dssgabriel / sve-string-routines-benchmarks

Comparative performance benchmarks for hand-optimized Arm SVE implementations of C standard library string routines.

Comparative benchmarks for SVE-optimized string routines

About

This repository benchmarks the current implementations of some AArch64 SVE string routines from Arm's optimized-routines library against newly proposed implementations.

This effort was carried out as part of the EMOPASS French Project and conducted at the Laboratory of Parallel Networks and Distributed Algorithms (LI-PaRAD) of the University of Versailles Saint-Quentin-en-Yvelines (UVSQ — Paris-Saclay University), in collaboration with industrial partners: SiPearl and Eviden.

In the context of benchmarking HPC applications on Arm Neoverse V1 microarchitectures (in particular the AWS Graviton3 and Graviton3E processors), we have shown that using INCx SVE instructions to increment the loop offset can introduce significant performance degradation. Instead, it is possible to replace this instruction with a combination of CNTx to retrieve the SVE register width in the loop prelude, and ADD to increment the loop offset accordingly. This change is only used in paths where the whole vector is known to be valid, i.e., when no elements in the register are masked through a predicate. Conveniently, this happens to be the critical path in most scenarios.

The benchmarks presented here show that this change never causes any performance regression compared to the current implementations in the ARM-software/optimized-routines repository. Moreover, some of the proposed implementations are competitive with their GNU libc counterparts, namely: strcpy, strncpy, strcmp and strncmp. Tests have been conducted on AWS Graviton3 and Graviton3E instances. Detailed results are presented in the Results section at the end of this README.

Build & run

Pre-requisites

• Arm SVE enabled CPU
• CMake 3.16+
• C11 conforming compiler

Building

cmake -S . -B build
cmake --build build

By default, this will compile the code to compare against AOR implementations. You can enable comparison against GNU libc by passing -DCMP_LIBC to the CMAKE_C_FLAGS variable.

Running

All program options can be listed with the following command:

./build/bench-sve-string-routines --help

Run specific routines with their name as option. E.g. for strncmp and strcpy:

./build/bench-sve-string-routines --strncmp --strcpy

Results

Please check out the dedicated README in the results/ directory.