Home | History | Annotate | Download | only in include

Lines Matching refs:__m128

31 typedef float __m128 __attribute__((__vector_size__(16)));
60 static __inline__ __m128 __DEFAULT_FN_ATTRS
61 _mm_add_ss(__m128 __a, __m128 __b)
80 static __inline__ __m128 __DEFAULT_FN_ATTRS
81 _mm_add_ps(__m128 __a, __m128 __b)
83 return (__m128)((__v4sf)__a + (__v4sf)__b);
102 static __inline__ __m128 __DEFAULT_FN_ATTRS
103 _mm_sub_ss(__m128 __a, __m128 __b)
123 static __inline__ __m128 __DEFAULT_FN_ATTRS
124 _mm_sub_ps(__m128 __a, __m128 __b)
126 return (__m128)((__v4sf)__a - (__v4sf)__b);
145 static __inline__ __m128 __DEFAULT_FN_ATTRS
146 _mm_mul_ss(__m128 __a, __m128 __b)
165 static __inline__ __m128 __DEFAULT_FN_ATTRS
166 _mm_mul_ps(__m128 __a, __m128 __b)
168 return (__m128)((__v4sf)__a * (__v4sf)__b);
187 static __inline__ __m128 __DEFAULT_FN_ATTRS
188 _mm_div_ss(__m128 __a, __m128 __b)
206 static __inline__ __m128 __DEFAULT_FN_ATTRS
207 _mm_div_ps(__m128 __a, __m128 __b)
209 return (__m128)((__v4sf)__a / (__v4sf)__b);
224 static __inline__ __m128 __DEFAULT_FN_ATTRS
225 _mm_sqrt_ss(__m128 __a)
227 __m128 __c = __builtin_ia32_sqrtss((__v4sf)__a);
228 return (__m128) { __c[0], __a[1], __a[2], __a[3] };
242 static __inline__ __m128 __DEFAULT_FN_ATTRS
243 _mm_sqrt_ps(__m128 __a)
260 static __inline__ __m128 __DEFAULT_FN_ATTRS
261 _mm_rcp_ss(__m128 __a)
263 __m128 __c = __builtin_ia32_rcpss((__v4sf)__a);
264 return (__m128) { __c[0], __a[1], __a[2], __a[3] };
278 static __inline__ __m128 __DEFAULT_FN_ATTRS
279 _mm_rcp_ps(__m128 __a)
297 static __inline__ __m128 __DEFAULT_FN_ATTRS
298 _mm_rsqrt_ss(__m128 __a)
300 __m128 __c = __builtin_ia32_rsqrtss((__v4sf)__a);
301 return (__m128) { __c[0], __a[1], __a[2], __a[3] };
315 static __inline__ __m128 __DEFAULT_FN_ATTRS
316 _mm_rsqrt_ps(__m128 __a)
338 static __inline__ __m128 __DEFAULT_FN_ATTRS
339 _mm_min_ss(__m128 __a, __m128 __b)
357 static __inline__ __m128 __DEFAULT_FN_ATTRS
358 _mm_min_ps(__m128 __a, __m128 __b)
380 static __inline__ __m128 __DEFAULT_FN_ATTRS
381 _mm_max_ss(__m128 __a, __m128 __b)
399 static __inline__ __m128 __DEFAULT_FN_ATTRS
400 _mm_max_ps(__m128 __a, __m128 __b)
417 static __inline__ __m128 __DEFAULT_FN_ATTRS
418 _mm_and_ps(__m128 __a, __m128 __b)
420 return (__m128)((__v4su)__a & (__v4su)__b);
439 static __inline__ __m128 __DEFAULT_FN_ATTRS
440 _mm_andnot_ps(__m128 __a, __m128 __b)
442 return (__m128)(~(__v4su)__a & (__v4su)__b);
457 static __inline__ __m128 __DEFAULT_FN_ATTRS
458 _mm_or_ps(__m128 __a, __m128 __b)
460 return (__m128)((__v4su)__a | (__v4su)__b);
476 static __inline__ __m128 __DEFAULT_FN_ATTRS
477 _mm_xor_ps(__m128 __a, __m128 __b)
