Lines Matching refs:__b
43 _mm_add_sd(__m128d __a, __m128d __b)
45 __a[0] += __b[0];
50 _mm_add_pd(__m128d __a, __m128d __b)
52 return __a + __b;
56 _mm_sub_sd(__m128d __a, __m128d __b)
58 __a[0] -= __b[0];
63 _mm_sub_pd(__m128d __a, __m128d __b)
65 return __a - __b;
69 _mm_mul_sd(__m128d __a, __m128d __b)
71 __a[0] *= __b[0];
76 _mm_mul_pd(__m128d __a, __m128d __b)
78 return __a * __b;
82 _mm_div_sd(__m128d __a, __m128d __b)
84 __a[0] /= __b[0];
89 _mm_div_pd(__m128d __a, __m128d __b)
91 return __a / __b;
95 _mm_sqrt_sd(__m128d __a, __m128d __b)
97 __m128d __c = __builtin_ia32_sqrtsd(__b);
108 _mm_min_sd(__m128d __a, __m128d __b)
110 return __builtin_ia32_minsd(__a, __b);
114 _mm_min_pd(__m128d __a, __m128d __b)
116 return __builtin_ia32_minpd(__a, __b);
120 _mm_max_sd(__m128d __a, __m128d __b)
122 return __builtin_ia32_maxsd(__a, __b);
126 _mm_max_pd(__m128d __a, __m128d __b)
128 return __builtin_ia32_maxpd(__a, __b);
132 _mm_and_pd(__m128d __a, __m128d __b)
134 return (__m128d)((__v4si)__a & (__v4si)__b);
138 _mm_andnot_pd(__m128d __a, __m128d __b)
140 return (__m128d)(~(__v4si)__a & (__v4si)__b);
144 _mm_or_pd(__m128d __a, __m128d __b)
146 return (__m128d)((__v4si)__a | (__v4si)__b);
150 _mm_xor_pd(__m128d __a, __m128d __b)
152 return (__m128d)((__v4si)__a ^ (__v4si)__b);
156 _mm_cmpeq_pd(__m128d __a, __m128d __b)
158 return (__m128d)__builtin_ia32_cmpeqpd(__a, __b);
162 _mm_cmplt_pd(__m128d __a, __m128d __b)
164 return (__m128d)__builtin_ia32_cmpltpd(__a, __b);
168 _mm_cmple_pd(__m128d __a, __m128d __b)
170 return (__m128d)__builtin_ia32_cmplepd(__a, __b);
174 _mm_cmpgt_pd(__m128d __a, __m128d __b)
176 return (__m128d)__builtin_ia32_cmpltpd(__b, __a);
180 _mm_cmpge_pd(__m128d __a, __m128d __b)
182 return (__m128d)__builtin_ia32_cmplepd(__b, __a);
186 _mm_cmpord_pd(__m128d __a, __m128d __b)
188 return (__m128d)__builtin_ia32_cmpordpd(__a, __b);
192 _mm_cmpunord_pd(__m128d __a, __m128d __b)
194 return (__m128d)__builtin_ia32_cmpunordpd(__a, __b);
198 _mm_cmpneq_pd(__m128d __a, __m128d __b)
200 return (__m128d)__builtin_ia32_cmpneqpd(__a, __b);
204 _mm_cmpnlt_pd(__m128d __a, __m128d __b)
206 return (__m128d)__builtin_ia32_cmpnltpd(__a, __b);
210 _mm_cmpnle_pd(__m128d __a, __m128d __b)
212 return (__m128d)__builtin_ia32_cmpnlepd(__a, __b);
216 _mm_cmpngt_pd(__m128d __a, __m128d __b)
218 return (__m128d)__builtin_ia32_cmpnltpd(__b, __a);
222 _mm_cmpnge_pd(__m128d __a, __m128d __b)
224 return (__m128d)__builtin_ia32_cmpnlepd(__b, __a);
228 _mm_cmpeq_sd(__m128d __a, __m128d __b)
230 return (__m128d)__builtin_ia32_cmpeqsd(__a, __b);
234 _mm_cmplt_sd(__m128d __a, __m128d __b)
236 return (__m128d)__builtin_ia32_cmpltsd(__a, __b);
240 _mm_cmple_sd(__m128d __a, __m128d __b)
242 return (__m128d)__builtin_ia32_cmplesd(__a, __b);
246 _mm_cmpgt_sd(__m128d __a, __m128d __b)
248 __m128d __c = __builtin_ia32_cmpltsd(__b, __a);
