im2col_1d_f32

void im2col_1d_f32(const float * x, float * cols, int C, int L, int KL, int OL, int stride_l, int pad_l, int dilation_l)

source edit

Single-precision 1-D im2col unfold.

Unfolds each length- $KL$ receptive field of a 1-D feature map into a column of a 2-D matrix so that conv1d can be expressed as a single GEMM. Out-of-bounds positions (from padding or dilation) are written as zero.

Math

For each output index $o \in [0, OL)$ , channel $c$ and kernel offset $k$ : $\text{cols}[c \cdot KL + k, o] = x[c, o \cdot \text{stride} + k \cdot \text{dilation} - \text{pad}]$ with zero for out-of-bounds positions.

Parameters

xconst float*

Input tensor, shape

(C, L)

(row-major).

colsfloat*

Output column matrix, shape

(C \cdot KL, OL)

Cint

Channel count.

Lint

Input spatial length.

KLint

Kernel length.

OLint

Output spatial length.

stride_l, pad_l, dilation_lint

Convolution stride / zero-padding / dilation along the spatial axis.

Notes

Memory layout assumes the standard NCHW-style channel-first convention, with the batch dimension handled at the caller (loop over $N$ ).