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 "shared/va_display.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         VAConfigID config_id_422, config_id_420;
 225         string error;
 226         va_dpy = try_open_va(va_display, { VAProfileJPEGBaseline }, VAEntrypointEncPicture,
 227                 {
 228                         { "4:2:2", VA_RT_FORMAT_YUV422, VA_FOURCC_UYVY, &config_id_422, &uyvy_format },
 229                         // We'd prefer VA_FOURCC_I420, but it's not supported by Intel's driver.
 230                         { "4:2:0", VA_RT_FORMAT_YUV420, VA_FOURCC_NV12, &config_id_420, &nv12_format }
 231                 },
 232                 /*chosen_profile=*/nullptr, &error);
 233         if (va_dpy == nullptr) {
 234                 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
 235         }
 236
 237         va_pool.reset(new VAResourcePool(va_dpy->va_dpy, uyvy_format, nv12_format, config_id_422, config_id_420, /*with_data_buffer=*/true));
 238
 239         encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
 240         if (va_dpy != nullptr) {
 241                 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
 242         }
 243
 244         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
 245         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
 246         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
 247         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
 248         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
 249         global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
 250
 251         running = true;
 252 }
 253
 254 MJPEGEncoder::~MJPEGEncoder()
 255 {
 256         for (auto &id_and_stream : streams) {
 257                 av_free(id_and_stream.second.avctx->pb->buffer);
 258         }
 259
 260         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
 261         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
 262         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
 263         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
 264         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
 265         global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
 266 }
 267
 268 void MJPEGEncoder::stop()
 269 {
 270         if (!running) {
 271                 return;
 272         }
 273         running = false;
 274         should_quit = true;
 275         any_frames_to_be_encoded.notify_all();
 276         any_frames_encoding.notify_all();
 277         encoder_thread.join();
 278         if (va_dpy != nullptr) {
 279                 va_receiver_thread.join();
 280         }
 281 }
 282
 283 namespace {
 284
 285 bool is_uyvy(RefCountedFrame frame)
 286 {
 287         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 288         return userdata->pixel_format == PixelFormat_8BitYCbCr && frame->interleaved;
 289 }
 290
 291 bool is_i420(RefCountedFrame frame)
 292 {
 293         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
 294         return userdata->pixel_format == PixelFormat_8BitYCbCrPlanar &&
 295                 userdata->ycbcr_format.chroma_subsampling_x == 2 &&
 296                 userdata->ycbcr_format.chroma_subsampling_y == 2;
 297 }
 298
 299 }  // namespace
 300
 301 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)
 302 {
 303         if (video_format.width == 0 || video_format.height == 0) {
 304                 ++metric_mjpeg_frames_zero_size_dropped;
 305                 return;
 306         }
 307         if (video_format.interlaced) {
 308                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
 309                 ++metric_mjpeg_frames_interlaced_dropped;
 310                 return;
 311         }
 312         if (!is_uyvy(frame) && !is_i420(frame)) {
 313                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
 314                 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
 315                 return;
 316         }
 317         if (video_format.width > 4096 || video_format.height > 4096) {
 318                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
 319                 ++metric_mjpeg_frames_oversized_dropped;
 320                 return;
 321         }
 322
 323         lock_guard<mutex> lock(mu);
 324         if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
 325                 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
 326                 ++metric_mjpeg_overrun_dropped;
 327                 return;
 328         }
 329         ++metric_mjpeg_overrun_submitted;
 330         frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio), white_balance });
 331         any_frames_to_be_encoded.notify_all();
 332 }
 333
 334 bool MJPEGEncoder::should_encode_mjpeg_for_card(unsigned card_index)
 335 {
 336         // Only bother doing MJPEG encoding if there are any connected clients
 337         // that want the stream.
 338         if (httpd->get_num_connected_multicam_clients() == 0 &&
 339             httpd->get_num_connected_siphon_clients(card_index) == 0) {
 340                 return false;
 341         }
 342
 343         auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
 344         return (it != global_flags.card_to_mjpeg_stream_export.end());
 345 }
 346
 347 void MJPEGEncoder::encoder_thread_func()
 348 {
 349         pthread_setname_np(pthread_self(), "MJPEG_Encode");
 350         posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
 351         posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
 352         posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
 353         posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
 354
 355         for (;;) {
 356                 QueuedFrame qf;
 357                 {
 358                         unique_lock<mutex> lock(mu);
 359                         any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
 360                         if (should_quit) break;
 361                         qf = move(frames_to_be_encoded.front());
 362                         frames_to_be_encoded.pop();
 363                 }
 364
 365                 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index));  // Or should_encode_mjpeg_for_card() would have returned false.
