git.sesse.net Git - nageru/blob - nageru/av1_encoder.cpp

   1 #include "av1_encoder.h"
   2
   3 #include <assert.h>
   4 #include <dlfcn.h>
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <unistd.h>
   9 #include <atomic>
  10 #include <cstdint>
  11 #include <functional>
  12 #include <mutex>
  13
  14 #include <EbSvtAv1.h>
  15 #include <EbSvtAv1Enc.h>
  16
  17 #include "defs.h"
  18 #include "flags.h"
  19 #include "shared/metrics.h"
  20 #include "shared/mux.h"
  21 #include "print_latency.h"
  22 #include "shared/timebase.h"
  23 #include "shared/memcpy_interleaved.h"
  24
  25 extern "C" {
  26 #include <libavcodec/avcodec.h>
  27 #include <libavformat/avformat.h>
  28 }
  29
  30 using namespace movit;
  31 using namespace std;
  32 using namespace std::chrono;
  33 using namespace std::placeholders;
  34
  35 namespace {
  36
  37 // AV1Encoder can be restarted if --record-av1-video is set, so make these
  38 // metrics global.
  39 atomic<int64_t> metric_av1_queued_frames{0};
  40 atomic<int64_t> metric_av1_max_queued_frames{AV1_QUEUE_LENGTH};
  41 atomic<int64_t> metric_av1_dropped_frames{0};
  42 atomic<int64_t> metric_av1_output_frames_i{0};
  43 atomic<int64_t> metric_av1_output_frames_p{0};
  44 Histogram metric_av1_qp;
  45 LatencyHistogram av1_latency_histogram;
  46
  47 once_flag av1_metrics_inited;
  48
  49 }  // namespace
  50
  51 AV1Encoder::AV1Encoder(const AVOutputFormat *oformat)
  52         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER)
  53 {
  54                 call_once(av1_metrics_inited, []{
  55                         global_metrics.add("av1_queued_frames",  {}, &metric_av1_queued_frames, Metrics::TYPE_GAUGE);
  56                         global_metrics.add("av1_max_queued_frames", {},  &metric_av1_max_queued_frames, Metrics::TYPE_GAUGE);
  57                         global_metrics.add("av1_dropped_frames", {},  &metric_av1_dropped_frames);
  58                         global_metrics.add("av1_output_frames", {{ "type", "i" }}, &metric_av1_output_frames_i);
  59                         global_metrics.add("av1_output_frames", {{ "type", "p" }}, &metric_av1_output_frames_p);
  60
  61                         metric_av1_qp.init_uniform(50);
  62                         global_metrics.add("av1_qp", {}, &metric_av1_qp);
  63                         av1_latency_histogram.init("av1");
  64                 });
  65
  66         const size_t bytes_per_pixel = global_flags.bit_depth > 8 ? 2 : 1;
  67         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * bytes_per_pixel * AV1_QUEUE_LENGTH]);
  68         for (unsigned i = 0; i < AV1_QUEUE_LENGTH; ++i) {
  69                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2 * bytes_per_pixel));
  70         }
  71         encoder_thread = thread(&AV1Encoder::encoder_thread_func, this);
  72 }
  73
  74 AV1Encoder::~AV1Encoder()
  75 {
  76         should_quit = true;
  77         queued_frames_nonempty.notify_all();
  78         encoder_thread.join();
  79 }
  80
  81 void AV1Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts)
  82 {
  83         assert(!should_quit);
  84
  85         QueuedFrame qf;
  86         qf.pts = pts;
  87         qf.duration = duration;
  88         qf.ycbcr_coefficients = ycbcr_coefficients;
  89         qf.received_ts = received_ts;
  90
  91         {
  92                 lock_guard<mutex> lock(mu);
  93                 if (free_frames.empty()) {
  94                         fprintf(stderr, "WARNING: AV1 queue full, dropping frame with pts %" PRId64 "\n", pts);
  95                         ++metric_av1_dropped_frames;
  96                         return;
  97                 }
  98
  99                 qf.data = free_frames.front();
 100                 free_frames.pop();
 101         }
 102
 103         // Since we're copying anyway, we can unpack from NV12 to fully planar on the fly.
 104         // SVT-AV1 makes its own copy, though, and it would have been nice to avoid the
 105         // double-copy (and also perhaps let the GPU do the 10-bit compression SVT-AV1
 106         // wants, instead of doing it on the CPU).
