im2col_f32

void im2col_f32(const float * x, float * cols, int C, int H, int W, int KH, int KW, int OH, int OW, int stride_h, int stride_w, int pad_h, int pad_w, int dilation_h, int dilation_w)

source edit

Single-precision 2-D im2col unfold.

Unfolds the 2-D feature map $x$ of shape $(C, H, W)$ into a matrix where each column is a $K_H \times K_W$ receptive-field patch flattened across channels. The output column matrix has shape $(C \cdot K_H \cdot K_W, O_H \cdot O_W)$ so that conv2d = sgemm(W_reshape, cols).

Out-of-bounds positions (those reaching outside the input due to padding, stride, or dilation) are written as zero.

Math

For each output position $(o_h, o_w)$ , kernel offset $(k_h, k_w)$ and channel $c$ , the corresponding input position is $i_h = o_h \cdot s_h + k_h \cdot d_h - p_h, \quad i_w = o_w \cdot s_w + k_w \cdot d_w - p_w$ and cols[c \cdot K_H K_W + k_h K_W + k_w, o_h \cdot O_W + o_w] = x[c, i_h, i_w] (zero if OOB).

Parameters

xconst float*

Input tensor, shape

(C, H, W)

row-major.

colsfloat*

Output column matrix, shape

(C \cdot K_H \cdot K_W, O_H \cdot O_W)

Cint

Channel count.

H, Wint

Input spatial dimensions.

KH, KWint

Kernel height and width.

OH, OWint

Output spatial dimensions.

stride_h, stride_wint

Per-axis stride.

pad_h, pad_wint

Per-axis zero-padding.

dilation_h, dilation_wint

Per-axis dilation.

Notes

Memory layout is NCHW (channel-first) — the batch dimension is iterated at the caller, with one im2col per batch element.