 366                 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
 367
 368                 if (va_dpy != nullptr) {
 369                         // Will call back in the receiver thread.
 370                         encode_jpeg_va(move(qf));
 371                 } else {
 372                         update_siphon_streams();
 373
 374                         HTTPD::StreamID multicam_id{ HTTPD::MULTICAM_STREAM, 0 };
 375                         HTTPD::StreamID siphon_id{ HTTPD::SIPHON_STREAM, qf.card_index };
 376                         assert(streams.count(multicam_id));
 377
 378                         // Write audio before video, since Futatabi expects it.
 379                         if (qf.audio.size() > 0) {
 380                                 write_audio_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index + global_flags.card_to_mjpeg_stream_export.size(), qf.audio);
 381                                 if (streams.count(siphon_id)) {
 382                                         write_audio_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/1, qf.audio);
 383                                 }
 384                         }
 385
 386                         // Encode synchronously, in the same thread.
 387                         vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
 388                         write_mjpeg_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index, jpeg.data(), jpeg.size());
 389                         if (streams.count(siphon_id)) {
 390                                 write_mjpeg_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/0, jpeg.data(), jpeg.size());
 391                         }
 392                 }
 393         }
 394
 395         free(tmp_y);
 396         free(tmp_cbcr);
 397         free(tmp_cb);
 398         free(tmp_cr);
 399 }
 400
 401 void MJPEGEncoder::write_mjpeg_packet(AVFormatContext *avctx, int64_t pts, unsigned stream_index, const uint8_t *jpeg, size_t jpeg_size)
 402 {
 403         AVPacket pkt;
 404         memset(&pkt, 0, sizeof(pkt));
 405         pkt.buf = nullptr;
 406         pkt.data = const_cast<uint8_t *>(jpeg);
 407         pkt.size = jpeg_size;
 408         pkt.stream_index = stream_index;
 409         pkt.flags = AV_PKT_FLAG_KEY;
 410         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
 411         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
 412         pkt.duration = 0;
 413
 414         if (av_write_frame(avctx, &pkt) < 0) {
 415                 fprintf(stderr, "av_write_frame() failed\n");
 416                 abort();
 417         }
 418 }
 419
 420 void MJPEGEncoder::write_audio_packet(AVFormatContext *avctx, int64_t pts, unsigned stream_index, const vector<int32_t> &audio)
 421 {
 422         AVPacket pkt;
 423         memset(&pkt, 0, sizeof(pkt));
 424         pkt.buf = nullptr;
 425         pkt.data = reinterpret_cast<uint8_t *>(const_cast<int32_t *>(&audio[0]));
 426         pkt.size = audio.size() * sizeof(audio[0]);
 427         pkt.stream_index = stream_index;
 428         pkt.flags = AV_PKT_FLAG_KEY;
 429         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
 430         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
 431         size_t num_stereo_samples = audio.size() / 2;
 432         pkt.duration = av_rescale_q(num_stereo_samples, AVRational{ 1, OUTPUT_FREQUENCY }, time_base);
 433
 434         if (av_write_frame(avctx, &pkt) < 0) {
 435                 fprintf(stderr, "av_write_frame() failed\n");
 436                 abort();
 437         }
 438 }
 439
 440 class VABufferDestroyer {
 441 public:
 442         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 443                 : dpy(dpy), buf(buf) {}
 444
 445         ~VABufferDestroyer() {
 446                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 447                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 448         }
 449
 450 private:
 451         VADisplay dpy;
 452         VABufferID buf;
 453 };
 454
 455 namespace {
 456
 457 void push16(uint16_t val, string *str)
 458 {
 459         str->push_back(val >> 8);
 460         str->push_back(val & 0xff);
