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

   1 #include "x264_encoder.h"
   2
   3 #include <stdio.h>
   4 #include <stdlib.h>
   5 #include <string.h>
   6 #include <unistd.h>
   7 #include <x264.h>
   8 #include <cstdint>
   9
  10 #include "defs.h"
  11 #include "flags.h"
  12 #include "mux.h"
  13 #include "timebase.h"
  14 #include "x264_speed_control.h"
  15
  16 extern "C" {
  17 #include <libavcodec/avcodec.h>
  18 #include <libavformat/avformat.h>
  19 }
  20
  21 using namespace std;
  22
  23 namespace {
  24
  25 void update_vbv_settings(x264_param_t *param)
  26 {
  27         if (global_flags.x264_vbv_buffer_size < 0) {
  28                 param->rc.i_vbv_buffer_size = param->rc.i_bitrate;  // One-second VBV.
  29         } else {
  30                 param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size;
  31         }
  32         if (global_flags.x264_vbv_max_bitrate < 0) {
  33                 param->rc.i_vbv_max_bitrate = param->rc.i_bitrate;  // CBR.
  34         } else {
  35                 param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate;
  36         }
  37 }
  38
  39 }  // namespace
  40
  41 X264Encoder::X264Encoder(AVOutputFormat *oformat)
  42         : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER)
  43 {
  44         frame_pool.reset(new uint8_t[WIDTH * HEIGHT * 2 * X264_QUEUE_LENGTH]);
  45         for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) {
  46                 free_frames.push(frame_pool.get() + i * (WIDTH * HEIGHT * 2));
  47         }
  48         encoder_thread = thread(&X264Encoder::encoder_thread_func, this);
  49 }
  50
  51 X264Encoder::~X264Encoder()
  52 {
  53         should_quit = true;
  54         queued_frames_nonempty.notify_all();
  55         encoder_thread.join();
  56 }
  57
  58 void X264Encoder::add_frame(int64_t pts, int64_t duration, const uint8_t *data)
  59 {
  60         QueuedFrame qf;
  61         qf.pts = pts;
  62         qf.duration = duration;
  63
  64         {
  65                 lock_guard<mutex> lock(mu);
  66                 if (free_frames.empty()) {
  67                         fprintf(stderr, "WARNING: x264 queue full, dropping frame with pts %ld\n", pts);
  68                         return;
  69                 }
  70
  71                 qf.data = free_frames.front();
  72                 free_frames.pop();
  73         }
  74
  75         memcpy(qf.data, data, WIDTH * HEIGHT * 2);
  76
  77         {
  78                 lock_guard<mutex> lock(mu);
  79                 queued_frames.push(qf);
  80                 queued_frames_nonempty.notify_all();
  81         }
  82 }
  83
  84 void X264Encoder::init_x264()
  85 {
  86         x264_param_t param;
  87         x264_param_default_preset(&param, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str());
  88
  89         param.i_width = WIDTH;
  90         param.i_height = HEIGHT;
  91         param.i_csp = X264_CSP_NV12;
  92         param.b_vfr_input = 1;
  93         param.i_timebase_num = 1;
  94         param.i_timebase_den = TIMEBASE;
  95         param.i_keyint_max = 50; // About one second.
  96         if (global_flags.x264_speedcontrol) {
  97                 param.i_frame_reference = 16;  // Because speedcontrol is never allowed to change this above what we set at start.
  98         }
  99
 100         // NOTE: These should be in sync with the ones in h264encode.cpp (sbs_rbsp()).
 101         param.vui.i_vidformat = 5;  // Unspecified.
 102         param.vui.b_fullrange = 0;
 103         param.vui.i_colorprim = 1;  // BT.709.
 104         param.vui.i_transfer = 2;  // Unspecified (since we use sRGB).
 105         param.vui.i_colmatrix = 6;  // BT.601/SMPTE 170M.
