tiny.simd.loadInterleaved4

1pub fn loadInterleaved4(comptime D: type, input: []const D.Lane) [4]D.Vector {2    std.debug.assert(input.len >= D.lane_count * 4);3    var result: [4]D.Vector = undefined;4    inline for (0..D.lane_count) |lane_index| {5        inline for (0..4) |vector_index| {6            result[vector_index][lane_index] = input[lane_index * 4 + vector_index];7        }8    }9    return result;10}

1pub const loadInterleaved4 = memory.loadInterleaved4;