21 static void FUNC(ff_hevc_idct_4x4, 
BIT_DEPTH)(int16_t *coeffs, 
int col_limit)
 
   29     src_01 = vec_ld(0, coeffs);
 
   30     src_23 = vec_ld(16, coeffs);
 
   32     transform4x4(src_01, src_23, res, 
shift, coeffs);
 
   33     src_01 = vec_packs(res[0], res[1]);
 
   34     src_23 = vec_packs(res[2], res[3]);
 
   37     src_01 = vec_perm(res_packed[0], res_packed[1], 
mask[0]);
 
   38     src_23 = vec_perm(res_packed[0], res_packed[1], 
mask[1]);
 
   40     transform4x4(src_01, src_23, res, 
shift2, coeffs);
 
   43     src_01 = vec_perm(res_packed[0], res_packed[1], 
mask[0]);
 
   44     src_23 = vec_perm(res_packed[0], res_packed[1], 
mask[1]);
 
   46     vec_st(src_01, 0, coeffs);
 
   47     vec_st(src_23, 16, coeffs);