479 return (__m128)((__v4su)__a ^ (__v4su)__b);
498 static __inline__ __m128 __DEFAULT_FN_ATTRS
499 _mm_cmpeq_ss(__m128 __a, __m128 __b)
501 return (__m128)__builtin_ia32_cmpeqss((__v4sf)__a, (__v4sf)__b);
516 static __inline__ __m128 __DEFAULT_FN_ATTRS
517 _mm_cmpeq_ps(__m128 __a, __m128 __b)
519 return (__m128)__builtin_ia32_cmpeqps((__v4sf)__a, (__v4sf)__b);
539 static __inline__ __m128 __DEFAULT_FN_ATTRS
540 _mm_cmplt_ss(__m128 __a, __m128 __b)
542 return (__m128)__builtin_ia32_cmpltss((__v4sf)__a, (__v4sf)__b);
558 static __inline__ __m128 __DEFAULT_FN_ATTRS
559 _mm_cmplt_ps(__m128 __a, __m128 __b)
561 return (__m128)__builtin_ia32_cmpltps((__v4sf)__a, (__v4sf)__b);
582 static __inline__ __m128 __DEFAULT_FN_ATTRS
583 _mm_cmple_ss(__m128 __a, __m128 __b)
585 return (__m128)__builtin_ia32_cmpless((__v4sf)__a, (__v4sf)__b);
601 static __inline__ __m128 __DEFAULT_FN_ATTRS
602 _mm_cmple_ps(__m128 __a, __m128 __b)
604 return (__m128)__builtin_ia32_cmpleps((__v4sf)__a, (__v4sf)__b);
624 static __inline__ __m128 __DEFAULT_FN_ATTRS
625 _mm_cmpgt_ss(__m128 __a, __m128 __b)
627 return (__m128)__builtin_shufflevector((__v4sf)__a,
645 static __inline__ __m128 __DEFAULT_FN_ATTRS
646 _mm_cmpgt_ps(__m128 __a, __m128 __b)
648 return (__m128)__builtin_ia32_cmpltps((__v4sf)__b, (__v4sf)__a);
669 static __inline__ __m128 __DEFAULT_FN_ATTRS
670 _mm_cmpge_ss(__m128 __a, __m128 __b)
672 return (__m128)__builtin_shufflevector((__v4sf)__a,
690 static __inline__ __m128 __DEFAULT_FN_ATTRS
691 _mm_cmpge_ps(__m128 __a, __m128 __b)
693 return (__m128)__builtin_ia32_cmpleps((__v4sf)__b, (__v4sf)__a);
713 static __inline__ __m128 __DEFAULT_FN_ATTRS
714 _mm_cmpneq_ss(__m128 __a, __m128 __b)
716 return (__m128)__builtin_ia32_cmpneqss((__v4sf)__a, (__v4sf)__b);
732 static __inline__ __m128 __DEFAULT_FN_ATTRS
733 _mm_cmpneq_ps(__m128 __a, __m128 __b)
735 return (__m128)__builtin_ia32_cmpneqps((__v4sf)__a, (__v4sf)__b);
756 static __inline__ __m128 __DEFAULT_FN_ATTRS
757 _mm_cmpnlt_ss(__m128 __a, __m128 __b)
759 return (__m128)__builtin_ia32_cmpnltss((__v4sf)__a, (__v4sf)__b);
776 static __inline__ __m128 __DEFAULT_FN_ATTRS
777 _mm_cmpnlt_ps(__m128 __a, __m128 __b)
779 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__a, (__v4sf)__b);
801 static __inline__ __m128 __DEFAULT_FN_ATTRS
802 _mm_cmpnle_ss(__m128 __a, __m128 __b)
804 return (__m128)__builtin_ia32_cmpnless((__v4sf)__a, (__v4sf)__b);
821 static __inline__ __m128 __DEFAULT_FN_ATTRS
822 _mm_cmpnle_ps(__m128 __a, __m128 __b)
824 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__a, (__v4sf)__b);
846 static __inline__ __m128 __DEFAULT_FN_ATTRS
847 _mm_cmpngt_ss(__m128 __a, __m128 __b)
