Ashwa 🐎

Hardware accelerated routines for single substring search

Index

• Language Ecosystems
• Supported Targets
• Installation
• API
• Benchmarks
  • search_one
  • search_two
  • search_three
  • search_n

Language Ecosystems

• rust
• npm
• pypi

Supported Targets

Architecture	Target Platform	Target ISA	Fallback
x86_64	Linux, macOS, Windows, Android, FreeBSD	AVX-512BW, AVX2, SSE4.2, SSSE3, SSE2	64-bit SWAR
AArch64 (ARM64)	Apple Silicon, Linux ARM64, Android, FreeBSD	128-bit ARM NEON	64-bit SWAR
x86 (i686)	Windows, Linux	SSE2	32-bit SWAR
ARMv7	Linux ARM, Android	128-bit ARM NEON	32-bit SWAR
WebAssembly	Browsers, Node.js (wasm32)	WASM SIMD128 (simd128)	32-bit SWAR

Installation

Add ashwa to your Cargo.toml:

[dependencies]
ashwa = "1.0.1"

Important

Minimum Supported Rust Version (MSRV) is 1.89.0.

ashwa needs Rust 1.89.0 or newer to use stable AVX-512BW support. This allows it to automatically choose the best instructions for each x86_64 processor without needing a nightly Rust version.

API

Refer to docs.rs for the complete crate documentation and API reference.

Benchmarks

• search_one
• search_two
• search_three
• search_n

Note

Benchmarks are evaluated across dedicated AWS EC2 hardware environments,

• x86_64 (x64)

  • Instance: Intel(R) Xeon(R) Platinum 8488C (8C/16T)
  • ISA: AVX512BW
  • Cache: L1d: 384 KiB · L2: 16 MiB · L3: 105 MiB
  • STREAM Triad (Read + Write): 25.76 GiB/s
  • Max Single-Core Read Bandwidth: ~12.5 GiB/s

• AArch64 (arm64)

  • Instance: AWS Graviton3 ARM Neoverse-V1 (16C/16T)
  • ISA: NEON
  • Cache: L1d: 1 MiB · L2: 16 MiB · L3: 32 MiB
  • STREAM Triad (Read + Write): 75.48 GiB/s
  • Max Single-Core Read Bandwidth: ~25.5 GiB/s

Note

All benchmarks execute while pinned to a single physical core (mostly to 0th core). For out-of-cache RAM-bound payloads (>256 MiB), every search routine fully saturates the hardwares maximum single-core memory bandwidth ceiling (~12 GiB/s on x64, ~25 GiB/s on arm64). Execution is 100% memory-bandwidth bound rather than compute-bound searching data as fast as the physical memory bus can deliver bytes into the CPU registers.

search_one

Level	Payload	Latency (x64)	Latency (arm64)	Throughput (x64)	Throughput (arm64)
L1 Cache	32 KiB	203.82 ns	668.40 ns	149.73 GiB/s	45.66 GiB/s
L2 Cache	512 KiB	3.31 µs	11.29 µs	147.74 GiB/s	43.24 GiB/s
L3 Cache	16 MiB	465.93 µs	461.74 µs	33.54 GiB/s	33.84 GiB/s
RAM	256 MiB	21.44 ms	10.87 ms	11.66 GiB/s	23.00 GiB/s
RAM	512 MiB	44.49 ms	20.91 ms	11.24 GiB/s	23.92 GiB/s
RAM	1 GiB	89.23 ms	40.30 ms	11.21 GiB/s	24.82 GiB/s

search_two

Level	Payload	Latency (x64)	Latency (arm64)	Throughput (x64)	Throughput (arm64)
L1 Cache	32 KiB	388.76 ns	1.49 µs	78.50 GiB/s	20.55 GiB/s
L2 Cache	512 KiB	9.32 µs	23.99 µs	52.41 GiB/s	20.35 GiB/s
L3 Cache	16 MiB	505.33 µs	768.56 µs	30.92 GiB/s	20.33 GiB/s
RAM	256 MiB	19.85 ms	13.00 ms	12.60 GiB/s	19.23 GiB/s
RAM	512 MiB	41.67 ms	26.09 ms	12.00 GiB/s	19.16 GiB/s
RAM	1 GiB	83.70 ms	52.12 ms	11.95 GiB/s	19.19 GiB/s

search_three

Level	Payload	Latency (x64)	Latency (arm64)	Throughput (x64)	Throughput (arm64)
L1 Cache	32 KiB	427.32 ns	2.17 µs	71.42 GiB/s	14.07 GiB/s
L2 Cache	512 KiB	7.87 µs	34.69 µs	62.01 GiB/s	14.08 GiB/s
L3 Cache	16 MiB	474.87 µs	1.12 ms	32.90 GiB/s	13.91 GiB/s
RAM	256 MiB	20.33 ms	17.92 ms	12.30 GiB/s	13.95 GiB/s
RAM	512 MiB	41.13 ms	35.96 ms	12.16 GiB/s	13.91 GiB/s
RAM	1 GiB	87.32 ms	71.86 ms	11.45 GiB/s	13.92 GiB/s

search_n (where len(N) = 8)

Level	Payload	Latency (x64)	Latency (arm64)	Throughput (x64)	Throughput (arm64)
L1 Cache	32 KiB	706.40 ns	2.47 µs	43.20 GiB/s	12.37 GiB/s
L2 Cache	512 KiB	11.51 µs	37.63 µs	42.43 GiB/s	12.98 GiB/s
L3 Cache	16 MiB	532.58 µs	1.22 ms	29.34 GiB/s	12.79 GiB/s
RAM	256 MiB	20.28 ms	19.24 ms	12.33 GiB/s	12.99 GiB/s
RAM	512 MiB	43.17 ms	38.46 ms	11.58 GiB/s	13.00 GiB/s
RAM	1 GiB	87.43 ms	76.88 ms	11.44 GiB/s	13.01 GiB/s