Official Go PR: golang/go#71618
This repository demonstrates a potential optimization of decoding JPEG images
in the Go standard library. This is achieved by unrolling
unzig and shift-clamp loops in scan.go.
On average across benchmarks (Go 1.23.6), this is ~13% faster on AMD64 and ~14%/11% faster on Apple Silicon ARM64 (M1/M4) at the cost of slightly larger binary size (~16KiB).
patch:
--- /usr/lib/go/src/image/jpeg/reader.go 2025-02-05 01:18:32.000000000 +0100
+++ ./optimized/jpeg/reader.go 2025-02-07 18:33:34.900861895 +0100
@@ -75,7 +76,7 @@
// unzig maps from the zig-zag ordering to the natural ordering. For example,
// unzig[3] is the column and row of the fourth element in zig-zag order. The
// value is 16, which means first column (16%8 == 0) and third row (16/8 == 2).
-var unzig = [blockSize]int{
+var unzig = [blockSize]uint8{
0, 1, 8, 16, 9, 2, 3, 10,
17, 24, 32, 25, 18, 11, 4, 5,
12, 19, 26, 33, 40, 48, 41, 34,
diff -ruN /usr/lib/go/src/image/jpeg/scan.go ./optimized/jpeg/scan.go
--- /usr/lib/go/src/image/jpeg/scan.go 2025-02-05 01:18:32.000000000 +0100
+++ ./optimized/jpeg/scan.go 2025-02-07 22:39:23.342471482 +0100
@@ -465,9 +465,73 @@
// to the image.
func (d *decoder) reconstructBlock(b *block, bx, by, compIndex int) error {
qt := &d.quant[d.comp[compIndex].tq]
- for zig := 0; zig < blockSize; zig++ {
- b[unzig[zig]] *= qt[zig]
- }
+
+ // This sequence exactly follows the indexes of the unzig mapping.
+ b[0] *= qt[0]
+ b[1] *= qt[1]
+ b[8] *= qt[2]
+ b[16] *= qt[3]
+ b[9] *= qt[4]
+ b[2] *= qt[5]
+ b[3] *= qt[6]
+ b[10] *= qt[7]
+ b[17] *= qt[8]
+ b[24] *= qt[9]
+ b[32] *= qt[10]
+ b[25] *= qt[11]
+ b[18] *= qt[12]
+ b[11] *= qt[13]
+ b[4] *= qt[14]
+ b[5] *= qt[15]
+ b[12] *= qt[16]
+ b[19] *= qt[17]
+ b[26] *= qt[18]
+ b[33] *= qt[19]
+ b[40] *= qt[20]
+ b[48] *= qt[21]
+ b[41] *= qt[22]
+ b[34] *= qt[23]
+ b[27] *= qt[24]
+ b[20] *= qt[25]
+ b[13] *= qt[26]
+ b[6] *= qt[27]
+ b[7] *= qt[28]
+ b[14] *= qt[29]
+ b[21] *= qt[30]
+ b[28] *= qt[31]
+ b[35] *= qt[32]
+ b[42] *= qt[33]
+ b[49] *= qt[34]
+ b[56] *= qt[35]
+ b[57] *= qt[36]
+ b[50] *= qt[37]
+ b[43] *= qt[38]
+ b[36] *= qt[39]
+ b[29] *= qt[40]
+ b[22] *= qt[41]
+ b[15] *= qt[42]
+ b[23] *= qt[43]
+ b[30] *= qt[44]
+ b[37] *= qt[45]
+ b[44] *= qt[46]
+ b[51] *= qt[47]
+ b[58] *= qt[48]
+ b[59] *= qt[49]
+ b[52] *= qt[50]
+ b[45] *= qt[51]
+ b[38] *= qt[52]
+ b[31] *= qt[53]
+ b[39] *= qt[54]
+ b[46] *= qt[55]
+ b[53] *= qt[56]
+ b[60] *= qt[57]
+ b[61] *= qt[58]
+ b[54] *= qt[59]
+ b[47] *= qt[60]
+ b[55] *= qt[61]
+ b[62] *= qt[62]
+ b[63] *= qt[63]
+
idct(b)
dst, stride := []byte(nil), 0
if d.nComp == 1 {
@@ -486,22 +550,82 @@
return UnsupportedError("too many components")
}
}
+
// Level shift by +128, clip to [0, 255], and write to dst.
