Module simd 

SIMD-optimized microkernels for tropical GEMM.

This module provides architecture-specific SIMD implementations of the microkernel, which is the innermost loop of the BLIS-style GEMM algorithm.

Supported Architectures

Architecture	Instruction Set	Register Width	Supported Types
x86_64	AVX-512	512-bit	f32, f64
x86_64	AVX2	256-bit	f32, f64
x86_64	SSE4.1	128-bit	f32, f64
aarch64	NEON	128-bit	f32
Any	Portable	Scalar	All types

Runtime Dispatch

At runtime, tropical_gemm_dispatch selects the best kernel:

// Automatically uses AVX2 on supported CPUs
tropical_gemm_dispatch::<MaxPlus<f32>>(...);

The dispatch mechanism:

1. simd_level() detects CPU features at runtime
2. KernelDispatch trait routes to the appropriate implementation
3. Falls back to portable kernel if no SIMD available

Microkernel Design

For tropical MaxPlus f32 with AVX2 (8-wide vectors):

// MR×NR = 8×8 output tile
for k in 0..KC:
    a_vec = load_8xf32(packed_a)     // 8 elements from A column
    for j in 0..8:
        b_scalar = broadcast(packed_b[j])  // 1 element from B row
        prod = a_vec + b_scalar            // tropical multiply
        c[j] = max(c[j], prod)             // tropical accumulate

Module Contents

• detect: CPU feature detection (SimdLevel)
• dispatch: Runtime kernel selection (KernelDispatch)
• kernels: Architecture-specific microkernel implementations

Re-exports

pub use dispatch::tropical_gemm_dispatch;

pub use dispatch::KernelDispatch;

pub use kernels::*;

Modules

dispatch

kernels

SIMD microkernel implementations.

Enums

SimdLevel

CPU feature detection for runtime SIMD dispatch. Available SIMD instruction sets.

Functions

simd_level

Get the detected SIMD level (cached).