Numerics, SIMD and tensors¶
fib.tensor provides typed dense tensors, strided views, broadcasting,
reductions, GEMM and dense-layer operations. Ordered reductions preserve order;
explicit fast variants permit reassociation. See the tensor API.
fib.simd exposes lanes; native widths depend on the target. simd/fma needs
actual fused multiply-add, while simd/muladd is portable. fib.autodiff
provides reverse-mode tape operations and optimizers over tensors; compiler AD
is still a proposal. Measurements in the shootout describe their dated
hardware and workload, not universal performance guarantees.