44 #define G1(row1,row2,row3,row4,buf) \  argument
45   row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \
46   row4 = _mm_xor_si128( row4, row1 ); \
52 #define G2(row1,row2,row3,row4,buf) \  argument
53   row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \
54   row4 = _mm_xor_si128( row4, row1 ); \
60 #define DIAGONALIZE(row1,row2,row3,row4) \  argument
65 #define UNDIAGONALIZE(row1,row2,row3,row4) \  argument
80   G1(row1,row2,row3,row4,buf1); \
82   G2(row1,row2,row3,row4,buf2); \
83   DIAGONALIZE(row1,row2,row3,row4); \
85   G1(row1,row2,row3,row4,buf3); \
87   G2(row1,row2,row3,row4,buf4); \
88   UNDIAGONALIZE(row1,row2,row3,row4); \