Function conv2d_raw 

pub fn conv2d_raw(
    input: &[f64],
    filters: &[f64],
    bias: &[f64],
    out: &mut [f64],
    n: usize,
    c_in: usize,
    h_in: usize,
    w_in: usize,
    c_out: usize,
    kh: usize,
    kw: usize,
    stride: usize,
)

2D convolution – NCHW layout, valid mode (no padding), configurable stride.

Layout

• input: [N, C_in, H_in, W_in] row-major contiguous
• filters: [C_out, C_in, kH, kW] row-major contiguous
• bias: [C_out]
• out: [N, C_out, H_out, W_out] pre-allocated by caller

where H_out = (H_in - kH) / stride + 1 and W_out = (W_in - kW) / stride + 1.

Numerical contract

Every kernel-to-patch dot product uses BinnedAccumulatorF64, guaranteeing bit-identical results regardless of stride, batch size, or channel count.

NoGC guarantee

All index arithmetic uses u64 before narrowing to usize, preventing overflow for high-resolution inputs (e.g., 8192×8192). The output buffer is caller-allocated; this function performs zero heap allocations.