Module simd 

SIMD Backend Dispatch Module

Provides architecture-specific SIMD implementations with automatic dispatch:

• AVX-512 for modern Intel/AMD (16 floats per iteration)
• AVX2 with FMA for Intel Haswell+ / AMD Zen+ (8 floats per iteration)
• NEON for ARM64/Apple Silicon (4 floats per iteration)
• WASM SIMD for WebAssembly (4 floats per iteration)
• Winograd F(2,3) for 2.25x faster 3x3 convolutions
• Scalar fallback for all other platforms

Re-exports

pub use winograd::conv_3x3_winograd;

pub use winograd::transform_filter;

pub use winograd::transform_input;

pub use winograd::transform_output;

pub use winograd::WinogradFilterCache;

pub use quantize::QuantParams;

pub use quantize::QuantizedTensor;

pub use quantize::QuantizationType;

pub use quantize::PerChannelQuantParams;

pub use quantize::quantize_simd;

pub use quantize::dequantize_simd;

pub use quantize::quantize_batch;

pub use quantize::dequantize_batch;

pub use quantize::pi_constants;

pub use avx2::*;

pub use scalar::*;

Modules

avx2

AVX2 and AVX-512 SIMD Implementations

neon

NEON SIMD Implementations for ARM64/Apple Silicon

quantize

INT8 Quantization with π-Based Calibration

scalar

Scalar Fallback Implementations

winograd

Winograd F(2,3) Convolution Implementation

Functions

batch_norm_simd

SIMD-accelerated batch normalization with automatic architecture dispatch

conv_3x3_simd

SIMD-accelerated 3x3 convolution with automatic architecture dispatch

depthwise_conv_3x3_simd

SIMD-accelerated depthwise 3x3 convolution

dot_product_simd

SIMD-accelerated dot product with automatic architecture dispatch

global_avg_pool_simd

SIMD-accelerated global average pooling

max_pool_2x2_simd

SIMD-accelerated max pooling 2x2

relu6_simd

SIMD-accelerated ReLU6 activation with automatic architecture dispatch

relu_simd

SIMD-accelerated ReLU activation with automatic architecture dispatch