Lines Matching refs:__m128
17 typedef float __m128 __attribute__((__vector_size__(16), __aligned__(16)));
49 static __inline__ __m128 __DEFAULT_FN_ATTRS
50 _mm_add_ss(__m128 __a, __m128 __b)
69 static __inline__ __m128 __DEFAULT_FN_ATTRS
70 _mm_add_ps(__m128 __a, __m128 __b)
72 return (__m128)((__v4sf)__a + (__v4sf)__b);
91 static __inline__ __m128 __DEFAULT_FN_ATTRS
92 _mm_sub_ss(__m128 __a, __m128 __b)
112 static __inline__ __m128 __DEFAULT_FN_ATTRS
113 _mm_sub_ps(__m128 __a, __m128 __b)
115 return (__m128)((__v4sf)__a - (__v4sf)__b);
134 static __inline__ __m128 __DEFAULT_FN_ATTRS
135 _mm_mul_ss(__m128 __a, __m128 __b)
154 static __inline__ __m128 __DEFAULT_FN_ATTRS
155 _mm_mul_ps(__m128 __a, __m128 __b)
157 return (__m128)((__v4sf)__a * (__v4sf)__b);
176 static __inline__ __m128 __DEFAULT_FN_ATTRS
177 _mm_div_ss(__m128 __a, __m128 __b)
195 static __inline__ __m128 __DEFAULT_FN_ATTRS
196 _mm_div_ps(__m128 __a, __m128 __b)
198 return (__m128)((__v4sf)__a / (__v4sf)__b);
213 static __inline__ __m128 __DEFAULT_FN_ATTRS
214 _mm_sqrt_ss(__m128 __a)
216 return (__m128)__builtin_ia32_sqrtss((__v4sf)__a);
230 static __inline__ __m128 __DEFAULT_FN_ATTRS
231 _mm_sqrt_ps(__m128 __a)
248 static __inline__ __m128 __DEFAULT_FN_ATTRS
249 _mm_rcp_ss(__m128 __a)
251 return (__m128)__builtin_ia32_rcpss((__v4sf)__a);
265 static __inline__ __m128 __DEFAULT_FN_ATTRS
266 _mm_rcp_ps(__m128 __a)
268 return (__m128)__builtin_ia32_rcpps((__v4sf)__a);
284 static __inline__ __m128 __DEFAULT_FN_ATTRS
285 _mm_rsqrt_ss(__m128 __a)
301 static __inline__ __m128 __DEFAULT_FN_ATTRS
302 _mm_rsqrt_ps(__m128 __a)
324 static __inline__ __m128 __DEFAULT_FN_ATTRS
325 _mm_min_ss(__m128 __a, __m128 __b)
343 static __inline__ __m128 __DEFAULT_FN_ATTRS
344 _mm_min_ps(__m128 __a, __m128 __b)
366 static __inline__ __m128 __DEFAULT_FN_ATTRS
367 _mm_max_ss(__m128 __a, __m128 __b)
385 static __inline__ __m128 __DEFAULT_FN_ATTRS
386 _mm_max_ps(__m128 __a, __m128 __b)
403 static __inline__ __m128 __DEFAULT_FN_ATTRS
404 _mm_and_ps(__m128 __a, __m128 __b)
406 return (__m128)((__v4su)__a & (__v4su)__b);
425 static __inline__ __m128 __DEFAULT_FN_ATTRS
426 _mm_andnot_ps(__m128 __a, __m128 __b)
428 return (__m128)(~(__v4su)__a & (__v4su)__b);
443 static __inline__ __m128 __DEFAULT_FN_ATTRS
444 _mm_or_ps(__m128 __a, __m128 __b)
446 return (__m128)((__v4su)__a | (__v4su)__b);
462 static __inline__ __m128 __DEFAULT_FN_ATTRS
463 _mm_xor_ps(__m128 __a, __m128 __b)
465 return (__m128)((__v4su)__a ^ (__v4su)__b);
484 static __inline__ __m128 __DEFAULT_FN_ATTRS
485 _mm_cmpeq_ss(__m128 __a, __m128 __b)
487 return (__m128)__builtin_ia32_cmpeqss((__v4sf)__a, (__v4sf)__b);
