git.sesse.net Git - nageru/blob - x264_encoder.cpp

   1 #include "x264_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 <x264.h>
  10 #include <cstdint>
  11
  12 #include "defs.h"
  13 #include "flags.h"
  14 #include "metrics.h"
  15 #include "mux.h"
  16 #include "print_latency.h"
  17 #include "timebase.h"
  18 #include "x264_dynamic.h"
  19 #include "x264_speed_control.h"
  20
  21 extern "C" {
  22 #include <libavcodec/avcodec.h>
  23 #include <libavformat/avformat.h>
  24 }
  25
  26 using namespace movit;
  27 using namespace std;
  28 using namespace std::chrono;
  29
  30 namespace {
  31
  32 void update_vbv_settings(x264_param_t *param)
  33 {
  34         if (global_flags.x264_bitrate == -1) {
  35                 return;
  36         }
  37         if (global_flags.x264_vbv_buffer_size < 0) {
  38                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  39         } else {
  40                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  41         }
  42         if (global_flags.x264_vbv_max_bitrate < 0) {
  43                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  44         } else {
  45                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  46         }
  47 }
  48
  49 }  // namespace
  50
  51 X264Encoder::X264Encoder(AVOutputFormat *oformat)
  52         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER),
  53           dyn(load_x264_for_bit_depth(global_flags.x264_bit_depth))
  54 {
  55         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
  56         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * bytes_per_pixel * X264_QUEUE_LENGTH]);
  57         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
  58                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2 * bytes_per_pixel));
  59         }
  60         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
  61
  62         global_metrics.add("x264_queued_frames", &metric_x264_queued_frames, Metrics::TYPE_GAUGE);
  63         global_metrics.add("x264_max_queued_frames", &metric_x264_max_queued_frames, Metrics::TYPE_GAUGE);
  64         global_metrics.add("x264_dropped_frames", &metric_x264_dropped_frames);
  65         global_metrics.add("x264_output_frames", {{ "type", "i" }}, &metric_x264_output_frames_i);
  66         global_metrics.add("x264_output_frames", {{ "type", "p" }}, &metric_x264_output_frames_p);
  67         global_metrics.add("x264_output_frames", {{ "type", "b" }}, &metric_x264_output_frames_b);
  68
  69         metric_x264_crf.init_uniform(50);
  70         global_metrics.add("x264_crf", &metric_x264_crf);
  71 }
  72
  73 X264Encoder::~X264Encoder()
  74 {
  75         should_quit = true;
  76         queued_frames_nonempty.notify_all();
  77         encoder_thread.join();
  78         if (dyn.handle) {
  79                 dlclose(dyn.handle);
  80         }
  81 }
  82
  83 void X264Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts)
  84 {
  85         assert(!should_quit);
  86
  87         QueuedFrame qf;
  88         qf.pts = pts;
  89         qf.duration = duration;
  90         qf.ycbcr_coefficients = ycbcr_coefficients;
  91         qf.received_ts = received_ts;
  92
  93         {
  94                 lock_guard<mutex> lock(mu);
  95                 if (free_frames.empty()) {
  96                         fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %ld\n", pts);
  97                         ++metric_x264_dropped_frames;
  98                         return;
  99                 }
 100
 101                 qf.data = free_frames.front();
 102                 free_frames.pop();
 103         }
 104
 105         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
 106         memcpy(qf.data, data, global_flags.width * global_flags.height * 2 * bytes_per_pixel);
 107
 108         {
 109                 lock_guard<mutex> lock(mu);
 110                 queued_frames.push(qf);
 111                 queued_frames_nonempty.notify_all();
 112                 metric_x264_queued_frames = queued_frames.size();
 113         }
 114 }
 115
 116 void X264Encoder::init_x264()
 117 {
 118         x264_param_t param;
 119         dyn.x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
 120
 121         param.i_width = global_flags.width;
 122         param.i_height = global_flags.height;
 123         param.i_csp = X264_CSP_NV12;
 124         if (global_flags.x264_bit_depth > 8) {
 125                 param.i_csp |= X264_CSP_HIGH_DEPTH;
 126         }
 127         param.b_vfr_input = 1;
 128         param.i_timebase_num = 1;
 129         param.i_timebase_den = TIMEBASE;
 130         param.i_keyint_max = 50; // About one second.
 131         if (global_flags.x264_speedcontrol) {
 132                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 133         }
 134
 135         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 136         param.vui.i_vidformat = 5;  // Unspecified.
 137         param.vui.b_fullrange = 0;
 138         param.vui.i_colorprim = 1;  // BT.709.
 139         param.vui.i_transfer = 2;  // Unspecified (since we use sRGB).
 140         if (global_flags.ycbcr_rec709_coefficients) {
 141                 param.vui.i_colmatrix = 1;  // BT.709.
