Module conv1d_depthwise_causal 

ADR-020 iter-11h-b — depthwise causal 1D convolution forward + backward kernels for the GpuTape autograd pipeline.

Distinct from ssm_conv (which fuses SiLU + handles autoregressive decode state). This module is for TRAINING-MODE backward pass:

Forward shape contract: x : [n_tokens, channels] row-major (f32) kernel_w : [channels, K] row-major (f32) y : [n_tokens, channels] row-major (f32)

Math (per output element (t, c)): y[t, c] = Σ_{k=0..K-1, t+k-(K-1)>=0} kernel_w[c, k] · x[t+k-(K-1), c]

Zero-pad on the past: outputs at t < K-1 see fewer than K input taps (the missing taps default to 0 — equivalent to “no prior decode state”, which is the training-time invariant).

Statics

CONV1D_DEPTHWISE_CAUSAL_SHADER_SOURCE

Functions

dispatch_conv1d_depthwise_causal_backward_dw_f32

dispatch_conv1d_depthwise_causal_backward_dx_f32

dispatch_conv1d_depthwise_causal_forward_f32

register