- for y := 0; y < 8; y++ {
- y8 := y * 8
- yStride := y * stride
- for x := 0; x < 8; x++ {
- c := b[y8+x]
- if c < -128 {
- c = 0
- } else if c > 127 {
- c = 255
- } else {
- c += 128
- }
- dst[yStride+x] = uint8(c)
+ writeDst := func(index int) {
+ c := (*b)[index] + 128
+ if c < 0 {
+ c = 0
+ } else if c > 255 {
+ c = 255
}
+ dst[(index/8)*stride+(index%8)] = uint8(c)
}
+ writeDst(0)
+ writeDst(1)
+ writeDst(2)
+ writeDst(3)
+ writeDst(4)
+ writeDst(5)
+ writeDst(6)
+ writeDst(7)
+ writeDst(8)
+ writeDst(9)
+ writeDst(10)
+ writeDst(11)
+ writeDst(12)
+ writeDst(13)
+ writeDst(14)
+ writeDst(15)
+ writeDst(16)
+ writeDst(17)
+ writeDst(18)
+ writeDst(19)
+ writeDst(20)
+ writeDst(21)
+ writeDst(22)
+ writeDst(23)
+ writeDst(24)
+ writeDst(25)
+ writeDst(26)
+ writeDst(27)
+ writeDst(28)
+ writeDst(29)
+ writeDst(30)
+ writeDst(31)
+ writeDst(32)
+ writeDst(33)
+ writeDst(34)
+ writeDst(35)
+ writeDst(36)
+ writeDst(37)
+ writeDst(38)
+ writeDst(39)
+ writeDst(40)
+ writeDst(41)
+ writeDst(42)
+ writeDst(43)
+ writeDst(44)
+ writeDst(45)
+ writeDst(46)
+ writeDst(47)
+ writeDst(48)
+ writeDst(49)
+ writeDst(50)
+ writeDst(51)
+ writeDst(52)
+ writeDst(53)
+ writeDst(54)
+ writeDst(55)
+ writeDst(56)
+ writeDst(57)
+ writeDst(58)
+ writeDst(59)
+ writeDst(60)
+ writeDst(61)
+ writeDst(62)
+ writeDst(63)
+
return nil
}
diff -ruN /usr/lib/go/src/image/jpeg/writer_test.go ./optimized/jpeg/writer_test.go
--- /usr/lib/go/src/image/jpeg/writer_test.go 2025-02-05 01:18:32.000000000 +0100
+++ ./optimized/jpeg/writer_test.go 2025-02-07 20:16:25.281699218 +0100
@@ -20,7 +20,7 @@
// zigzag maps from the natural ordering to the zig-zag ordering. For example,
// zigzag[0*8 + 3] is the zig-zag sequence number of the element in the fourth
// column and first row.