502 static __inline__ __m128 __DEFAULT_FN_ATTRS
503 _mm_cmpeq_ps(__m128 __a, __m128 __b)
505 return (__m128)__builtin_ia32_cmpeqps((__v4sf)__a, (__v4sf)__b);
525 static __inline__ __m128 __DEFAULT_FN_ATTRS
526 _mm_cmplt_ss(__m128 __a, __m128 __b)
528 return (__m128)__builtin_ia32_cmpltss((__v4sf)__a, (__v4sf)__b);
544 static __inline__ __m128 __DEFAULT_FN_ATTRS
545 _mm_cmplt_ps(__m128 __a, __m128 __b)
547 return (__m128)__builtin_ia32_cmpltps((__v4sf)__a, (__v4sf)__b);
568 static __inline__ __m128 __DEFAULT_FN_ATTRS
569 _mm_cmple_ss(__m128 __a, __m128 __b)
571 return (__m128)__builtin_ia32_cmpless((__v4sf)__a, (__v4sf)__b);
587 static __inline__ __m128 __DEFAULT_FN_ATTRS
588 _mm_cmple_ps(__m128 __a, __m128 __b)
590 return (__m128)__builtin_ia32_cmpleps((__v4sf)__a, (__v4sf)__b);
610 static __inline__ __m128 __DEFAULT_FN_ATTRS
611 _mm_cmpgt_ss(__m128 __a, __m128 __b)
613 return (__m128)__builtin_shufflevector((__v4sf)__a,
631 static __inline__ __m128 __DEFAULT_FN_ATTRS
632 _mm_cmpgt_ps(__m128 __a, __m128 __b)
634 return (__m128)__builtin_ia32_cmpltps((__v4sf)__b, (__v4sf)__a);
655 static __inline__ __m128 __DEFAULT_FN_ATTRS
656 _mm_cmpge_ss(__m128 __a, __m128 __b)
658 return (__m128)__builtin_shufflevector((__v4sf)__a,
676 static __inline__ __m128 __DEFAULT_FN_ATTRS
677 _mm_cmpge_ps(__m128 __a, __m128 __b)
679 return (__m128)__builtin_ia32_cmpleps((__v4sf)__b, (__v4sf)__a);
699 static __inline__ __m128 __DEFAULT_FN_ATTRS
700 _mm_cmpneq_ss(__m128 __a, __m128 __b)
702 return (__m128)__builtin_ia32_cmpneqss((__v4sf)__a, (__v4sf)__b);
718 static __inline__ __m128 __DEFAULT_FN_ATTRS
719 _mm_cmpneq_ps(__m128 __a, __m128 __b)
721 return (__m128)__builtin_ia32_cmpneqps((__v4sf)__a, (__v4sf)__b);
742 static __inline__ __m128 __DEFAULT_FN_ATTRS
743 _mm_cmpnlt_ss(__m128 __a, __m128 __b)
745 return (__m128)__builtin_ia32_cmpnltss((__v4sf)__a, (__v4sf)__b);
762 static __inline__ __m128 __DEFAULT_FN_ATTRS
763 _mm_cmpnlt_ps(__m128 __a, __m128 __b)
765 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__a, (__v4sf)__b);
787 static __inline__ __m128 __DEFAULT_FN_ATTRS
788 _mm_cmpnle_ss(__m128 __a, __m128 __b)
790 return (__m128)__builtin_ia32_cmpnless((__v4sf)__a, (__v4sf)__b);
807 static __inline__ __m128 __DEFAULT_FN_ATTRS
808 _mm_cmpnle_ps(__m128 __a, __m128 __b)
810 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__a, (__v4sf)__b);
832 static __inline__ __m128 __DEFAULT_FN_ATTRS
833 _mm_cmpngt_ss(__m128 __a, __m128 __b)
835 return (__m128)__builtin_shufflevector((__v4sf)__a,
854 static __inline__ __m128 __DEFAULT_FN_ATTRS
855 _mm_cmpngt_ps(__m128 __a, __m128 __b)