 107         const size_t bytes_per_pixel = global_flags.bit_depth > 8 ? 2 : 1;
 108         size_t frame_size = global_flags.width * global_flags.height * bytes_per_pixel;
 109         assert(global_flags.width % 2 == 0);
 110         assert(global_flags.height % 2 == 0);
 111         uint8_t *y = qf.data;
 112         uint8_t *cb = y + frame_size;
 113         uint8_t *cr = cb + frame_size / 4;
 114         memcpy(y, data, frame_size);
 115         if (global_flags.bit_depth == 8) {
 116                 memcpy_interleaved(cb, cr, data + frame_size, frame_size / 2);
 117         } else {
 118                 const uint16_t *src = reinterpret_cast<const uint16_t *>(data + frame_size);
 119                 uint16_t *cb16 = reinterpret_cast<uint16_t *>(cb);
 120                 uint16_t *cr16 = reinterpret_cast<uint16_t *>(cr);
 121                 memcpy_interleaved_word(cb16, cr16, src, frame_size / 4);
 122         }
 123
 124         {
 125                 lock_guard<mutex> lock(mu);
 126                 queued_frames.push(qf);
 127                 queued_frames_nonempty.notify_all();
 128                 metric_av1_queued_frames = queued_frames.size();
 129         }
 130 }
 131
 132 void AV1Encoder::init_av1()
 133 {
 134         EbSvtAv1EncConfiguration config;
 135         EbErrorType ret = svt_av1_enc_init_handle(&encoder, nullptr, &config);
 136         if (ret != EB_ErrorNone) {
 137                 fprintf(stderr, "Error initializing SVT-AV1 handle (error %08x)\n", ret);
 138                 exit(EXIT_FAILURE);
 139         }
 140
 141         config.enc_mode = global_flags.av1_preset;
 142         config.intra_period_length = 63;  // Approx. one second, conforms to the (n % 8) - 1 == 0 rule.
 143         config.source_width = global_flags.width;
 144         config.source_height = global_flags.height;
 145         config.frame_rate_numerator = global_flags.av1_fps_num;
 146         config.frame_rate_denominator = global_flags.av1_fps_den;
 147         config.encoder_bit_depth = global_flags.bit_depth;
 148         config.rate_control_mode = 2;  // CBR.
 149         config.pred_structure = 1;  // PRED_LOW_DELAY_B (needed for CBR).
 150         config.target_bit_rate = global_flags.av1_bitrate * 1000;
 151
 152         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 153         config.color_primaries = EB_CICP_CP_BT_709;
 154         config.transfer_characteristics = EB_CICP_TC_SRGB;
 155         if (global_flags.ycbcr_rec709_coefficients) {
 156                 config.matrix_coefficients = EB_CICP_MC_BT_709;
 157         } else {
 158                 config.matrix_coefficients = EB_CICP_MC_BT_601;
 159         }
 160         config.color_range = EB_CR_STUDIO_RANGE;
 161 #if SVT_AV1_CHECK_VERSION(1, 0, 0)
 162         config.chroma_sample_position = EB_CSP_VERTICAL;
 163 #endif
 164
 165         const vector<string> &extra_param = global_flags.av1_extra_param;
 166         for (const string &str : extra_param) {
 167                 const size_t pos = str.find(',');
 168                 if (pos == string::npos) {
 169                         if (svt_av1_enc_parse_parameter(&config, str.c_str(), nullptr) != EB_ErrorNone) {
 170                                 fprintf(stderr, "ERROR: SVT-AV1 rejected parameter '%s' with no value\n", str.c_str());
 171                                 exit(EXIT_FAILURE);
 172                         }
 173                 } else {
 174                         const string key = str.substr(0, pos);
 175                         const string value = str.substr(pos + 1);
 176                         if (svt_av1_enc_parse_parameter(&config, key.c_str(), value.c_str()) != EB_ErrorNone) {
 177                                 fprintf(stderr, "ERROR: SVT-AV1 rejected parameter '%s' set to '%s'\n",
 178                                         key.c_str(), value.c_str());
 179                                 exit(EXIT_FAILURE);
 180                         }
 181                 }
 182         }
 183
 184         ret = svt_av1_enc_set_parameter(encoder, &config);
 185         if (ret != EB_ErrorNone) {
 186                 fprintf(stderr, "Error configuring SVT-AV1 (error %08x)\n", ret);
 187                 exit(EXIT_FAILURE);
 188         }
 189
 190         ret = svt_av1_enc_init(encoder);
 191         if (ret != EB_ErrorNone) {
 192                 fprintf(stderr, "Error initializing SVT-AV1 (error %08x)\n", ret);
 193                 exit(EXIT_FAILURE);
 194         }
 195
 196         if (wants_global_headers) {
 197                 EbBufferHeaderType *header = NULL;
 198
 199                 ret = svt_av1_enc_stream_header(encoder, &header);
 200                 if (ret != EB_ErrorNone) {
 201                         fprintf(stderr, "Error building SVT-AV1 header (error %08x)\n", ret);
 202                         exit(EXIT_FAILURE);
 203                 }
 204
 205                 global_headers = string(reinterpret_cast<const char *>(header->p_buffer), header->n_filled_len);
 206
 207                 svt_av1_enc_stream_header_release(header);  // Don't care about errors.
