Skip to content

Commit 8031bc6

Browse files
onethumbonethumb
committed
Update README
VPCLMULQDQ support is no longer experimental as of Rust 1.89.0.
1 parent bda483e commit 8031bc6

File tree

1 file changed

+14
-50
lines changed

1 file changed

+14
-50
lines changed

‎README.md‎

Lines changed: 14 additions & 50 deletions
Original file line numberDiff line numberDiff line change
@@ -50,12 +50,10 @@ the local system. Specifying the `DESTDIR` environment variable will allow you t
5050
DESTDIR=/my/custom/path make install
5151
```
5252

53-
You'll need to adjust if you want to optimize with [feature flags](Cargo.toml).
54-
5553
## Usage
5654

57-
Add `crc-fast = version = "1.3"` to your `Cargo.toml` dependencies, which will enable every available optimization for
58-
the `stable` toolchain. Adjust as necessary for your desired [acceleration targets](#acceleration-targets).
55+
Add `crc-fast = version = "1.5"` to your `Cargo.toml` dependencies, which will enable every available optimization for
56+
the `stable` toolchain.
5957

6058
### Digest
6159

@@ -312,22 +310,7 @@ but all known public & private implementations agree on the correct value, which
312310
# Acceleration targets
313311

314312
This library has baseline support for accelerating all known `CRC-32` and `CRC-64` variants on `aarch64`, `x86_64`, and
315-
`x86` internally in pure `Rust`. It's extremely fast (up to dozens of GiB/s) by default if no feature flags are
316-
used.
317-
318-
### tl;dr: Just tell me how to turn it up to 11! 🤘
319-
320-
For `aarch64` and older `x86_64` systems, the release build will use the best available acceleration:
321-
322-
```
323-
cargo build --release
324-
```
325-
326-
For modern `x86_64` systems, you can enable [experimental VPCLMULQDQ support](#experimental-vpclmulqdq-support-in-rust)
327-
for a ~2X performance boost.
328-
329-
At [Awesome](https://awesome.co/), we use these 👆 at large scale in production at [Flickr](https://flickr.com/) and
330-
[SmugMug](https://www.smugmug.com/).
313+
`x86` internally in pure `Rust`.
331314

332315
### Checking your platform capabilities
333316

@@ -344,41 +327,23 @@ cargo run arch-check
344327
cargo build --release
345328
```
346329

347-
### Experimental VPCLMULQDQ support in Rust
348-
349-
This library also supports [VPCLMULQDQ](https://en.wikichip.org/wiki/x86/vpclmulqdq) for accelerating all `CRC-32` and
350-
`CRC-64` variants on modern `x86_64`
351-
platforms which support it when using `nightly` builds and the `vpclmulqdq` feature flag.
352-
353-
Typical performance boosts are ~2X, and they apply to CPUs beginning with Intel
354-
[Ice Lake](https://en.wikipedia.org/wiki/Ice_Lake_%28microprocessor%29) (Sep 2019) and
355-
AMD [Zen4](https://en.wikipedia.org/wiki/Zen_4) (Sep 2022).
356-
357-
```
358-
rustup toolchain install nightly
359-
cargo +nightly build --release --features=vpclmulqdq
360-
```
361-
362-
`AVX512` support with `VPCLMULQDQ` is stabilized on [1.89.0](https://releases.rs/docs/1.89.0/), so once that becomes
363-
stable in August 2025, this library will be updated to use it by default without needing the `nightly` toolchain.
364-
365330
## Performance
366331

367332
Modern systems can exceed 100 GiB/s for calculating `CRC-32/ISCSI`, `CRC-32/ISO-HDLC`,
368333
`CRC-64/NVME`, and all other reflected variants. (Forward variants are slower, due to the extra shuffle-masking, but
369334
are still extremely fast in this library).
370335

371-
This is a summary of the best [targets](#acceleration-targets) for the most important and popular CRC checksums.
336+
This is a summary of the performance for the most important and popular CRC checksums.
372337

373338
### CRC-32/ISCSI (reflected)
374339

375340
AKA `crc32c` in many, but not all, implementations.
376341

377342
| Arch | Brand | CPU | System | Target | 1KiB (GiB/s) | 1MiB (GiB/s) |
378343
|:--------|:------|:----------------|:--------------------------|:--------------------|-------------:|-------------:|
379-
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | avx512-vpclmulqdq* | ~49 | ~111 |
344+
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | avx512-vpclmulqdq | ~49 | ~111 |
380345
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | sse-pclmulqdq | ~18 | ~52 |
381-
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq* | ~23 | ~54 |
346+
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq | ~23 | ~54 |
382347
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | sse-pclmulqdq | ~11 | ~20 |
383348
| aarch64 | AWS | Graviton4 | EC2 c8g.metal-48xl | neon-eor3-pclmulqdq | ~19 | ~39 |
384349
| aarch64 | AWS | Graviton2 | EC2 c6g.metal | neon-pclmulqdq | ~10 | ~17 |
@@ -391,9 +356,9 @@ AKA `crc32` in many, but not all, implementations.
391356