857 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__b, (__v4sf)__a);
879 static __inline__ __m128 __DEFAULT_FN_ATTRS
880 _mm_cmpnge_ss(__m128 __a, __m128 __b)
882 return (__m128)__builtin_shufflevector((__v4sf)__a,
901 static __inline__ __m128 __DEFAULT_FN_ATTRS
902 _mm_cmpnge_ps(__m128 __a, __m128 __b)
904 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__b, (__v4sf)__a);
926 static __inline__ __m128 __DEFAULT_FN_ATTRS
927 _mm_cmpord_ss(__m128 __a, __m128 __b)
929 return (__m128)__builtin_ia32_cmpordss((__v4sf)__a, (__v4sf)__b);
946 static __inline__ __m128 __DEFAULT_FN_ATTRS
947 _mm_cmpord_ps(__m128 __a, __m128 __b)
949 return (__m128)__builtin_ia32_cmpordps((__v4sf)__a, (__v4sf)__b);
971 static __inline__ __m128 __DEFAULT_FN_ATTRS
972 _mm_cmpunord_ss(__m128 __a, __m128 __b)
974 return (__m128)__builtin_ia32_cmpunordss((__v4sf)__a, (__v4sf)__b);
991 static __inline__ __m128 __DEFAULT_FN_ATTRS
992 _mm_cmpunord_ps(__m128 __a, __m128 __b)
994 return (__m128)__builtin_ia32_cmpunordps((__v4sf)__a, (__v4sf)__b);
1016 _mm_comieq_ss(__m128 __a, __m128 __b)
1041 _mm_comilt_ss(__m128 __a, __m128 __b)
1065 _mm_comile_ss(__m128 __a, __m128 __b)
1089 _mm_comigt_ss(__m128 __a, __m128 __b)
1113 _mm_comige_ss(__m128 __a, __m128 __b)
1137 _mm_comineq_ss(__m128 __a, __m128 __b)
1161 _mm_ucomieq_ss(__m128 __a, __m128 __b)
1185 _mm_ucomilt_ss(__m128 __a, __m128 __b)
1210 _mm_ucomile_ss(__m128 __a, __m128 __b)
1235 _mm_ucomigt_ss(__m128 __a, __m128 __b)
1260 _mm_ucomige_ss(__m128 __a, __m128 __b)
1284 _mm_ucomineq_ss(__m128 __a, __m128 __b)
1302 _mm_cvtss_si32(__m128 __a)
1320 _mm_cvt_ss2si(__m128 __a)
1340 _mm_cvtss_si64(__m128 __a)
1358 _mm_cvtps_pi32(__m128 __a)
1374 _mm_cvt_ps2pi(__m128 __a)
1393 _mm_cvttss_si32(__m128 __a)
1412 _mm_cvtt_ss2si(__m128 __a)
1432 _mm_cvttss_si64(__m128 __a)
1451 _mm_cvttps_pi32(__m128 __a)
1468 _mm_cvtt_ps2pi(__m128 __a)
1489 static __inline__ __m128 __DEFAULT_FN_ATTRS
1490 _mm_cvtsi32_ss(__m128 __a, int __b)
1512 static __inline__ __m128 __DEFAULT_FN_ATTRS
1513 _mm_cvt_si2ss(__m128 __a, int __b)
1536 static __inline__ __m128 __DEFAULT_FN_ATTRS
1537 _mm_cvtsi64_ss(__m128 __a, long long __b)
1562 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
1563 _mm_cvtpi32_ps(__m128 __a, __m64 __b)
1585 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
1586 _mm_cvt_pi2ps(__m128 __a, __m64 __b)
1603 _mm_cvtss_f32(__m128 __a)
1623 static __inline__ __m128 __DEFAULT_FN_ATTRS
1624 _mm_loadh_pi(__m128 __a, const __m64 *__p)
1631 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
1650 static __inline__ __m128 __DEFAULT_FN_ATTRS
1651 _mm_loadl_pi(__m128 __a, const __m64 *__p)
