5× faster fast_blur in image-rs

19 hours ago

The author optimized the 'fast_blur' method in the Rust image crate, achieving up to 5.9x faster performance for images with u8 pixels.
Blur algorithms include Gaussian Blur (O(k) per pixel with separation), Box Blur (O(1) per pixel using sliding window), and Fast Almost-Gaussian Filtering (approximates Gaussian using multiple box blurs at O(1) cost).
The hot path in 'fast_blur' was dominated by float conversions and operations (roundf, to_f32, min/max), which were replaced with integer arithmetic using u32 accumulators to eliminate these overheads.
A 'BlurAccumulator<T>' trait was designed to generically support multiple pixel types (e.g., u8 uses u32, others use f32), enabling compile-time optimization without runtime branches.
To avoid expensive integer divisions, a reciprocal multiplication technique (based on Granlund & Montgomery, 1994) was implemented, replacing divisions with faster multiplications and bit-shifts.
Combining integer accumulators and reciprocal multiplication resulted in a 5.9x speedup, reducing processing time from ~52ms to ~8ms for a 1920x1080 RGBA image, enabling real-time applications.
The optimizations were merged into image-rs and will be available in the next release, with acknowledgments to contributors for arithmetic and design feedback.

Hasty Briefsbeta