Lines Matching refs:__m128
21 typedef float __m128 __attribute__((__vector_size__(16), __aligned__(16))); typedef
57 static __inline__ __m128 __DEFAULT_FN_ATTRS
58 _mm_add_ss(__m128 __a, __m128 __b) in _mm_add_ss()
77 static __inline__ __m128 __DEFAULT_FN_ATTRS
78 _mm_add_ps(__m128 __a, __m128 __b) in _mm_add_ps()
80 return (__m128)((__v4sf)__a + (__v4sf)__b); in _mm_add_ps()
99 static __inline__ __m128 __DEFAULT_FN_ATTRS
100 _mm_sub_ss(__m128 __a, __m128 __b) in _mm_sub_ss()
120 static __inline__ __m128 __DEFAULT_FN_ATTRS
121 _mm_sub_ps(__m128 __a, __m128 __b) in _mm_sub_ps()
123 return (__m128)((__v4sf)__a - (__v4sf)__b); in _mm_sub_ps()
142 static __inline__ __m128 __DEFAULT_FN_ATTRS
143 _mm_mul_ss(__m128 __a, __m128 __b) in _mm_mul_ss()
162 static __inline__ __m128 __DEFAULT_FN_ATTRS
163 _mm_mul_ps(__m128 __a, __m128 __b) in _mm_mul_ps()
165 return (__m128)((__v4sf)__a * (__v4sf)__b); in _mm_mul_ps()
184 static __inline__ __m128 __DEFAULT_FN_ATTRS
185 _mm_div_ss(__m128 __a, __m128 __b) in _mm_div_ss()
203 static __inline__ __m128 __DEFAULT_FN_ATTRS
204 _mm_div_ps(__m128 __a, __m128 __b) in _mm_div_ps()
206 return (__m128)((__v4sf)__a / (__v4sf)__b); in _mm_div_ps()
221 static __inline__ __m128 __DEFAULT_FN_ATTRS
222 _mm_sqrt_ss(__m128 __a) in _mm_sqrt_ss()
224 return (__m128)__builtin_ia32_sqrtss((__v4sf)__a); in _mm_sqrt_ss()
238 static __inline__ __m128 __DEFAULT_FN_ATTRS
239 _mm_sqrt_ps(__m128 __a) in _mm_sqrt_ps()
256 static __inline__ __m128 __DEFAULT_FN_ATTRS
257 _mm_rcp_ss(__m128 __a) in _mm_rcp_ss()
259 return (__m128)__builtin_ia32_rcpss((__v4sf)__a); in _mm_rcp_ss()
273 static __inline__ __m128 __DEFAULT_FN_ATTRS
274 _mm_rcp_ps(__m128 __a) in _mm_rcp_ps()
276 return (__m128)__builtin_ia32_rcpps((__v4sf)__a); in _mm_rcp_ps()
292 static __inline__ __m128 __DEFAULT_FN_ATTRS
293 _mm_rsqrt_ss(__m128 __a) in _mm_rsqrt_ss()
309 static __inline__ __m128 __DEFAULT_FN_ATTRS
310 _mm_rsqrt_ps(__m128 __a) in _mm_rsqrt_ps()
334 static __inline__ __m128 __DEFAULT_FN_ATTRS
335 _mm_min_ss(__m128 __a, __m128 __b) in _mm_min_ss()
355 static __inline__ __m128 __DEFAULT_FN_ATTRS
356 _mm_min_ps(__m128 __a, __m128 __b) in _mm_min_ps()
380 static __inline__ __m128 __DEFAULT_FN_ATTRS
381 _mm_max_ss(__m128 __a, __m128 __b) in _mm_max_ss()
401 static __inline__ __m128 __DEFAULT_FN_ATTRS
402 _mm_max_ps(__m128 __a, __m128 __b) in _mm_max_ps()
419 static __inline__ __m128 __DEFAULT_FN_ATTRS
420 _mm_and_ps(__m128 __a, __m128 __b) in _mm_and_ps()
422 return (__m128)((__v4su)__a & (__v4su)__b); in _mm_and_ps()
