don't
perf(types): accelerate did plc validation
Use SIMD where available to validate PLC DIDs.
- Stable rust: AVX2 impl for x86-64 targets. Inspired by @metaflame.dev's
ratproto-did crate. <https://tangled.org/metaflame.dev/ratproto-did>.
- Nightly rust: `std::simd` based impl for all targets. Requires the
`nightly` feature to be enabled.
Improves validation from ~68 million DID/sec to ~475 million DID/sec
on an AMD 5950x.
Signed-off-by: tjh <x@tjh.dev>
+407
-1
+167
Cargo.lock
+167
Cargo.lock
···
30
30
]
31
31
32
32
[[package]]
33
+
name = "alloca"
34
+
version = "0.4.0"
35
+
source = "registry+https://github.com/rust-lang/crates.io-index"
36
+
checksum = "e5a7d05ea6aea7e9e64d25b9156ba2fee3fdd659e34e41063cd2fc7cd020d7f4"
37
+
dependencies = [
38
+
"cc",
39
+
]
40
+
41
+
[[package]]
33
42
name = "allocator-api2"
34
43
version = "0.2.21"
35
44
source = "registry+https://github.com/rust-lang/crates.io-index"
36
45
checksum = "683d7910e743518b0e34f1186f92494becacb047c7b6bf616c96772180fef923"
46
+
47
+
[[package]]
48
+
name = "anes"
49
+
version = "0.1.6"
50
+
source = "registry+https://github.com/rust-lang/crates.io-index"
51
+
checksum = "4b46cbb362ab8752921c97e041f5e366ee6297bd428a31275b9fcf1e380f7299"
37
52
38
53
[[package]]
39
54
name = "anstream"
···
364
349
checksum = "6bd91ee7b2422bcb158d90ef4d14f75ef67f340943fc4149891dcce8f8b972a3"
365
350
366
351
[[package]]
352
+
name = "cast"
353
+
version = "0.3.0"
354
+
source = "registry+https://github.com/rust-lang/crates.io-index"
355
+
checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
356
+
357
+
[[package]]
367
358
name = "cc"
368
359
version = "1.2.55"
369
360
source = "registry+https://github.com/rust-lang/crates.io-index"
···
398
377
version = "0.2.1"
399
378
source = "registry+https://github.com/rust-lang/crates.io-index"
400
379
checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
380
+
381
+
[[package]]
382
+
name = "ciborium"
383
+
version = "0.2.2"
384
+
source = "registry+https://github.com/rust-lang/crates.io-index"
385
+
checksum = "42e69ffd6f0917f5c029256a24d0161db17cea3997d185db0d35926308770f0e"
386
+
dependencies = [
387
+
"ciborium-io",
388
+
"ciborium-ll",
389
+
"serde",
390
+
]
391
+
392
+
[[package]]
393
+
name = "ciborium-io"
394
+
version = "0.2.2"
395
+
source = "registry+https://github.com/rust-lang/crates.io-index"
396
+
checksum = "05afea1e0a06c9be33d539b876f1ce3692f4afea2cb41f740e7743225ed1c757"
397
+
398
+
[[package]]
399
+
name = "ciborium-ll"
400
+
version = "0.2.2"
401
+
source = "registry+https://github.com/rust-lang/crates.io-index"
402
+
checksum = "57663b653d948a338bfb3eeba9bb2fd5fcfaecb9e199e87e1eda4d9e8b240fd9"
403
+
dependencies = [
404
+
"ciborium-io",
405
+
"half",
406
+
]
401
407
402
408
[[package]]
403
409
name = "cipher"
···
614
566
]
615
567
616
568
[[package]]
569
+
name = "criterion"
570
+
version = "0.8.2"
571
+
source = "registry+https://github.com/rust-lang/crates.io-index"
572
+
checksum = "950046b2aa2492f9a536f5f4f9a3de7b9e2476e575e05bd6c333371add4d98f3"
573
+
dependencies = [
574
+
"alloca",
575
+
"anes",
576
+
"cast",
577
+
"ciborium",
578
+
"clap",
579
+
"criterion-plot",
580
+
"itertools",
581
+
"num-traits",
582
+
"oorandom",
583
+
"page_size",
584
+
"plotters",
585
+
"rayon",
586
+
"regex",
587
+
"serde",
588
+
"serde_json",
589
+
"tinytemplate",
590
+
"walkdir",
591
+
]
592
+
593
+
[[package]]
594
+
name = "criterion-plot"
595
+
version = "0.8.2"
596
+
source = "registry+https://github.com/rust-lang/crates.io-index"
