...
https://caslab.csl.yale.edu/workshops/hasp2016/HASP16-10_slides.pdf
论文AVX512相关代码
https://github.com/rbCabral/SHA-3
用了2个优化方式
1)基于avx2指令的seq sha3优化，每个msg计算hash使用avx2来加速
2) 基于par计算msg hash,多条消息放入msg数组，基于avx2并行运算hash数组
目前cloudflare circlhttps://github.com/cloudflare/circl/blob/master/simd/keccakf1600/f1600x4_amd64.go
使用的是2)par计算
2.使用库对比
目前我们的accumulator使用sha3计算hash使用的tiny-keccak库，
...
这是2年前情况(tiny-keccack一直没有迭代)， 我在跳板机上测试结论使用最新的sha3库，
修改tiny-keccak下comparison/Cargo.toml文件
| 代码块 |
|---|
% git diff comparison/Cargo.toml
diff --git a/comparison/Cargo.toml b/comparison/Cargo.toml
index 1b88b6a..b0dd061 100644
--- a/comparison/Cargo.toml
+++ b/comparison/Cargo.toml
@@ -6,4 +6,4 @@ authors = ["debris <marek.kotewicz@gmail.com>"]
[dependencies]
tiny-keccak = { path = "../", features = ["sha3"] }
-sha3 = "0.8.2"
+sha3 = "0.10.0"
|
| 代码块 |
|---|
+++ b/comparison/benches/sha3.rs
@@ -40,8 +40,8 @@ fn rust_crypto_sha3_256_input_32_bytes(b: &mut Bencher) {
b.iter(|| {
let mut sha3 = Sha3_256::default();
- sha3.input(&data);
- sha3.result();
+ sha3.update(&data);
+ sha3.finalize();
});
}
@@ -53,7 +53,7 @@ fn rust_crypto_sha3_256_input_4096_bytes(b: &mut Bencher) {
b.iter(|| {
let mut sha3 = Sha3_256::default();
- sha3.input(&data);
- sha3.result();
+ sha3.update(&data);
+ sha3.finalize();
});
}
|
| 代码块 |
|---|
Intel(R) Xeon(R) CPU E5-2666 v3 @ 2.90GHz 7G Memory |
我在跳板机上测试结论
| 代码块 |
|---|
rustup run nightly cargo bench
running 4 tests
test rust_crypto_sha3_256_input_32_bytes ... bench: 599 ns/iter (+/- 12) = 53 MB/s
test rust_crypto_sha3_256_input_4096_bytes ... bench: 17,665 ns/iter (+/- 96) = 231 MB/s
test tiny_keccak_sha3_256_input_32_bytes ... bench: 595 ns/iter (+/- 5) = 53 MB/s
test tiny_keccak_sha3_256_input_4096_bytes ... bench: 17,737 ns/iter (+/- 158) = 230 MB/s
|
在macOs上
| 代码块 |
|---|
running 4 tests test rust_crypto_sha3_256_input_32_bytes ... bench: 441 ns/iter (+/- 47) = 72 MB/s test rust_crypto_sha3_256_input_4096_bytes ... bench: 12,875 ns/iter (+/- 775) = 318 MB/s test tiny_keccak_sha3_256_input_32_bytes ... bench: 452 ns/iter (+/- 23) = 70 MB/s test tiny_keccak_sha3_256_input_4096_bytes ... bench: 13,073 ns/iter (+/- 880) = 313 MB/s |
目前看起来crypto hash效率慢慢赶上来了，可能会超过tiny-kecccak,前者一直在开发维护，后者两年都没有更新了
目前有一份从XKCP的C实现 ffi 给rust调用的实现
https://github.com/starcoinorg/starcoin-crypto/tree/main/crates/diem-crypto/ext