441 static __inline__ __m128 __DEFAULT_FN_ATTRS
442 _mm_andnot_ps(__m128 __a, __m128 __b) in _mm_andnot_ps()
444 return (__m128)(~(__v4su)__a & (__v4su)__b); in _mm_andnot_ps()
459 static __inline__ __m128 __DEFAULT_FN_ATTRS
460 _mm_or_ps(__m128 __a, __m128 __b) in _mm_or_ps()
462 return (__m128)((__v4su)__a | (__v4su)__b); in _mm_or_ps()
478 static __inline__ __m128 __DEFAULT_FN_ATTRS
479 _mm_xor_ps(__m128 __a, __m128 __b) in _mm_xor_ps()
481 return (__m128)((__v4su)__a ^ (__v4su)__b); in _mm_xor_ps()
503 static __inline__ __m128 __DEFAULT_FN_ATTRS
504 _mm_cmpeq_ss(__m128 __a, __m128 __b) in _mm_cmpeq_ss()
506 return (__m128)__builtin_ia32_cmpeqss((__v4sf)__a, (__v4sf)__b); in _mm_cmpeq_ss()
524 static __inline__ __m128 __DEFAULT_FN_ATTRS
525 _mm_cmpeq_ps(__m128 __a, __m128 __b) in _mm_cmpeq_ps()
527 return (__m128)__builtin_ia32_cmpeqps((__v4sf)__a, (__v4sf)__b); in _mm_cmpeq_ps()
550 static __inline__ __m128 __DEFAULT_FN_ATTRS
551 _mm_cmplt_ss(__m128 __a, __m128 __b) in _mm_cmplt_ss()
553 return (__m128)__builtin_ia32_cmpltss((__v4sf)__a, (__v4sf)__b); in _mm_cmplt_ss()
572 static __inline__ __m128 __DEFAULT_FN_ATTRS
573 _mm_cmplt_ps(__m128 __a, __m128 __b) in _mm_cmplt_ps()
575 return (__m128)__builtin_ia32_cmpltps((__v4sf)__a, (__v4sf)__b); in _mm_cmplt_ps()
598 static __inline__ __m128 __DEFAULT_FN_ATTRS
599 _mm_cmple_ss(__m128 __a, __m128 __b) in _mm_cmple_ss()
601 return (__m128)__builtin_ia32_cmpless((__v4sf)__a, (__v4sf)__b); in _mm_cmple_ss()
620 static __inline__ __m128 __DEFAULT_FN_ATTRS
621 _mm_cmple_ps(__m128 __a, __m128 __b) in _mm_cmple_ps()
623 return (__m128)__builtin_ia32_cmpleps((__v4sf)__a, (__v4sf)__b); in _mm_cmple_ps()
646 static __inline__ __m128 __DEFAULT_FN_ATTRS
647 _mm_cmpgt_ss(__m128 __a, __m128 __b) in _mm_cmpgt_ss()
649 return (__m128)__builtin_shufflevector((__v4sf)__a, in _mm_cmpgt_ss()
670 static __inline__ __m128 __DEFAULT_FN_ATTRS
671 _mm_cmpgt_ps(__m128 __a, __m128 __b) in _mm_cmpgt_ps()
673 return (__m128)__builtin_ia32_cmpltps((__v4sf)__b, (__v4sf)__a); in _mm_cmpgt_ps()
696 static __inline__ __m128 __DEFAULT_FN_ATTRS
697 _mm_cmpge_ss(__m128 __a, __m128 __b) in _mm_cmpge_ss()
699 return (__m128)__builtin_shufflevector((__v4sf)__a, in _mm_cmpge_ss()
720 static __inline__ __m128 __DEFAULT_FN_ATTRS
721 _mm_cmpge_ps(__m128 __a, __m128 __b) in _mm_cmpge_ps()
723 return (__m128)__builtin_ia32_cmpleps((__v4sf)__b, (__v4sf)__a); in _mm_cmpge_ps()
746 static __inline__ __m128 __DEFAULT_FN_ATTRS
747 _mm_cmpneq_ss(__m128 __a, __m128 __b) in _mm_cmpneq_ss()
