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         global_metrics.add_histogram("x264_crf", {}, metric_x264_crf, &metric_x264_crf_sum, crf_buckets);
  69 }
  70
  71 X264Encoder::~X264Encoder()
  72 {
  73         should_quit = true;
  74         queued_frames_nonempty.notify_all();
  75         encoder_thread.join();
  76         if (dyn.handle) {
  77                 dlclose(dyn.handle);
  78         }
  79 }
  80
  81 void X264Encoder::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: x264 queue full, dropping frame with pts %ld\n", pts);
  95                         ++metric_x264_dropped_frames;
  96                         return;
  97                 }
  98
  99                 qf.data = free_frames.front();
 100                 free_frames.pop();
 101         }
 102
 103         size_t bytes_per_pixel = global_flags.x264_bit_depth > 8 ? 2 : 1;
 104         memcpy(qf.data, data, global_flags.width * global_flags.height * 2 * bytes_per_pixel);
 105
 106         {
 107                 lock_guard<mutex> lock(mu);
 108                 queued_frames.push(qf);
 109                 queued_frames_nonempty.notify_all();
 110                 metric_x264_queued_frames = queued_frames.size();
 111         }
 112 }
 113
 114 void X264Encoder::init_x264()
 115 {
 116         x264_param_t param;
 117         dyn.x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
 118
 119         param.i_width = global_flags.width;
 120         param.i_height = global_flags.height;
 121         param.i_csp = X264_CSP_NV12;
 122         if (global_flags.x264_bit_depth > 8) {
 123                 param.i_csp |= X264_CSP_HIGH_DEPTH;
 124         }
 125         param.b_vfr_input = 1;
 126         param.i_timebase_num = 1;
 127         param.i_timebase_den = TIMEBASE;
 128         param.i_keyint_max = 50; // About one second.
 129         if (global_flags.x264_speedcontrol) {
 130                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 131         }
 132
 133         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 134         param.vui.i_vidformat = 5;  // Unspecified.
 135         param.vui.b_fullrange = 0;
 136         param.vui.i_colorprim = 1;  // BT.709.
 137         param.vui.i_transfer = 2;  // Unspecified (since we use sRGB).
 138         if (global_flags.ycbcr_rec709_coefficients) {
 139                 param.vui.i_colmatrix = 1;  // BT.709.
 140         } else {
 141                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 142         }
 143
 144         if (!isinf(global_flags.x264_crf)) {
 145                 param.rc.i_rc_method = X264_RC_CRF;
 146                 param.rc.f_rf_constant = global_flags.x264_crf;
 147         } else {
 148                 param.rc.i_rc_method = X264_RC_ABR;
 149                 param.rc.i_bitrate = global_flags.x264_bitrate;
 150         }
 151         update_vbv_settings(&param);
 152         if (param.rc.i_vbv_max_bitrate > 0) {
 153                 // If the user wants VBV control to cap the max rate, it is
 154                 // also reasonable to assume that they are fine with the stream
 155                 // constantly being around that rate even for very low-complexity
 156                 // content; the obvious and extreme example being a static
 157                 // black picture.
 158                 //
 159                 // One would think it's fine to have low-complexity content use
 160                 // less bitrate, but it seems to cause problems in practice;
 161                 // e.g. VLC seems to often drop the stream (similar to a buffer
 162                 // underrun) in such cases, but only when streaming from Nageru,
 163                 // not when reading a dump of the same stream from disk.
 164                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 165                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 166                 // but it's a typical case of problems that can arise. Similarly,
 167                 // TCP's congestion control is not always fond of the rate staying
 168                 // low for a while and then rising quickly -- a variation on the same
 169                 // problem.
 170                 //
 171                 // We solve this by simply asking x264 to fill in dummy bits
 172                 // in these cases, so that the bitrate stays reasonable constant.
 173                 // It's a waste of bandwidth, but it makes things go much more
 174                 // smoothly in these cases. (We don't do it if VBV control is off
 175                 // in general, not the least because it makes no sense and x264
 176                 // thus ignores the parameter.)
