tiny.simd.storeInterleaved4

1pub fn storeInterleaved4(2    comptime D: type,3    a: D.Vector,4    b: D.Vector,5    c: D.Vector,6    d: D.Vector,7    output: []D.Lane,8) void {9    std.debug.assert(output.len >= D.lane_count * 4);10    const values = [4]D.Vector{ a, b, c, d };11    inline for (0..D.lane_count) |lane_index| {12        inline for (0..4) |vector_index| {13            output[lane_index * 4 + vector_index] = values[vector_index][lane_index];14        }15    }16}

1pub const storeInterleaved4 = memory.storeInterleaved4;