 208           }
 209 }
 210
 211 void AV1Encoder::encoder_thread_func()
 212 {
 213         if (nice(5) == -1) {
 214                 perror("nice()");
 215                 // No exit; it's not fatal.
 216         }
 217         pthread_setname_np(pthread_self(), "AV1_encode");
 218         init_av1();
 219         av1_init_done = true;
 220
 221         bool frames_left;
 222
 223         do {
 224                 QueuedFrame qf;
 225
 226                 // Wait for a queued frame, then dequeue it.
 227                 {
 228                         unique_lock<mutex> lock(mu);
 229                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 230                         if (!queued_frames.empty()) {
 231                                 qf = queued_frames.front();
 232                                 queued_frames.pop();
 233                         } else {
 234                                 qf.pts = -1;
 235                                 qf.duration = -1;
 236                                 qf.data = nullptr;
 237                         }
 238
 239                         metric_av1_queued_frames = queued_frames.size();
 240                         frames_left = !queued_frames.empty();
 241                 }
 242
 243                 encode_frame(qf);
 244
 245                 {
 246                         lock_guard<mutex> lock(mu);
 247                         free_frames.push(qf.data);
 248                 }
 249
 250                 // We should quit only if the should_quit flag is set _and_ we have nothing
 251                 // in our queue.
 252         } while (!should_quit || frames_left);
 253
 254         // Signal end of stream.
 255         EbBufferHeaderType hdr;
 256         hdr.n_alloc_len   = 0;
 257         hdr.n_filled_len  = 0;
 258         hdr.n_tick_count  = 0;
 259         hdr.p_app_private = nullptr;
 260         hdr.pic_type      = EB_AV1_INVALID_PICTURE;
 261         hdr.p_buffer      = nullptr;
 262         hdr.metadata      = nullptr;
 263         hdr.flags         = EB_BUFFERFLAG_EOS;
 264         svt_av1_enc_send_picture(encoder, &hdr);
 265
 266         bool seen_eof = false;
 267         do {
 268                 EbBufferHeaderType *buf;
 269                 EbErrorType ret = svt_av1_enc_get_packet(encoder, &buf, /*pic_send_done=*/true);
 270                 if (ret == EB_NoErrorEmptyQueue) {
 271                         assert(false);
 272                 }
 273                 seen_eof = (buf->flags & EB_BUFFERFLAG_EOS);
 274                 process_packet(buf);
 275         } while (!seen_eof);
 276
 277         svt_av1_enc_deinit(encoder);
 278         svt_av1_enc_deinit_handle(encoder);
 279 }
 280
 281 void AV1Encoder::encode_frame(AV1Encoder::QueuedFrame qf)
 282 {
 283         if (qf.data) {
 284                 const size_t bytes_per_pixel = global_flags.bit_depth > 8 ? 2 : 1;
 285
 286                 EbSvtIOFormat pic;
 287                 pic.luma = qf.data;
 288                 pic.cb = pic.luma + global_flags.width * global_flags.height * bytes_per_pixel;
 289                 pic.cr = pic.cb + (global_flags.width * global_flags.height / 4) * bytes_per_pixel;
 290                 pic.y_stride = global_flags.width;  // In pixels, so no bytes_per_pixel.
 291                 pic.cb_stride = global_flags.width / 2;  // Likewise.
 292                 pic.cr_stride = global_flags.width / 2;  // Likewise.
