oxibonsai-kernels

1-bit Q1_0_g128 compute kernels for OxiBonsai — dequantization, GEMV, GEMM.

Implements the full compute stack for 1-bit inference: scalar reference kernels, SIMD-accelerated tiers (AVX2+FMA, AVX-512, NEON), tiled cache-blocked GEMM, parallel Rayon dispatch, and GPU backend stubs (CUDA, Metal).

Part of the OxiBonsai project.

Features

• dequant_1bit_g128 — dequantize Q1_0_g128 blocks to f32
• gemv_1bit_g128 — fused 1-bit GEMV (matrix-vector multiply)
• gemm_1bit_g128 — fused 1-bit GEMM (batched matrix multiply)
• KernelDispatcher::auto_detect() — selects the best SIMD tier at runtime
• Tiled GEMM with cache-line alignment and software prefetch hints
• Parallel dispatch via Rayon (gemv_1bit_g128_par, gemm_1bit_g128_par)
• Platform tuning: PlatformProfile, TunedThresholds
• GPU backend trait (GpuBackend) with CUDA and Metal stubs

SIMD Tiers

Usage

[dependencies]
# Auto-detect at runtime:
oxibonsai-kernels = { version = "0.1.0", features = ["simd-avx2"] }

use oxibonsai_kernels::KernelDispatcher;

let dispatcher = KernelDispatcher::auto_detect();
// dispatcher selects AVX2, AVX-512, NEON, or scalar automatically

License

Apache-2.0 — COOLJAPAN OU