mirror of
https://github.com/zeebo/xxh3.git
synced 2026-10-25 07:10:23 +00:00
XXH3 algorithm in Go
- Go 90.8%
- Assembly 9%
- C 0.1%
- Makefile 0.1%
|
History
|
||
|---|---|---|
| .github/workflows | ||
| avo | ||
| internal | ||
| .gitignore | ||
| _compat.c | ||
| accum_generic.go | ||
| accum_stubs_amd64.go | ||
| accum_stubs_arm64.go | ||
| accum_stubs_other.go | ||
| accum_vector_avx512_amd64.s | ||
| accum_vector_avx_amd64.s | ||
| accum_vector_neon_arm64.s | ||
| accum_vector_sse_amd64.s | ||
| compat_test.go | ||
| compat_vector_test.go | ||
| consts.go | ||
| escape_test.go | ||
| go.mod | ||
| go.sum | ||
| hash64.go | ||
| hash64_seed.go | ||
| hash64_test.go | ||
| hash128.go | ||
| hash128_seed.go | ||
| hash128_test.go | ||
| hasher.go | ||
| hasher128.go | ||
| hasher_seed_bench_test.go | ||
| hasher_seed_test.go | ||
| hasher_test.go | ||
| LICENSE | ||
| Makefile | ||
| README.md | ||
| utils.go | ||
XXH3
This package is a port of the xxh3 library to Go.
Upstream has fixed the output as of v0.8.0, and this package matches that.
Benchmarks
AMD Ryzen 9 9950X
Small Sizes
| Bytes | 64-bit | 64-bit Seed | 128-bit | 128-bit Seed |
|---|---|---|---|---|
0 |
1.16 ns |
1.72 ns |
1.31 ns |
2.13 ns |
1-3 |
1.58 ns (0.63-1.90 GB/s) |
1.79 ns (0.56-1.68 GB/s) |
1.92 ns (0.52-1.56 GB/s) |
2.32 ns (0.43-1.29 GB/s) |
4-8 |
1.56 ns (2.56-5.13 GB/s) |
1.80 ns (2.22-4.44 GB/s) |
1.81 ns (2.21-4.42 GB/s) |
2.15 ns (1.86-3.72 GB/s) |
9-16 |
1.38 ns (6.52-11.6 GB/s) |
1.71 ns (5.26-9.36 GB/s) |
2.25 ns (4.00-7.11 GB/s) |
2.50 ns (3.60-6.40 GB/s) |
17-32 |
1.72 ns (9.88-18.6 GB/s) |
2.07 ns (8.21-15.5 GB/s) |
2.69 ns (6.32-11.9 GB/s) |
3.05 ns (5.57-10.5 GB/s) |
33-64 |
2.39 ns (13.8-26.8 GB/s) |
2.76 ns (12.0-23.2 GB/s) |
3.47 ns (9.51-18.4 GB/s) |
3.87 ns (8.53-16.5 GB/s) |
65-96 |
3.16 ns (20.6-30.4 GB/s) |
3.63 ns (17.9-26.4 GB/s) |
4.46 ns (14.6-21.5 GB/s) |
4.95 ns (13.1-19.4 GB/s) |
97-128 |
3.89 ns (24.9-32.9 GB/s) |
4.41 ns (22.0-29.0 GB/s) |
5.22 ns (18.6-24.5 GB/s) |
5.75 ns (16.9-22.3 GB/s) |
Large Sizes
64-bit default is shown. The 128-bit variant adds only a small finalization (within a few % here). A non-zero seed re-derives the 192-byte secret on every call, so seeded hashing is markedly slower at these sizes (e.g. 512 B SSE2: 20.3 vs 11.2 ns/op) and only amortizes to <1% by ~100KB.
| Bytes | Rate | SSE2 Rate | AVX2 Rate | AVX512 Rate |
|---|---|---|---|---|
129 |
4.52 ns/op (28.5 GB/s) |
|||
240 |
7.39 ns/op (32.5 GB/s) |
|||
241 |
13.6 ns/op (17.7 GB/s) |
7.95 ns/op (30.3 GB/s) |
11.3 ns/op (21.4 GB/s) |
|
512 |
23.1 ns/op (22.2 GB/s) |
11.2 ns/op (45.7 GB/s) |
14.1 ns/op (36.4 GB/s) |
|
1024 |
42.1 ns/op (24.3 GB/s) |
20.7 ns/op (49.5 GB/s) |
20.9 ns/op (49.0 GB/s) |
23.9 ns/op (42.9 GB/s) |
100KB |
4.03 us/op (25.4 GB/s) |
2.10 us/op (48.7 GB/s) |
1.07 us/op (96.0 GB/s) |
685 ns/op (149 GB/s) |
Intel i7-8850H @ 2.60GHz
Small Sizes
| Bytes | Rate |
|---|---|
0 |
0.74 ns/op |
1-3 |
4.19 ns/op (0.24 GB/s - 0.71 GB/s) |
4-8 |
4.16 ns/op (0.97 GB/s - 1.98 GB/s) |
9-16 |
4.46 ns/op (2.02 GB/s - 3.58 GB/s) |
17-32 |
6.22 ns/op (2.76 GB/s - 5.15 GB/s) |
33-64 |
8.00 ns/op (4.13 GB/s - 8.13 GB/s) |
65-96 |
11.0 ns/op (5.91 GB/s - 8.84 GB/s) |
97-128 |
12.8 ns/op (7.68 GB/s - 10.0 GB/s) |
Large Sizes
| Bytes | Rate | SSE2 Rate | AVX2 Rate |
|---|---|---|---|
129 |
13.6 ns/op (9.45 GB/s) |
||
240 |
23.8 ns/op (10.1 GB/s) |
||
241 |
40.5 ns/op (5.97 GB/s) |
23.3 ns/op (10.4 GB/s) |
20.1 ns/op (12.0 GB/s) |
512 |
69.8 ns/op (7.34 GB/s) |
30.4 ns/op (16.9 GB/s) |
24.7 ns/op (20.7 GB/s) |
1024 |
132 ns/op (7.77 GB/s) |
48.9 ns/op (20.9 GB/s) |
37.7 ns/op (27.2 GB/s) |
100KB |
13.0 us/op (7.88 GB/s) |
4.05 us/op (25.3 GB/s) |
2.31 us/op (44.3 GB/s) |