-var zigzag = [blockSize]int{
+var zigzag = [blockSize]uint8{
0, 1, 5, 6, 14, 15, 27, 28,
2, 4, 7, 13, 16, 26, 29, 42,
3, 8, 12, 17, 25, 30, 41, 43,
@@ -32,7 +32,7 @@
}
func TestZigUnzig(t *testing.T) {
- for i := 0; i < blockSize; i++ {
+ for i := range uint8(blockSize) {
if unzig[zigzag[i]] != i {
t.Errorf("unzig[zigzag[%d]] == %d", i, unzig[zigzag[i]])
}goos: linux
goarch: amd64
pkg: github.com/romshark/jpegbench
cpu: AMD Ryzen 7 5700X 8-Core Processor
│ std.txt │ opt.txt │
│ sec/op │ sec/op vs base │
Decode/11375x8992_6mb.jpg-16 598.9m ± 0% 501.1m ± 0% -16.32% (p=0.000 n=12)
Decode/1280x719_84kb.jpg-16 8.092m ± 0% 6.549m ± 0% -19.07% (p=0.000 n=12)
Decode/15400x6940_20mb.jpg-16 1.950 ± 0% 1.626 ± 0% -16.60% (p=0.000 n=12)
Decode/1920x1193_600kb.jpg-16 33.36m ± 0% 29.45m ± 0% -11.74% (p=0.000 n=12)
Decode/32x32_.jpg-16 22.41µ ± 0% 21.31µ ± 0% -4.92% (p=0.000 n=12)
Decode/6000x4000_2mb.jpg-16 358.0m ± 0% 316.3m ± 0% -11.64% (p=0.000 n=12)
Decode/600x239_35kb.jpg-16 2.709m ± 0% 2.330m ± 0% -13.98% (p=0.000 n=12)
Decode/9319x5792_6480kb.jpg-16 939.6m ± 0% 844.3m ± 1% -10.15% (p=0.000 n=12)
geomean 39.91m 34.66m -13.15%
│ std.txt │ opt.txt │
│ B/op │ B/op vs base │
Decode/11375x8992_6mb.jpg-16 97.58Mi ± 0% 97.58Mi ± 0% ~ (p=0.748 n=12)
Decode/1280x719_84kb.jpg-16 1.335Mi ± 0% 1.335Mi ± 0% ~ (p=0.572 n=12)
Decode/15400x6940_20mb.jpg-16 306.0Mi ± 0% 306.0Mi ± 0% ~ (p=1.000 n=12) ¹
Decode/1920x1193_600kb.jpg-16 3.310Mi ± 0% 3.310Mi ± 0% 0.00% (p=0.040 n=12)
Decode/32x32_.jpg-16 15.02Ki ± 0% 15.02Ki ± 0% ~ (p=1.000 n=12) ¹
Decode/6000x4000_2mb.jpg-16 171.7Mi ± 0% 171.7Mi ± 0% ~ (p=1.000 n=12)
Decode/600x239_35kb.jpg-16 437.5Ki ± 0% 437.5Ki ± 0% ~ (p=1.000 n=12)
Decode/9319x5792_6480kb.jpg-16 386.5Mi ± 0% 386.5Mi ± 0% ~ (p=1.000 n=12)
geomean 9.276Mi 9.276Mi +0.00%
¹ all samples are equal
│ std.txt │ opt.txt │
│ allocs/op │ allocs/op vs base │
Decode/11375x8992_6mb.jpg-16 626.0 ± 0% 626.0 ± 0% ~ (p=1.000 n=12) ¹
Decode/1280x719_84kb.jpg-16 71.00 ± 0% 71.00 ± 0% ~ (p=1.000 n=12) ¹
Decode/15400x6940_20mb.jpg-16 1.248k ± 0% 1.248k ± 0% ~ (p=1.000 n=12) ¹
Decode/1920x1193_600kb.jpg-16 8.000 ± 0% 8.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/32x32_.jpg-16 5.000 ± 0% 5.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/6000x4000_2mb.jpg-16 11.00 ± 0% 11.00 ± 0% ~ (p=1.000 n=12) ¹
