17 #define LOAD_MSG_0_1(b0, b1) \  argument
20 b0 = _mm_unpacklo_epi64(m0, m1); \
25 #define LOAD_MSG_0_2(b0, b1) \  argument
28 b0 = _mm_unpackhi_epi64(m0, m1); \
33 #define LOAD_MSG_0_3(b0, b1) \  argument
36 b0 = _mm_unpacklo_epi64(m4, m5); \
41 #define LOAD_MSG_0_4(b0, b1) \  argument
44 b0 = _mm_unpackhi_epi64(m4, m5); \
49 #define LOAD_MSG_1_1(b0, b1) \  argument
52 b0 = _mm_unpacklo_epi64(m7, m2); \
57 #define LOAD_MSG_1_2(b0, b1) \  argument
60 b0 = _mm_unpacklo_epi64(m5, m4); \
65 #define LOAD_MSG_1_3(b0, b1) \  argument
68 b0 = _mm_shuffle_epi32(m0, _MM_SHUFFLE(1,0,3,2)); \
73 #define LOAD_MSG_1_4(b0, b1) \  argument
76 b0 = _mm_unpacklo_epi64(m6, m1); \
81 #define LOAD_MSG_2_1(b0, b1) \  argument
84 b0 = _mm_alignr_epi8(m6, m5, 8); \
89 #define LOAD_MSG_2_2(b0, b1) \  argument
92 b0 = _mm_unpacklo_epi64(m4, m0); \
97 #define LOAD_MSG_2_3(b0, b1) \  argument
100 b0 = _mm_blend_epi16(m5, m1, 0xF0); \
105 #define LOAD_MSG_2_4(b0, b1) \  argument
108 b0 = _mm_unpacklo_epi64(m7, m3); \
113 #define LOAD_MSG_3_1(b0, b1) \  argument
116 b0 = _mm_unpackhi_epi64(m3, m1); \
121 #define LOAD_MSG_3_2(b0, b1) \  argument
124 b0 = _mm_unpackhi_epi64(m4, m0); \
129 #define LOAD_MSG_3_3(b0, b1) \  argument
132 b0 = _mm_blend_epi16(m1, m2, 0xF0); \
137 #define LOAD_MSG_3_4(b0, b1) \  argument
140 b0 = _mm_unpacklo_epi64(m3, m5); \
145 #define LOAD_MSG_4_1(b0, b1) \  argument
148 b0 = _mm_unpackhi_epi64(m4, m2); \
153 #define LOAD_MSG_4_2(b0, b1) \  argument
156 b0 = _mm_blend_epi16(m0, m3, 0xF0); \
161 #define LOAD_MSG_4_3(b0, b1) \  argument
164 b0 = _mm_blend_epi16(m7, m5, 0xF0); \
169 #define LOAD_MSG_4_4(b0, b1) \  argument
172 b0 = _mm_alignr_epi8(m6, m0, 8); \
177 #define LOAD_MSG_5_1(b0, b1) \  argument
180 b0 = _mm_unpacklo_epi64(m1, m3); \
185 #define LOAD_MSG_5_2(b0, b1) \  argument
188 b0 = _mm_unpacklo_epi64(m6, m5); \
193 #define LOAD_MSG_5_3(b0, b1) \  argument
196 b0 = _mm_blend_epi16(m2, m3, 0xF0); \
201 #define LOAD_MSG_5_4(b0, b1) \  argument
204 b0 = _mm_unpackhi_epi64(m6, m2); \
209 #define LOAD_MSG_6_1(b0, b1) \  argument
212 b0 = _mm_blend_epi16(m6, m0, 0xF0); \
217 #define LOAD_MSG_6_2(b0, b1) \  argument
220 b0 = _mm_unpackhi_epi64(m2, m7); \
225 #define LOAD_MSG_6_3(b0, b1) \  argument
228 b0 = _mm_unpacklo_epi64(m0, m3); \
233 #define LOAD_MSG_6_4(b0, b1) \  argument
236 b0 = _mm_unpackhi_epi64(m3, m1); \
241 #define LOAD_MSG_7_1(b0, b1) \  argument
244 b0 = _mm_unpackhi_epi64(m6, m3); \
249 #define LOAD_MSG_7_2(b0, b1) \  argument
252 b0 = _mm_alignr_epi8(m7, m5, 8); \
257 #define LOAD_MSG_7_3(b0, b1) \  argument
260 b0 = _mm_unpackhi_epi64(m2, m7); \
265 #define LOAD_MSG_7_4(b0, b1) \  argument
268 b0 = _mm_unpacklo_epi64(m0, m2); \
273 #define LOAD_MSG_8_1(b0, b1) \  argument
276 b0 = _mm_unpacklo_epi64(m3, m7); \
281 #define LOAD_MSG_8_2(b0, b1) \  argument
284 b0 = _mm_unpackhi_epi64(m7, m4); \
289 #define LOAD_MSG_8_3(b0, b1) \  argument
292 b0 = m6; \
297 #define LOAD_MSG_8_4(b0, b1) \  argument
300 b0 = _mm_blend_epi16(m1, m3, 0xF0); \
305 #define LOAD_MSG_9_1(b0, b1) \  argument
308 b0 = _mm_unpacklo_epi64(m5, m4); \
313 #define LOAD_MSG_9_2(b0, b1) \  argument
316 b0 = _mm_unpacklo_epi64(m1, m2); \
321 #define LOAD_MSG_9_3(b0, b1) \  argument
324 b0 = _mm_unpackhi_epi64(m7, m4); \
329 #define LOAD_MSG_9_4(b0, b1) \  argument
332 b0 = _mm_alignr_epi8(m7, m5, 8); \
337 #define LOAD_MSG_10_1(b0, b1) \  argument
340 b0 = _mm_unpacklo_epi64(m0, m1); \
345 #define LOAD_MSG_10_2(b0, b1) \  argument
348 b0 = _mm_unpackhi_epi64(m0, m1); \
353 #define LOAD_MSG_10_3(b0, b1) \  argument
356 b0 = _mm_unpacklo_epi64(m4, m5); \
361 #define LOAD_MSG_10_4(b0, b1) \  argument
364 b0 = _mm_unpackhi_epi64(m4, m5); \
369 #define LOAD_MSG_11_1(b0, b1) \  argument
372 b0 = _mm_unpacklo_epi64(m7, m2); \
377 #define LOAD_MSG_11_2(b0, b1) \  argument
380 b0 = _mm_unpacklo_epi64(m5, m4); \
385 #define LOAD_MSG_11_3(b0, b1) \  argument
388 b0 = _mm_shuffle_epi32(m0, _MM_SHUFFLE(1,0,3,2)); \
393 #define LOAD_MSG_11_4(b0, b1) \  argument
396 b0 = _mm_unpacklo_epi64(m6, m1); \