#include "shared/timebase.h"
#include "va_display_with_cleanup.h"
+#include <movit/colorspace_conversion_effect.h>
+
#include <va/va.h>
#include <va/va_drm.h>
#include <va/va_x11.h>
+using namespace Eigen;
using namespace bmusb;
+using namespace movit;
using namespace std;
static VAImageFormat uyvy_format;
return buf_size;
}
+namespace {
+
+void add_video_stream(AVFormatContext *avctx)
+{
+ AVStream *stream = avformat_new_stream(avctx, nullptr);
+ if (stream == nullptr) {
+ fprintf(stderr, "avformat_new_stream() failed\n");
+ abort();
+ }
+
+ // 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;
+
+ // Used for aspect ratio only. Can change without notice (the mux won't care).
+ stream->codecpar->width = global_flags.width;
+ stream->codecpar->height = global_flags.height;
+
+ // TODO: We could perhaps use the interpretation for each card here
+ // (or at least the command-line flags) instead of the defaults,
+ // but what would we do when they change?
+ stream->codecpar->color_primaries = AVCOL_PRI_BT709;
+ stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
+ stream->codecpar->color_space = AVCOL_SPC_BT709;
+ stream->codecpar->color_range = AVCOL_RANGE_MPEG;
+ stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
+ stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
+}
+
+void add_audio_stream(AVFormatContext *avctx)
+{
+ AVStream *stream = avformat_new_stream(avctx, 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;
+}
+
+void finalize_mux(AVFormatContext *avctx)
+{
+ AVDictionary *options = NULL;
+ vector<pair<string, string>> opts = MUX_OPTS;
+ for (pair<string, string> opt : opts) {
+ av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
+ }
+ if (avformat_write_header(avctx, &options) < 0) {
+ fprintf(stderr, "avformat_write_header() failed\n");
+ abort();
+ }
+}
+
+} // namespace
+
MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
: httpd(httpd)
{
avctx->flags = AVFMT_FLAG_CUSTOM_IO;
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();
- }
-
- // 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;
-
- // Used for aspect ratio only. Can change without notice (the mux won't care).
- stream->codecpar->width = global_flags.width;
- stream->codecpar->height = global_flags.height;
-
- // TODO: We could perhaps use the interpretation for each card here
- // (or at least the command-line flags) instead of the defaults,
- // but what would we do when they change?
- stream->codecpar->color_primaries = AVCOL_PRI_BT709;
- stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
- stream->codecpar->color_space = AVCOL_SPC_BT709;
- stream->codecpar->color_range = AVCOL_RANGE_MPEG;
- stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
- stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
+ add_video_stream(avctx.get());
}
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<pair<string, string>> opts = MUX_OPTS;
- for (pair<string, string> opt : opts) {
- av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
- }
- if (avformat_write_header(avctx.get(), &options) < 0) {
- fprintf(stderr, "avformat_write_header() failed\n");
- abort();
+ add_audio_stream(avctx.get());
}
+ finalize_mux(avctx.get());
// Initialize VA-API.
string error;
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, vector<int32_t> audio)
+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<int32_t> audio, const RGBTriplet &white_balance)
{
PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
if (video_format.width == 0 || video_format.height == 0) {
return;
}
++metric_mjpeg_overrun_submitted;
- frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio) });
+ 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();
}
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);
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<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
+vector<uint8_t> 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.
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<unsigned, unsigned> key(width, height);
if (va_data_for_resolution.count(key)) {
// 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<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
+ vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, &cinfo);
// Picture parameters.
VAEncPictureParameterBufferJPEG pic_param;
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;
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;
projects / nageru / blobdiff