layer_norm_forward_f64

void layer_norm_forward_f64(const double * x, const double * gamma, const double * beta, double * y, double * saved_mean, double * saved_rstd, int outer, int N, double eps)

source edit

Double-precision LayerNorm forward — generic template implementation that computes mean and variance with a two-pass loop (no vDSP equivalent exists for double at the same fusion level).

Math

$y_i = \gamma_i \frac{x_i - \mu}{\sqrt{\sigma^2 + \varepsilon}} + \beta_i.$

Parameters

xconst double*

Input buffer of shape (outer, N).

gamma, betaconst double*

Per-feature affine parameters of shape (N,).

ydouble*

Output buffer of shape (outer, N).

saved_mean, saved_rstddouble*

Per-row mean and reciprocal standard deviation, both shape (outer,).

outer, Nstd::size_t

Number of rows and length of the normalised axis.

epsdouble

Variance stabiliser.