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 = 1;  // TODO: 10-bit support.
  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.
 106         size_t bytes_per_pixel = 1;  // TODO: 10-bit support.
 107         size_t frame_size = global_flags.width * global_flags.height * bytes_per_pixel;
 108         assert(global_flags.width % 2 == 0);
 109         assert(global_flags.height % 2 == 0);
 110         uint8_t *y = qf.data;
 111         uint8_t *cb = y + frame_size;
 112         uint8_t *cr = cb + frame_size / 4;
 113         memcpy(y, data, frame_size);
 114         memcpy_interleaved(cb, cr, data + frame_size, frame_size / 2);
 115
 116         {
 117                 lock_guard<mutex> lock(mu);
 118                 queued_frames.push(qf);
 119                 queued_frames_nonempty.notify_all();
 120                 metric_av1_queued_frames = queued_frames.size();
 121         }
 122 }
 123
 124 void AV1Encoder::init_av1()
 125 {
 126         EbSvtAv1EncConfiguration config;
 127         EbErrorType ret = svt_av1_enc_init_handle(&encoder, nullptr, &config);
 128         if (ret != EB_ErrorNone) {
 129                 fprintf(stderr, "Error initializing SVT-AV1 handle (error %08x)\n", ret);
 130                 exit(EXIT_FAILURE);
 131         }
 132
 133         config.enc_mode = global_flags.av1_preset;
 134         config.intra_period_length = 63;  // Approx. one second, conforms to the (n % 8) - 1 == 0 rule.
 135         config.source_width = global_flags.width;
 136         config.source_height = global_flags.height;
 137         config.frame_rate_numerator = global_flags.av1_fps_num;
 138         config.frame_rate_denominator = global_flags.av1_fps_den;
 139         config.encoder_bit_depth = 8;  // TODO: 10-bit support.
 140         config.rate_control_mode = 2;  // CBR.
 141         config.pred_structure = 1;  // PRED_LOW_DELAY_B (needed for CBR).
 142         config.target_bit_rate = global_flags.av1_bitrate * 1000;
 143
 144         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 145         config.color_primaries = EB_CICP_CP_BT_709;
 146         config.transfer_characteristics = EB_CICP_TC_SRGB;
 147         if (global_flags.ycbcr_rec709_coefficients) {
 148                 config.matrix_coefficients = EB_CICP_MC_BT_709;
 149         } else {
 150                 config.matrix_coefficients = EB_CICP_MC_BT_601;
 151         }
 152         config.color_range = EB_CR_STUDIO_RANGE;
 153 #if SVT_AV1_CHECK_VERSION(1, 0, 0)
 154         config.chroma_sample_position = EB_CSP_VERTICAL;
 155 #endif
 156
 157         const vector<string> &extra_param = global_flags.av1_extra_param;
 158         for (const string &str : extra_param) {
 159                 const size_t pos = str.find(',');
 160                 if (pos == string::npos) {
 161                         if (svt_av1_enc_parse_parameter(&config, str.c_str(), nullptr) != EB_ErrorNone) {
 162                                 fprintf(stderr, "ERROR: SVT-AV1 rejected parameter '%s' with no value\n", str.c_str());
 163                                 exit(EXIT_FAILURE);
 164                         }
 165                 } else {
 166                         const string key = str.substr(0, pos);
 167                         const string value = str.substr(pos + 1);
 168                         if (svt_av1_enc_parse_parameter(&config, key.c_str(), value.c_str()) != EB_ErrorNone) {
 169                                 fprintf(stderr, "ERROR: SVT-AV1 rejected parameter '%s' set to '%s'\n",
 170                                         key.c_str(), value.c_str());
 171                                 exit(EXIT_FAILURE);
 172                         }
 173                 }
 174         }
 175
 176         ret = svt_av1_enc_set_parameter(encoder, &config);
 177         if (ret != EB_ErrorNone) {
 178                 fprintf(stderr, "Error configuring SVT-AV1 (error %08x)\n", ret);
 179                 exit(EXIT_FAILURE);
 180         }
 181
 182         ret = svt_av1_enc_init(encoder);
 183         if (ret != EB_ErrorNone) {
 184                 fprintf(stderr, "Error initializing SVT-AV1 (error %08x)\n", ret);
 185                 exit(EXIT_FAILURE);
 186         }
 187
 188         if (wants_global_headers) {
 189                 EbBufferHeaderType *header = NULL;
 190
 191                 ret = svt_av1_enc_stream_header(encoder, &header);
 192                 if (ret != EB_ErrorNone) {
 193                         fprintf(stderr, "Error building SVT-AV1 header (error %08x)\n", ret);
 194                         exit(EXIT_FAILURE);
 195                 }
 196
 197                 global_headers = string(reinterpret_cast<const char *>(header->p_buffer), header->n_filled_len);
 198
 199                 svt_av1_enc_stream_header_release(header);  // Don't care about errors.