 293                 pic.width = global_flags.width;
 294                 pic.height = global_flags.height;
 295                 pic.origin_x = 0;
 296                 pic.origin_y = 0;
 297                 pic.color_fmt = EB_YUV420;
 298                 pic.bit_depth = global_flags.bit_depth > 8 ? EB_TEN_BIT : EB_EIGHT_BIT;
 299
 300                 EbBufferHeaderType hdr;
 301                 hdr.p_buffer      = reinterpret_cast<uint8_t *>(&pic);
 302                 hdr.n_alloc_len   = (global_flags.width * global_flags.height * 3 / 2) * bytes_per_pixel;
 303                 hdr.n_filled_len  = hdr.n_alloc_len;
 304                 hdr.n_tick_count  = 0;
 305                 hdr.p_app_private = reinterpret_cast<void *>(intptr_t(qf.duration));
 306                 hdr.pic_type      = EB_AV1_INVALID_PICTURE;  // Actually means auto, according to FFmpeg.
 307                 hdr.metadata      = nullptr;
 308                 hdr.flags         = 0;
 309                 hdr.pts           = av_rescale_q(qf.pts, AVRational{ 1, TIMEBASE }, AVRational{ global_flags.av1_fps_den, global_flags.av1_fps_num });
 310                 if (hdr.pts <= last_pts) {
 311                         fprintf(stderr, "WARNING: Receiving frames faster than given --av1-fps value (%d/%d); dropping frame.\n",
 312                                 global_flags.av1_fps_num, global_flags.av1_fps_den);
 313                 } else {
 314                         svt_av1_enc_send_picture(encoder, &hdr);
 315                         frames_being_encoded[hdr.pts] = qf.received_ts;
 316                         last_pts = hdr.pts;
 317                 }
 318         }
 319
 320         for ( ;; ) {
 321                 EbBufferHeaderType *buf;
 322                 EbErrorType ret = svt_av1_enc_get_packet(encoder, &buf, /*pic_send_done=*/false);
 323                 if (ret == EB_NoErrorEmptyQueue) {
 324                         return;
 325                 }
 326                 process_packet(buf);
 327         }
 328 }
 329
 330 void AV1Encoder::process_packet(EbBufferHeaderType *buf)
 331 {
 332         if (buf->n_filled_len == 0) {
 333                 // TODO: Can this ever happen?
 334                 svt_av1_enc_release_out_buffer(&buf);
 335                 return;
 336         }
 337
 338         switch (buf->pic_type) {
 339                 case EB_AV1_KEY_PICTURE:
 340                 case EB_AV1_INTRA_ONLY_PICTURE:
 341                         ++metric_av1_output_frames_i;
 342                         break;
 343                 case EB_AV1_INTER_PICTURE:  // We don't really know whether it's P or B.
 344                         ++metric_av1_output_frames_p;
 345                         break;
 346                 default:
 347                         break;
 348         }
 349         metric_av1_qp.count_event(buf->qp);
 350
 351         if (frames_being_encoded.count(buf->pts)) {
 352                 ReceivedTimestamps received_ts = frames_being_encoded[buf->pts];
 353                 frames_being_encoded.erase(buf->pts);
 354
 355                 static int frameno = 0;
 356                 print_latency("Current AV1 latency (video inputs → network mux):",
 357                                 received_ts, /*b_frame=*/false, &frameno, &av1_latency_histogram);
 358         } else {
 359                 assert(false);
 360         }
 361
 362         AVPacket pkt;
 363         memset(&pkt, 0, sizeof(pkt));
 364         pkt.buf = nullptr;
 365         pkt.data = buf->p_buffer;
 366         pkt.size = buf->n_filled_len;
 367         pkt.stream_index = 0;
 368         if (buf->pic_type == EB_AV1_KEY_PICTURE) {
 369                 pkt.flags = AV_PKT_FLAG_KEY;
 370         } else if (buf->pic_type == EB_AV1_NON_REF_PICTURE) {
 371                 // I have no idea if this does anything in practice,
 372                 // but the libavcodec plugin does it.
 373                 pkt.flags = AV_PKT_FLAG_DISPOSABLE;
 374         } else {
 375                 pkt.flags = 0;
 376         }
 377         pkt.pts = av_rescale_q(buf->pts, AVRational{ global_flags.av1_fps_den, global_flags.av1_fps_num }, AVRational{ 1, TIMEBASE });
 378         pkt.dts = av_rescale_q(buf->dts, AVRational{ global_flags.av1_fps_den, global_flags.av1_fps_num }, AVRational{ 1, TIMEBASE });
 379
 380         for (Mux *mux : muxes) {
 381                 mux->add_packet(pkt, pkt.pts, pkt.dts);
 382         }
 383
 384         svt_av1_enc_release_out_buffer(&buf);
 385 }