597
+
checksum = "d8d80a2f4f5b554395e47b5d8305bc3d27813bacb73493eb1001e8f76dae29ea"
598
+
dependencies = [
599
+
"cast",
600
+
"itertools",
601
+
]
602
+
603
+
[[package]]
617
604
name = "critical-section"
618
605
version = "1.2.0"
619
606
source = "registry+https://github.com/rust-lang/crates.io-index"
···
696
613
version = "0.8.21"
697
614
source = "registry+https://github.com/rust-lang/crates.io-index"
698
615
checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28"
616
+
617
+
[[package]]
618
+
name = "crunchy"
619
+
version = "0.2.4"
620
+
source = "registry+https://github.com/rust-lang/crates.io-index"
621
+
checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5"
699
622
700
623
[[package]]
701
624
name = "crypto-bigint"
···
2321
2232
name = "gordian-types"
2322
2233
version = "0.0.0"
2323
2234
dependencies = [
2235
+
"criterion",
2236
+
"data-encoding",
2237
+
"data-encoding-macro",
2238
+
"rand 0.9.2",
2324
2239
"serde",
2325
2240
"serde_json",
2326
2241
"smallstr",
···
2361
2268
"tokio",
2362
2269
"tokio-util",
2363
2270
"tracing",
2271
+
]
2272
+
2273
+
[[package]]
2274
+
name = "half"
2275
+
version = "2.7.1"
2276
+
source = "registry+https://github.com/rust-lang/crates.io-index"
2277
+
checksum = "6ea2d84b969582b4b1864a92dc5d27cd2b77b622a8d79306834f1be5ba20d84b"
2278
+
dependencies = [
2279
+
"cfg-if",
2280
+
"crunchy",
2281
+
"zerocopy",
2364
2282
]
2365
2283
2366
2284
[[package]]
···
2860
2756
checksum = "a6cb138bb79a146c1bd460005623e142ef0181e3d0219cb493e02f7d08a35695"
2861
2757
2862
2758
[[package]]
2759
+
name = "itertools"
2760
+
version = "0.13.0"
2761
+
source = "registry+https://github.com/rust-lang/crates.io-index"
2762
+
checksum = "413ee7dfc52ee1a4949ceeb7dbc8a33f2d6c088194d9f922fb8318faf1f01186"
2763
+
dependencies = [
2764
+
"either",
2765
+
]
2766
+
2767
+
[[package]]
2863
2768
name = "itoa"
2864
2769
version = "1.0.17"
2865
2770
source = "registry+https://github.com/rust-lang/crates.io-index"
···
3340
3227
checksum = "384b8ab6d37215f3c5301a95a4accb5d64aa607f1fcb26a11b5303878451b4fe"
3341
3228
3342
3229
[[package]]
3230
+
name = "oorandom"
3231
+
version = "11.1.5"
3232
+
source = "registry+https://github.com/rust-lang/crates.io-index"
3233
+
checksum = "d6790f58c7ff633d8771f42965289203411a5e5c68388703c06e14f24770b41e"
3234
+
3235
+
[[package]]
3343
3236
name = "openssl-probe"
3344
3237
version = "0.2.1"
3345
3238
source = "registry+https://github.com/rust-lang/crates.io-index"
···
3403
3284
"primeorder",
3404
3285
"rand_core 0.6.4",
3405
3286
"sha2",
3287
+
]
3288
+
3289
+
[[package]]
3290
+
name = "page_size"
3291
+
version = "0.6.0"
3292
+
source = "registry+https://github.com/rust-lang/crates.io-index"
3293
+
checksum = "30d5b2194ed13191c1999ae0704b7839fb18384fa22e49b57eeaa97d79ce40da"
3294
+
dependencies = [
3295
+
"libc",
3296
+
"winapi",
3406
3297
]
3407
3298
3408
3299
[[package]]
···
3497
3368
version = "0.3.32"
3498
3369
source = "registry+https://github.com/rust-lang/crates.io-index"
3499
3370
checksum = "7edddbd0b52d732b21ad9a5fab5c704c14cd949e5e9a1ec5929a24fded1b904c"
3371
+
3372
+
[[package]]
3373
+
name = "plotters"
3374
+
version = "0.3.7"
3375
+
source = "registry+https://github.com/rust-lang/crates.io-index"
3376
+
checksum = "5aeb6f403d7a4911efb1e33402027fc44f29b5bf6def3effcc22d7bb75f2b747"
3377
+
dependencies = [
3378
+
"num-traits",
3379
+
"plotters-backend",
3380
+
"plotters-svg",
3381
+
"wasm-bindgen",
3382
+
"web-sys",
3383
+
]
3384
