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