Struct Conv2d 

pub struct Conv2d {
    pub w: Vec<f32>,
    pub b: Vec<f32>,
    pub oc: usize,
    pub ic: usize,
    pub k: usize,
}

2-D convolution, stride 1, square kernel, symmetric padding (pad = k/2). Parallel over output channels.

Fields

w: Vec<f32>

b: Vec<f32>

oc: usize

ic: usize

k: usize

Implementations

impl Conv2d

pub fn apply(&self, x: &[f32], h: usize, w: usize) -> Vec<f32>

x: [ic, h, w] → [oc, h, w]. GPU first (implicit-GEMM Metal kernel — no im2col matrix at all), gated to shapes where the transfer is amortized; otherwise banded im2col + GEMM: bands of output rows are lowered to a [rows·w, ic·k²] patch matrix and hit fcd_ops::gemm_nt (Accelerate/AMX on macOS, the portable blocked kernel elsewhere) — the band cap keeps the patch matrix ≤ ~128 MB at any image size.