X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=x264_encoder.cpp;h=8e3b567abe1684801315cd86dec9721aade3be9a;hb=bd5b2de9a277b87c75d71d94bd8c5095ab14ecf7;hp=bbbb1ba33df907babd72c9f2a9c91d1596950936;hpb=3be00c8dd8b841cecc44f57234b9fc2d3a94cb45;p=nageru diff --git a/x264_encoder.cpp b/x264_encoder.cpp index bbbb1ba..8e3b567 100644 --- a/x264_encoder.cpp +++ b/x264_encoder.cpp @@ -1,24 +1,53 @@ +#include "x264_encoder.h" + +#include +#include +#include #include #include +#include +#include #include "defs.h" #include "flags.h" #include "mux.h" +#include "print_latency.h" #include "timebase.h" -#include "x264_encoder.h" +#include "x264_speed_control.h" extern "C" { +#include #include } +using namespace movit; using namespace std; +using namespace std::chrono; + +namespace { + +void update_vbv_settings(x264_param_t *param) +{ + if (global_flags.x264_vbv_buffer_size < 0) { + param->rc.i_vbv_buffer_size = param->rc.i_bitrate; // One-second VBV. + } else { + param->rc.i_vbv_buffer_size = global_flags.x264_vbv_buffer_size; + } + if (global_flags.x264_vbv_max_bitrate < 0) { + param->rc.i_vbv_max_bitrate = param->rc.i_bitrate; // CBR. + } else { + param->rc.i_vbv_max_bitrate = global_flags.x264_vbv_max_bitrate; + } +} + +} // namespace X264Encoder::X264Encoder(AVOutputFormat *oformat) : wants_global_headers(oformat->flags & AVFMT_GLOBALHEADER) { - frame_pool.reset(new uint8_t[WIDTH * HEIGHT * 2 * X264_QUEUE_LENGTH]); + frame_pool.reset(new uint8_t[global_flags.width * global_flags.height * 2 * X264_QUEUE_LENGTH]); for (unsigned i = 0; i < X264_QUEUE_LENGTH; ++i) { - free_frames.push(frame_pool.get() + i * (WIDTH * HEIGHT * 2)); + free_frames.push(frame_pool.get() + i * (global_flags.width * global_flags.height * 2)); } encoder_thread = thread(&X264Encoder::encoder_thread_func, this); } @@ -30,11 +59,15 @@ X264Encoder::~X264Encoder() encoder_thread.join(); } -void X264Encoder::add_frame(int64_t pts, int64_t duration, const uint8_t *data) +void X264Encoder::add_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts) { + assert(!should_quit); + QueuedFrame qf; qf.pts = pts; qf.duration = duration; + qf.ycbcr_coefficients = ycbcr_coefficients; + qf.received_ts = received_ts; { lock_guard lock(mu); @@ -47,7 +80,7 @@ void X264Encoder::add_frame(int64_t pts, int64_t duration, const uint8_t *data) free_frames.pop(); } - memcpy(qf.data, data, WIDTH * HEIGHT * 2); + memcpy(qf.data, data, global_flags.width * global_flags.height * 2); { lock_guard lock(mu); @@ -61,30 +94,78 @@ void X264Encoder::init_x264() x264_param_t param; x264_param_default_preset(¶m, global_flags.x264_preset.c_str(), global_flags.x264_tune.c_str()); - param.i_width = WIDTH; - param.i_height = HEIGHT; + param.i_width = global_flags.width; + param.i_height = global_flags.height; param.i_csp = X264_CSP_NV12; param.b_vfr_input = 1; param.i_timebase_num = 1; param.i_timebase_den = TIMEBASE; param.i_keyint_max = 50; // About one second. + if (global_flags.x264_speedcontrol) { + param.i_frame_reference = 16; // Because speedcontrol is never allowed to change this above what we set at start. + } - // NOTE: These should be in sync with the ones in h264encode.cpp (sbs_rbsp()). + // NOTE: These should be in sync with the ones in quicksync_encoder.cpp (sps_rbsp()). param.vui.i_vidformat = 5; // Unspecified. param.vui.b_fullrange = 0; param.vui.i_colorprim = 1; // BT.709. param.vui.i_transfer = 2; // Unspecified (since we use sRGB). - param.vui.i_colmatrix = 6; // BT.601/SMPTE 170M. + if (global_flags.ycbcr_rec709_coefficients) { + param.vui.i_colmatrix = 1; // BT.709. + } else { + param.vui.i_colmatrix = 6; // BT.601/SMPTE 170M. + } - // 4.5 Mbit/sec, CBR. param.rc.i_rc_method = X264_RC_ABR; - param.rc.i_bitrate = 4500; + param.rc.i_bitrate = global_flags.x264_bitrate; + update_vbv_settings(¶m); + if (param.rc.i_vbv_max_bitrate > 0) { + // If the user wants VBV control to cap the max rate, it is + // also reasonable to assume that they are fine with the stream + // constantly being around that rate even for very low-complexity + // content; the obvious and extreme example being a static + // black picture. + // + // One would think it's fine to have low-complexity content use + // less bitrate, but it seems to cause problems in practice; + // e.g. VLC seems to often drop the stream (similar to a buffer + // underrun) in such cases, but only when streaming from Nageru, + // not when reading a dump of the same stream from disk. + // I'm not 100% sure whether it's in VLC (possibly some buffering + // in the HTTP layer), in microhttpd or somewhere in Nageru itself, + // but it's a typical case of problems that can arise. Similarly, + // TCP's congestion control is not always fond of the rate staying + // low for a while and then rising quickly -- a variation on the same + // problem. + // + // We solve this by