Crate kodegen_simd

Crate kodegen_simd 

Source
Expand description

Ultra-High-Performance SIMD Operations for cyrup Ecosystem

Production-quality vectorized implementations shared across cyrup packages:

  • Vector similarity operations (from memory package)
  • Platform-specific optimizations and fallbacks

§Core Features

  • Vectorized Similarity: Parallel cosine similarity with runtime CPU feature detection
  • Zero Allocation: Pre-allocated buffers and stack-based temporary storage
  • Adaptive Selection: Automatic SIMD vs scalar selection based on vector size
  • Platform Support: x86_64 AVX2, ARM64 NEON with portable fallbacks

§Usage Examples

Similarity Operations:

use kodegen_simd::similarity::smart_cosine_similarity;

let a = vec![1.0, 2.0, 3.0, 4.0];
let b = vec![4.0, 3.0, 2.0, 1.0];
let similarity = smart_cosine_similarity(&a, &b);

Re-exports§

pub use constants::SIMD_WIDTH_8;
pub use constants::VERSION;
pub use error::SimdError;
pub use error::SimdResult;
pub use logits::apply_penalties_simd;
pub use logits::prepare_nucleus_sampling_simd;
pub use logits::topk_filtering_simd;
pub use ops::argmax;
pub use ops::argmax;
pub use ops::scale_temperature;
pub use ops::softmax;
pub use ops::softmax;
pub use runtime::CpuFeatures;
pub use runtime::CpuInfo;
pub use runtime::get_cpu_features;
pub use runtime::get_cpu_info;
pub use runtime::should_use_simd;
pub use similarity::cosine_similarity;
pub use similarity::simd_cosine_similarity;
pub use similarity::smart_cosine_similarity;
pub use utils::simd_available;

Modules§

benchmark
Benchmarking utilities for SIMD operations
config
Configuration types for SIMD-accelerated operations
constants
Constants used throughout the SIMD operations.
context
Processing context for SIMD-accelerated operations
error
Error types for SIMD operations
logits
Logits processing module for SIMD-accelerated operations
ops
SIMD-optimized operations for machine learning workloads
runtime
Runtime CPU feature detection and dispatch for SIMD operations
serde_constraints
Serde type constraints for structured generation Serde Type Constraint API for Structured Generation
similarity
High-performance vector similarity operations with automatic SIMD acceleration
utils
Utility functions for SIMD-accelerated operations