 106
 107
 108         param.rc.i_rc_method = X264_RC_ABR;
 109         param.rc.i_bitrate = global_flags.x264_bitrate;
 110         update_vbv_settings(&param);
 111         if (param.rc.i_vbv_max_bitrate > 0) {
 112                 // If the user wants VBV control to cap the max rate, it is
 113                 // also reasonable to assume that they are fine with the stream
 114                 // constantly being around that rate even for very low-complexity
 115                 // content; the obvious and extreme example being a static
 116                 // black picture.
 117                 //
 118                 // One would think it's fine to have low-complexity content use
 119                 // less bitrate, but it seems to cause problems in practice;
 120                 // e.g. VLC seems to often drop the stream (similar to a buffer
 121                 // underrun) in such cases, but only when streaming from Nageru,
 122                 // not when reading a dump of the same stream from disk.
 123                 // I'm not 100% sure whether it's in VLC (possibly some buffering
 124                 // in the HTTP layer), in microhttpd or somewhere in Nageru itself,
 125                 // but it's a typical case of problems that can arise. Similarly,
 126                 // TCP's congestion control is not always fond of the rate staying
 127                 // low for a while and then rising quickly -- a variation on the same
 128                 // problem.
 129                 //
 130                 // We solve this by simply asking x264 to fill in dummy bits
 131                 // in these cases, so that the bitrate stays reasonable constant.
 132                 // It's a waste of bandwidth, but it makes things go much more
 133                 // smoothly in these cases. (We don't do it if VBV control is off
 134                 // in general, not the least because it makes no sense and x264
 135                 // thus ignores the parameter.)
 136                 param.rc.b_filler = 1;
 137         }
 138
 139         // Occasionally players have problem with extremely low quantizers;
 140         // be on the safe side. Shouldn't affect quality in any meaningful way.
 141         param.rc.i_qp_min = 5;
 142
 143         for (const string &str : global_flags.x264_extra_param) {
 144                 const size_t pos = str.find(',');
 145                 if (pos == string::npos) {
 146                         if (x264_param_parse(&param, str.c_str(), nullptr) != 0) {
 147                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str());
 148                         }
 149                 } else {
 150                         const string key = str.substr(0, pos);
 151                         const string value = str.substr(pos + 1);
 152                         if (x264_param_parse(&param, key.c_str(), value.c_str()) != 0) {
 153                                 fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n",
 154                                         key.c_str(), value.c_str());
 155                         }
 156                 }
 157         }
 158
 159         x264_param_apply_profile(&param, "high");
 160
 161         param.b_repeat_headers = !wants_global_headers;
 162
 163         x264 = x264_encoder_open(&param);
 164         if (x264 == nullptr) {
 165                 fprintf(stderr, "ERROR: x264 initialization failed.\n");
 166                 exit(1);
 167         }
 168
 169         if (global_flags.x264_speedcontrol) {
 170                 speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f));
 171         }
 172
 173         if (wants_global_headers) {
 174                 x264_nal_t *nal;
 175                 int num_nal;
 176
 177                 x264_encoder_headers(x264, &nal, &num_nal);
 178
 179                 for (int i = 0; i < num_nal; ++i) {
 180                         if (nal[i].i_type == NAL_SEI) {
 181                                 // Don't put the SEI in extradata; make it part of the first frame instead.
 182                                 buffered_sei += string((const char *)nal[i].p_payload, nal[i].i_payload);
 183                         } else {
 184                                 global_headers += string((const char *)nal[i].p_payload, nal[i].i_payload);
 185                         }
 186                 }
 187         }
 188 }
 189
 190 void X264Encoder::encoder_thread_func()
 191 {
 192         if (nice(5) == -1) {  // Note that x264 further nices some of its threads.
 193                 perror("nice()");
 194                 // No exit; it's not fatal.
 195         }
 196         init_x264();
 197
 198         bool frames_left;
 199
 200         do {
 201                 QueuedFrame qf;
 202
 203                 // Wait for a queued frame, then dequeue it.
