EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Lobsters·Sep 2·5 min read
Implementing FMA and finding bugs in C and Rust standard libraries
TL;DR
A bug was discovered in the implementation of fused multiply-add (FMA) in Rust and musl libc, which affects precision in calculations. This finding highlights the ongoing issues with Intel's hardware support for FMA, despite its importance in mathematical computations.
✦ Why It Matters
Check your project's reliance on FMA and ensure compatibility with hardware specifications.
Key Takeaways
- Implement vectorized FMA for performance improvements in Rust.
- Test FMA implementations against known inputs to catch rounding errors.
- Report bugs in standard libraries to improve community tools.
- Stay informed about hardware limitations affecting FMA support.
How It Works
The new FMA implementation computes a * b + c in f64, then rounds to f32, ensuring correct handling of rounding errors, especially for subnormal values. This approach leverages SIMD to process multiple values simultaneously, significantly enhancing performance.