Module elementwise 

Element-wise activation and bias shaders.

All shaders use workgroup size 256 (1D) and take a push constant N: u32 for bounds checking. Each thread processes one element.

Constants

BIAS_ADD

Bias add: out[i] = z[i] + bias[i % cols].

RELU

ReLU activation: out[i] = max(0, in[i]).

SIGMOID

Sigmoid activation: out[i] = 1 / (1 + exp(-in[i])).

TANH

Tanh activation: out[i] = tanh(in[i]).