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

   1 #include "x264_encoder.h"
   2
   3 #include <assert.h>
   4 #include <stdio.h>
   5 #include <stdlib.h>
   6 #include <string.h>
   7 #include <unistd.h>
   8 #include <x264.h>
   9 #include <cstdint>
  10
  11 #include "defs.h"
  12 #include "flags.h"
  13 #include "mux.h"
  14 #include "print_latency.h"
  15 #include "timebase.h"
  16 #include "x264_speed_control.h"
  17
  18 extern "C" {
  19 #include <libavcodec/avcodec.h>
  20 #include <libavformat/avformat.h>
  21 }
  22
  23 using namespace std;
  24 using namespace std::chrono;
  25
  26 namespace {
  27
  28 void update_vbv_settings(x264_param_t *param)
  29 {
  30         if (global_flags.x264_vbv_buffer_size < 0) {
  31                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  32         } else {
  33                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  34         }
  35         if (global_flags.x264_vbv_max_bitrate < 0) {
  36                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  37         } else {
  38                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  39         }
  40 }
  41
  42 }  // namespace
  43
  44 X264Encoder::X264Encoder(AVOutputFormat *oformat)
  45         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER)
  46 {
  47         frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * X264_QUEUE_LENGTH]);
  48         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
  49                 free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2));
  50         }
  51         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
  52 }
  53
  54 X264Encoder::~X264Encoder()
  55 {
  56         should_quit = true;
  57         queued_frames_nonempty.notify_all();
  58         encoder_thread.join();
  59 }
  60
  61 void X264Encoder::add_frame(int64_t pts, int64_t duration, const uint8_t *data, const ReceivedTimestamps &received_ts)
  62 {
  63         QueuedFrame qf;
  64         qf.pts = pts;
  65         qf.duration = duration;
  66         qf.received_ts = received_ts;
  67
  68         {
  69                 lock_guard<mutex> lock(mu);
  70                 if (free_frames.empty()) {
  71                         fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %ld\n", pts);
  72                         return;
  73                 }
  74
  75                 qf.data = free_frames.front();
  76                 free_frames.pop();
  77         }
  78
  79         memcpy(qf.data, data, global_flags.width * global_flags.height * 2);
  80
  81         {
  82                 lock_guard<mutex> lock(mu);
  83                 queued_frames.push(qf);
  84                 queued_frames_nonempty.notify_all();
  85         }
  86 }
  87
  88 void X264Encoder::init_x264()
  89 {
  90         x264_param_t param;
  91         x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
  92
  93         param.i_width = global_flags.width;
  94         param.i_height = global_flags.height;
  95         param.i_csp = X264_CSP_NV12;
  96         param.b_vfr_input = 1;
  97         param.i_timebase_num = 1;
  98         param.i_timebase_den = TIMEBASE;
  99         param.i_keyint_max = 50; // About one second.
 100         if (global_flags.x264_speedcontrol) {
 101                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
 102         }
 103
 104         // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()).
 105         param.vui.i_vidformat = 5;  // Unspecified.
 106         param.vui.b_fullrange = 0;
 107         param.vui.i_colorprim = 1;  // BT.709.
 108         param.vui.i_transfer = 2;  // Unspecified (since we use sRGB).
 109         if (global_flags.ycbcr_rec709_coefficients) {
 110                 param.vui.i_colmatrix = 1;  // BT.709.
 111         } else {
 112                 param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 113         }
 114
 115         param.rc.i_rc_method = X264_RC_ABR;
 116         param.rc.i_bitrate = global_flags.x264_bitrate;
 117         update_vbv_settings(&param);
 118         if (param.rc.i_vbv_max_bitrate > 0) {
 119                 // If the user wants VBV control to cap the max rate, it is
 120                 // also reasonable to assume that they are fine with the stream
 121                 // constantly being around that rate even for very low-complexity
 122                 // content; the obvious and extreme example being a static
 123                 // black picture.
 124                 //
 125                 // One would think it's fine to have low-complexity content use
 126                 // less bitrate, but it seems to cause problems in practice;
 127                 // e.g. VLC seems to often drop the stream (similar to a buffer
 128                 // underrun) in such cases, but only when streaming from Nageru,
 129                 // not when reading a dump of the same stream from disk.
 130                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 131                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 132                 // but it's a typical case of problems that can arise. Similarly,
 133                 // TCP's congestion control is not always fond of the rate staying
 134                 // low for a while and then rising quickly -- a variation on the same
 135                 // problem.
 136                 //
 137                 // We solve this by simply asking x264 to fill in dummy bits
 138                 // in these cases, so that the bitrate stays reasonable constant.
 139                 // It's a waste of bandwidth, but it makes things go much more
 140                 // smoothly in these cases. (We don't do it if VBV control is off
 141                 // in general, not the least because it makes no sense and x264
 142                 // thus ignores the parameter.)
 143                 param.rc.b_filler = 1;
 144         }
 145
 146         // Occasionally players have problem with extremely low quantizers;
 147         // be on the safe side. Shouldn't affect quality in any meaningful way.
