Lines Matching refs:__b
50 /// \param __b
56 _mm_add_sd(__m128d __a, __m128d __b)
58 __a[0] += __b[0];
70 /// \param __b
75 _mm_add_pd(__m128d __a, __m128d __b)
77 return (__m128d)((__v2df)__a + (__v2df)__b);
92 /// \param __b
98 _mm_sub_sd(__m128d __a, __m128d __b)
100 __a[0] -= __b[0];
112 /// \param __b
117 _mm_sub_pd(__m128d __a, __m128d __b)
119 return (__m128d)((__v2df)__a - (__v2df)__b);
133 /// \param __b
139 _mm_mul_sd(__m128d __a, __m128d __b)
141 __a[0] *= __b[0];
153 /// \param __b
158 _mm_mul_pd(__m128d __a, __m128d __b)
160 return (__m128d)((__v2df)__a * (__v2df)__b);
175 /// \param __b
181 _mm_div_sd(__m128d __a, __m128d __b)
183 __a[0] /= __b[0];
196 /// \param __b
201 _mm_div_pd(__m128d __a, __m128d __b)
203 return (__m128d)((__v2df)__a / (__v2df)__b);
219 /// \param __b
223 /// square root of the lower 64 bits of operand \a __b, and whose upper 64
226 _mm_sqrt_sd(__m128d __a, __m128d __b)
228 __m128d __c = __builtin_ia32_sqrtsd((__v2df)__b);
261 /// \param __b
268 _mm_min_sd(__m128d __a, __m128d __b)
270 return __builtin_ia32_minsd((__v2df)__a, (__v2df)__b);
283 /// \param __b
288 _mm_min_pd(__m128d __a, __m128d __b)
290 return __builtin_ia32_minpd((__v2df)__a, (__v2df)__b);
305 /// \param __b
312 _mm_max_sd(__m128d __a, __m128d __b)
314 return __builtin_ia32_maxsd((__v2df)__a, (__v2df)__b);
327 /// \param __b
332 _mm_max_pd(__m128d __a, __m128d __b)
334 return __builtin_ia32_maxpd((__v2df)__a, (__v2df)__b);
345 /// \param __b
350 _mm_and_pd(__m128d __a, __m128d __b)
352 return (__m128d)((__v2du)__a & (__v2du)__b);
365 /// \param __b
371 _mm_andnot_pd(__m128d __a, __m128d __b)
373 return (__m128d)(~(__v2du)__a & (__v2du)__b);
384 /// \param __b
389 _mm_or_pd(__m128d __a, __m128d __b)
391 return (__m128d)((__v2du)__a | (__v2du)__b);
402 /// \param __b
407 _mm_xor_pd(__m128d __a, __m128d __b)
409 return (__m128d)((__v2du)__a ^ (__v2du)__b);
422 /// \param __b
426 _mm_cmpeq_pd(__m128d __a, __m128d __b)
428 return (__m128d)__builtin_ia32_cmpeqpd((__v2df)__a, (__v2df)__b);
442 /// \param __b
446 _mm_cmplt_pd(__m128d __a, __m128d __b)
448 return (__m128d)__builtin_ia32_cmpltpd((__v2df)__a, (__v2df)__b);
463 /// \param __b
467 _mm_cmple_pd(__m128d __a, __m128d __b)
469 return (__m128d)__builtin_ia32_cmplepd((__v2df)__a, (__v2df)__b);
484 /// \param __b
488 _mm_cmpgt_pd(__m128d __a, __m128d __b)
490 return (__m128d)__builtin_ia32_cmpltpd((__v2df)__b, (__v2df)__a);
505 /// \param __b
509 _mm_cmpge_pd(__m128d __a, __m128d __b)
511 return (__m128d)__builtin_ia32_cmplepd((__v2df)__b, (__v2df)__a);
528 /// \param __b
532 _mm_cmpord_pd(__m128d __a, __m128d __b)
534 return (__m128d)__builtin_ia32_cmpordpd((__v2df)__a, (__v2df)__b);
552 /// \param __b
556 _mm_cmpunord_pd(__m128d __a, __m128d __b)