 461 }
 462
 463 void push32(uint32_t val, string *str)
 464 {
 465         str->push_back(val >> 24);
 466         str->push_back((val >> 16) & 0xff);
 467         str->push_back((val >> 8) & 0xff);
 468         str->push_back(val & 0xff);
 469 }
 470
 471 }  // namespace
 472
 473 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)
 474 {
 475         jpeg_error_mgr jerr;
 476         cinfo->err = jpeg_std_error(&jerr);
 477         jpeg_create_compress(cinfo);
 478
 479         cinfo->dest = (jpeg_destination_mgr *)dest;
 480
 481         cinfo->input_components = 3;
 482         jpeg_set_defaults(cinfo);
 483         jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
 484
 485         cinfo->image_width = width;
 486         cinfo->image_height = height;
 487         cinfo->raw_data_in = true;
 488         jpeg_set_colorspace(cinfo, JCS_YCbCr);
 489         cinfo->comp_info[0].h_samp_factor = y_h_samp_factor;
 490         cinfo->comp_info[0].v_samp_factor = y_v_samp_factor;
 491         cinfo->comp_info[1].h_samp_factor = 1;
 492         cinfo->comp_info[1].v_samp_factor = 1;
 493         cinfo->comp_info[2].h_samp_factor = 1;
 494         cinfo->comp_info[2].v_samp_factor = 1;
 495         cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
 496         jpeg_start_compress(cinfo, true);
 497
 498         if (fabs(white_balance.r - 1.0f) > 1e-3 ||
 499             fabs(white_balance.g - 1.0f) > 1e-3 ||
 500             fabs(white_balance.b - 1.0f) > 1e-3) {
 501                 // Convert from (linear) RGB to XYZ.
 502                 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
 503                 Vector3d xyz = rgb_to_xyz_matrix * Vector3d(white_balance.r, white_balance.g, white_balance.b);
 504
 505                 // Convert from XYZ to xyz by normalizing.
 506                 xyz /= (xyz[0] + xyz[1] + xyz[2]);
 507
 508                 // Create a very rudimentary EXIF header to hold our white point.
 509                 string exif;
 510
 511                 // Exif header, followed by some padding.
 512                 exif = "Exif";
 513                 push16(0, &exif);
 514
 515                 // TIFF header first:
 516                 exif += "MM";  // Big endian.
 517
 518                 // Magic number.
 519                 push16(42, &exif);
 520
 521                 // Offset of first IFD (relative to the MM, immediately after the header).
 522                 push32(exif.size() - 6 + 4, &exif);
 523
 524                 // Now the actual IFD.
 525
 526                 // One entry.
 527                 push16(1, &exif);
 528
 529                 // WhitePoint tag ID.
 530                 push16(0x13e, &exif);
 531
 532                 // Rational type.
 533                 push16(5, &exif);
 534
 535                 // Two values (x and y; z is implicit due to normalization).
 536                 push32(2, &exif);
 537
 538                 // Offset (relative to the MM, immediately after the last IFD).
 539                 push32(exif.size() - 6 + 8, &exif);
 540
 541                 // No more IFDs.
 542                 push32(0, &exif);
 543
 544                 // The actual values.
 545                 push32(lrintf(xyz[0] * 10000.0f), &exif);
 546                 push32(10000, &exif);
 547                 push32(lrintf(xyz[1] * 10000.0f), &exif);
 548                 push32(10000, &exif);
 549
 550                 jpeg_write_marker(cinfo, JPEG_APP0 + 1, (const JOCTET *)exif.data(), exif.size());
 551         }
 552
 553         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
 554         // (and nothing else).
 555         jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 556 }
 557
 558 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)
 559 {
 560         VectorDestinationManager dest;
 561         init_jpeg(width, height, white_balance, &dest, cinfo, y_h_samp_factor, y_v_samp_factor);
 562
 563         // Make a dummy black image; there's seemingly no other easy way of
 564         // making libjpeg outputting all of its headers.
 565         assert(y_v_samp_factor <= 2);  // Or we'd need larger JSAMPROW arrays below.
