tiny.simd.sumsOf4

1pub fn sumsOf4(comptime D: type, value: D.Vector) D.repartition(wideLane(wideLane(D.Lane))).Vector {2    const W = wideLane(wideLane(D.Lane));3    const R = D.repartition(W);4    if (D.lane_count < 4) @compileError("sumsOf4 requires at least four lanes");5    var result: R.Vector = undefined;6    inline for (0..R.lane_count) |index| {7        const base = index * 4;8        result[index] = widen(W, value[base]) + widen(W, value[base + 1]) +9            widen(W, value[base + 2]) + widen(W, value[base + 3]);10    }11    return result;12}

1pub const sumsOf4 = reduce.sumsOf4;