849 return (__m128)__builtin_shufflevector((__v4sf)__a,
868 static __inline__ __m128 __DEFAULT_FN_ATTRS
869 __m128 __a, __m128 __b)
871 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__b, (__v4sf)__a);
893 static __inline__ __m128 __DEFAULT_FN_ATTRS
894 _mm_cmpnge_ss(__m128 __a, __m128 __b)
896 return (__m128)__builtin_shufflevector((__v4sf)__a,
915 static __inline__ __m128 __DEFAULT_FN_ATTRS
916 _mm_cmpnge_ps(__m128 __a, __m128 __b)
918 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__b, (__v4sf)__a);
940 static __inline__ __m128 __DEFAULT_FN_ATTRS
941 _mm_cmpord_ss(__m128 __a, __m128 __b)
943 return (__m128)__builtin_ia32_cmpordss((__v4sf)__a, (__v4sf)__b);
960 static __inline__ __m128 __DEFAULT_FN_ATTRS
961 _mm_cmpord_ps(__m128 __a, __m128 __b)
963 return (__m128)__builtin_ia32_cmpordps((__v4sf)__a, (__v4sf)__b);
985 static __inline__ __m128 __DEFAULT_FN_ATTRS
986 _mm_cmpunord_ss(__m128 __a, __m128 __b)
988 return (__m128)__builtin_ia32_cmpunordss((__v4sf)__a, (__v4sf)__b);
1005 static __inline__ __m128 __DEFAULT_FN_ATTRS
1006 _mm_cmpunord_ps(__m128 __a, __m128 __b)
1008 return (__m128)__builtin_ia32_cmpunordps((__v4sf)__a, (__v4sf)__b);
1027 _mm_comieq_ss(__m128 __a, __m128 __b)
1049 _mm_comilt_ss(__m128 __a, __m128 __b)
1070 _mm_comile_ss(__m128 __a, __m128 __b)
1091 _mm_comigt_ss(__m128 __a, __m128 __b)
1112 _mm_comige_ss(__m128 __a, __m128 __b)
1133 _mm_comineq_ss(__m128 __a, __m128 __b)
1154 _mm_ucomieq_ss(__m128 __a, __m128 __b)
1175 _mm_ucomilt_ss(__m128 __a, __m128 __b)
1197 _mm_ucomile_ss(__m128 __a, __m128 __b)
1219 _mm_ucomigt_ss(__m128 __a, __m128 __b)
1241 _mm_ucomige_ss(__m128 __a, __m128 __b)
1262 _mm_ucomineq_ss(__m128 __a, __m128 __b)
1280 _mm_cvtss_si32(__m128 __a)
1298 _mm_cvt_ss2si(__m128 __a)
1318 _mm_cvtss_si64(__m128 __a)
1336 _mm_cvtps_pi32(__m128 __a)
1352 _mm_cvt_ps2pi(__m128 __a)
1371 _mm_cvttss_si32(__m128 __a)
1390 _mm_cvtt_ss2si(__m128 __a)
1410 _mm_cvttss_si64(__m128 __a)
1429 _mm_cvttps_pi32(__m128 __a)
1446 _mm_cvtt_ps2pi(__m128 __a)
1467 static __inline__ __m128 __DEFAULT_FN_ATTRS
1468 _mm_cvtsi32_ss(__m128 __a, int __b)
1490 static __inline__ __m128 __DEFAULT_FN_ATTRS
1491 _mm_cvt_si2ss(__m128 __a, int __b)
1514 static __inline__ __m128 __DEFAULT_FN_ATTRS
1515 _mm_cvtsi64_ss(__m128 __a, long long __b)
1540 static __inline__ __m128 __DEFAULT_FN_ATTRS
1541 _mm_cvtpi32_ps(__m128 __a, __m64 __b)
1563 static __inline__ __m128 __DEFAULT_FN_ATTRS
1564 _mm_cvt_pi2ps(__m128 __a, __m64 __b)
1581 _mm_cvtss_f32(__m128 __a)
1601 static __inline__ __m128 __DEFAULT_FN_ATTRS
1602 _mm_loadh_pi(__m128 __a, const __m64 *__p)
1609 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