 566         size_t block_height_y = 8 * y_v_samp_factor;
 567         size_t block_height_cbcr = 8;
 568
 569         JSAMPROW yptr[16], cbptr[16], crptr[16];
 570         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 571         memset(tmp_y, 0, 4096);
 572         memset(tmp_cb, 0, 4096);
 573         memset(tmp_cr, 0, 4096);
 574         for (unsigned yy = 0; yy < block_height_y; ++yy) {
 575                 yptr[yy] = tmp_y;
 576         }
 577         for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
 578                 cbptr[yy] = tmp_cb;
 579                 crptr[yy] = tmp_cr;
 580         }
 581         for (unsigned y = 0; y < height; y += block_height_y) {
 582                 jpeg_write_raw_data(cinfo, data, block_height_y);
 583         }
 584         jpeg_finish_compress(cinfo);
 585
 586         // We're only interested in the header, not the data after it.
 587         dest.term_destination();
 588         for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
 589                 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
 590                         unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
 591                         dest.dest.resize(i + len + 2);
 592                         break;
 593                 }
 594         }
 595
 596         return dest.dest;
 597 }
 598
 599 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)
 600 {
 601         VAKey key{width, height, y_h_samp_factor, y_v_samp_factor, white_balance};
 602         if (va_data_for_parameters.count(key)) {
 603                 return va_data_for_parameters[key];
 604         }
 605
 606         // Use libjpeg to generate a header and set sane defaults for e.g.
 607         // quantization tables. Then do the actual encode with VA-API.
 608         jpeg_compress_struct cinfo;
 609         vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, y_h_samp_factor, y_v_samp_factor, &cinfo);
 610
 611         // Picture parameters.
 612         VAEncPictureParameterBufferJPEG pic_param;
 613         memset(&pic_param, 0, sizeof(pic_param));
 614         pic_param.reconstructed_picture = VA_INVALID_ID;
 615         pic_param.picture_width = cinfo.image_width;
 616         pic_param.picture_height = cinfo.image_height;
 617         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 618                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 619                 pic_param.component_id[component_idx] = comp->component_id;
 620                 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
 621         }
 622         pic_param.num_components = cinfo.num_components;
 623         pic_param.num_scan = 1;
 624         pic_param.sample_bit_depth = 8;
 625         pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
 626         pic_param.pic_flags.bits.huffman = 1;
 627         pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
 628
 629         // Quantization matrices.
 630         VAQMatrixBufferJPEG q;
 631         memset(&q, 0, sizeof(q));
 632
 633         q.load_lum_quantiser_matrix = true;
 634         q.load_chroma_quantiser_matrix = true;
 635         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 636                 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
 637                 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
 638                 if (qtbl == nullptr) continue;
 639
 640                 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
 641                 for (int i = 0; i < 64; ++i) {
 642                         if (qtbl->quantval[i] > 255) {
 643                                 fprintf(stderr, "Baseline JPEG only!\n");
 644                                 abort();
 645                         }
 646                         qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
 647                 }
 648         }
 649
 650         // Huffman tables (arithmetic is not supported).
 651         VAHuffmanTableBufferJPEGBaseline huff;
 652         memset(&huff, 0, sizeof(huff));
 653
 654         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 655                 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 656                 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 657                 if (ac_hufftbl == nullptr) {
 658                         assert(dc_hufftbl == nullptr);
 659                         huff.load_huffman_table[huff_tbl_idx] = 0;
 660                 } else {
 661                         assert(dc_hufftbl != nullptr);
 662                         huff.load_huffman_table[huff_tbl_idx] = 1;
 663
 664                         for (int i = 0; i < 16; ++i) {
 665                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 666                         }
 667                         for (int i = 0; i < 12; ++i) {
 668                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 669                         }
 670                         for (int i = 0; i < 16; ++i) {
 671                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 672                         }
 673                         for (int i = 0; i < 162; ++i) {
 674                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 675                         }
 676                 }
 677         }
 678
 679         // Slice parameters (metadata about the slice).
