pub fn conv2d_reference(
h: usize,
w: usize,
kh: usize,
kw: usize,
input: &[f32],
kernel: &[f32],
) -> Result<Vec<f32>, ReferenceEvalError>Expand description
Standalone 2D convolution reference (single channel, stride 1, no
padding). Input H x W, kernel Kh x Kw; output (H-Kh+1) x (W-Kw+1). Mirrors the rewriter’s Conv2d lowering, which the
PTX emitter expresses as a sequence of LoadUnified +
VecFma + StoreUnified ops over a sliding window.