 148         param.rc.i_qp_min = 5;
 149
 150         for (const string &str : global_flags.x264_extra_param) {
 151                 const size_t pos = str.find(',');
 152                 if (pos == string::npos) {
 153                         if (x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 154                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 155                         }
 156                 } else {
 157                         const string key = str.substr(0, pos);
 158                         const string value = str.substr(pos + 1);
 159                         if (x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 160                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 161                                         key.c_str(), value.c_str());
 162                         }
 163                 }
 164         }
 165
 166         x264_param_apply_profile(&param, "high");
 167
 168         param.b_repeat_headers = !wants_global_headers;
 169
 170         x264 = x264_encoder_open(&param);
 171         if (x264 == nullptr) {
 172                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 173                 exit(1);
 174         }
 175
 176         if (global_flags.x264_speedcontrol) {
 177                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 178         }
 179
 180         if (wants_global_headers) {
 181                 x264_nal_t *nal;
 182                 int num_nal;
 183
 184                 x264_encoder_headers(x264, &nal, &num_nal);
 185
 186                 for (int i = 0; i < num_nal; ++i) {
 187                         if (nal[i].i_type == NAL_SEI) {
 188                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 189                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 190                         } else {
 191                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 192                         }
 193                 }
 194         }
 195 }
 196
 197 void X264Encoder::encoder_thread_func()
 198 {
 199         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 200                 perror("nice()");
 201                 // No exit; it's not fatal.
 202         }
 203         init_x264();
 204         x264_init_done = true;
 205
 206         bool frames_left;
 207
 208         do {
 209                 QueuedFrame qf;
 210
 211                 // Wait for a queued frame, then dequeue it.
 212                 {
 213                         unique_lock<mutex> lock(mu);
 214                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 215                         if (!queued_frames.empty()) {
 216                                 qf = queued_frames.front();
 217                                 queued_frames.pop();
 218                         } else {
 219                                 qf.pts = -1;
 220                                 qf.duration = -1;
 221                                 qf.data = nullptr;
 222                         }
 223
 224                         frames_left = !queued_frames.empty();
 225                 }
 226
 227                 encode_frame(qf);
 228
 229                 {
 230                         lock_guard<mutex> lock(mu);
 231                         free_frames.push(qf.data);
 232                 }
 233
 234                 // We should quit only if the should_quit flag is set _and_ we have nothing
 235                 // in either queue.
 236         } while (!should_quit || frames_left || x264_encoder_delayed_frames(x264) > 0);
 237
 238         x264_encoder_close(x264);
 239 }
 240
 241 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 242 {
 243         x264_nal_t *nal = nullptr;
 244         int num_nal = 0;
 245         x264_picture_t pic;
 246         x264_picture_t *input_pic = nullptr;
 247
 248         if (qf.data) {
 249                 x264_picture_init(&pic);
 250
 251                 pic.i_pts = qf.pts;
 252                 pic.img.i_csp = X264_CSP_NV12;
 253                 pic.img.i_plane = 2;
 254                 pic.img.plane[0] = qf.data;
 255                 pic.img.i_stride[0] = global_flags.width;
 256                 pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 257                 pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 258                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 259
 260                 input_pic = &pic;
 261
 262                 frames_being_encoded[qf.pts] = qf.received_ts;
 263         }
 264
 265         // See if we have a new bitrate to change to.
 266         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 267         if (new_rate != 0) {
 268                 if (speed_control) {
 269                         speed_control->set_config_override_function([new_rate](x264_param_t *param) {
 270                                 param->rc.i_bitrate = new_rate;
 271                                 update_vbv_settings(param);
 272                         });
 273                 } else {
 274                         x264_param_t param;
 275                         x264_encoder_parameters(x264, &param);
 276                         param.rc.i_bitrate = new_rate;
 277                         update_vbv_settings(&param);
 278                         x264_encoder_reconfig(x264, &param);
 279                 }
 280         }
 281
 282         if (speed_control) {
 283                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 284         }
 285         x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 286         if (speed_control) {
 287                 speed_control->after_frame();
 288         }
 289
 290         if (num_nal == 0) return;
 291
 292         if (frames_being_encoded.count(pic.i_pts)) {
 293                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 294                 frames_being_encoded.erase(pic.i_pts);
 295
 296                 static int frameno = 0;
 297                 print_latency("Current x264 latency (video inputs → network mux):",
 298                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 299                         &frameno);
 300         } else {
 301                 assert(false);
 302         }
 303
 304         // We really need one AVPacket for the entire frame, it seems,
 305         // so combine it all.
 306         size_t num_bytes = buffered_sei.size();
 307         for (int i = 0; i < num_nal; ++i) {
 308                 num_bytes += nal[i].i_payload;
 309         }
 310
 311         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 312         uint8_t *ptr = data.get();
 313
 314         if (!buffered_sei.empty()) {
 315                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 316                 ptr += buffered_sei.size();
 317                 buffered_sei.clear();
 318         }
 319         for (int i = 0; i < num_nal; ++i) {
 320                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 321                 ptr += nal[i].i_payload;
 322         }
 323
 324         AVPacket pkt;
 325         memset(&pkt, 0, sizeof(pkt));
 326         pkt.buf = nullptr;
 327         pkt.data = data.get();
 328         pkt.size = num_bytes;
 329         pkt.stream_index = 0;
 330         if (pic.b_keyframe) {
 331                 pkt.flags = AV_PKT_FLAG_KEY;
 332         } else {
 333                 pkt.flags = 0;
 334         }
 335         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 336
 337         mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 338 }