 177                 param.rc.b_filler = 1;
 178         }
 179
 180         // Occasionally players have problem with extremely low quantizers;
 181         // be on the safe side. Shouldn't affect quality in any meaningful way.
 182         param.rc.i_qp_min = 5;
 183
 184         for (const string &str : global_flags.x264_extra_param) {
 185                 const size_t pos = str.find(',');
 186                 if (pos == string::npos) {
 187                         if (dyn.x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 188                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 189                         }
 190                 } else {
 191                         const string key = str.substr(0, pos);
 192                         const string value = str.substr(pos + 1);
 193                         if (dyn.x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 194                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 195                                         key.c_str(), value.c_str());
 196                         }
 197                 }
 198         }
 199
 200         if (global_flags.x264_bit_depth > 8) {
 201                 dyn.x264_param_apply_profile(&param, "high10");
 202         } else {
 203                 dyn.x264_param_apply_profile(&param, "high");
 204         }
 205
 206         param.b_repeat_headers = !wants_global_headers;
 207
 208         x264 = dyn.x264_encoder_open(&param);
 209         if (x264 == nullptr) {
 210                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 211                 exit(1);
 212         }
 213
 214         if (global_flags.x264_speedcontrol) {
 215                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 216         }
 217
 218         if (wants_global_headers) {
 219                 x264_nal_t *nal;
 220                 int num_nal;
 221
 222                 dyn.x264_encoder_headers(x264, &nal, &num_nal);
 223
 224                 for (int i = 0; i < num_nal; ++i) {
 225                         if (nal[i].i_type == NAL_SEI) {
 226                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 227                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 228                         } else {
 229                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 230                         }
 231                 }
 232         }
 233 }
 234
 235 void X264Encoder::encoder_thread_func()
 236 {
 237         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 238                 perror("nice()");
 239                 // No exit; it's not fatal.
 240         }
 241         pthread_setname_np(pthread_self(), "x264_encode");
 242         init_x264();
 243         x264_init_done = true;
 244
 245         bool frames_left;
 246
 247         do {
 248                 QueuedFrame qf;
 249
 250                 // Wait for a queued frame, then dequeue it.
 251                 {
 252                         unique_lock<mutex> lock(mu);
 253                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 254                         if (!queued_frames.empty()) {
 255                                 qf = queued_frames.front();
 256                                 queued_frames.pop();
 257                         } else {
 258                                 qf.pts = -1;
 259                                 qf.duration = -1;
 260                                 qf.data = nullptr;
 261                         }
 262
 263                         metric_x264_queued_frames = queued_frames.size();
 264                         frames_left = !queued_frames.empty();
 265                 }
 266
 267                 encode_frame(qf);
 268
 269                 {
 270                         lock_guard<mutex> lock(mu);
 271                         free_frames.push(qf.data);
 272                 }
 273
 274                 // We should quit only if the should_quit flag is set _and_ we have nothing
 275                 // in either queue.
 276         } while (!should_quit || frames_left || dyn.x264_encoder_delayed_frames(x264) > 0);
 277
 278         dyn.x264_encoder_close(x264);
 279 }
 280
 281 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 282 {
 283         x264_nal_t *nal = nullptr;
 284         int num_nal = 0;
 285         x264_picture_t pic;
 286         x264_picture_t *input_pic = nullptr;
 287
 288         if (qf.data) {
 289                 dyn.x264_picture_init(&pic);
 290
 291                 pic.i_pts = qf.pts;
 292                 if (global_flags.x264_bit_depth > 8) {
 293                         pic.img.i_csp = X264_CSP_NV12 | X264_CSP_HIGH_DEPTH;
 294                         pic.img.i_plane = 2;
 295                         pic.img.plane[0] = qf.data;
 296                         pic.img.i_stride[0] = global_flags.width * sizeof(uint16_t);
 297                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height * sizeof(uint16_t);
 298                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint32_t);
 299                 } else {
 300                         pic.img.i_csp = X264_CSP_NV12;
 301                         pic.img.i_plane = 2;
 302                         pic.img.plane[0] = qf.data;
 303                         pic.img.i_stride[0] = global_flags.width;
 304                         pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 305                         pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 306                 }
 307                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 308
 309                 input_pic = &pic;
 310
 311                 frames_being_encoded[qf.pts] = qf.received_ts;
 312         }
 313
 314         // See if we have a new bitrate to change to.