253 _mm_cmpge_sd(__m128d __a, __m128d __b)
255 __m128d __c = __builtin_ia32_cmplesd(__b, __a);
260 _mm_cmpord_sd(__m128d __a, __m128d __b)
262 return (__m128d)__builtin_ia32_cmpordsd(__a, __b);
266 _mm_cmpunord_sd(__m128d __a, __m128d __b)
268 return (__m128d)__builtin_ia32_cmpunordsd(__a, __b);
272 _mm_cmpneq_sd(__m128d __a, __m128d __b)
274 return (__m128d)__builtin_ia32_cmpneqsd(__a, __b);
278 _mm_cmpnlt_sd(__m128d __a, __m128d __b)
280 return (__m128d)__builtin_ia32_cmpnltsd(__a, __b);
284 _mm_cmpnle_sd(__m128d __a, __m128d __b)
286 return (__m128d)__builtin_ia32_cmpnlesd(__a, __b);
290 _mm_cmpngt_sd(__m128d __a, __m128d __b)
292 __m128d __c = __builtin_ia32_cmpnltsd(__b, __a);
297 _mm_cmpnge_sd(__m128d __a, __m128d __b)
299 __m128d __c = __builtin_ia32_cmpnlesd(__b, __a);
304 _mm_comieq_sd(__m128d __a, __m128d __b)
306 return __builtin_ia32_comisdeq(__a, __b);
310 _mm_comilt_sd(__m128d __a, __m128d __b)
312 return __builtin_ia32_comisdlt(__a, __b);
316 _mm_comile_sd(__m128d __a, __m128d __b)
318 return __builtin_ia32_comisdle(__a, __b);
322 _mm_comigt_sd(__m128d __a, __m128d __b)
324 return __builtin_ia32_comisdgt(__a, __b);
328 _mm_comige_sd(__m128d __a, __m128d __b)
330 return __builtin_ia32_comisdge(__a, __b);
334 _mm_comineq_sd(__m128d __a, __m128d __b)
336 return __builtin_ia32_comisdneq(__a, __b);
340 _mm_ucomieq_sd(__m128d __a, __m128d __b)
342 return __builtin_ia32_ucomisdeq(__a, __b);
346 _mm_ucomilt_sd(__m128d __a, __m128d __b)
348 return __builtin_ia32_ucomisdlt(__a, __b);
352 _mm_ucomile_sd(__m128d __a, __m128d __b)
354 return __builtin_ia32_ucomisdle(__a, __b);
358 _mm_ucomigt_sd(__m128d __a, __m128d __b)
360 return __builtin_ia32_ucomisdgt(__a, __b);
364 _mm_ucomige_sd(__m128d __a, __m128d __b)
366 return __builtin_ia32_ucomisdge(__a, __b);
370 _mm_ucomineq_sd(__m128d __a, __m128d __b)
372 return __builtin_ia32_ucomisdneq(__a, __b);
406 _mm_cvtsd_ss(__m128 __a, __m128d __b)
408 __a[0] = __b[0];
413 _mm_cvtsi32_sd(__m128d __a, int __b)
415 __a[0] = __b;
420 _mm_cvtss_sd(__m128d __a, __m128 __b)
422 __a[0] = __b[0];
557 _mm_move_sd(__m128d __a, __m128d __b)
559 return (__m128d){ __b[0], __a[1] };
619 _mm_add_epi8(__m128i __a, __m128i __b)
621 return (__m128i)((__v16qi)__a + (__v16qi)__b);
625 _mm_add_epi16(__m128i __a, __m128i __b)
627 return (__m128i)((__v8hi)__a + (__v8hi)__b);
631 _mm_add_epi32(__m128i __a, __m128i __b)
633 return (__m128i)((__v4si)__a + (__v4si)__b);
637 _mm_add_si64(__m64 __a, __m64 __b)
639 return __a + __b;
643 _mm_add_epi64(__m128i __a, __m128i __b)
645 return __a + __b;
649 _mm_adds_epi8(__m128i __a, __m128i __b)
651 return (__m128i)__builtin_ia32_paddsb128((__v16qi)__a, (__v16qi)__b);
655 _mm_adds_epi16(__m128i __a, __m128i __b)
657 return (__m128i)__builtin_ia32_paddsw128((__v8hi)__a, (__v8hi)__b);
661 _mm_adds_epu8(__m128i __a, __m128i __b)