 680         VAEncSliceParameterBufferJPEG parms;
 681         memset(&parms, 0, sizeof(parms));
 682         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
 683                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
 684                 parms.components[component_idx].component_selector = comp->component_id;
 685                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 686                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 687                 if (parms.components[component_idx].dc_table_selector > 1 ||
 688                     parms.components[component_idx].ac_table_selector > 1) {
 689                         fprintf(stderr, "Uses too many Huffman tables\n");
 690                         abort();
 691                 }
 692         }
 693         parms.num_components = cinfo.num_components;
 694         parms.restart_interval = cinfo.restart_interval;
 695
 696         jpeg_destroy_compress(&cinfo);
 697
 698         VAData ret;
 699         ret.jpeg_header = move(jpeg_header);
 700         ret.pic_param = pic_param;
 701         ret.q = q;
 702         ret.huff = huff;
 703         ret.parms = parms;
 704         va_data_for_parameters[key] = ret;
 705         return ret;
 706 }
 707
 708 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
 709 {
 710         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
 711         unsigned width = qf.video_format.width;
 712         unsigned height = qf.video_format.height;
 713
 714         VAResourcePool::VAResources resources;
 715         ReleaseVAResources release;
 716         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 717                 assert(is_uyvy(qf.frame));
 718                 resources = move(userdata->va_resources);
 719                 release = move(userdata->va_resources_release);
 720         } else {
 721                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 722                 if (is_uyvy(qf.frame)) {
 723                         resources = va_pool->get_va_resources(width, height, VA_FOURCC_UYVY);
 724                 } else {
 725                         assert(is_i420(qf.frame));
 726                         resources = va_pool->get_va_resources(width, height, VA_FOURCC_NV12);
 727                 }
 728                 release = ReleaseVAResources(va_pool.get(), resources);
 729         }
 730
 731         int y_h_samp_factor, y_v_samp_factor;
 732         if (is_uyvy(qf.frame)) {
 733                 // 4:2:2 (sample Y' twice as often horizontally as Cb or Cr, vertical is left alone).
 734                 y_h_samp_factor = 2;
 735                 y_v_samp_factor = 1;
 736         } else {
 737                 // 4:2:0 (sample Y' twice as often as Cb or Cr, in both directions)
 738                 assert(is_i420(qf.frame));
 739                 y_h_samp_factor = 2;
 740                 y_v_samp_factor = 2;
 741         }
 742
 743         VAData va_data = get_va_data_for_parameters(width, height, y_h_samp_factor, y_v_samp_factor, qf.white_balance);
 744         va_data.pic_param.coded_buf = resources.data_buffer;
 745
 746         VABufferID pic_param_buffer;
 747         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
 748         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 749         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 750
 751         VABufferID q_buffer;
 752         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
 753         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 754         VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
 755
 756         VABufferID huff_buffer;
 757         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
 758         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 759         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 760
 761         VABufferID slice_param_buffer;
 762         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
 763         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 764         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 765
 766         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
 767                 // The pixel data is already put into the image by the caller.
 768                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 769                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 770         } else {
 771                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 772
 773                 // Upload the pixel data.
 774                 uint8_t *surface_p = nullptr;
 775                 vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
 776
 777                 if (is_uyvy(qf.frame)) {
 778                         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 779                         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 780
 781                         const uint8_t *src = qf.frame->data_copy + field_start;
 782                         uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
 783                         memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
 784                 } else {
 785                         assert(is_i420(qf.frame));
 786                         assert(!qf.frame->interleaved);  // Makes no sense for I420.
 787
 788                         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 789                         const uint8_t *y_src = qf.frame->data + qf.video_format.width * field_start_line;
 790                         const uint8_t *cb_src = y_src + width * height;
 791                         const uint8_t *cr_src = cb_src + (width / 2) * (height / 2);
 792
 793                         uint8_t *y_dst = (unsigned char *)surface_p + resources.image.offsets[0];
 794                         uint8_t *cbcr_dst = (unsigned char *)surface_p + resources.image.offsets[1];
 795
 796                         memcpy_with_pitch(y_dst, y_src, qf.video_format.width, resources.image.pitches[0], qf.video_format.height);
 797                         interleave_with_pitch(cbcr_dst, cb_src, cr_src, qf.video_format.width / 2, resources.image.pitches[1], qf.video_format.height / 2);
 798                 }
 799
 800                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 801                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 802         }
 803
 804         qf.frame->data_copy = nullptr;
 805
 806         // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
 807         // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
 808         // Exactly why is unclear, but it seems to involve L3 cache usage when there
 809         // are many high-res (1080p+) images in play.
 810         va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
 811         CHECK_VASTATUS(va_status, "vaPutImage");
 812
 813         // Finally, stick in the JPEG header.