+
3385
+
[[package]]
3386
+
name = "plotters-backend"
3387
+
version = "0.3.7"
3388
+
source = "registry+https://github.com/rust-lang/crates.io-index"
3389
+
checksum = "df42e13c12958a16b3f7f4386b9ab1f3e7933914ecea48da7139435263a4172a"
3390
+
3391
+
[[package]]
3392
+
name = "plotters-svg"
3393
+
version = "0.3.7"
3394
+
source = "registry+https://github.com/rust-lang/crates.io-index"
3395
+
checksum = "51bae2ac328883f7acdfea3d66a7c35751187f870bc81f94563733a154d7a670"
3396
+
dependencies = [
3397
+
"plotters-backend",
3398
+
]
3500
3399
3501
3400
[[package]]
3502
3401
name = "portable-atomic"
···
4854
4697
dependencies = [
4855
4698
"displaydoc",
4856
4699
"zerovec",
4700
+
]
4701
+
4702
+
[[package]]
4703
+
name = "tinytemplate"
4704
+
version = "1.2.1"
4705
+
source = "registry+https://github.com/rust-lang/crates.io-index"
4706
+
checksum = "be4d6b5f19ff7664e8c98d03e2139cb510db9b0a60b55f8e8709b689d939b6bc"
4707
+
dependencies = [
4708
+
"serde",
4709
+
"serde_json",
4857
4710
]
4858
4711
4859
4712
[[package]]
+11
crates/gordian-types/Cargo.toml
+11
crates/gordian-types/Cargo.toml
···
7
7
edition.workspace = true
8
8
publish.workspace = true
9
9
10
+
[features]
11
+
nightly = []
12
+
10
13
[dependencies]
11
14
smallstr = { version = "0.3.1" }
12
15
thiserror.workspace = true
···
19
16
time = { workspace = true, optional = true }
20
17
21
18
[dev-dependencies]
19
+
criterion = "0.8.2"
20
+
data-encoding.workspace = true
21
+
data-encoding-macro = "0.1.19"
22
+
rand.workspace = true
22
23
serde_json.workspace = true
24
+
25
+
[[bench]]
26
+
name = "parse_did"
27
+
harness = false
+56
crates/gordian-types/benches/parse_did.rs
+56
crates/gordian-types/benches/parse_did.rs
···
1
+
use std::iter;
2
+
3
+
use criterion::BatchSize;
4
+
use criterion::Criterion;
5
+
use criterion::Throughput;
6
+
use criterion::criterion_group;
7
+
use criterion::criterion_main;
8
+
use data_encoding::Encoding;
9
+
use gordian_types::Did;
10
+
use gordian_types::DidBuf;
11
+
use rand::Rng;
12
+
use rand::SeedableRng;
13
+
use rand::rngs::StdRng;
14
+
15
+
const BASE32: Encoding = data_encoding_macro::new_encoding! {
16
+
symbols: "234567abcdefghijklmnopqrstuvwxyz",
17
+
padding: None,
18
+
};
19
+
20
+
fn parse_did(c: &mut Criterion) {
21
+
let mut group = c.benchmark_group("DID:PLC");
22
+
23
+
let mut rng = StdRng::seed_from_u64(8612);
24
+
let items: Vec<_> = iter::repeat_with(|| {
25
+
let bytes: [u8; 15] = rng.random();
26
+
format!("did:plc:{}", BASE32.encode(&bytes))
27
+
})
28
+
.take(1000)
29
+
.collect();
30
+
31
+
group.throughput(Throughput::Elements(items.len() as u64));
32
+
group.bench_with_input("Did::parse", items.as_slice(), |b, dids: &[String]| {
33
+
b.iter(|| {
34
+
for did in dids {
35
+
assert!(Did::parse(&did).is_ok());
36
+
}
37
+
});
38
+
});
39
+
40
+
group.throughput(Throughput::Elements(items.len() as u64));
41
+
group.bench_with_input("String->DidBuf", items.as_slice(), |b, dids: &[String]| {
42
+
b.iter_batched(
43
+
|| dids.to_vec(),
44
+
|data| {
45
+
let dids: Vec<DidBuf> = data.into_iter().map(|s| s.try_into().unwrap()).collect();
46
+
dids
47
+
},
48
+
BatchSize::SmallInput,
49
+
);
50
+
});
51
+
52
+
group.finish();
53
+
}
54
+
55
+
criterion_group!(benches, parse_did);
56
+
criterion_main!(benches);
+170
crates/gordian-types/src/did.rs
+170
crates/gordian-types/src/did.rs
···
136
136
}
137
137
138
138
fn validate_did(did: &str) -> Result<(), Error> {
139
+
// Optimized validation path for PLC DIDs.
