X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=nageru%2Fmjpeg_encoder.cpp;h=39a5959ed3bfeeedbd6b5715ba93b180239f1457;hb=f9024d141398e69e7b4011becd3ebbe37eaa1776;hp=46bb94c7639112f76b031c96c63acab9560b99ce;hpb=9ffd4f03f314cc6e0254449593def95c9bc203d6;p=nageru diff --git a/nageru/mjpeg_encoder.cpp b/nageru/mjpeg_encoder.cpp index 46bb94c..39a5959 100644 --- a/nageru/mjpeg_encoder.cpp +++ b/nageru/mjpeg_encoder.cpp @@ -21,11 +21,15 @@ extern "C" { #include "shared/timebase.h" #include "va_display_with_cleanup.h" +#include + #include #include #include +using namespace Eigen; using namespace bmusb; +using namespace movit; using namespace std; static VAImageFormat uyvy_format; @@ -120,7 +124,7 @@ MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display) // a situation with only one video stream (and possibly one audio stream) // with known width/height, and we don't need the extra functionality it provides. avctx.reset(avformat_alloc_context()); - avctx->oformat = av_guess_format("mp4", nullptr, nullptr); + avctx->oformat = av_guess_format("nut", nullptr, nullptr); uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE); avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr); @@ -133,7 +137,11 @@ MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display) fprintf(stderr, "avformat_new_stream() failed\n"); abort(); } - stream->time_base = AVRational{ 1, TIMEBASE }; + + // FFmpeg is very picky about having audio at 1/48000 timebase, + // no matter what we write. Even though we'd prefer our usual 1/120000, + // put the video on the same one, so that we can have locked audio. + stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY }; stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO; stream->codecpar->codec_id = AV_CODEC_ID_MJPEG; @@ -151,6 +159,19 @@ MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display) stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT; stream->codecpar->field_order = AV_FIELD_PROGRESSIVE; } + for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) { + AVStream *stream = avformat_new_stream(avctx.get(), nullptr); + if (stream == nullptr) { + fprintf(stderr, "avformat_new_stream() failed\n"); + abort(); + } + stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY }; + stream->codecpar->codec_type = AVMEDIA_TYPE_AUDIO; + stream->codecpar->codec_id = AV_CODEC_ID_PCM_S32LE; + stream->codecpar->channel_layout = AV_CH_LAYOUT_STEREO; + stream->codecpar->channels = 2; + stream->codecpar->sample_rate = OUTPUT_FREQUENCY; + } AVDictionary *options = NULL; vector> opts = MUX_OPTS; @@ -269,7 +290,7 @@ unique_ptr MJPEGEncoder::try_open_va(const string &va_disp return va_dpy; } -void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFrame frame, const bmusb::VideoFormat &video_format, size_t y_offset, size_t cbcr_offset) +void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFrame frame, const bmusb::VideoFormat &video_format, size_t y_offset, size_t cbcr_offset, vector audio, const RGBTriplet &white_balance) { PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata; if (video_format.width == 0 || video_format.height == 0) { @@ -300,7 +321,7 @@ void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFram return; } ++metric_mjpeg_overrun_submitted; - frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset }); + frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio), white_balance }); any_frames_to_be_encoded.notify_all(); } @@ -341,6 +362,11 @@ void MJPEGEncoder::encoder_thread_func() // Will call back in the receiver thread. encode_jpeg_va(move(qf)); } else { + // Write audio before video, since Futatabi expects it. + if (qf.audio.size() > 0) { + write_audio_packet(qf.pts, qf.card_index, qf.audio); + } + // Encode synchronously, in the same thread. vector jpeg = encode_jpeg_libjpeg(qf); write_mjpeg_packet(qf.pts, qf.card_index, jpeg.data(), jpeg.size()); @@ -364,6 +390,27 @@ void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const ui pkt.flags = AV_PKT_FLAG_KEY; AVRational time_base = avctx->streams[pkt.stream_index]->time_base; pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base); + pkt.duration = 0; + + if (av_write_frame(avctx.get(), &pkt) < 0) { + fprintf(stderr, "av_write_frame() failed\n"); + abort(); + } +} + +void MJPEGEncoder::write_audio_packet(int64_t pts, unsigned card_index, const vector &audio) +{ + AVPacket pkt; + memset(&pkt, 0, sizeof(pkt)); + pkt.buf = nullptr; + pkt.data = reinterpret_cast(const_cast(&audio[0])); + pkt.size = audio.size() * sizeof(audio[0]); + pkt.stream_index = card_index + global_flags.card_to_mjpeg_stream_export.size(); + pkt.flags = AV_PKT_FLAG_KEY; + AVRational time_base = avctx->streams[pkt.stream_index]->time_base; + pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base); + size_t num_stereo_samples = audio.size() / 2; + pkt.duration = av_rescale_q(num_stereo_samples, AVRational{ 1, OUTPUT_FREQUENCY }, time_base); if (av_write_frame(avctx.get(), &pkt) < 0) { fprintf(stderr, "av_write_frame() failed\n"); @@ -452,7 +499,25 @@ void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources) va_resources_freelist.push_front(resources); } -void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo) +namespace { + +void push16(uint16_t val, string *str) +{ + str->push_back(val >> 8); + str->push_back(val & 0xff); +} + +void push32(uint32_t val, string *str) +{ + str->push_back(val >> 24); + str->push_back((val >> 16) & 0xff); + str->push_back((val >> 8) & 0xff); + str->push_back(val & 0xff); +} + +} // namespace + +void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, const RGBTriplet &white_balance, VectorDestinationManager *dest, jpeg_compress_struct *cinfo) { jpeg_error_mgr jerr; cinfo->err = jpeg_std_error(&jerr); @@ -477,15 +542,70 @@ void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinat cinfo->CCIR601_sampling = true; // Seems to be mostly ignored by libjpeg, though. jpeg_start_compress(cinfo, true); + if (fabs(white_balance.r - 1.0f) > 1e-3 || + fabs(white_balance.g - 1.0f) > 1e-3 || + fabs(white_balance.b - 1.0f) > 1e-3) { + // Convert from (linear) RGB to XYZ. + Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB); + Vector3d xyz = rgb_to_xyz_matrix * Vector3d(white_balance.r, white_balance.g, white_balance.b); + + // Convert from XYZ to xyz by normalizing. + xyz /= (xyz[0] + xyz[1] + xyz[2]); + + // Create a very rudimentary EXIF header to hold our white point. + string exif; + + // Exif header, followed by some padding. + exif = "Exif"; + push16(0, &exif); + + // TIFF header first: + exif += "MM"; // Big endian. + + // Magic number. + push16(42, &exif); + + // Offset of first IFD (relative to the MM, immediately after the header). + push32(exif.size() - 6 + 4, &exif); + + // Now the actual IFD. + + // One entry. + push16(1, &exif); + + // WhitePoint tag ID. + push16(0x13e, &exif); + + // Rational type. + push16(5, &exif); + + // Two values (x and y; z is implicit due to normalization). + push32(2, &exif); + + // Offset (relative to the MM, immediately after the last IFD). + push32(exif.size() - 6 + 8, &exif); + + // No more IFDs. + push32(0, &exif); + + // The actual values. + push32(lrintf(xyz[0] * 10000.0f), &exif); + push32(10000, &exif); + push32(lrintf(xyz[1] * 10000.0f), &exif); + push32(10000, &exif); + + jpeg_write_marker(cinfo, JPEG_APP0 + 1, (const JOCTET *)exif.data(), exif.size()); + } + // This comment marker is private to FFmpeg. It signals limited Y'CbCr range // (and nothing else). jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601")); } -vector MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo) +vector MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, const RGBTriplet &white_balance, jpeg_compress_struct *cinfo) { VectorDestinationManager dest; - init_jpeg_422(width, height, &dest, cinfo); + init_jpeg_422(width, height, white_balance, &dest, cinfo); // Make a dummy black image; there's seemingly no other easy way of // making libjpeg outputting all of its headers. @@ -517,7 +637,7 @@ vector MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, j return dest.dest; } -MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height) +MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height, const RGBTriplet &white_balance) { pair key(width, height); if (va_data_for_resolution.count(key)) { @@ -527,7 +647,7 @@ MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, un // Use libjpeg to generate a header and set sane defaults for e.g. // quantization tables. Then do the actual encode with VA-API. jpeg_compress_struct cinfo; - vector jpeg_header = get_jpeg_header(width, height, &cinfo); + vector jpeg_header = get_jpeg_header(width, height, white_balance, &cinfo); // Picture parameters. VAEncPictureParameterBufferJPEG pic_param; @@ -643,7 +763,7 @@ void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf) release = ReleaseVAResources(this, resources); } - VAData va_data = get_va_data_for_resolution(width, height); + VAData va_data = get_va_data_for_resolution(width, height, qf.white_balance); va_data.pic_param.coded_buf = resources.data_buffer; VABufferID pic_param_buffer; @@ -752,6 +872,11 @@ void MJPEGEncoder::va_receiver_thread_func() frames_encoding.pop(); } + // Write audio before video, since Futatabi expects it. + if (qf.audio.size() > 0) { + write_audio_packet(qf.pts, qf.card_index, qf.audio); + } + VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface); CHECK_VASTATUS(va_status, "vaSyncSurface"); @@ -774,7 +899,7 @@ vector MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf) VectorDestinationManager dest; jpeg_compress_struct cinfo; - init_jpeg_422(width, height, &dest, &cinfo); + init_jpeg_422(width, height, qf.white_balance, &dest, &cinfo); size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support. size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2; @@ -784,7 +909,7 @@ vector MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf) for (unsigned y = 0; y < qf.video_format.height; y += 8) { const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2; - memcpy_interleaved(tmp_y, tmp_cbcr, src, qf.video_format.width * 8 * 2); + memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2); memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8); for (unsigned yy = 0; yy < 8; ++yy) { yptr[yy] = tmp_y + yy * width;