simply asking x264 to fill in dummy bits + // in these cases, so that the bitrate stays reasonable constant. + // It's a waste of bandwidth, but it makes things go much more + // smoothly in these cases. (We don't do it if VBV control is off + // in general, not the least because it makes no sense and x264 + // thus ignores the parameter.) + param.rc.b_filler = 1; + } - // One-second VBV. - param.rc.i_vbv_max_bitrate = 4500; - param.rc.i_vbv_buffer_size = 4500; + // Occasionally players have problem with extremely low quantizers; + // be on the safe side. Shouldn't affect quality in any meaningful way. + param.rc.i_qp_min = 5; - // TODO: more flags here, via x264_param_parse(). + for (const string &str : global_flags.x264_extra_param) { + const size_t pos = str.find(','); + if (pos == string::npos) { + if (x264_param_parse(¶m, str.c_str(), nullptr) != 0) { + fprintf(stderr, "ERROR: x264 rejected parameter '%s'\n", str.c_str()); + } + } else { + const string key = str.substr(0, pos); + const string value = str.substr(pos + 1); + if (x264_param_parse(¶m, key.c_str(), value.c_str()) != 0) { + fprintf(stderr, "ERROR: x264 rejected parameter '%s' set to '%s'\n", + key.c_str(), value.c_str()); + } + } + } x264_param_apply_profile(¶m, "high"); @@ -96,6 +177,10 @@ void X264Encoder::init_x264() exit(1); } + if (global_flags.x264_speedcontrol) { + speed_control.reset(new X264SpeedControl(x264, /*f_speed=*/1.0f, X264_QUEUE_LENGTH, /*f_buffer_init=*/1.0f)); + } + if (wants_global_headers) { x264_nal_t *nal; int num_nal; @@ -115,8 +200,12 @@ void X264Encoder::init_x264() void X264Encoder::encoder_thread_func() { - nice(5); // Note that x264 further nices some of its threads. + if (nice(5) == -1) { // Note that x264 further nices some of its threads. + perror("nice()"); + // No exit; it's not fatal. + } init_x264(); + x264_init_done = true; bool frames_left; @@ -158,6 +247,7 @@ void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf) x264_nal_t *nal = nullptr; int num_nal = 0; x264_picture_t pic; + x264_picture_t *input_pic = nullptr; if (qf.data) { x264_picture_init(&pic); @@ -166,14 +256,71 @@ void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf) pic.img.i_csp = X264_CSP_NV12; pic.img.i_plane = 2; pic.img.plane[0] = qf.data; - pic.img.i_stride[0] = WIDTH; - pic.img.plane[1] = qf.data + WIDTH * HEIGHT; - pic.img.i_stride[1] = WIDTH / 2 * sizeof(uint16_t); + pic.img.i_stride[0] = global_flags.width; + pic.img.plane[1] = qf.data + global_flags.width * global_flags.height; + pic.img.i_stride[1] = global_flags.width / 2 * sizeof(uint16_t); pic.opaque = reinterpret_cast(intptr_t(qf.duration)); - x264_encoder_encode(x264, &nal, &num_nal, &pic, &pic); + input_pic = &pic; + + frames_being_encoded[qf.pts] = qf.received_ts; + } + + // See if we have a new bitrate to change to. + unsigned new_rate = new_bitrate_kbit.exchange(0); // Read and clear. + if (new_rate != 0) { + bitrate_override_func = [new_rate](x264_param_t *param) { + param->rc.i_bitrate = new_rate; + update_vbv_settings(param); + }; + } + + auto ycbcr_coefficients_override_func = [qf](x264_param_t *param) { + if (qf.ycbcr_coefficients == YCBCR_REC_709) { + param->vui.i_colmatrix = 1; // BT.709. + } else { + assert(qf.ycbcr_coefficients == YCBCR_REC_601); + param->vui.i_colmatrix = 6; // BT.601/SMPTE 170M. + } + }; + + if (speed_control) { + speed_control->set_config_override_function([this, ycbcr_coefficients_override_func](x264_param_t *param) { + if (bitrate_override_func) { + bitrate_override_func(param); + } + ycbcr_coefficients_override_func(param); + }); } else { - x264_encoder_encode(x264, &nal, &num_nal, nullptr, &pic); + x264_param_t param; + x264_encoder_parameters(x264, ¶m); + if (bitrate_override_func) { + bitrate_override_func(¶m); + } + ycbcr_coefficients_override_func(¶m); + x264_encoder_reconfig(x264, ¶m); + } + + if (speed_control) { + speed_control->before_frame(float(free_frames.size()) / X264_QUEUE_LENGTH, X264_QUEUE_LENGTH, 1e6 * qf.duration / TIMEBASE); + } + x264_encoder_encode(x264, &nal, &num_nal, input_pic, &pic); + if (speed_control) { + speed_control->after_frame(); + } + + if (num_nal == 0) return; + + if (frames_being_encoded.count(pic.i_pts)) { + ReceivedTimestamps received_ts = frames_being_encoded[pic.i_pts]; + frames_being_encoded.erase(pic.i_pts); + + static int frameno = 0; + print_latency("Current x264 latency (video inputs → network mux):", + received_ts, (pic.i_type == X264_TYPE_B || pic.i_type == X264_TYPE_BREF), + &frameno); + } else { + assert(false); } // We really need one AVPacket for the entire frame, it seems, @@ -209,5 +356,7 @@ void X264Encoder::encode_frame(X264Encoder::QueuedFrame qf) } pkt.duration = reinterpret_cast(pic.opaque); - mux->add_packet(pkt, pic.i_pts, pic.i_dts); + for (Mux *mux : muxes) { + mux->add_packet(pkt, pic.i_pts, pic.i_dts); + } }