Crate rlx_cpu 

RLX CPU backend — executes optimized IR graphs on CPU.

Takes a fused + memory-planned IR graph and executes it using:

• BLAS (Accelerate/MKL/OpenBLAS) for matmul
• NEON/AVX SIMD kernels for element-wise ops
• Persistent Rayon thread pool for parallelism
• Arena allocator for zero per-call allocation

Modules

arena

Arena allocator — ONE allocation, zero per-call overhead.

asm_check

FileCheck-style disassembly regression tests (plan #10).

attention_bwd

Scaled dot-product attention backward (recomputes scores + softmax).

autotune

Auto-tuner — finds the optimal RuntimeConfig for a model on current hardware.

blas

Direct BLAS FFI — zero abstraction overhead.

calibrate

Activation-scale calibration for post-training INT8 quantization.

config

Runtime configuration — compile-time platform defaults + runtime hardware detection.

cost

Cost model — estimates execution time for kernel dispatch decisions.

dequant_cache

Cache dequantized GGUF weight bytes for static params.

dispatch

Dispatch table — calibration-aware kernel selection (plan #2).

executor

Graph executor — runs a fused IR graph on CPU using the arena + kernels.

gdn

Gated-DeltaNet BLAS micro-kernels (Tier C.10).

gguf_matmul

Fused GGUF K-quant dequant + matmul without materializing full F32 weights (Tier C.11).

intrinsics

ISA-split intrinsics layer (plan #85).

kernel_config

Compile-time kernel-config tables (plan #14).

kernels

SIMD kernels for fused operations.

llada2_gate

lm_head

Greedy tied-LM-head argmax without materializing full vocab logits.

moe_residency

Per-forward MoE expert residency mask (TIDE placement) for CPU dispatch.

moe_topk_capture

Capture MoE router [Op::TopK] outputs during CPU forward (TIDE refresh input).

naive

Naive reference implementations — for correctness testing and benchmarking.

op_registry

Per-backend (CPU) kernel registry for Op::Custom.

pool

Rayon-backed parallel for: par_for(total, grain, |off, cnt| …).

splat

CPU dispatch hooks for rlx_ir::Op::GaussianSplatRender — bodies registered from rlx-splat.

thunk

Thunks — pre-compiled kernel dispatch with zero per-call overhead.

tile

CPU TileIO impls (plans #23 + #27).

training_bwd

umap_knn

Reference k-NN from a row-major [n, n] pairwise distance matrix.