663 return (__m128i)__builtin_ia32_paddusb128((__v16qi)__a, (__v16qi)__b);
667 _mm_adds_epu16(__m128i __a, __m128i __b)
669 return (__m128i)__builtin_ia32_paddusw128((__v8hi)__a, (__v8hi)__b);
673 _mm_avg_epu8(__m128i __a, __m128i __b)
675 return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b);
679 _mm_avg_epu16(__m128i __a, __m128i __b)
681 return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b);
685 _mm_madd_epi16(__m128i __a, __m128i __b)
687 return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)__a, (__v8hi)__b);
691 _mm_max_epi16(__m128i __a, __m128i __b)
693 return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)__a, (__v8hi)__b);
697 _mm_max_epu8(__m128i __a, __m128i __b)
699 return (__m128i)__builtin_ia32_pmaxub128((__v16qi)__a, (__v16qi)__b);
703 _mm_min_epi16(__m128i __a, __m128i __b)
705 return (__m128i)__builtin_ia32_pminsw128((__v8hi)__a, (__v8hi)__b);
709 _mm_min_epu8(__m128i __a, __m128i __b)
711 return (__m128i)__builtin_ia32_pminub128((__v16qi)__a, (__v16qi)__b);
715 _mm_mulhi_epi16(__m128i __a, __m128i __b)
717 return (__m128i)__builtin_ia32_pmulhw128((__v8hi)__a, (__v8hi)__b);
721 _mm_mulhi_epu16(__m128i __a, __m128i __b)
723 return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)__a, (__v8hi)__b);
727 _mm_mullo_epi16(__m128i __a, __m128i __b)
729 return (__m128i)((__v8hi)__a * (__v8hi)__b);
733 _mm_mul_su32(__m64 __a, __m64 __b)
735 return __builtin_ia32_pmuludq((__v2si)__a, (__v2si)__b);
739 _mm_mul_epu32(__m128i __a, __m128i __b)
741 return __builtin_ia32_pmuludq128((__v4si)__a, (__v4si)__b);
745 _mm_sad_epu8(__m128i __a, __m128i __b)
747 return __builtin_ia32_psadbw128((__v16qi)__a, (__v16qi)__b);
751 _mm_sub_epi8(__m128i __a, __m128i __b)
753 return (__m128i)((__v16qi)__a - (__v16qi)__b);
757 _mm_sub_epi16(__m128i __a, __m128i __b)
759 return (__m128i)((__v8hi)__a - (__v8hi)__b);
763 _mm_sub_epi32(__m128i __a, __m128i __b)
765 return (__m128i)((__v4si)__a - (__v4si)__b);
769 _mm_sub_si64(__m64 __a, __m64 __b)
771 return __a - __b;
775 _mm_sub_epi64(__m128i __a, __m128i __b)
777 return __a - __b;
781 _mm_subs_epi8(__m128i __a, __m128i __b)
783 return (__m128i)__builtin_ia32_psubsb128((__v16qi)__a, (__v16qi)__b);
787 _mm_subs_epi16(__m128i __a, __m128i __b)
789 return (__m128i)__builtin_ia32_psubsw128((__v8hi)__a, (__v8hi)__b);
793 _mm_subs_epu8(__m128i __a, __m128i __b)
795 return (__m128i)__builtin_ia32_psubusb128((__v16qi)__a, (__v16qi)__b);
799 _mm_subs_epu16(__m128i __a, __m128i __b)
801 return (__m128i)__builtin_ia32_psubusw128((__v8hi)__a, (__v8hi)__b);
805 _mm_and_si128(__m128i __a, __m128i __b)
807 return __a & __b;
811 _mm_andnot_si128(__m128i __a, __m128i __b)
813 return ~__a & __b;
817 _mm_or_si128(__m128i __a, __m128i __b)
819 return __a | __b;
823 _mm_xor_si128(__m128i __a, __m128i __b)
825 return __a ^ __b;
971 _mm_cmpeq_epi8(__m128i __a, __m128i __b)
