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 h264encode.cpp (sbs_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         param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 110
 111
 112         param.rc.i_rc_method = X264_RC_ABR;
 113         param.rc.i_bitrate = global_flags.x264_bitrate;
 114         update_vbv_settings(&param);
 115         if (param.rc.i_vbv_max_bitrate > 0) {
 116                 // If the user wants VBV control to cap the max rate, it is
 117                 // also reasonable to assume that they are fine with the stream
 118                 // constantly being around that rate even for very low-complexity
 119                 // content; the obvious and extreme example being a static
 120                 // black picture.
 121                 //
 122                 // One would think it's fine to have low-complexity content use
 123                 // less bitrate, but it seems to cause problems in practice;
 124                 // e.g. VLC seems to often drop the stream (similar to a buffer
 125                 // underrun) in such cases, but only when streaming from Nageru,
 126                 // not when reading a dump of the same stream from disk.
 127                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 128                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 129                 // but it's a typical case of problems that can arise. Similarly,
 130                 // TCP's congestion control is not always fond of the rate staying
 131                 // low for a while and then rising quickly -- a variation on the same
 132                 // problem.
 133                 //
 134                 // We solve this by simply asking x264 to fill in dummy bits
 135                 // in these cases, so that the bitrate stays reasonable constant.
 136                 // It's a waste of bandwidth, but it makes things go much more
 137                 // smoothly in these cases. (We don't do it if VBV control is off
 138                 // in general, not the least because it makes no sense and x264
 139                 // thus ignores the parameter.)
 140                 param.rc.b_filler = 1;
 141         }
 142
 143         // Occasionally players have problem with extremely low quantizers;
 144         // be on the safe side. Shouldn't affect quality in any meaningful way.
 145         param.rc.i_qp_min = 5;
 146
 147         for (const string &str : global_flags.x264_extra_param) {
 148                 const size_t pos = str.find(',');
 149                 if (pos == string::npos) {
 150                         if (x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 151                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 152                         }
 153                 } else {
 154                         const string key = str.substr(0, pos);
 155                         const string value = str.substr(pos + 1);
 156                         if (x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 157                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 158                                         key.c_str(), value.c_str());
 159                         }
 160                 }
 161         }
 162
 163         x264_param_apply_profile(&param, "high");
 164
 165         param.b_repeat_headers = !wants_global_headers;
 166
 167         x264 = x264_encoder_open(&param);
 168         if (x264 == nullptr) {
 169                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 170                 exit(1);
 171         }
 172
 173         if (global_flags.x264_speedcontrol) {
 174                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 175         }
 176
 177         if (wants_global_headers) {
 178                 x264_nal_t *nal;
 179                 int num_nal;
 180
 181                 x264_encoder_headers(x264, &nal, &num_nal);
 182
 183                 for (int i = 0; i < num_nal; ++i) {
 184                         if (nal[i].i_type == NAL_SEI) {
 185                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 186                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 187                         } else {
 188                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 189                         }
 190                 }
 191         }
 192 }
 193
 194 void X264Encoder::encoder_thread_func()
 195 {
 196         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 197                 perror("nice()");
 198                 // No exit; it's not fatal.
 199         }
 200         init_x264();
 201         x264_init_done = true;
 202
 203         bool frames_left;
 204
 205         do {
 206                 QueuedFrame qf;
 207
 208                 // Wait for a queued frame, then dequeue it.
 209                 {
 210                         unique_lock<mutex> lock(mu);
 211                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 212                         if (!queued_frames.empty()) {
 213                                 qf = queued_frames.front();
 214                                 queued_frames.pop();
 215                         } else {
 216                                 qf.pts = -1;
 217                                 qf.duration = -1;
 218                                 qf.data = nullptr;
 219                         }
 220
 221                         frames_left = !queued_frames.empty();
 222                 }
 223
 224                 encode_frame(qf);
 225
 226                 {
 227                         lock_guard<mutex> lock(mu);
 228                         free_frames.push(qf.data);
 229                 }
 230
 231                 // We should quit only if the should_quit flag is set _and_ we have nothing
 232                 // in either queue.
 233         } while (!should_quit || frames_left || x264_encoder_delayed_frames(x264) > 0);
 234
 235         x264_encoder_close(x264);
 236 }
 237
 238 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 239 {
 240         x264_nal_t *nal = nullptr;
 241         int num_nal = 0;
 242         x264_picture_t pic;
 243         x264_picture_t *input_pic = nullptr;
 244
 245         if (qf.data) {
 246                 x264_picture_init(&pic);
 247
 248                 pic.i_pts = qf.pts;
 249                 pic.img.i_csp = X264_CSP_NV12;
 250                 pic.img.i_plane = 2;
 251                 pic.img.plane[0] = qf.data;
 252                 pic.img.i_stride[0] = global_flags.width;
 253                 pic.img.plane[1] = qf.data + global_flags.width * global_flags.height;
 254                 pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t);
 255                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 256
 257                 input_pic = &pic;
 258
 259                 frames_being_encoded[qf.pts] = qf.received_ts;
 260         }
 261
 262         // See if we have a new bitrate to change to.
 263         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 264         if (new_rate != 0) {
 265                 if (speed_control) {
 266                         speed_control->set_config_override_function([new_rate](x264_param_t *param) {
 267                                 param->rc.i_bitrate = new_rate;
 268                                 update_vbv_settings(param);
 269                         });
 270                 } else {
 271                         x264_param_t param;
 272                         x264_encoder_parameters(x264, &param);
 273                         param.rc.i_bitrate = new_rate;
 274                         update_vbv_settings(&param);
 275                         x264_encoder_reconfig(x264, &param);
 276                 }
 277         }
 278
 279         if (speed_control) {
 280                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 281         }
 282         x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 283         if (speed_control) {
 284                 speed_control->after_frame();
 285         }
 286
 287         if (num_nal == 0) return;
 288
 289         if (frames_being_encoded.count(pic.i_pts)) {
 290                 ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts];
 291                 frames_being_encoded.erase(pic.i_pts);
 292
 293                 static int frameno = 0;
 294                 print_latency("Current x264 latency (video inputs → network mux):",
 295                         received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF),
 296                         &frameno);
 297         } else {
 298                 assert(false);
 299         }
 300
 301         // We really need one AVPacket for the entire frame, it seems,
 302         // so combine it all.
 303         size_t num_bytes = buffered_sei.size();
 304         for (int i = 0; i < num_nal; ++i) {
 305                 num_bytes += nal[i].i_payload;
 306         }
 307
 308         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 309         uint8_t *ptr = data.get();
 310
 311         if (!buffered_sei.empty()) {
 312                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 313                 ptr += buffered_sei.size();
 314                 buffered_sei.clear();
 315         }
 316         for (int i = 0; i < num_nal; ++i) {
 317                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 318                 ptr += nal[i].i_payload;
 319         }
 320
 321         AVPacket pkt;
 322         memset(&pkt, 0, sizeof(pkt));
 323         pkt.buf = nullptr;
 324         pkt.data = data.get();
 325         pkt.size = num_bytes;
 326         pkt.stream_index = 0;
 327         if (pic.b_keyframe) {
 328                 pkt.flags = AV_PKT_FLAG_KEY;
 329         } else {
 330                 pkt.flags = 0;
 331         }
 332         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 333
 334         mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 335 }