Module kernels 

SIMD kernels for fused operations.

These are the production kernels extracted from burnembed’s ndarray_fused.rs. Each kernel processes data in-place or into a pre-allocated output buffer (from the arena). No allocation.

Functions

bias_gelu

conv_transpose2d_nchw

NCHW transposed convolution (PyTorch ConvTranspose2d, no bias). Weight layout [C_in, C_out/groups, kH, kW].

gelu_approx_inplace

gelu_inplace

group_norm_nchw

NCHW group normalization: normalizes each (C/G)×H×W group.

layer_norm2d_nchw

NCHW LayerNorm2d (candle / SAM semantics): normalize across channels at each spatial position. gamma/beta are per-channel [C].

layer_norm_row

neon_sgemm_bias_small

neon_sgemm_small

neon_softmax

par_bias_gelu

Parallel bias + GELU across thread pool.

par_gelu_approx_inplace

par_gelu_inplace

par_residual_bias_ln

Parallel residual + bias + LayerNorm.

par_silu_inplace

Parallel SiLU in-place. Same threshold reasoning as par_gelu_inplace.

residual_bias_layer_norm

Fused residual + bias + LayerNorm on [n, h] buffers. Computes: output[row] = LN(a[row] + b[row] + bias, gamma, beta)

residual_bias_rms_norm

Fused residual + bias + RMSNorm on [n, h] buffers. Computes: output[row] = RmsNorm(a[row] + b[row] + bias, gamma, beta)

resize_nearest_2x_nchw

Nearest-neighbor 2× upsample on planar NCHW.

scalar_gelu_approx

Tanh-approximation GELU (matches PyTorch/candle Tensor::gelu): y = 0.5 x (1 + tanh(√(2/π) · (x + 0.044715 x³)))

silu_inplace