140
+
if validate_did_plc(did) {
141
+
return Ok(());
142
+
}
143
+
139
144
let mut parts = did.splitn(3, ':');
140
145
match (parts.next(), parts.next(), parts.next()) {
141
146
(Some("did"), Some(""), _) => Err(Error::EmptyMethod),
···
176
171
}
177
172
178
173
// @TODO Validate percent encoding in ident.
174
+
}
175
+
176
+
fn validate_did_plc(did: &str) -> bool {
177
+
if did.len() == 32 {
178
+
// x86_64 specific impl for stable rust.
179
+
#[cfg(target_arch = "x86_64")]
180
+
#[cfg(not(feature = "nightly"))]
181
+
if is_x86_feature_detected!("avx2") {
182
+
// SAFETY: AVX2 support has been detected and the DID candidate is 32 bytes
183
+
// long.
184
+
return unsafe { validate_x86_64::validate_did_plc(did) };
185
+
}
186
+
187
+
#[cfg(feature = "nightly")]
188
+
return validate_simd::validate_did_plc(did);
189
+
}
190
+
191
+
let Some(ident) = did.strip_prefix("did:plc:") else {
192
+
return false;
193
+
};
194
+
195
+
ident
196
+
.as_bytes()
197
+
.iter()
198
+
.all(|val: &u8| (b'2'..=b'7').contains(val) || val.is_ascii_lowercase())
199
+
}
200
+
201
+
#[cfg(any(feature = "nightly", target_arch = "x86_64"))]
202
+
#[repr(align(32))]
203
+
#[derive(Clone, Copy)]
204
+
struct Aligned32([u8; 32]);
205
+
206
+
#[cfg(any(feature = "nightly", target_arch = "x86_64"))]
207
+
impl Aligned32 {
208
+
const fn new(bytes: &'static [u8; 32]) -> Self {
209
+
Self(*bytes)
210
+
}
211
+
}
212
+
213
+
#[cfg(any(feature = "nightly", target_arch = "x86_64"))]
214
+
const PLC_ALPHA_MIN: Aligned32 = Aligned32::new(b"did:plc:aaaaaaaaaaaaaaaaaaaaaaaa");
215
+
#[cfg(any(feature = "nightly", target_arch = "x86_64"))]
216
+
const PLC_ALPHA_MAX: Aligned32 = Aligned32::new(b"did:plc:zzzzzzzzzzzzzzzzzzzzzzzz");
217
+
#[cfg(any(feature = "nightly", target_arch = "x86_64"))]
218
+
const PLC_NUM_MIN: Aligned32 = Aligned32::new(b"did:plc:222222222222222222222222");
219
+
#[cfg(any(feature = "nightly", target_arch = "x86_64"))]
220
+
const PLC_NUM_MAX: Aligned32 = Aligned32::new(b"did:plc:777777777777777777777777");
221
+
222
+
#[cfg(feature = "nightly")]
223
+
mod validate_simd {
224
+
//! Portable simd impl. Requires a nightly compiler.
225
+
226
+
use std::ops::BitAnd;
227
+
use std::ops::BitOr;
228
+
use std::simd::prelude::*;
229
+
230
+
use super::PLC_ALPHA_MAX;
231
+
use super::PLC_ALPHA_MIN;
232
+
use super::PLC_NUM_MAX;
233
+
use super::PLC_NUM_MIN;
234
+
235
+
impl super::Aligned32 {
236
+
#[inline]
237
+
const fn as_u8x32(self) -> u8x32 {
238
+
u8x32::from_array(self.0)
239
+
}
240
+
}
241
+
242
+
/// SIMD acclerated DID PLC validation.
243
+
pub fn validate_did_plc(did: &str) -> bool {
244
+
debug_assert_eq!(did.len(), 32);
245
+
246
+
// Use the faster avx2 impl when available.
247
+
#[cfg(target_arch = "x86_64")]
248
+
if is_x86_feature_detected!("avx2") {
249
+
// SAFETY: AVX2 support has been detected.
