Skip to content

Commit c5c16f4

Browse files
twmbtwmb
committed
README: benchmark table by Go release, numbers only
1 parent fe7a6f3 commit c5c16f4

1 file changed

Lines changed: 25 additions & 32 deletions

File tree

‎README.md‎

Lines changed: 25 additions & 32 deletions
Original file line numberDiff line numberDiff line change
@@ -60,39 +60,32 @@ Benchmarks
6060
==========
6161

6262
Cycles per call from perf (`cycles:u` at a fixed iteration count, minimum of
63-
five runs) on Go 1.27.1, Comet Lake i7-10710U, so CPU frequency and thermal
64-
throttling drop out. `asm` is the amd64 assembly this library used to ship,
65-
`old Go` the pure Go path other architectures ran, `new Go` the current code.
66-
`Sizes` rows exercise the block loop, `Branches` rows the 0 to 16 byte tail,
63+
three runs) on a Comet Lake i7-10710U, so CPU frequency and thermal
64+
throttling drop out. `asm` is the amd64 assembly this library shipped through
65+
v1.1.8; the other columns are this code built by that Go release. `Sizes`
66+
rows exercise the block loop, `Branches` rows the 0 to 16 byte tail, and
6767
`RandomLengths` rows draw an xorshift length in [0, 64) on every call.
6868

6969
```
70-
benchmark asm old Go new Go new/asm new/old
71-
128Sizes/8192 4217 4479 4364 +3.5% -2.6%
72-
128Sizes/1024 534 576 557 +4.3% -3.4%
73-
128Sizes/256 140 159 149 +6.8% -6.0%
74-
128Sizes/64 48 52 49 +1.5% -7.1%
75-
128Sizes/32 35 36 34 -2.1% -3.9%
76-
128Branches/16 26 23 26 -1.4% +10.8%
77-
128Branches/13 25 34 21 -14.5% -37.1%
78-
128Branches/7 26 26 21 -18.5% -19.7%
79-
128Branches/3 25 22 21 -16.4% -3.1%
80-
64Sizes/8192 4216 4533 4362 +3.5% -3.8%
81-
64Sizes/32 35 38 34 -3.5% -10.7%
82-
32Sizes/8192 10216 10191 9585 -6.2% -5.9%
83-
32Sizes/64 89 88 76 -13.6% -13.6%
84-
32Sizes/32 50 49 41 -18.1% -18.0%
85-
32Branches/3 15 15 16 +5.6% +5.8%
86-
RandomLengths128 110 96 90 -18.1% -5.8%
87-
RandomLengths32 95 96 84 -11.6% -12.0%
70+
benchmark asm go1.25 go1.26 go1.27 1.26/asm 1.27/asm
71+
128Branches/3 25 25 21 21 -16.7% -16.7%
72+
128Branches/7 26 26 21 21 -18.6% -18.5%
73+
128Branches/13 25 27 22 22 -12.5% -14.5%
74+
128Branches/16 27 35 26 26 -1.4% -1.3%
75+
128Sizes/32 35 46 34 34 -2.3% -2.2%
76+
128Sizes/64 48 70 48 49 +1.2% +1.2%
77+
128Sizes/256 140 209 149 149 +6.9% +6.8%
78+
128Sizes/1024 534 769 567 557 +6.2% +4.3%
79+
128Sizes/8192 4215 6212 4361 4363 +3.5% +3.5%
80+
64Sizes/32 35 46 34 34 -3.7% -3.5%
81+
64Sizes/1024 535 788 575 557 +7.4% +4.1%
82+
64Sizes/8192 4221 6326 4360 4366 +3.3% +3.4%
83+
32Branches/3 15 19 16 16 +5.2% +5.6%
84+
32Sizes/32 50 56 41 41 -18.2% -18.0%
85+
32Sizes/64 89 97 77 77 -13.6% -13.6%
86+
32Sizes/256 322 339 296 302 -7.9% -6.2%
87+
32Sizes/1024 1284 1335 1243 1211 -3.2% -5.7%
88+
32Sizes/8192 10080 10405 9896 9582 -1.8% -4.9%
89+
RandomLengths128 110 88 91 90 -18.0% -18.2%
90+
RandomLengths32 96 88 83 85 -13.5% -11.4%
8891
```
89-
90-
Per 16 byte block the assembly is 8.2 cycles and the Go loop 8.5, on both Go
91-
1.26 and 1.27. The loop indexes data[i:i+16] with a hoisted bound rather than
92-
reslicing, which is three fewer instructions per block. Holding the two
93-
multipliers in registers instead of constants measured 8.1 on Go 1.27 and 9.3
94-
on Go 1.26, the difference being one register name and its REX prefix, so
95-
they stay constants. The tail is read out of the input's last 16 bytes with
96-
two overlapping loads rather than a switch on its length, which is a jump
97-
table and mispredicts on varying key lengths. The 32 bit sum sits on its four
98-
cycle per block dependency chain.

0 commit comments

Comments
 (0)