392357
| Arch | Brand | CPU | System | Target | 1KiB (GiB/s) | 1MiB (GiB/s) |
393358
|:--------|:------|:----------------|:--------------------------|:--------------------|-------------:|-------------:|
394-
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-248xl | avx512-vpclmulqdq* | ~24 | ~110 |
359+
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-248xl | avx512-vpclmulqdq | ~24 | ~110 |
395360
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-248xl | sse-pclmulqdq | ~21 | ~28 |
396-
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq* | ~24 | ~55 |
361+
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq | ~24 | ~55 |
397362
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | sse-pclmulqdq | ~12 | ~14 |
398363
| aarch64 | AWS | Graviton4 | EC2 c8g.metal-48xl | neon-eor3-pclmulqdq | ~19 | ~39 |
399364
| aarch64 | AWS | Graviton2 | EC2 c6g.metal | neon-pclmulqdq | ~10 | ~17 |
@@ -406,9 +371,9 @@ AKA `crc32` in many, but not all, implementations.
406371

407372
| Arch | Brand | CPU | System | Target | 1KiB (GiB/s) | 1MiB (GiB/s) |
408373
|:--------|:------|:----------------|:--------------------------|:--------------------|-------------:|-------------:|
409-
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | avx512-vpclmulqdq* | ~25 | ~110 |
374+
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | avx512-vpclmulqdq | ~25 | ~110 |
410375
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | sse-pclmulqdq | ~21 | ~28 |
411-
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq* | ~25 | ~55 |
376+
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq | ~25 | ~55 |
412377
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | sse-pclmulqdq | ~11 | ~14 |
413378
| aarch64 | AWS | Graviton4 | EC2 c8g.metal-48xl | neon-eor3-pclmulqdq | ~20 | ~37 |
414379
| aarch64 | AWS | Graviton2 | EC2 c6g.metal | neon-pclmulqdq | ~10 | ~16 |
@@ -419,9 +384,9 @@ AKA `crc32` in many, but not all, implementations.
419384

420385
| Arch | Brand | CPU | System | Target | 1KiB (GiB/s) | 1MiB (GiB/s) |
421386
|:--------|:------|:----------------|:--------------------------|:--------------------|-------------:|-------------:|
422-
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | avx512-vpclmulqdq* | ~23 | ~56 |
387+
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | avx512-vpclmulqdq | ~23 | ~56 |
423388
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | sse-pclmulqdq | ~19 | ~28 |
424-
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq* | ~21 | ~43 |
389+
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq | ~21 | ~43 |
425390
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | sse-pclmulqdq | ~11 | ~13 |
426391
| aarch64 | AWS | Graviton4 | EC2 c8g.metal-48xl | neon-eor3-pclmulqdq | ~16 | ~32 |
427392
| aarch64 | AWS | Graviton2 | EC2 c6g.metal | neon-pclmulqdq | ~9 | ~14 |
@@ -432,16 +397,15 @@ AKA `crc32` in many, but not all, implementations.
432397

433398
| Arch | Brand | CPU | System | Target | 1KiB (GiB/s) | 1MiB (GiB/s) |
434399
|:--------|:------|:----------------|:--------------------------|:--------------------|-------------:|-------------:|
435-
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | avx512-vpclmulqdq* | ~24 | ~56 |
400+
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | avx512-vpclmulqdq | ~24 | ~56 |
436401
| x86_64 | Intel | Sapphire Rapids | EC2 c7i.metal-24xl | sse-pclmulqdq | ~19 | ~28 |
437-
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq* | ~21 | ~43 |
402+
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | avx512-vpclmulqdq | ~21 | ~43 |
438403
| x86_64 | AMD | Genoa | EC2 c7a.metal-48xl | sse-pclmulqdq | ~11 | ~13 |
439404
| aarch64 | AWS | Graviton4 | EC2 c8g.metal-48xl | neon-eor3-pclmulqdq | ~18 | ~31 |
440405
| aarch64 | AWS | Graviton2 | EC2 c6g.metal | neon-pclmulqdq | ~9 | ~14 |
441406
| aarch64 | Apple | M3 Ultra | Mac Studio (32 core) | neon-eor3-pclmulqdq | ~40 | ~59 |
442407
| aarch64 | Apple | M4 Max | MacBook Pro 16" (16 core) | neon-eor3-pclmulqdq | ~46 | ~61 |
443408

444-
\* = [Experimental VPCLMULQDQ support in Rust](#experimental-vpclmulqdq-support-in-rust) is enabled.
445409

446410
## Other CRC widths
447411

0 commit comments

Comments
 (0)