Sep 0.11.0 – 9.5 GB/S CSV Parsing Using ARM Neon SIMD on Apple M1

a year ago

Sep 0.11.0 was released with a new ARM NEON SIMD parser optimized for ARM64 CPUs like Apple M1 and Microsoft Cobalt 100.
Performance improvements: 9.5 GB/s on Apple M1 (up from ~7 GB/s) and ~6 GB/s on Cobalt 100 (up from ~4 GB/s).
Sep is now ~14x faster than CsvHelper and 6.4x faster than Sylvan.Data.Csv on Apple M1.
The new parser uses ARM NEON SIMD (AdvSimd in .NET) for efficient CSV parsing, leveraging saturated conversion and optimized move mask operations.
Performance benchmarks were conducted on GitHub runners, showing higher variance due to virtual machine environments.
The blog post details the technical approach, including disassembling ARM64 code using NativeAOT and dumpbin on Windows.
Key optimizations include handling 1024 bits (64 chars) at a time and using Geoff Langdale's efficient bulk move mask technique for ARM NEON.
Links to additional resources on ARM SIMD optimizations and techniques are provided for further reading.

Hasty Briefsbeta