973 return (__m128i)((__v16qi)__a == (__v16qi)__b);
977 _mm_cmpeq_epi16(__m128i __a, __m128i __b)
979 return (__m128i)((__v8hi)__a == (__v8hi)__b);
983 _mm_cmpeq_epi32(__m128i __a, __m128i __b)
985 return (__m128i)((__v4si)__a == (__v4si)__b);
989 _mm_cmpgt_epi8(__m128i __a, __m128i __b)
994 return (__m128i)((__v16qs)__a > (__v16qs)__b);
998 _mm_cmpgt_epi16(__m128i __a, __m128i __b)
1000 return (__m128i)((__v8hi)__a > (__v8hi)__b);
1004 _mm_cmpgt_epi32(__m128i __a, __m128i __b)
1006 return (__m128i)((__v4si)__a > (__v4si)__b);
1010 _mm_cmplt_epi8(__m128i __a, __m128i __b)
1012 return _mm_cmpgt_epi8(__b, __a);
1016 _mm_cmplt_epi16(__m128i __a, __m128i __b)
1018 return _mm_cmpgt_epi16(__b, __a);
1022 _mm_cmplt_epi32(__m128i __a, __m128i __b)
1024 return _mm_cmpgt_epi32(__b, __a);
1029 _mm_cvtsi64_sd(__m128d __a, long long __b)
1031 __a[0] = __b;
1083 __v4si __b = (__v4si)__a;
1084 return __b[0];
1174 _mm_set1_epi8(char __b)
1176 return (__m128i)(__v16qi){ __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b };
1210 _mm_store_si128(__m128i *__p, __m128i __b)
1212 *__p = __b;
1216 _mm_storeu_si128(__m128i *__p, __m128i __b)
1218 __builtin_ia32_storedqu((char *)__p, (__v16qi)__b);
1281 _mm_packs_epi16(__m128i __a, __m128i __b)
1283 return (__m128i)__builtin_ia32_packsswb128((__v8hi)__a, (__v8hi)__b);
1287 _mm_packs_epi32(__m128i __a, __m128i __b)
1289 return (__m128i)__builtin_ia32_packssdw128((__v4si)__a, (__v4si)__b);
1293 _mm_packus_epi16(__m128i __a, __m128i __b)
1295 return (__m128i)__builtin_ia32_packuswb128((__v8hi)__a, (__v8hi)__b);
1301 __v8hi __b = (__v8hi)__a;
1302 return (unsigned short)__b[__imm & 7];
1306 _mm_insert_epi16(__m128i __a, int __b, int __imm)
1309 __c[__imm & 7] = __b;
1342 _mm_unpackhi_epi8(__m128i __a, __m128i __b)
1344 return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15);
1348 _mm_unpackhi_epi16(__m128i __a, __m128i __b)
1350 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7);
1354 _mm_unpackhi_epi32(__m128i __a, __m128i __b)
1356 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 2, 4+2, 3, 4+3);
1360 _mm_unpackhi_epi64(__m128i __a, __m128i __b)
1362 return (__m128i)__builtin_shufflevector(__a, __b, 1, 2+1);
1366 _mm_unpacklo_epi8(__m128i __a, __m128i __b)
1368 return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7);
1372 _mm_unpacklo_epi16(__m128i __a, __m128i __b)
1374 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3);
1378 _mm_unpacklo_epi32(__m128i __a, __m128i __b)
1380 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 0, 4+0, 1, 4+1);
1384 _mm_unpacklo_epi64(__m128i __a, __m128i __b)
1386 return (__m128i)__builtin_shufflevector(__a, __b, 0, 2+0);
1408 _mm_unpackhi_pd(__m128d __a, __m128d __b)
1410 return __builtin_shufflevector(__a, __b, 1, 2+1);
1414 _mm_unpacklo_pd(__m128d __a, __m128d __b)
1416 return __builtin_shufflevector(__a, __b, 0, 2+0);