558 return (__m128d)__builtin_ia32_cmpunordpd((__v2df)__a, (__v2df)__b);
573 /// \param __b
577 _mm_cmpneq_pd(__m128d __a, __m128d __b)
579 return (__m128d)__builtin_ia32_cmpneqpd((__v2df)__a, (__v2df)__b);
594 /// \param __b
598 _mm_cmpnlt_pd(__m128d __a, __m128d __b)
600 return (__m128d)__builtin_ia32_cmpnltpd((__v2df)__a, (__v2df)__b);
615 /// \param __b
619 _mm_cmpnle_pd(__m128d __a, __m128d __b)
621 return (__m128d)__builtin_ia32_cmpnlepd((__v2df)__a, (__v2df)__b);
636 /// \param __b
640 _mm_cmpngt_pd(__m128d __a, __m128d __b)
642 return (__m128d)__builtin_ia32_cmpnltpd((__v2df)__b, (__v2df)__a);
657 /// \param __b
661 _mm_cmpnge_pd(__m128d __a, __m128d __b)
663 return (__m128d)__builtin_ia32_cmpnlepd((__v2df)__b, (__v2df)__a);
677 /// compared to the lower double-precision value of \a __b.
678 /// \param __b
684 _mm_cmpeq_sd(__m128d __a, __m128d __b)
686 return (__m128d)__builtin_ia32_cmpeqsd((__v2df)__a, (__v2df)__b);
702 /// compared to the lower double-precision value of \a __b.
703 /// \param __b
709 _mm_cmplt_sd(__m128d __a, __m128d __b)
711 return (__m128d)__builtin_ia32_cmpltsd((__v2df)__a, (__v2df)__b);
727 /// compared to the lower double-precision value of \a __b.
728 /// \param __b
734 _mm_cmple_sd(__m128d __a, __m128d __b)
736 return (__m128d)__builtin_ia32_cmplesd((__v2df)__a, (__v2df)__b);
752 /// compared to the lower double-precision value of \a __b.
753 /// \param __b
759 _mm_cmpgt_sd(__m128d __a, __m128d __b)
761 __m128d __c = __builtin_ia32_cmpltsd((__v2df)__b, (__v2df)__a);
778 /// compared to the lower double-precision value of \a __b.
779 /// \param __b
785 _mm_cmpge_sd(__m128d __a, __m128d __b)
787 __m128d __c = __builtin_ia32_cmplesd((__v2df)__b, (__v2df)__a);
806 /// compared to the lower double-precision value of \a __b.
807 /// \param __b
813 _mm_cmpord_sd(__m128d __a, __m128d __b)
815 return (__m128d)__builtin_ia32_cmpordsd((__v2df)__a, (__v2df)__b);
834 /// compared to the lower double-precision value of \a __b.
835 /// \param __b
841 _mm_cmpunord_sd(__m128d __a, __m128d __b)
843 return (__m128d)__builtin_ia32_cmpunordsd((__v2df)__a, (__v2df)__b);
859 /// compared to the lower double-precision value of \a __b.
860 /// \param __b
866 _mm_cmpneq_sd(__m128d __a, __m128d __b)
868 return (__m128d)__builtin_ia32_cmpneqsd((__v2df)__a, (__v2df)__b);
884 /// compared to the lower double-precision value of \a __b.
885 /// \param __b
891 _mm_cmpnlt_sd(__m128d __a, __m128d __b)
893 return (__m128d)__builtin_ia32_cmpnltsd((__v2df)__a, (__v2df)__b);
909 /// compared to the lower double-precision value of \a __b.
910 /// \param __b
916 _mm_cmpnle_sd(__m128d __a, __m128d __b)
918 return (__m128d)__builtin_ia32_cmpnlesd((__v2df)__a, (__v2df)__b);
934 /// compared to the lower double-precision value of \a __b.
