Module batchnorm 

Batch normalization kernel.

Matches batchnorm-kernel-v1.yaml.

Training mode: computes per-channel mean/variance from the batch, normalizes, applies affine transform, and updates running statistics via EMA.

Inference mode: uses running mean/variance directly for normalization.

Input layout: N*C flattened, where N = batch size, C = channels. Element (n, c) is at index n * c_count + c.

Functions

batchnorm_avx2^⚠

AVX2 BatchNorm – delegates to scalar.

batchnorm_ptx

PTX assembly for BatchNorm kernel (training mode).

batchnorm_scalar

Scalar reference implementation of BatchNorm.