CPU kernels (oracle). Optional Accelerate GEMM via feature accelerate.

This crate implements every g primitive on the CPU. It is the reference backend: fast parallel kernels over [g_core::Tensor] views, plus fused ops (embedding, softmax, cross-entropy, linear recurrence, normalization) that keep training memory and autodiff graph size low.

Enable the accelerate feature to link Apple's Accelerate framework, which upgrades GEMM to BLAS and transcendentals to vForce. Without it the same functions run in portable Rust fallbacks.

Parallelism

Kernels use [rayon] and split work above an element-count threshold. The thread pool size comes from the process's global rayon configuration; see [thread_count] for the value kernels observe.