935 /// \param __b
941 _mm_cmpngt_sd(__m128d __a, __m128d __b)
943 __m128d __c = __builtin_ia32_cmpnltsd((__v2df)__b, (__v2df)__a);
960 /// compared to the lower double-precision value of \a __b.
961 /// \param __b
967 _mm_cmpnge_sd(__m128d __a, __m128d __b)
969 __m128d __c = __builtin_ia32_cmpnlesd((__v2df)__b, (__v2df)__a);
985 /// compared to the lower double-precision value of \a __b.
986 /// \param __b
992 _mm_comieq_sd(__m128d __a, __m128d __b)
994 return __builtin_ia32_comisdeq((__v2df)__a, (__v2df)__b);
1011 /// compared to the lower double-precision value of \a __b.
1012 /// \param __b
1018 _mm_comilt_sd(__m128d __a, __m128d __b)
1020 return __builtin_ia32_comisdlt((__v2df)__a, (__v2df)__b);
1037 /// compared to the lower double-precision value of \a __b.
1038 /// \param __b
1044 _mm_comile_sd(__m128d __a, __m128d __b)
1046 return __builtin_ia32_comisdle((__v2df)__a, (__v2df)__b);
1063 /// compared to the lower double-precision value of \a __b.
1064 /// \param __b
1070 _mm_comigt_sd(__m128d __a, __m128d __b)
1072 return __builtin_ia32_comisdgt((__v2df)__a, (__v2df)__b);
1089 /// compared to the lower double-precision value of \a __b.
1090 /// \param __b
1096 _mm_comige_sd(__m128d __a, __m128d __b)
1098 return __builtin_ia32_comisdge((__v2df)__a, (__v2df)__b);
1115 /// compared to the lower double-precision value of \a __b.
1116 /// \param __b
1122 _mm_comineq_sd(__m128d __a, __m128d __b)
1124 return __builtin_ia32_comisdneq((__v2df)__a, (__v2df)__b);
1139 /// compared to the lower double-precision value of \a __b.
1140 /// \param __b
1146 _mm_ucomieq_sd(__m128d __a, __m128d __b)
1148 return __builtin_ia32_ucomisdeq((__v2df)__a, (__v2df)__b);
1165 /// compared to the lower double-precision value of \a __b.
1166 /// \param __b
1172 _mm_ucomilt_sd(__m128d __a, __m128d __b)
1174 return __builtin_ia32_ucomisdlt((__v2df)__a, (__v2df)__b);
1191 /// compared to the lower double-precision value of \a __b.
1192 /// \param __b
1198 _mm_ucomile_sd(__m128d __a, __m128d __b)
1200 return __builtin_ia32_ucomisdle((__v2df)__a, (__v2df)__b);
1217 /// compared to the lower double-precision value of \a __b.
1218 /// \param __b
1224 _mm_ucomigt_sd(__m128d __a, __m128d __b)
1226 return __builtin_ia32_ucomisdgt((__v2df)__a, (__v2df)__b);
1243 /// compared to the lower double-precision value of \a __b.
1244 /// \param __b
1250 _mm_ucomige_sd(__m128d __a, __m128d __b)
1252 return __builtin_ia32_ucomisdge((__v2df)__a, (__v2df)__b);
1269 /// compared to the lower double-precision value of \a __b.