 814         VAEncPackedHeaderParameterBuffer header_parm;
 815         header_parm.type = VAEncPackedHeaderRawData;
 816         header_parm.bit_length = 8 * va_data.jpeg_header.size();
 817
 818         VABufferID header_parm_buffer;
 819         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
 820         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 821         VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
 822
 823         VABufferID header_data_buffer;
 824         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
 825         CHECK_VASTATUS(va_status, "vaCreateBuffer");
 826         VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
 827
 828         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 829         CHECK_VASTATUS(va_status, "vaBeginPicture");
 830         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 831         CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
 832         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
 833         CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
 834         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 835         CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
 836         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 837         CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
 838         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
 839         CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
 840         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
 841         CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
 842         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 843         CHECK_VASTATUS(va_status, "vaEndPicture");
 844
 845         qf.resources = move(resources);
 846         qf.resource_releaser = move(release);
 847
 848         lock_guard<mutex> lock(mu);
 849         frames_encoding.push(move(qf));
 850         any_frames_encoding.notify_all();
 851 }
 852
 853 void MJPEGEncoder::va_receiver_thread_func()
 854 {
 855         pthread_setname_np(pthread_self(), "MJPEG_Receive");
 856         for (;;) {
 857                 QueuedFrame qf;
 858                 {
 859                         unique_lock<mutex> lock(mu);
 860                         any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
 861                         if (should_quit) return;
 862                         qf = move(frames_encoding.front());
 863                         frames_encoding.pop();
 864                 }
 865
 866                 update_siphon_streams();
 867
 868                 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index));  // Or should_encode_mjpeg_for_card() would have returned false.
 869                 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
 870
 871                 HTTPD::StreamID multicam_id{ HTTPD::MULTICAM_STREAM, 0 };
 872                 HTTPD::StreamID siphon_id{ HTTPD::SIPHON_STREAM, qf.card_index };
 873                 assert(streams.count(multicam_id));
 874                 assert(streams[multicam_id].avctx != nullptr);
 875
 876                 // Write audio before video, since Futatabi expects it.
 877                 if (qf.audio.size() > 0) {
 878                         write_audio_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index + global_flags.card_to_mjpeg_stream_export.size(), qf.audio);
 879                         if (streams.count(siphon_id)) {
 880                                 write_audio_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/1, qf.audio);
 881                         }
 882                 }
 883
 884                 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
 885                 CHECK_VASTATUS(va_status, "vaSyncSurface");
 886
 887                 VACodedBufferSegment *segment;
 888                 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
 889                 CHECK_VASTATUS(va_status, "vaMapBuffer");
 890
 891                 const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
 892                 write_mjpeg_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index, coded_buf, segment->size);
 893                 if (streams.count(siphon_id)) {
 894                         write_mjpeg_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/0, coded_buf, segment->size);
 895                 }
 896
 897                 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
 898                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
 899         }
 900 }
 901
 902 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
 903 {
 904         unsigned width = qf.video_format.width;
 905         unsigned height = qf.video_format.height;
 906
 907         VectorDestinationManager dest;
 908         jpeg_compress_struct cinfo;
 909
 910         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 911
 912         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
 913         if (userdata->pixel_format == PixelFormat_8BitYCbCr) {
 914                 init_jpeg(width, height, qf.white_balance, &dest, &cinfo, /*y_h_samp_factor=*/2, /*y_v_samp_factor=*/1);
 915
 916                 assert(qf.frame->interleaved);
 917                 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 918
 919                 JSAMPROW yptr[8], cbptr[8], crptr[8];
 920                 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 921                 for (unsigned y = 0; y < qf.video_format.height; y += 8) {
 922                         const uint8_t *src;
 923                         src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
 924
 925                         memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
 926                         memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
 927                         for (unsigned yy = 0; yy < 8; ++yy) {
 928                                 yptr[yy] = tmp_y + yy * width;
 929                                 cbptr[yy] = tmp_cb + yy * width / 2;
 930                                 crptr[yy] = tmp_cr + yy * width / 2;
 931                         }
 932                         jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
 933                 }
 934         } else {
 935                 assert(userdata->pixel_format == PixelFormat_8BitYCbCrPlanar);
 936
 937                 const movit::YCbCrFormat &ycbcr = userdata->ycbcr_format;
 938                 init_jpeg(width, height, qf.white_balance, &dest, &cinfo, ycbcr.chroma_subsampling_x, ycbcr.chroma_subsampling_y);
 939                 assert(ycbcr.chroma_subsampling_y <= 2);  // Or we'd need larger JSAMPROW arrays below.