250
+
return unsafe { validate_did_plc_avx2(did) };
251
+
}
252
+
253
+
let data = u8x32::from_slice(did.as_bytes());
254
+
data.simd_ge(const { PLC_ALPHA_MIN.as_u8x32() })
255
+
.bitand(data.simd_le(const { PLC_ALPHA_MAX.as_u8x32() }))
256
+
.bitor(
257
+
data.simd_ge(const { PLC_NUM_MIN.as_u8x32() })
258
+
.bitand(data.simd_le(const { PLC_NUM_MAX.as_u8x32() })),
259
+
)
260
+
.all()
261
+
}
262
+
263
+
// AVX2 extensions are not enabled by the default x86_64 profile. Setting the
264
+
// avx2 target feature forces the compilier to generate avx2 instructions
265
+
// for this function only.
266
+
#[cfg(target_arch = "x86_64")]
267
+
#[target_feature(enable = "avx2")]
268
+
fn validate_did_plc_avx2(did: &str) -> bool {
269
+
let data = u8x32::from_slice(did.as_bytes());
270
+
data.simd_ge(const { PLC_ALPHA_MIN.as_u8x32() })
271
+
.bitand(data.simd_le(const { PLC_ALPHA_MAX.as_u8x32() }))
272
+
.bitor(
273
+
data.simd_ge(const { PLC_NUM_MIN.as_u8x32() })
274
+
.bitand(data.simd_le(const { PLC_NUM_MAX.as_u8x32() })),
275
+
)
276
+
.all()
277
+
}
278
+
}
279
+
280
+
#[cfg(not(feature = "nightly"))]
281
+
#[cfg(target_arch = "x86_64")]
282
+
mod validate_x86_64 {
283
+
//! x86_64-specific impl. This is usable with stable rust. Remove when
284
+
//! `std::simd` is stablized.
285
+
286
+
use std::arch::x86_64::__m256i;
287
+
288
+
use super::PLC_ALPHA_MAX;
289
+
use super::PLC_ALPHA_MIN;
290
+
use super::PLC_NUM_MAX;
291
+
use super::PLC_NUM_MIN;
292
+
293
+
impl super::Aligned32 {
294
+
#[inline]
295
+
pub fn as_mm256i(self) -> __m256i {
296
+
unsafe {
297
+
// SAFETY self.0 is aligned to 32 bytes.
298
+
std::arch::x86_64::_mm256_load_si256(self.0.as_ptr().cast())
299
+
}
300
+
}
301
+
}
302
+
303
+
#[target_feature(enable = "avx2")]
304
+
pub fn validate_did_plc(did: &str) -> bool {
305
+
use std::arch::x86_64;
306
+
307
+
debug_assert_eq!(did.len(), 32);
308
+
309
+
let data = unsafe {
310
+
// SAFETY: did_plc is 32 bytes (256 bits) long.
311
+
x86_64::_mm256_loadu_si256(did.as_ptr().cast())
312
+
};
313
+
314
+
let mask = x86_64::_mm256_movemask_epi8(x86_64::_mm256_or_si256(
315
+
x86_64::_mm256_and_si256(
316
+
x86_64::_mm256_cmpeq_epi8(
317
+
x86_64::_mm256_min_epu8(data, PLC_ALPHA_MIN.as_mm256i()),
318
+
PLC_ALPHA_MIN.as_mm256i(),
319
+
),
320
+
x86_64::_mm256_cmpeq_epi8(
321
+
x86_64::_mm256_max_epu8(data, PLC_ALPHA_MAX.as_mm256i()),
322
+
PLC_ALPHA_MAX.as_mm256i(),
323
+
),
324
+
),
325
+
x86_64::_mm256_and_si256(
326
+
x86_64::_mm256_cmpeq_epi8(
327
+
x86_64::_mm256_min_epu8(data, PLC_NUM_MIN.as_mm256i()),
328
+
PLC_NUM_MIN.as_mm256i(),
329
+
),
330
+
x86_64::_mm256_cmpeq_epi8(
331
+
x86_64::_mm256_max_epu8(data, PLC_NUM_MAX.as_mm256i()),
332
+
PLC_NUM_MAX.as_mm256i(),
333
+
),
334
+
),
335
+
));
336
+
337
+
mask == !0
338
+
}
179
339
}
180
340
181
341
/// Length in bytes an [`DidBuf`] can be before being allocated on the heap.