Use 'crc-fast' crate for better CRC performance

[onethumb]

Motivation and Context

The crc-fast crate provides better performance (>100GiB/s for CRC32 and CRC32C) and a unified solution for calculating CRC32, CRC32C, and CRC64NVME, the three supported CRC checksums.

Description

Replaces the current three separate implementations with a single implementation.

Testing

Drop-in minimal change, existing test coverage is comprehensive and still passes.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[onethumb]

Based on the PR template, Contributing Guidelines, and reviewing some other open PRs, it wasn't obvious to me what I should do with a possible .changelog entry, so I omitted that portion. Happy to add it if someone can clarify what, if anything, I should be doing there.

Also not clear how some of the other Cargo.lock files get updated appropriately to pick up this change from aws-smithy-checksum. I'm happy to update these, if necessary, and someone can point me at docs for how. I can see that simply assembling the SDK with Gradle updates many, but not all, of these, so perhaps it's some other build process?

The remaining untouched lock files which reference the older packages are:

• aws/rust-runtime/Cargo.lock
• aws/sdk/Cargo.lock
• aws/sdk/benchmarks/previous-release-comparison/Cargo.lock (which seems likely ok?)
• aws/sdk/benchmarks/s3-express/Cargo.lock

Tagging @landonxjames since we've worked on checksums for this project in the past.

[landonxjames]

Thanks for the contribution! Massaged a few things that were impacting CI (version bump in aws-smithy-checksums crate, updated the Cargo.lock files with ./gradlew aws:sdk:cargoUpdateAllLockfiles, ran rustfmt in the checksums crate). Assuming that all works out it should take care of most of the issues.

I think all that remains would be the failures on our exotic platform tests. I suspect those probably require a software-only fallback implementation for platforms that don't support any of the expected intrinsics:

error: Unsupported architecture for SIMD CRC calculation
  --> /cargo/registry/src/index.crates.io-6f17d22bba15001f/crc-fast-1.1.0/src/arch/mod.rs:75:9
   |
75 |         compile_error!("Unsupported architecture for SIMD CRC calculation");
   |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

error: Unsupported architecture for SIMD CRC calculation
  --> /cargo/registry/src/index.crates.io-6f17d22bba15001f/crc-fast-1.1.0/src/arch/mod.rs:91:5
   |
91 |     compile_error!("Unsupported architecture for SIMD CRC calculation");
   |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

warning: unused import: `crate::algorithm`
 --> /cargo/registry/src/index.crates.io-6f17d22bba15001f/crc-fast-1.1.0/src/arch/mod.rs:8:5
  |
8 | use crate::algorithm;
  |     ^^^^^^^^^^^^^^^^
  |
  = note: `#[warn(unused_imports)]` on by default

warning: unused imports: `Width32` and `Width64`
 --> /cargo/registry/src/index.crates.io-6f17d22bba15001f/crc-fast-1.1.0/src/arch/mod.rs:9:33
  |
9 | use crate::structs::{CrcParams, Width32, Width64};
  |                                 ^^^^^^^  ^^^^^^^

error[E0308]: mismatched types
  --> /cargo/registry/src/index.crates.io-6f17d22bba15001f/crc-fast-1.1.0/src/arch/mod.rs:74:5
   |
74 | /     {
75 | |         compile_error!("Unsupported architecture for SIMD CRC calculation");
76 | |     }
   | |_____^ expected `u64`, found `()`

error[E0308]: mismatched types
  --> /cargo/registry/src/index.crates.io-6f17d22bba15001f/crc-fast-1.1.0/src/arch/mod.rs:79:24
   |
79 | pub fn get_target() -> String {
   |        ----------      ^^^^^^ expected `String`, found `()`
   |        |
   |        implicitly returns `()` as its body has no tail or `return` expression

For more information about this error, try `rustc --explain E0308`.
warning: `crc-fast` (lib) generated 2 warnings
error: could not compile `crc-fast` (lib) due to 4 previous errors; 2 warnings emitte

[onethumb]

Great, thanks for the info. Is it easiest if I include a software fallback in the in crc_fast crate (which would be trivial to implement)?

It's possible the POWER8 VPMSUMW and VPMSUMD intrinsics are available for PowerPC/PPC64 in Rust, I'd have to check. They were introduced in POWER8 (2013), but I don't know what your exotic baseline is and whether it's earlier than that?

[landonxjames]

Great, thanks for the info. Is it easiest if I include a software fallback in the in crc_fast crate (which would be trivial to implement)?

I think that would be best for us. Although both of our exotic platform tests are for powerpc variants that is really just to cover that we support 32bit architectures and a less common 64bit one, they aren't really meant to be fully representational of all the platforms we strive to support.