 142         } else {
 143                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 144         }
 145
 146         if (!isinf(global_flags.x264_crf)) {
 147                 param.rc.i_rc_method = X264_RC_CRF;
 148                 param.rc.f_rf_constant = global_flags.x264_crf;
 149         } else {
 150                 param.rc.i_rc_method = X264_RC_ABR;
 151                 param.rc.i_bitrate = global_flags.x264_bitrate;
 152         }
 153         update_vbv_settings(&param);
 154         if (param.rc.i_vbv_max_bitrate > 0) {
 155                 // If the user wants VBV control to cap the max rate, it is
 156                 // also reasonable to assume that they are fine with the stream
 157                 // constantly being around that rate even for very low-complexity
 158                 // content; the obvious and extreme example being a static
 159                 // black picture.
 160                 //
 161                 // One would think it's fine to have low-complexity content use
 162                 // less bitrate, but it seems to cause problems in practice;
 163                 // e.g. VLC seems to often drop the stream (similar to a buffer
 164                 // underrun) in such cases, but only when streaming from Nageru,
 165                 // not when reading a dump of the same stream from disk.
 166                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 167                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 168                 // but it's a typical case of problems that can arise. Similarly,
 169                 // TCP's congestion control is not always fond of the rate staying
 170                 // low for a while and then rising quickly -- a variation on the same
 171                 // problem.
 172                 //
 173                 // We solve this by simply asking x264 to fill in dummy bits
 174                 // in these cases, so that the bitrate stays reasonable constant.
 175                 // It's a waste of bandwidth, but it makes things go much more
 176                 // smoothly in these cases. (We don't do it if VBV control is off
 177                 // in general, not the least because it makes no sense and x264
 178                 // thus ignores the parameter.)
 179                 param.rc.b_filler = 1;
 180         }
 181
 182         // Occasionally players have problem with extremely low quantizers;
 183         // be on the safe side. Shouldn't affect quality in any meaningful way.
 184         param.rc.i_qp_min = 5;
 185
 186         for (const string &str : global_flags.x264_extra_param) {
 187                 const size_t pos = str.find(',');
 188                 if (pos == string::npos) {
 189                         if (dyn.x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 190                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 191                         }
 192                 } else {
 193                         const string key = str.substr(0, pos);
 194                         const string value = str.substr(pos + 1);
 195                         if (dyn.x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 196                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 197                                         key.c_str(), value.c_str());
 198                         }
 199                 }
 200         }
 201
 202         if (global_flags.x264_bit_depth > 8) {
 203                 dyn.x264_param_apply_profile(&param, "high10");
 204         } else {
 205                 dyn.x264_param_apply_profile(&param, "high");
 206         }
 207
 208         param.b_repeat_headers = !wants_global_headers;
 209
 210         x264 = dyn.x264_encoder_open(&param);
 211         if (x264 == nullptr) {
 212                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 213                 exit(1);
 214         }
 215
 216         if (global_flags.x264_speedcontrol) {
 217                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 218         }
 219
 220         if (wants_global_headers) {
 221                 x264_nal_t *nal;
 222                 int num_nal;
 223
 224                 dyn.x264_encoder_headers(x264, &nal, &num_nal);
 225
 226                 for (int i = 0; i < num_nal; ++i) {
 227                         if (nal[i].i_type == NAL_SEI) {
 228                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 229                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 230                         } else {
 231                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 232                         }
 233                 }
 234         }
 235 }
 236
 237 void X264Encoder::encoder_thread_func()
 238 {
 239         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 240                 perror("nice()");
 241                 // No exit; it's not fatal.
 242         }
 243         pthread_setname_np(pthread_self(), "x264_encode");
 244         init_x264();
 245         x264_init_done = true;
 246
 247         bool frames_left;
 248
 249         do {
 250                 QueuedFrame qf;
 251
 252                 // Wait for a queued frame, then dequeue it.
 253                 {
 254                         unique_lock<mutex> lock(mu);
 255                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 256                         if (!queued_frames.empty()) {
 257                                 qf = queued_frames.front();
 258                                 queued_frames.pop();
 259                         } else {
 260                                 qf.pts = -1;
 261                                 qf.duration = -1;
 262                                 qf.data = nullptr;
 263                         }
 264
 265                         metric_x264_queued_frames = queued_frames.size();
 266                         frames_left = !queued_frames.empty();
 267                 }
 268
 269                 encode_frame(qf);
 270
 271                 {
 272                         lock_guard<mutex> lock(mu);
 273                         free_frames.push(qf.data);
 274                 }
 275
 276                 // We should quit only if the should_quit flag is set _and_ we have nothing
 277                 // in either queue.