1270 /// \param __b
1276 _mm_ucomineq_sd(__m128d __a, __m128d __b)
1278 return __builtin_ia32_ucomisdneq((__v2df)__a, (__v2df)__b);
1393 /// \param __b
1400 _mm_cvtsd_ss(__m128 __a, __m128d __b)
1402 return (__m128)__builtin_ia32_cvtsd2ss((__v4sf)__a, (__v2df)__b);
1417 /// \param __b
1423 _mm_cvtsi32_sd(__m128d __a, int __b)
1425 __a[0] = __b;
1442 /// \param __b
1449 _mm_cvtss_sd(__m128d __a, __m128 __b)
1451 __a[0] = __b[0];
1927 /// \param __b
1932 _mm_move_sd(__m128d __a, __m128d __b)
1934 __a[0] = __b[0];
2112 /// \param __b
2117 _mm_add_epi8(__m128i __a, __m128i __b)
2119 return (__m128i)((__v16qu)__a + (__v16qu)__b);
2134 /// \param __b
2139 _mm_add_epi16(__m128i __a, __m128i __b)
2141 return (__m128i)((__v8hu)__a + (__v8hu)__b);
2156 /// \param __b
2161 _mm_add_epi32(__m128i __a, __m128i __b)
2163 return (__m128i)((__v4su)__a + (__v4su)__b);
2175 /// \param __b
2179 _mm_add_si64(__m64 __a, __m64 __b)
2181 return (__m64)__builtin_ia32_paddq((__v1di)__a, (__v1di)__b);
2196 /// \param __b
2201 _mm_add_epi64(__m128i __a, __m128i __b)
2203 return (__m128i)((__v2du)__a + (__v2du)__b);
2217 /// \param __b
2222 _mm_adds_epi8(__m128i __a, __m128i __b)
2224 return (__m128i)__builtin_ia32_paddsb128((__v16qi)__a, (__v16qi)__b);
2239 /// \param __b
2244 _mm_adds_epi16(__m128i __a, __m128i __b)
2246 return (__m128i)__builtin_ia32_paddsw128((__v8hi)__a, (__v8hi)__b);
2260 /// \param __b
2265 _mm_adds_epu8(__m128i __a, __m128i __b)
2267 return (__m128i)__builtin_ia32_paddusb128((__v16qi)__a, (__v16qi)__b);
2281 /// \param __b
2286 _mm_adds_epu16(__m128i __a, __m128i __b)
2288 return (__m128i)__builtin_ia32_paddusw128((__v8hi)__a, (__v8hi)__b);
2301 /// \param __b
2306 _mm_avg_epu8(__m128i __a, __m128i __b)
2308 return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b);
2321 /// \param __b
2326 _mm_avg_epu16(__m128i __a, __m128i __b)
2328 return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b);
2347 /// \param __b
2352 _mm_madd_epi16(__m128i __a, __m128i __b)
2354 return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)__a, (__v8hi)__b);
2367 /// \param __b
2372 _mm_max_epi16(__m128i __a, __m128i __b)
2374 return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)__a, (__v8hi)__b);
2387 /// \param __b
2392 _mm_max_epu8(__m128i __a, __m128i __b)
2394 return (__m128i)__builtin_ia32_pmaxub128((__v16qi)__a, (__v16qi)__b);
2407 /// \param __b
2412 _mm_min_epi16(__m128i __a, __m128i __b)
2414 return (__m128i)__builtin_ia32_pminsw128((__v8hi)__a, (__v8hi)__b);
2427 /// \param __b
2432 _mm_min_epu8(__m128i __a, __m128i __b)
2434 return (__m128i)__builtin_ia32_pminub128((__v16qi)__a, (__v16qi)__b);
2447 /// \param __b
2452 _mm_mulhi_epi16(__m128i __a, __m128i __b)
2454 return (__m128i)__builtin_ia32_pmulhw128((__v8hi)__a, (__v8hi)__b);
2467 /// \param __b
2472 _mm_mulhi_epu16(__m128i __a, __m128i __b)
2474 return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)__a, (__v8hi)__b);
2487 /// \param __b