 204                 {
 205                         unique_lock<mutex> lock(mu);
 206                         queued_frames_nonempty.wait(lock, [this]() { return !queued_frames.empty() || should_quit; });
 207                         if (!queued_frames.empty()) {
 208                                 qf = queued_frames.front();
 209                                 queued_frames.pop();
 210                         } else {
 211                                 qf.pts = -1;
 212                                 qf.duration = -1;
 213                                 qf.data = nullptr;
 214                         }
 215
 216                         frames_left = !queued_frames.empty();
 217                 }
 218
 219                 encode_frame(qf);
 220
 221                 {
 222                         lock_guard<mutex> lock(mu);
 223                         free_frames.push(qf.data);
 224                 }
 225
 226                 // We should quit only if the should_quit flag is set _and_ we have nothing
 227                 // in either queue.
 228         } while (!should_quit || frames_left || x264_encoder_delayed_frames(x264) > 0);
 229
 230         x264_encoder_close(x264);
 231 }
 232
 233 void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf)
 234 {
 235         x264_nal_t *nal = nullptr;
 236         int num_nal = 0;
 237         x264_picture_t pic;
 238         x264_picture_t *input_pic = nullptr;
 239
 240         if (qf.data) {
 241                 x264_picture_init(&pic);
 242
 243                 pic.i_pts = qf.pts;
 244                 pic.img.i_csp = X264_CSP_NV12;
 245                 pic.img.i_plane = 2;
 246                 pic.img.plane[0] = qf.data;
 247                 pic.img.i_stride[0] = WIDTH;
 248                 pic.img.plane[1] = qf.data + WIDTH * HEIGHT;
 249                 pic.img.i_stride[1] = WIDTH / 2 * sizeof(uint16_t);
 250                 pic.opaque = reinterpret_cast<void *>(intptr_t(qf.duration));
 251
 252                 input_pic = &pic;
 253         }
 254
 255         // See if we have a new bitrate to change to.
 256         unsigned new_rate = new_bitrate_kbit.exchange(0);  // Read and clear.
 257         if (new_rate != 0) {
 258                 if (speed_control) {
 259                         speed_control->set_config_override_function([new_rate](x264_param_t *param) {
 260                                 param->rc.i_bitrate = new_rate;
 261                                 update_vbv_settings(param);
 262                         });
 263                 } else {
 264                         x264_param_t param;
 265                         x264_encoder_parameters(x264, &param);
 266                         param.rc.i_bitrate = new_rate;
 267                         update_vbv_settings(&param);
 268                         x264_encoder_reconfig(x264, &param);
 269                 }
 270         }
 271
 272         if (speed_control) {
 273                 speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE);
 274         }
 275         x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic);
 276         if (speed_control) {
 277                 speed_control->after_frame();
 278         }
 279
 280         // We really need one AVPacket for the entire frame, it seems,
 281         // so combine it all.
 282         size_t num_bytes = buffered_sei.size();
 283         for (int i = 0; i < num_nal; ++i) {
 284                 num_bytes += nal[i].i_payload;
 285         }
 286
 287         unique_ptr<uint8_t[]> data(new uint8_t[num_bytes]);
 288         uint8_t *ptr = data.get();
 289
 290         if (!buffered_sei.empty()) {
 291                 memcpy(ptr, buffered_sei.data(), buffered_sei.size());
 292                 ptr += buffered_sei.size();
 293                 buffered_sei.clear();
 294         }
 295         for (int i = 0; i < num_nal; ++i) {
 296                 memcpy(ptr, nal[i].p_payload, nal[i].i_payload);
 297                 ptr += nal[i].i_payload;
 298         }
 299
 300         AVPacket pkt;
 301         memset(&pkt, 0, sizeof(pkt));
 302         pkt.buf = nullptr;
 303         pkt.data = data.get();
 304         pkt.size = num_bytes;
 305         pkt.stream_index = 0;
 306         if (pic.b_keyframe) {
 307                 pkt.flags = AV_PKT_FLAG_KEY;
 308         } else {
 309                 pkt.flags = 0;
 310         }
 311         pkt.duration = reinterpret_cast<intptr_t>(pic.opaque);
 312
 313         mux->add_packet(pkt, pic.i_pts, pic.i_dts);
 314 }