 278         } while (!should_quit || frames_left || dyn.x264_encoder_delayed_frames(x264) > 0);
 279
 280         dyn.x264_encoder_close(x264);
 281 }
 282
 283 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 284 {
 285         x264_nal_t *nal = nullptr;
 286         int num_nal = 0;
 287         x264_picture_t pic;
 288         x264_picture_t *input_pic = nullptr;
 289
 290         if (qf.data) {
 291                 dyn.x264_picture_init(&pic);
 292
 293                 pic.i_pts = qf.pts;
 294                 if (global_flags.x264_bit_depth > 8) {
 295                         pic.img.i_csp = X264_CSP_NV12 | X264_CSP_HIGH_DEPTH;
 296                         pic.img.i_plane = 2;
 297                         pic.img.plane[0] = qf.data;
 298                         pic.img.i_stride[0] = global_flags.width * sizeof(uint16_t);
 299                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height * sizeof(uint16_t);
 300                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint32_t);
 301                 } else {
 302                         pic.img.i_csp = X264_CSP_NV12;
 303                         pic.img.i_plane = 2;
 304                         pic.img.plane[0] = qf.data;
 305                         pic.img.i_stride[0] = global_flags.width;
 306                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 307                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 308                 }
 309                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 310
 311                 input_pic = &pic;
 312
 313                 frames_being_encoded[qf.pts] = qf.received_ts;
 314         }
 315
 316         // See if we have a new bitrate to change to.
 317         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 318         if (new_rate != 0) {
 319                 bitrate_override_func = [new_rate](x264_param_t *param) {
 320                         param->rc.i_bitrate = new_rate;
 321                         update_vbv_settings(param);
 322                 };
 323         }
 324
 325         auto ycbcr_coefficients_override_func = [qf](x264_param_t *param) {
 326                 if (qf.ycbcr_coefficients == YCBCR_REC_709) {
 327                         param->vui.i_colmatrix = 1;  // BT.709.
 328                 } else {
 329                         assert(qf.ycbcr_coefficients == YCBCR_REC_601);
 330                         param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 331                 }
 332         };
 333
 334         if (speed_control) {
 335                 speed_control->set_config_override_function([this, ycbcr_coefficients_override_func](x264_param_t *param) {
 336                         if (bitrate_override_func) {
 337                                 bitrate_override_func(param);
 338                         }
 339                         ycbcr_coefficients_override_func(param);
 340                 });
 341         } else {
 342                 x264_param_t param;
 343                 dyn.x264_encoder_parameters(x264, &param);
 344                 if (bitrate_override_func) {
 345                         bitrate_override_func(&param);
 346                 }
 347                 ycbcr_coefficients_override_func(&param);
 348                 dyn.x264_encoder_reconfig(x264, &param);
 349         }
 350
 351         if (speed_control) {
 352                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 353         }
 354         dyn.x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 355         if (speed_control) {
 356                 speed_control->after_frame();
 357         }
 358
 359         if (num_nal == 0) return;
 360
 361         if (IS_X264_TYPE_I(pic.i_type)) {
 362                 ++metric_x264_output_frames_i;
 363         } else if (IS_X264_TYPE_B(pic.i_type)) {
 364                 ++metric_x264_output_frames_b;
 365         } else {
 366                 ++metric_x264_output_frames_p;
 367         }
 368
 369         metric_x264_crf.count_event(pic.prop.f_crf_avg);
 370
 371         if (frames_being_encoded.count(pic.i_pts)) {
 372                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 373                 frames_being_encoded.erase(pic.i_pts);
 374
 375                 static int frameno = 0;
 376                 print_latency("Current x264 latency (video inputs → network mux):",
 377                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 378                         &frameno);
 379         } else {
 380                 assert(false);
 381         }
 382
 383         // We really need one AVPacket for the entire frame, it seems,
 384         // so combine it all.
 385         size_t num_bytes = buffered_sei.size();
 386         for (int i = 0; i < num_nal; ++i) {
 387                 num_bytes += nal[i].i_payload;
 388         }
 389
 390         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 391         uint8_t *ptr = data.get();
 392
 393         if (!buffered_sei.empty()) {
 394                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 395                 ptr += buffered_sei.size();
 396                 buffered_sei.clear();
 397         }
 398         for (int i = 0; i < num_nal; ++i) {
 399                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 400                 ptr += nal[i].i_payload;
 401         }
 402
 403         AVPacket pkt;
 404         memset(&pkt, 0, sizeof(pkt));
 405         pkt.buf = nullptr;
 406         pkt.data = data.get();
 407         pkt.size = num_bytes;
 408         pkt.stream_index = 0;
 409         if (pic.b_keyframe) {
 410                 pkt.flags = AV_PKT_FLAG_KEY;
 411         } else {
 412                 pkt.flags = 0;
 413         }
 414         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 415
 416         for (Mux *mux : muxes) {
 417                 mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 418         }
 419 }