 200           }
 201 }
 202
 203 void AV1Encoder::encoder_thread_func()
 204 {
 205         if (nice(5) == -1) {
 206                 perror("nice()");
 207                 // No exit; it's not fatal.
 208         }
 209         pthread_setname_np(pthread_self(), "AV1_encode");
 210         init_av1();
 211         av1_init_done = true;
 212
 213         bool frames_left;
 214
 215         do {
 216                 QueuedFrame qf;
 217
 218                 // Wait for a queued frame, then dequeue it.
 219                 {
 220                         unique_lock<mutex> lock(mu);
 221                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 222                         if (!queued_frames.empty()) {
 223                                 qf = queued_frames.front();
 224                                 queued_frames.pop();
 225                         } else {
 226                                 qf.pts = -1;
 227                                 qf.duration = -1;
 228                                 qf.data = nullptr;
 229                         }
 230
 231                         metric_av1_queued_frames = queued_frames.size();
 232                         frames_left = !queued_frames.empty();
 233                 }
 234
 235                 encode_frame(qf);
 236
 237                 {
 238                         lock_guard<mutex> lock(mu);
 239                         free_frames.push(qf.data);
 240                 }
 241
 242                 // We should quit only if the should_quit flag is set _and_ we have nothing
 243                 // in our queue.
 244         } while (!should_quit || frames_left);
 245
 246         // Signal end of stream.
 247         EbBufferHeaderType hdr;
 248         hdr.n_alloc_len   = 0;
 249         hdr.n_filled_len  = 0;
 250         hdr.n_tick_count  = 0;
 251         hdr.p_app_private = nullptr;
 252         hdr.pic_type      = EB_AV1_INVALID_PICTURE;
 253         hdr.p_buffer      = nullptr;
 254         hdr.metadata      = nullptr;
 255         hdr.flags         = EB_BUFFERFLAG_EOS;
 256         svt_av1_enc_send_picture(encoder, &hdr);
 257
 258         bool seen_eof = false;
 259         do {
 260                 EbBufferHeaderType *buf;
 261                 EbErrorType ret = svt_av1_enc_get_packet(encoder, &buf, /*pic_send_done=*/true);
 262                 if (ret == EB_NoErrorEmptyQueue) {
 263                         assert(false);
 264                 }
 265                 seen_eof = (buf->flags & EB_BUFFERFLAG_EOS);
 266                 process_packet(buf);
 267         } while (!seen_eof);
 268
 269         svt_av1_enc_deinit(encoder);
 270         svt_av1_enc_deinit_handle(encoder);
 271 }
 272
 273 void AV1Encoder::encode_frame(AV1Encoder::QueuedFrame qf)
 274 {
 275         if (qf.data) {
 276                 EbSvtIOFormat pic;
 277                 pic.luma = qf.data;
 278                 pic.cb = pic.luma + global_flags.width * global_flags.height;
 279                 pic.cr = pic.cb + global_flags.width * global_flags.height / 4;
 280                 pic.y_stride = global_flags.width;
 281                 pic.cb_stride = global_flags.width / 2;
 282                 pic.cr_stride = global_flags.width / 2;
 283                 pic.width = global_flags.width;
 284                 pic.height = global_flags.height;
 285                 pic.origin_x = 0;
 286                 pic.origin_y = 0;
 287                 pic.color_fmt = EB_YUV420;
 288                 pic.bit_depth = EB_EIGHT_BIT;  // TODO: 10-bit.
 289
 290                 EbBufferHeaderType hdr;
 291                 hdr.p_buffer      = reinterpret_cast<uint8_t *>(&pic);
 292                 hdr.n_alloc_len   = global_flags.width * global_flags.height * 3 / 2;  // TODO: 10-bit.
