JPEG decoder and encoder with SIMD optimizations.
The decoder reads baseline and progressive JPEG, including grayscale, RGB, CMYK and YCCK, and is resilient to truncated or corrupt scan data. The encoder writes baseline JPEG with a choice of chroma subsampling, optional optimized Huffman tables, restart intervals and embedded metadata.
Assembly paths are bit-identical to the pure Go ones, so output does not depend on the architecture
or on the noasm build tag.
opts := &jpegn.EncodeOptions{
Quality: 90,
Subsampling: jpegn.Subsample420, // or Auto, 444, 440, 422, Gray
OptimizeCoding: true, // smaller files, one extra pass
}
err := jpegn.Encode(w, img, opts)SubsampleAuto follows the source: an image.YCbCr keeps its native ratio and its planes are
used without resampling, an image.Gray is written as a single component, anything else becomes 4:2:0.
Metadata is carried through a decode and re-encode with RawExif:
raw, _ := jpegn.RawExif(bytes.NewReader(src))
err := jpegn.Encode(w, img, &jpegn.EncodeOptions{
Quality: 90,
Exif: raw,
ResetOrientation: true, // when the pixels were already rotated
Segments: []jpegn.Segment{{Marker: 0xE2, Data: iccProfile}},
})Decoding, compared to the standard library:
BenchmarkDecodeBaseline420-8 3058 791961 ns/op 426324 B/op 6 allocs/op
BenchmarkDecodeBaseline420StdLib-8 1106 2047137 ns/op 407090 B/op 5 allocs/op
BenchmarkDecodeProgressive420-8 991 2303827 ns/op 1999880 B/op 10 allocs/op
BenchmarkDecodeProgressive420StdLib-8 603 4017204 ns/op 1980104 B/op 17 allocs/op
BenchmarkDecodeConfig-8 4975368 474.3 ns/op 48 B/op 1 allocs/op
BenchmarkDecodeConfigStdLib-8 1219048 2059 ns/op 13616 B/op 2 allocs/op
BenchmarkDecodeToRGBANearestNeighbor-8 1822 1317247 ns/op 2000039 B/op 10 allocs/op
BenchmarkDecodeToRGBACatmullRom-8 1633 1423599 ns/op 2262183 B/op 12 allocs/op
BenchmarkDecodeToRGBAStdLib-8 740 3219690 ns/op 1455735 B/op 7 allocs/op
Encoding, compared to the standard library:
BenchmarkEncodeRGBA420-8 1651 1407864 ns/op 186.20 MB/s 32834 B/op 2 allocs/op
BenchmarkEncodeRGBA420Stdlib-8 456 5381613 ns/op 48.71 MB/s 66032 B/op 12 allocs/op
BenchmarkEncodeYCbCr420-8 2404 998885 ns/op 262.44 MB/s 33062 B/op 2 allocs/op
BenchmarkEncodeYCbCr420Stdlib-8 517 4809590 ns/op 54.50 MB/s 66032 B/op 12 allocs/op
BenchmarkEncodeGray-8 2581 931858 ns/op 281.31 MB/s 33160 B/op 2 allocs/op
BenchmarkEncodeGrayStdlib-8 750 3108625 ns/op 84.33 MB/s 65936 B/op 10 allocs/op
The standard library always writes 4:2:0 with the standard Huffman tables, so these have no counterpart:
BenchmarkEncodeRGBA444-8 1480 1624043 ns/op 161.41 MB/s 41675 B/op 2 allocs/op
BenchmarkEncodeRGBA422-8 1438 1658117 ns/op 158.10 MB/s 41514 B/op 2 allocs/op
BenchmarkEncodeRGBA420Optimize-8 1134 2226588 ns/op 117.73 MB/s 33331 B/op 2 allocs/op
File size is within 0.2% of the standard library at the same quality and subsampling, and identical
to libjpeg-turbo at quality 50. OptimizeCoding reduces it by 4% at quality 50, rising to 20% at
quality 95.
Difference with assembly optimizations (noasm vs asm):
benchmark old ns/op new ns/op delta
BenchmarkDecodeBaseline420-8 1703998 898566 -47.27%
BenchmarkDecodeProgressive420-8 2970034 2335332 -21.37%
BenchmarkDecodeToRGBANearestNeighbor-8 2969821 1277749 -56.98%
BenchmarkDecodeToRGBACatmullRom-8 4189137 1423913 -66.01%
BenchmarkIdct-8 161.3 42.81 -73.46%
BenchmarkUpsampleNearestNeighbor-8 352034 142694 -59.47%
BenchmarkUpsampleCatmullRom-8 2957886 397606 -86.56%
BenchmarkEncodeRGBA420-8 3987337 1542867 -61.31%
BenchmarkEncodeRGBA444-8 5651131 1699801 -69.92%
BenchmarkEncodeYCbCr420-8 2447185 1065880 -56.44%
BenchmarkEncodeGray-8 1881926 985907 -47.61%
BenchmarkFDCT-8 113.4 24.62 -78.29%
BenchmarkRGBToYCbCrRow-8 4841 367 -92.41%
BenchmarkQuantizeBlock-8 173.2 24.83 -85.66%
noasm- do not use SIMD optimizations