 315         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 316         if (new_rate != 0) {
 317                 bitrate_override_func = [new_rate](x264_param_t *param) {
 318                         param->rc.i_bitrate = new_rate;
 319                         update_vbv_settings(param);
 320                 };
 321         }
 322
 323         auto ycbcr_coefficients_override_func = [qf](x264_param_t *param) {
 324                 if (qf.ycbcr_coefficients == YCBCR_REC_709) {
 325                         param->vui.i_colmatrix = 1;  // BT.709.
 326                 } else {
 327                         assert(qf.ycbcr_coefficients == YCBCR_REC_601);
 328                         param->vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 329                 }
 330         };
 331
 332         if (speed_control) {
 333                 speed_control->set_config_override_function([this, ycbcr_coefficients_override_func](x264_param_t *param) {
 334                         if (bitrate_override_func) {
 335                                 bitrate_override_func(param);
 336                         }
 337                         ycbcr_coefficients_override_func(param);
 338                 });
 339         } else {
 340                 x264_param_t param;
 341                 dyn.x264_encoder_parameters(x264, &param);
 342                 if (bitrate_override_func) {
 343                         bitrate_override_func(&param);
 344                 }
 345                 ycbcr_coefficients_override_func(&param);
 346                 dyn.x264_encoder_reconfig(x264, &param);
 347         }
 348
 349         if (speed_control) {
 350                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 351         }
 352         dyn.x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 353         if (speed_control) {
 354                 speed_control->after_frame();
 355         }
 356
 357         if (num_nal == 0) return;
 358
 359         if (IS_X264_TYPE_I(pic.i_type)) {
 360                 ++metric_x264_output_frames_i;
 361         } else if (IS_X264_TYPE_B(pic.i_type)) {
 362                 ++metric_x264_output_frames_b;
 363         } else {
 364                 ++metric_x264_output_frames_p;
 365         }
 366
 367         if (pic.prop.f_crf_avg <= 0.0) {
 368                 ++metric_x264_crf[0];
 369         } else if (pic.prop.f_crf_avg <= crf_buckets - 1) {
 370                 ++metric_x264_crf[int(floor(pic.prop.f_crf_avg))];
 371         } else {
 372                 // Just clamp; this isn't ideal, but at least the total count will be right.
 373                 ++metric_x264_crf[crf_buckets - 1];
 374         }
 375         metric_x264_crf_sum = metric_x264_crf_sum + pic.prop.f_crf_avg;
 376
 377         if (frames_being_encoded.count(pic.i_pts)) {
 378                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 379                 frames_being_encoded.erase(pic.i_pts);
 380
 381                 static int frameno = 0;
 382                 print_latency("Current x264 latency (video inputs → network mux):",
 383                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 384                         &frameno);
 385         } else {
 386                 assert(false);
 387         }
 388
 389         // We really need one AVPacket for the entire frame, it seems,
 390         // so combine it all.
 391         size_t num_bytes = buffered_sei.size();
 392         for (int i = 0; i < num_nal; ++i) {
 393                 num_bytes += nal[i].i_payload;
 394         }
 395
 396         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 397         uint8_t *ptr = data.get();
 398
 399         if (!buffered_sei.empty()) {
 400                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 401                 ptr += buffered_sei.size();
 402                 buffered_sei.clear();
 403         }
 404         for (int i = 0; i < num_nal; ++i) {
 405                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 406                 ptr += nal[i].i_payload;
 407         }
 408
 409         AVPacket pkt;
 410         memset(&pkt, 0, sizeof(pkt));
 411         pkt.buf = nullptr;
 412         pkt.data = data.get();
 413         pkt.size = num_bytes;
 414         pkt.stream_index = 0;
 415         if (pic.b_keyframe) {
 416                 pkt.flags = AV_PKT_FLAG_KEY;
 417         } else {
 418                 pkt.flags = 0;
 419         }
 420         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 421
 422         for (Mux *mux : muxes) {
 423                 mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 424         }
 425 }