 293                 hdr.n_filled_len  = hdr.n_alloc_len;
 294                 hdr.n_tick_count  = 0;
 295                 hdr.p_app_private = reinterpret_cast<void *>(intptr_t(qf.duration));
 296                 hdr.pic_type      = EB_AV1_INVALID_PICTURE;  // Actually means auto, according to FFmpeg.
 297                 hdr.metadata      = nullptr;
 298                 hdr.flags         = 0;
 299                 hdr.pts           = av_rescale_q(qf.pts, AVRational{ 1, TIMEBASE }, AVRational{ global_flags.av1_fps_den, global_flags.av1_fps_num });
 300                 if (hdr.pts <= last_pts) {
 301                         fprintf(stderr, "WARNING: Receiving frames faster than given --av1-fps value (%d/%d); dropping frame.\n",
 302                                 global_flags.av1_fps_num, global_flags.av1_fps_den);
 303                 } else {
 304                         svt_av1_enc_send_picture(encoder, &hdr);
 305                         frames_being_encoded[hdr.pts] = qf.received_ts;
 306                         last_pts = hdr.pts;
 307                 }
 308         }
 309
 310         for ( ;; ) {
 311                 EbBufferHeaderType *buf;
 312                 EbErrorType ret = svt_av1_enc_get_packet(encoder, &buf, /*pic_send_done=*/false);
 313                 if (ret == EB_NoErrorEmptyQueue) {
 314                         return;
 315                 }
 316                 process_packet(buf);
 317         }
 318 }
 319
 320 void AV1Encoder::process_packet(EbBufferHeaderType *buf)
 321 {
 322         if (buf->n_filled_len == 0) {
 323                 // TODO: Can this ever happen?
 324                 svt_av1_enc_release_out_buffer(&buf);
 325                 return;
 326         }
 327
 328         switch (buf->pic_type) {
 329                 case EB_AV1_KEY_PICTURE:
 330                 case EB_AV1_INTRA_ONLY_PICTURE:
 331                         ++metric_av1_output_frames_i;
 332                         break;
 333                 case EB_AV1_INTER_PICTURE:  // We don't really know whether it's P or B.
 334                         ++metric_av1_output_frames_p;
 335                         break;
 336                 default:
 337                         break;
 338         }
 339         metric_av1_qp.count_event(buf->qp);
 340
 341         if (frames_being_encoded.count(buf->pts)) {
 342                 ReceivedTimestamps received_ts = frames_being_encoded[buf->pts];
 343                 frames_being_encoded.erase(buf->pts);
 344
 345                 static int frameno = 0;
 346                 print_latency("Current AV1 latency (video inputs → network mux):",
 347                                 received_ts, /*b_frame=*/false, &frameno, &av1_latency_histogram);
 348         } else {
 349                 assert(false);
 350         }
 351
 352         AVPacket pkt;
 353         memset(&pkt, 0, sizeof(pkt));
 354         pkt.buf = nullptr;
 355         pkt.data = buf->p_buffer;
 356         pkt.size = buf->n_filled_len;
 357         pkt.stream_index = 0;
 358         if (buf->pic_type == EB_AV1_KEY_PICTURE) {
 359                 pkt.flags = AV_PKT_FLAG_KEY;
 360         } else if (buf->pic_type == EB_AV1_NON_REF_PICTURE) {
 361                 // I have no idea if this does anything in practice,
 362                 // but the libavcodec plugin does it.
 363                 pkt.flags = AV_PKT_FLAG_DISPOSABLE;
 364         } else {
 365                 pkt.flags = 0;
 366         }
 367         pkt.pts = av_rescale_q(buf->pts, AVRational{ global_flags.av1_fps_den, global_flags.av1_fps_num }, AVRational{ 1, TIMEBASE });
 368         pkt.dts = av_rescale_q(buf->dts, AVRational{ global_flags.av1_fps_den, global_flags.av1_fps_num }, AVRational{ 1, TIMEBASE });
 369
 370         for (Mux *mux : muxes) {
 371                 mux->add_packet(pkt, pkt.pts, pkt.dts);
 372         }
 373
 374         svt_av1_enc_release_out_buffer(&buf);
 375 }