2492 _mm_mullo_epi16(__m128i __a, __m128i __b)
2494 return (__m128i)((__v8hu)__a * (__v8hu)__b);
2507 /// \param __b
2511 _mm_mul_su32(__m64 __a, __m64 __b)
2513 return __builtin_ia32_pmuludq((__v2si)__a, (__v2si)__b);
2526 /// \param __b
2530 _mm_mul_epu32(__m128i __a, __m128i __b)
2532 return __builtin_ia32_pmuludq128((__v4si)__a, (__v4si)__b);
2547 /// \param __b
2552 _mm_sad_epu8(__m128i __a, __m128i __b)
2554 return __builtin_ia32_psadbw128((__v16qi)__a, (__v16qi)__b);
2565 /// \param __b
2570 _mm_sub_epi8(__m128i __a, __m128i __b)
2572 return (__m128i)((__v16qu)__a - (__v16qu)__b);
2583 /// \param __b
2588 _mm_sub_epi16(__m128i __a, __m128i __b)
2590 return (__m128i)((__v8hu)__a - (__v8hu)__b);
2601 /// \param __b
2606 _mm_sub_epi32(__m128i __a, __m128i __b)
2608 return (__m128i)((__v4su)__a - (__v4su)__b);
2620 /// \param __b
2625 _mm_sub_si64(__m64 __a, __m64 __b)
2627 return (__m64)__builtin_ia32_psubq((__v1di)__a, (__v1di)__b);
2638 /// \param __b
2643 _mm_sub_epi64(__m128i __a, __m128i __b)
2645 return (__m128i)((__v2du)__a - (__v2du)__b);
2659 /// \param __b
2664 _mm_subs_epi8(__m128i __a, __m128i __b)
2666 return (__m128i)__builtin_ia32_psubsb128((__v16qi)__a, (__v16qi)__b);
2680 /// \param __b
2685 _mm_subs_epi16(__m128i __a, __m128i __b)
2687 return (__m128i)__builtin_ia32_psubsw128((__v8hi)__a, (__v8hi)__b);
2700 /// \param __b
2705 _mm_subs_epu8(__m128i __a, __m128i __b)
2707 return (__m128i)__builtin_ia32_psubusb128((__v16qi)__a, (__v16qi)__b);
2720 /// \param __b
2725 _mm_subs_epu16(__m128i __a, __m128i __b)
2727 return (__m128i)__builtin_ia32_psubusw128((__v8hi)__a, (__v8hi)__b);
2738 /// \param __b
2743 _mm_and_si128(__m128i __a, __m128i __b)
2745 return (__m128i)((__v2du)__a & (__v2du)__b);
2758 /// \param __b
2763 _mm_andnot_si128(__m128i __a, __m128i __b)
2765 return (__m128i)(~(__v2du)__a & (__v2du)__b);
2775 /// \param __b
2780 _mm_or_si128(__m128i __a, __m128i __b)
2782 return (__m128i)((__v2du)__a | (__v2du)__b);
2793 /// \param __b
2798 _mm_xor_si128(__m128i __a, __m128i __b)
2800 return (__m128i)((__v2du)__a ^ (__v2du)__b);
3167 /// \param __b
3171 _mm_cmpeq_epi8(__m128i __a, __m128i __b)
3173 return (__m128i)((__v16qi)__a == (__v16qi)__b);
3186 /// \param __b
3190 _mm_cmpeq_epi16(__m128i __a, __m128i __b)
3192 return (__m128i)((__v8hi)__a == (__v8hi)__b);
3205 /// \param __b
3209 _mm_cmpeq_epi32(__m128i __a, __m128i __b)
3211 return (__m128i)((__v4si)__a == (__v4si)__b);
3225 /// \param __b
3229 _mm_cmpgt_epi8(__m128i __a, __m128i __b)
3233 return (__m128i)((__v16qs)__a > (__v16qs)__b);
3248 /// \param __b
3252 _mm_cmpgt_epi16(__m128i __a, __m128i __b)
3254 return (__m128i)((__v8hi)__a > (__v8hi)__b);
3269 /// \param __b
3273 _mm_cmpgt_epi32(__m128i __a, __m128i __b)
3275 return (__m128i)((__v4si)__a > (__v4si)__b);
3290 /// \param __b
3294 _mm_cmplt_epi8(__m128i __a, __m128i __b)
3296 return _mm_cmpgt_epi8(__b, __a);
3311 /// \param __b
