tiny.simd.multiRotateRight

1pub fn multiRotateRight(2    comptime D: type,3    value: D.Vector,4    indices: D.repartition(u8).Vector,5) D.Vector {6    validate(D);7    if (@bitSizeOf(D.Lane) != 64) @compileError("multiRotateRight requires 64-bit lanes");8    const bits: @Vector(D.lane_count, u64) = @bitCast(value);9    const index_lanes: [D.lane_count * 8]u8 = indices;10    var result: @Vector(D.lane_count, u64) = @splat(0);11    inline for (0..D.lane_count) |lane_index| {12        var lane: u64 = 0;13        inline for (0..8) |byte_index| {14            const amount = index_lanes[lane_index * 8 + byte_index] & 63;15            const byte: u8 = @truncate(std.math.rotr(u64, bits[lane_index], amount));16            const destination = if (builtin.cpu.arch.endian() == .little)17                byte_index18            else19                byte_index ^ 7;20            lane |= @as(u64, byte) << @intCast(destination * 8);21        }22        result[lane_index] = lane;23    }24    return @bitCast(result);25}

1pub const multiRotateRight = rotate.multiRotateRight;