47 #define G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
49   row1h = _mm_add_epi64(_mm_add_epi64(row1h, b1), row2h); \
61   row2h = _mm_xor_si128(row2h, row3h); \
64   row2h = _mm_roti_epi64(row2h, -24); \
66 #define G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
68   row1h = _mm_add_epi64(_mm_add_epi64(row1h, b1), row2h); \
80   row2h = _mm_xor_si128(row2h, row3h); \
83   row2h = _mm_roti_epi64(row2h, -63); \
86 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
87   t0 = _mm_alignr_epi8(row2h, row2l, 8); \
88   t1 = _mm_alignr_epi8(row2l, row2h, 8); \
90   row2h = t1; \
101 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
102   t0 = _mm_alignr_epi8(row2l, row2h, 8); \
103   t1 = _mm_alignr_epi8(row2h, row2l, 8); \
105   row2h = t1; \
117 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
125   row2l = _mm_unpackhi_epi64(row2l, _mm_unpacklo_epi64(row2h, row2h)); \
126   row2h = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(t1, t1))
128 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
134   row2l = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(row2l, row2l)); \
135   row2h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row2h, row2h)); \
149   G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
151   G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
152   DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
154   G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
156   G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
157   UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h);