1178 {
1179         ffmpeg_contexts.emplace(stream_id, WritePacket2Context{ this, stream_id });
1180 }