Decode/600x239_35kb.jpg-16 6.000 ± 0% 6.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/9319x5792_6480kb.jpg-16 12.00 ± 0% 12.00 ± 0% ~ (p=1.000 n=12)
geomean 33.93 33.93 +0.00%
¹ all samples are equal
Binary sizes:
optimized: 1967278 (1.9M)
standard: 1951199 (1.9M)
diff: 16079 (16KiB)
goos: darwin
goarch: amd64
pkg: github.com/romshark/jpegbench
cpu: Intel(R) Core(TM) i7-4850HQ CPU @ 2.30GHz
│ std.txt │ opt.txt │
│ sec/op │ sec/op vs base │
Decode/11375x8992_6mb.jpg-8 765.7m ± 1% 673.4m ± 0% -12.05% (p=0.000 n=12)
Decode/1280x719_84kb.jpg-8 10.232m ± 2% 8.805m ± 1% -13.95% (p=0.000 n=12)
Decode/15400x6940_20mb.jpg-8 2.456 ± 0% 2.138 ± 0% -12.97% (p=0.000 n=12)
Decode/1920x1193_600kb.jpg-8 41.90m ± 1% 38.61m ± 1% -7.87% (p=0.000 n=12)
Decode/32x32_.jpg-8 32.06µ ± 0% 30.96µ ± 2% -3.45% (p=0.000 n=12)
Decode/6000x4000_2mb.jpg-8 481.9m ± 1% 442.9m ± 1% -8.10% (p=0.000 n=12)
Decode/600x239_35kb.jpg-8 3.442m ± 0% 3.020m ± 0% -12.24% (p=0.000 n=12)
Decode/9319x5792_6480kb.jpg-8 1.254 ± 1% 1.184 ± 1% -5.60% (p=0.000 n=12)
geomean 52.04m 47.04m -9.60%
│ std.txt │ opt.txt │
│ B/op │ B/op vs base │
Decode/11375x8992_6mb.jpg-8 97.58Mi ± 0% 97.58Mi ± 0% ~ (p=0.886 n=12)
Decode/1280x719_84kb.jpg-8 1.335Mi ± 0% 1.335Mi ± 0% +0.00% (p=0.003 n=12)
Decode/15400x6940_20mb.jpg-8 306.0Mi ± 0% 306.0Mi ± 0% ~ (p=1.000 n=12)
Decode/1920x1193_600kb.jpg-8 3.310Mi ± 0% 3.310Mi ± 0% +0.00% (p=0.036 n=12)
Decode/32x32_.jpg-8 15.02Ki ± 0% 15.02Ki ± 0% ~ (p=1.000 n=12) ¹
Decode/6000x4000_2mb.jpg-8 171.7Mi ± 0% 171.7Mi ± 0% ~ (p=0.056 n=12)
Decode/600x239_35kb.jpg-8 437.5Ki ± 0% 437.5Ki ± 0% ~ (p=0.119 n=12)
Decode/9319x5792_6480kb.jpg-8 386.5Mi ± 0% 386.5Mi ± 0% ~ (p=0.479 n=12)
geomean 9.276Mi 9.276Mi +0.00%
¹ all samples are equal
│ std.txt │ opt.txt │
│ allocs/op │ allocs/op vs base │
Decode/11375x8992_6mb.jpg-8 626.0 ± 0% 626.0 ± 0% ~ (p=1.000 n=12)
Decode/1280x719_84kb.jpg-8 71.00 ± 0% 71.00 ± 0% ~ (p=1.000 n=12) ¹
Decode/15400x6940_20mb.jpg-8 1.248k ± 0% 1.248k ± 0% ~ (p=1.000 n=12)
Decode/1920x1193_600kb.jpg-8 8.000 ± 0% 8.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/32x32_.jpg-8 5.000 ± 0% 5.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/6000x4000_2mb.jpg-8 11.00 ± 0% 11.00 ± 0% ~ (p=1.000 n=12) ¹
Decode/600x239_35kb.jpg-8 6.000 ± 0% 6.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/9319x5792_6480kb.jpg-8 12.00 ± 8% 12.00 ± 0% ~ (p=0.479 n=12)