 940
 941                 size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
 942                 const uint8_t *y_start = qf.frame->data + qf.video_format.width * field_start_line;
 943                 const uint8_t *cb_start = y_start + width * height;
 944                 const uint8_t *cr_start = cb_start + (width / ycbcr.chroma_subsampling_x) * (height / ycbcr.chroma_subsampling_y);
 945
 946                 size_t block_height_y = 8 * ycbcr.chroma_subsampling_y;
 947                 size_t block_height_cbcr = 8;
 948
 949                 JSAMPROW yptr[16], cbptr[16], crptr[16];
 950                 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
 951                 for (unsigned y = 0; y < qf.video_format.height; y += block_height_y) {
 952                         for (unsigned yy = 0; yy < block_height_y; ++yy) {
 953                                 yptr[yy] = const_cast<JSAMPROW>(y_start) + (y + yy) * width;
 954                         }
 955                         unsigned cbcr_y = y / ycbcr.chroma_subsampling_y;
 956                         for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
 957                                 cbptr[yy] = const_cast<JSAMPROW>(cb_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
 958                                 crptr[yy] = const_cast<JSAMPROW>(cr_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
 959                         }
 960                         jpeg_write_raw_data(&cinfo, data, block_height_y);
 961                 }
 962         }
 963         jpeg_finish_compress(&cinfo);
 964
 965         return dest.dest;
 966 }
 967
 968 void MJPEGEncoder::add_stream(HTTPD::StreamID stream_id)
 969 {
 970         AVFormatContextWithCloser avctx;
 971
 972         // Set up the mux. We don't use the Mux wrapper, because it's geared towards
 973         // a situation with only one video stream (and possibly one audio stream)
 974         // with known width/height, and we don't need the extra functionality it provides.
 975         avctx.reset(avformat_alloc_context());
 976         avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 977
 978         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
 979         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, &ffmpeg_contexts[stream_id], nullptr, nullptr, nullptr);
 980         avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
 981         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 982
 983         if (stream_id.type == HTTPD::MULTICAM_STREAM) {
 984                 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
 985                         add_video_stream(avctx.get());
 986                 }
 987                 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
 988                         add_audio_stream(avctx.get());
 989                 }
 990         } else {
 991                 assert(stream_id.type == HTTPD::SIPHON_STREAM);
 992                 add_video_stream(avctx.get());
 993                 add_audio_stream(avctx.get());
 994         }
 995         finalize_mux(avctx.get());
 996
 997         Stream s;
 998         s.avctx = move(avctx);
 999         streams[stream_id] = move(s);
1000 }
1001
1002 void MJPEGEncoder::update_siphon_streams()
1003 {
1004         // Bring the list of streams into sync with what the clients need.
1005         for (auto it = streams.begin(); it != streams.end(); ) {
1006                 if (it->first.type != HTTPD::SIPHON_STREAM) {
1007                         ++it;
1008                         continue;
1009                 }
1010                 if (httpd->get_num_connected_siphon_clients(it->first.index) == 0) {
1011                         av_free(it->second.avctx->pb->buffer);
1012                         streams.erase(it++);
1013                 } else {
1014                         ++it;
1015                 }
1016         }
1017         for (unsigned stream_idx = 0; stream_idx < MAX_VIDEO_CARDS; ++stream_idx) {
1018                 HTTPD::StreamID stream_id{ HTTPD::SIPHON_STREAM, stream_idx };
1019                 if (streams.count(stream_id) == 0 && httpd->get_num_connected_siphon_clients(stream_idx) > 0) {
1020                         add_stream(stream_id);
1021                 }
1022         }
1023 }
1024
1025 void MJPEGEncoder::create_ffmpeg_context(HTTPD::StreamID stream_id)
1026 {
1027         ffmpeg_contexts.emplace(stream_id, WritePacket2Context{ this, stream_id });
1028 }