Right shift for int4

Vectorial magic.

__device__ int4 right_shift_int4(int4 a_value, unsigned int a_num)
{
    int x_h = a_value.x >> (a_num % 32);
    int x_l = a_value.x << (32 - (a_num %32));

    int y_h = a_value.y >> (a_num % 32);
    int y_l = a_value.y << (32 - (a_num % 32));

    int z_h = a_value.z >> (a_num % 32);
    int z_l = a_value.z << (32 - (a_num % 32));

    int w_h = a_value.w >> (a_num % 32);

    int tmp[8] = {0};
    tmp[4] = x_h;
    tmp[5] = x_l | y_h;
    tmp[6] = y_l | z_h;
    tmp[7] = z_l | w_h;

    unsigned int offset = a_num / 32;

    return make_int4(tmp[4 - offset], tmp[5 - offset], tmp[6 - offset], tmp[7 - offset]);
}