1658 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
1677 static __inline__ __m128 __DEFAULT_FN_ATTRS
1684 return __extension__ (__m128){ __u, 0, 0, 0 };
1699 static __inline__ __m128 __DEFAULT_FN_ATTRS
1706 return __extension__ (__m128){ __u, __u, __u, __u };
1722 static __inline__ __m128 __DEFAULT_FN_ATTRS
1725 return *(const __m128*)__p;
1739 static __inline__ __m128 __DEFAULT_FN_ATTRS
1761 static __inline__ __m128 __DEFAULT_FN_ATTRS
1764 __m128 __a = _mm_load_ps(__p);
1775 static __inline__ __m128 __DEFAULT_FN_ATTRS
1778 return (__m128)__builtin_ia32_undef128();
1795 static __inline__ __m128 __DEFAULT_FN_ATTRS
1798 return __extension__ (__m128){ __w, 0, 0, 0 };
1813 static __inline__ __m128 __DEFAULT_FN_ATTRS
1816 return __extension__ (__m128){ __w, __w, __w, __w };
1832 static __inline__ __m128 __DEFAULT_FN_ATTRS
1859 static __inline__ __m128 __DEFAULT_FN_ATTRS
1862 return __extension__ (__m128){ __w, __x, __y, __z };
1887 static __inline__ __m128 __DEFAULT_FN_ATTRS
1890 return __extension__ (__m128){ __z, __y, __x, __w };
1902 static __inline__ __m128 __DEFAULT_FN_ATTRS
1905 return __extension__ (__m128){ 0, 0, 0, 0 };
1920 _mm_storeh_pi(__m64 *__p, __m128 __a)
1941 _mm_storel_pi(__m64 *__p, __m128 __a)
1962 _mm_store_ss(float *__p, __m128 __a)
1983 _mm_storeu_ps(float *__p, __m128 __a)
2004 _mm_store_ps(float *__p, __m128 __a)
2006 *(__m128*)__p = __a;
2023 _mm_store1_ps(float *__p, __m128 __a)
2043 _mm_store_ps1(float *__p, __m128 __a)
2062 _mm_storer_ps(float *__p, __m128 __a)
2139 _mm_stream_ps(float *__p, __m128 __a)
2573 /// __m128 _mm_shuffle_ps(__m128 a, __m128 b, const int mask);
2604 (__m128)__builtin_ia32_shufps((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), \
2623 static __inline__ __m128 __DEFAULT_FN_ATTRS
2624 _mm_unpackhi_ps(__m128 __a, __m128 __b)
2645 static __inline__ __m128 __DEFAULT_FN_ATTRS
2646 _mm_unpacklo_ps(__m128 __a, __m128 __b)
2667 static __inline__ __m128 __DEFAULT_FN_ATTRS
2668 _mm_move_ss(__m128 __a, __m128 __b)
2689 static __inline__ __m128 __DEFAULT_FN_ATTRS
2690 _mm_movehl_ps(__m128 __a, __m128 __b)
2710 static __inline__ __m128 __DEFAULT_FN_ATTRS
2711 _mm_movelh_ps(__m128 __a, __m128 __b)
2728 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2732 __m128 __r;
2758 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2762 __m128 __r;
2787 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2812 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2839 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2842 __m128 __c;
2869 _mm_cvtps_pi16(__m128 __a)
2899 _mm_cvtps_pi8(__m128 __a)
2924 _mm_movemask_ps(__m128 __a)
2972 __m128 tmp3, tmp2, tmp1, tmp0; \