I'll take a look at those two other failing CI checks in the coming days, they seem to be hitting some weird versioning stuff unrelated to your change.

[onethumb]

@landonxjames Ok, I added a software fallback in crc-fast v1.2.0 and bumped the version here. Should pass the exotic tests now. Let me know if there's anything else.

[landonxjames]

Cleaned up most of the CI issues and it looks like there is still one problem compiling the crc-fast crate for the wasm32-unknown-unknown target (which I believe doesn't have a libc). The error suggest using the c_char type from std::ffi, so that might fix it, but I'll have to do some testing.

https://github.com/smithy-lang/smithy-rs/actions/runs/14936949828/job/41969146796?pr=4111

error[E0412]: cannot find type `c_char` in crate `libc`
   --> /opt/cargo/registry/src/index.crates.io-6f17d22bba15001f/crc-fast-1.2.0/src/ffi.rs:240:58
    |
240 | pub extern "C" fn crc_fast_get_version() -> *const libc::c_char {
    |                                                          ^^^^^^ not found in `libc`
    |
help: consider importing one of these type aliases
    |
8   + use std::ffi::c_char;
    |
8   + use std::os::raw::c_char;
    |
8   + use core::ffi::c_char;
    |
help: if you import `c_char`, refer to it directly
    |
240 - pub extern "C" fn crc_fast_get_version() -> *const libc::c_char {
240 + pub extern "C" fn crc_fast_get_version() -> *const c_char {
    |

[onethumb]

@landonxjames Ok, I fixed the wasm32-unknown-unknown issue (building a C-compatible shared library for wasm targets isn't really helpful anyway) in crc-fast v1.2.1 and updated accordingly. Let me know if anything else comes up. 👍

[landonxjames]

All the CI is passing so this looks good (that Semver failure always happens for external contributions that make changes to things included in our build image). Running the Canary against this PR as a final check: https://github.com/smithy-lang/smithy-rs/actions/runs/14978881530

[landonxjames]

Merged, thanks for the contribution! (To anyone from the future looking at this, I had to override and merge from the CI account because of the previously mentioned and expected Semver failure)

[atwam]

While this looks like a great improvement, there is an issue with the use of the optimize_crc32_auto feature of crc-fast crate (see awslabs/aws-sdk-rust#1291 (comment))

In short, that breaks cross-compilation, where binaries built on a machine may not work on a different machine that has the same architecture but a different set of instructions.

[GPSnoopy]

@landonxjames I'm uncomfortable that:

• The author of this PR is also the author of the crc-fast crate that is being introduced as a dependency.
• The dependency has been added without properly reviewing the claims that it is faster.
• The crc-fast crate has virtually zero downloads prior to being added to smithy-rs (and IMHO should be considered as not ready for inclusion into the AWS SDK).

Best case scenario, it's immature and therefore has its share of bugs. See the awslabs/aws-lambda-rust-runtime#993 and awslabs/aws-sdk-rust#1291, which includes the program being killed with "Illegal Instruction".

Worse case scenario (albeit unlikely - and I do not claim that this is the case), the dependency is malicious and contains security backdoors (see https://www.reddit.com/r/sysadmin/comments/1bqu3zx/backdoor_in_upstream_xzliblzma_leading_to_ssh/).

I would vote for rolling back this PR, and reconsider crc-fast once it has been battle hardened, and considered production ready by the wider Rust community.

[GPSnoopy]

Alternatively, experimental dependencies should be behind opt-in feature flags

[landonxjames]

Hi @GPSnoopy, I understand your concern, and we regret the regression this has caused, but some of those claims are inaccurate.

The author of this PR is also the author of the crc-fast crate that is being introduced as a dependency.

The author was also the author of the crc64fast-nvme we were using previously. We have worked with them before and they are a trusted partner.

The dependency has been added without properly reviewing the claims that it is faster.

The implementation and performance was reviewed and benchmarked extensively internally (including by principal engineers with expertise in high performance checksum implementations)

The crc-fast crate has virtually zero downloads prior to being added to smithy-rs (and IMHO should be considered as not ready for inclusion into the AWS SDK).

We are aware of this and chose to take the dependency anyway after discussion with the author, careful review of the code, and consideration for lowering our total number of dependencies.

Best case scenario, it's immature and therefore has its share of bugs. See the awslabs/aws-lambda-rust-runtime#993 and awslabs/aws-sdk-rust#1291, which includes the program being killed with "Illegal Instruction".

We are aware of this issue and working to figure out a fix.

[GPSnoopy]

Hi @landonxjames. Thanks for clarifying and giving more context. 👍