Performance on Apple Silicon

[brookman]

Hi. Thanks for providing the awesome crypto crates! 🦀 I was experimenting with the ChaCha20Poly1305 crate and found it to run 4x slower on a Mac with M1 CPU compared to an older Intel i7 Mac. Now I‘m wondering if this is a general issue with Rust (it seems that the arm64/AArch64-Apple-Darwin target is not tier 1 yet) or if there are specific optimizations in the RustCrypto crates for Intel. Or maybe I‘m missing a feature flag? Could somebody give me a hint here?

Thanks in advance!

[tarcieri]

The chacha20 and poly1305 crates have AVX2-optimized implementations on Intel. On ARM64 we use a portable software implementation instead and have no SIMD support for these crates on ARM64 platforms other than what LLVM's auto-vectorization might be doing.

The aes-gcm crate, on the other hand, supports ARMv8 hardware acceleration for AES and GCM (i.e. PMULL) and will get you better results on M1.

[tripplet]

There is a neon/aarch64 backend for the underlying RustCrypto/stream-ciphers/chacha20 which is supported on modern Apple arm processors but it must be enabled explicitly chacha20/README.md chacha20/src/backends.rs#L18

[brookman]

Thank you for the hint @tripplet ! I'm trying to enable it by adding the following to my .cargo/config.toml file:

[target.aarch64-linux-android]
rustflags = ["-C", "target-feature=+neon", "--cfg", "chacha20_force_neon"]

[target.aarch64-apple-ios]
rustflags = ["-C", "target-feature=+neon", "--cfg", "chacha20_force_neon"]

[target.aarch64-apple-darwin]
rustflags = ["-C", "target-feature=+neon", "--cfg", "chacha20_force_neon"]

Am I doing this correctly?

[tripplet]

I'm not experienced in cross compilation, I only used rustflags = ["--cfg", "chacha20_force_neon"] to compile on Mac for Mac but looks good to me.

[brookman]

@tripplet it works. Thank you very much! 😊