geomean 33.93 33.93 +0.00%
¹ all samples are equal
Binary sizes:
optimized: 1975670 (1.9M)
standard: 1959607 (1.9M)
diff: 16063 (16063B)
goos: darwin
goarch: arm64
pkg: github.com/romshark/jpegbench
cpu: Apple M1 Max
│ std.txt │ opt.txt │
│ sec/op │ sec/op vs base │
Decode/11375x8992_6mb.jpg-10 486.6m ± 0% 398.5m ± 1% -18.11% (p=0.000 n=12)
Decode/1280x719_84kb.jpg-10 6.512m ± 0% 5.221m ± 0% -19.82% (p=0.000 n=12)
Decode/15400x6940_20mb.jpg-10 1.566 ± 0% 1.277 ± 0% -18.48% (p=0.000 n=12)
Decode/1920x1193_600kb.jpg-10 29.26m ± 0% 26.04m ± 0% -11.01% (p=0.000 n=12)
Decode/32x32_.jpg-10 18.19µ ± 0% 16.69µ ± 0% -8.27% (p=0.000 n=12)
Decode/6000x4000_2mb.jpg-10 279.8m ± 0% 252.1m ± 0% -9.89% (p=0.000 n=12)
Decode/600x239_35kb.jpg-10 2.187m ± 0% 1.800m ± 0% -17.71% (p=0.000 n=12)
Decode/9319x5792_6480kb.jpg-10 750.2m ± 0% 687.8m ± 0% -8.32% (p=0.000 n=12)
geomean 32.39m 27.83m -14.08%
│ std.txt │ opt.txt │
│ B/op │ B/op vs base │
Decode/11375x8992_6mb.jpg-10 97.58Mi ± 0% 97.58Mi ± 0% ~ (p=1.000 n=12)
Decode/1280x719_84kb.jpg-10 1.335Mi ± 0% 1.335Mi ± 0% ~ (p=0.261 n=12)
Decode/15400x6940_20mb.jpg-10 306.0Mi ± 0% 306.0Mi ± 0% ~ (p=1.000 n=12)
Decode/1920x1193_600kb.jpg-10 3.310Mi ± 0% 3.310Mi ± 0% +0.00% (p=0.002 n=12)
Decode/32x32_.jpg-10 15.02Ki ± 0% 15.02Ki ± 0% ~ (p=1.000 n=12) ¹
Decode/6000x4000_2mb.jpg-10 171.7Mi ± 0% 171.7Mi ± 0% ~ (p=1.000 n=12)
Decode/600x239_35kb.jpg-10 437.5Ki ± 0% 437.5Ki ± 0% +0.00% (p=0.026 n=12)
Decode/9319x5792_6480kb.jpg-10 386.5Mi ± 0% 386.5Mi ± 0% ~ (p=1.000 n=12)
geomean 9.276Mi 9.276Mi +0.00%
¹ all samples are equal
│ std.txt │ opt.txt │
│ allocs/op │ allocs/op vs base │
Decode/11375x8992_6mb.jpg-10 626.0 ± 0% 626.0 ± 0% ~ (p=1.000 n=12) ¹
Decode/1280x719_84kb.jpg-10 71.00 ± 0% 71.00 ± 0% ~ (p=1.000 n=12) ¹
Decode/15400x6940_20mb.jpg-10 1.248k ± 0% 1.248k ± 0% ~ (p=1.000 n=12)
Decode/1920x1193_600kb.jpg-10 8.000 ± 0% 8.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/32x32_.jpg-10 5.000 ± 0% 5.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/6000x4000_2mb.jpg-10 11.00 ± 0% 11.00 ± 0% ~ (p=1.000 n=12) ¹
Decode/600x239_35kb.jpg-10 6.000 ± 0% 6.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/9319x5792_6480kb.jpg-10 12.00 ± 0% 12.00 ± 0% ~ (p=1.000 n=12)
geomean 33.93 33.93 +0.00%
¹ all samples are equal
Binary sizes:
optimized: 1975694 (1.9M)
standard: 1959599 (1.9M)
diff: 16095 (16095B)
goos: darwin
goarch: arm64
pkg: github.com/romshark/jpegbench
cpu: Apple M4 Pro
│ std.txt │ opt.txt │
│ sec/op │ sec/op vs base │
Decode/11375x8992_6mb.jpg-14 321.0m ± 1% 277.6m ± 0% -13.52% (p=0.000 n=12)
Decode/1280x719_84kb.jpg-14 4.169m ± 0% 3.394m ± 0% -18.60% (p=0.000 n=12)
Decode/15400x6940_20mb.jpg-14 1064.5m ± 0% 900.9m ± 0% -15.37% (p=0.000 n=12)
Decode/1920x1193_600kb.jpg-14 20.94m ± 0% 19.44m ± 0% -7.19% (p=0.000 n=12)
Decode/32x32_.jpg-14 13.91µ ± 0% 13.16µ ± 0% -5.40% (p=0.000 n=12)
Decode/6000x4000_2mb.jpg-14 183.7m ± 1% 169.4m ± 1% -7.81% (p=0.000 n=12)
Decode/600x239_35kb.jpg-14 1.417m ± 0% 1.158m ± 1% -18.30% (p=0.000 n=12)
Decode/9319x5792_6480kb.jpg-14 504.3m ± 1% 475.6m ± 0% -5.70% (p=0.000 n=12)
geomean 21.98m 19.42m -11.64%
│ std.txt │ opt.txt │
│ B/op │ B/op vs base │
Decode/11375x8992_6mb.jpg-14 97.58Mi ± 0% 97.58Mi ± 0% ~ (p=0.120 n=12)
Decode/1280x719_84kb.jpg-14 1.335Mi ± 0% 1.335Mi ± 0% -0.00% (p=0.006 n=12)
Decode/15400x6940_20mb.jpg-14 306.0Mi ± 0% 306.0Mi ± 0% ~ (p=1.000 n=12)
Decode/1920x1193_600kb.jpg-14 3.310Mi ± 0% 3.310Mi ± 0% ~ (p=0.919 n=12)
Decode/32x32_.jpg-14 15.02Ki ± 0% 15.02Ki ± 0% ~ (p=1.000 n=12) ¹
Decode/6000x4000_2mb.jpg-14 171.7Mi ± 0% 171.7Mi ± 0% ~ (p=0.733 n=12)
Decode/600x239_35kb.jpg-14 437.5Ki ± 0% 437.5Ki ± 0% ~ (p=0.217 n=12)
Decode/9319x5792_6480kb.jpg-14 386.5Mi ± 0% 386.5Mi ± 0% ~ (p=0.590 n=12)
geomean 9.276Mi 9.276Mi -0.00%
¹ all samples are equal
│ std.txt │ opt.txt │
│ allocs/op │ allocs/op vs base │
Decode/11375x8992_6mb.jpg-14 626.0 ± 0% 626.0 ± 0% ~ (p=1.000 n=12) ¹
Decode/1280x719_84kb.jpg-14 71.00 ± 0% 71.00 ± 0% ~ (p=1.000 n=12) ¹
Decode/15400x6940_20mb.jpg-14 1.248k ± 0% 1.248k ± 0% ~ (p=1.000 n=12) ¹
Decode/1920x1193_600kb.jpg-14 8.000 ± 0% 8.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/32x32_.jpg-14 5.000 ± 0% 5.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/6000x4000_2mb.jpg-14 11.00 ± 0% 11.00 ± 0% ~ (p=1.000 n=12) ¹
Decode/600x239_35kb.jpg-14 6.000 ± 0% 6.000 ± 0% ~ (p=1.000 n=12) ¹
Decode/9319x5792_6480kb.jpg-14 12.00 ± 0% 12.00 ± 0% ~ (p=1.000 n=12) ¹
geomean 33.93 33.93 +0.00%
¹ all samples are equal
Binary sizes:
optimized: 1975662 (1.9M)
standard: 1959591 (1.9M)
diff: 16071 (16071B)
go test -v ./..../bench.sh 12 . && ./cmpbinsz.sh