43 #define G1(row1,row2,row3,row4,buf) \  argument
44   row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \
45   row4 = _mm_xor_si128( row4, row1 ); \
51 #define G2(row1,row2,row3,row4,buf) \  argument
52   row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \
53   row4 = _mm_xor_si128( row4, row1 ); \
59 #define DIAGONALIZE(row1,row2,row3,row4) \  argument
64 #define UNDIAGONALIZE(row1,row2,row3,row4) \  argument
79   G1(row1,row2,row3,row4,buf1); \
81   G2(row1,row2,row3,row4,buf2); \
82   DIAGONALIZE(row1,row2,row3,row4); \
84   G1(row1,row2,row3,row4,buf3); \
86   G2(row1,row2,row3,row4,buf4); \
87   UNDIAGONALIZE(row1,row2,row3,row4); \