1628 static __inline__ __m128 __DEFAULT_FN_ATTRS
1629 _mm_loadl_pi(__m128 __a, const __m64 *__p)
1636 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
1655 static __inline__ __m128 __DEFAULT_FN_ATTRS
1662 return (__m128){ __u, 0, 0, 0 };
1677 static __inline__ __m128 __DEFAULT_FN_ATTRS
1684 return (__m128){ __u, __u, __u, __u };
1700 static __inline__ __m128 __DEFAULT_FN_ATTRS
1703 return *(__m128*)__p;
1717 static __inline__ __m128 __DEFAULT_FN_ATTRS
1721 __m128 __v;
1739 static __inline__ __m128 __DEFAULT_FN_ATTRS
1742 __m128 __a = _mm_load_ps(__p);
1753 static __inline__ __m128 __DEFAULT_FN_ATTRS
1756 return (__m128)__builtin_ia32_undef128();
1773 static __inline__ __m128 __DEFAULT_FN_ATTRS
1776 return (__m128){ __w, 0, 0, 0 };
1791 static __inline__ __m128 __DEFAULT_FN_ATTRS
1794 return (__m128){ __w, __w, __w, __w };
1810 static __inline__ __m128 __DEFAULT_FN_ATTRS
1837 static __inline__ __m128 __DEFAULT_FN_ATTRS
1840 return (__m128){ __w, __x, __y, __z };
1865 static __inline__ __m128 __DEFAULT_FN_ATTRS
1868 return (__m128){ __z, __y, __x, __w };
1880 static __inline__ __m128 __DEFAULT_FN_ATTRS
1883 return (__m128){ 0, 0, 0, 0 };
1898 _mm_storeh_pi(__m64 *__p, __m128 __a)
1915 _mm_storel_pi(__m64 *__p, __m128 __a)
1932 _mm_store_ss(float *__p, __m128 __a)
1953 _mm_storeu_ps(float *__p, __m128 __a)
1956 __m128 __v;
1974 _mm_store_ps(float *__p, __m128 __a)
1976 *(__m128*)__p = __a;
1993 _mm_store1_ps(float *__p, __m128 __a)
2013 _mm_store_ps1(float *__p, __m128 __a)
2032 _mm_storer_ps(float *__p, __m128 __a)
2106 _mm_stream_ps(float *__p, __m128 __a)
2532 /// __m128 _mm_shuffle_ps(__m128 a, __m128 b, const int mask);
2563 (__m128)__builtin_shufflevector((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), \
2585 static __inline__ __m128 __DEFAULT_FN_ATTRS
2586 _mm_unpackhi_ps(__m128 __a, __m128 __b)
2607 static __inline__ __m128 __DEFAULT_FN_ATTRS
2608 _mm_unpacklo_ps(__m128 __a, __m128 __b)
2628 static __inline__ __m128 __DEFAULT_FN_ATTRS
2629 _mm_move_ss(__m128 __a, __m128 __b)
2649 static __inline__ __m128 __DEFAULT_FN_ATTRS
2650 _mm_movehl_ps(__m128 __a, __m128 __b)
2670 static __inline__ __m128 __DEFAULT_FN_ATTRS
2671 _mm_movelh_ps(__m128 __a, __m128 __b)
2689 static __inline__ __m128 __DEFAULT_FN_ATTRS
2693 __m128 __r;
2720 static __inline__ __m128 __DEFAULT_FN_ATTRS
2724 __m128 __r;
2750 static __inline__ __m128 __DEFAULT_FN_ATTRS
2776 static __inline__ __m128 __DEFAULT_FN_ATTRS
2804 static __inline__ __m128 __DEFAULT_FN_ATTRS
2807 __m128 __c;
2834 _mm_cvtps_pi16(__m128 __a)
2864 _mm_cvtps_pi8(__m128 __a)
2889 _mm_movemask_ps(__m128 __a)
2937 __m128 tmp3, tmp2, tmp1, tmp0; \