3315 _mm_cmplt_epi16(__m128i __a, __m128i __b)
3317 return _mm_cmpgt_epi16(__b, __a);
3332 /// \param __b
3336 _mm_cmplt_epi32(__m128i __a, __m128i __b)
3338 return _mm_cmpgt_epi32(__b, __a);
3354 /// \param __b
3360 _mm_cvtsi64_sd(__m128d __a, long long __b)
3362 __a[0] = __b;
3498 __v4si __b = (__v4si)__a;
3499 return __b[0];
3836 /// \param __b
3842 _mm_set1_epi8(char __b)
3844 return _mm_set_epi8(__b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b);
3992 /// \param __b
3995 _mm_store_si128(__m128i *__p, __m128i __b)
3997 *__p = __b;
4008 /// \param __b
4011 _mm_storeu_si128(__m128i_u *__p, __m128i __b)
4016 ((struct __storeu_si128*)__p)->__v = __b;
4029 /// \param __b
4032 _mm_storeu_si64(void *__p, __m128i __b)
4037 ((struct __storeu_si64*)__p)->__v = ((__v2di)__b)[0];
4050 /// \param __b
4053 _mm_storeu_si32(void *__p, __m128i __b)
4058 ((struct __storeu_si32*)__p)->__v = ((__v4si)__b)[0];
4071 /// \param __b
4074 _mm_storeu_si16(void *__p, __m128i __b)
4079 ((struct __storeu_si16*)__p)->__v = ((__v8hi)__b)[0];
4267 /// \param __b
4275 _mm_packs_epi16(__m128i __a, __m128i __b)
4277 return (__m128i)__builtin_ia32_packsswb128((__v8hi)__a, (__v8hi)__b);
4295 /// \param __b
4303 _mm_packs_epi32(__m128i __a, __m128i __b)
4305 return (__m128i)__builtin_ia32_packssdw128((__v4si)__a, (__v4si)__b);
4323 /// \param __b
4331 _mm_packus_epi16(__m128i __a, __m128i __b)
4333 return (__m128i)__builtin_ia32_packuswb128((__v8hi)__a, (__v8hi)__b);
4374 /// the lower 16 bits of \a __b.
4375 /// \param __b
4380 /// lower 16 bits of \a __b are written.
4513 /// \param __b
4525 _mm_unpackhi_epi8(__m128i __a, __m128i __b)
4527 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);
4544 /// \param __b
4552 _mm_unpackhi_epi16(__m128i __a, __m128i __b)
4554 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7);
4569 /// \param __b
4575 _mm_unpackhi_epi32(__m128i __a, __m128i __b)
4577 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 2, 4+2, 3, 4+3);
4591 /// \param __b
4596 _mm_unpackhi_epi64(__m128i __a, __m128i __b)
4598 return (__m128i)__builtin_shufflevector((__v2di)__a, (__v2di)__b, 1, 2+1);
4619 /// \param __b
4631 _mm_unpacklo_epi8(__m128i __a, __m128i __b)
4633 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);
4651 /// \param __b
4659 _mm_unpacklo_epi16(__m128i __a, __m128i __b)
4661 return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3);
4676 /// \param __b
4682 _mm_unpacklo_epi32(__m128i __a, __m128i __b)
4684 return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 0, 4+0, 1, 4+1);
4698 /// \param __b
4703 _mm_unpacklo_epi64(__m128i __a, __m128i __b)
4705 return (__m128i)__builtin_shufflevector((__v2di)__a, (__v2di)__b, 0, 2+0);
4771 /// \param __b
4776 _mm_unpackhi_pd(__m128d __a, __m128d __b)
4778 return __builtin_shufflevector((__v2df)__a, (__v2df)__b, 1, 2+1);
4792 /// \param __b
4797 _mm_unpacklo_pd(__m128d __a, __m128d __b)
4799 return __builtin_shufflevector((__v2df)__a, (__v2df)__b, 0, 2+0);