749 return (__m128)__builtin_ia32_cmpneqss((__v4sf)__a, (__v4sf)__b); in _mm_cmpneq_ss()
768 static __inline__ __m128 __DEFAULT_FN_ATTRS
769 _mm_cmpneq_ps(__m128 __a, __m128 __b) in _mm_cmpneq_ps()
771 return (__m128)__builtin_ia32_cmpneqps((__v4sf)__a, (__v4sf)__b); in _mm_cmpneq_ps()
795 static __inline__ __m128 __DEFAULT_FN_ATTRS
796 _mm_cmpnlt_ss(__m128 __a, __m128 __b) in _mm_cmpnlt_ss()
798 return (__m128)__builtin_ia32_cmpnltss((__v4sf)__a, (__v4sf)__b); in _mm_cmpnlt_ss()
818 static __inline__ __m128 __DEFAULT_FN_ATTRS
819 _mm_cmpnlt_ps(__m128 __a, __m128 __b) in _mm_cmpnlt_ps()
821 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__a, (__v4sf)__b); in _mm_cmpnlt_ps()
845 static __inline__ __m128 __DEFAULT_FN_ATTRS
846 _mm_cmpnle_ss(__m128 __a, __m128 __b) in _mm_cmpnle_ss()
848 return (__m128)__builtin_ia32_cmpnless((__v4sf)__a, (__v4sf)__b); in _mm_cmpnle_ss()
868 static __inline__ __m128 __DEFAULT_FN_ATTRS
869 _mm_cmpnle_ps(__m128 __a, __m128 __b) in _mm_cmpnle_ps()
871 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__a, (__v4sf)__b); in _mm_cmpnle_ps()
895 static __inline__ __m128 __DEFAULT_FN_ATTRS
896 _mm_cmpngt_ss(__m128 __a, __m128 __b) in _mm_cmpngt_ss()
898 return (__m128)__builtin_shufflevector((__v4sf)__a, in _mm_cmpngt_ss()
920 static __inline__ __m128 __DEFAULT_FN_ATTRS
921 _mm_cmpngt_ps(__m128 __a, __m128 __b) in _mm_cmpngt_ps()
923 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__b, (__v4sf)__a); in _mm_cmpngt_ps()
947 static __inline__ __m128 __DEFAULT_FN_ATTRS
948 _mm_cmpnge_ss(__m128 __a, __m128 __b) in _mm_cmpnge_ss()
950 return (__m128)__builtin_shufflevector((__v4sf)__a, in _mm_cmpnge_ss()
972 static __inline__ __m128 __DEFAULT_FN_ATTRS
973 _mm_cmpnge_ps(__m128 __a, __m128 __b) in _mm_cmpnge_ps()
975 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__b, (__v4sf)__a); in _mm_cmpnge_ps()
999 static __inline__ __m128 __DEFAULT_FN_ATTRS
1000 _mm_cmpord_ss(__m128 __a, __m128 __b) in _mm_cmpord_ss()
1002 return (__m128)__builtin_ia32_cmpordss((__v4sf)__a, (__v4sf)__b); in _mm_cmpord_ss()
1023 static __inline__ __m128 __DEFAULT_FN_ATTRS
1024 _mm_cmpord_ps(__m128 __a, __m128 __b) in _mm_cmpord_ps()
1026 return (__m128)__builtin_ia32_cmpordps((__v4sf)__a, (__v4sf)__b); in _mm_cmpord_ps()
1050 static __inline__ __m128 __DEFAULT_FN_ATTRS
1051 _mm_cmpunord_ss(__m128 __a, __m128 __b) in _mm_cmpunord_ss()
1053 return (__m128)__builtin_ia32_cmpunordss((__v4sf)__a, (__v4sf)__b); in _mm_cmpunord_ss()
1074 static __inline__ __m128 __DEFAULT_FN_ATTRS
1075 _mm_cmpunord_ps(__m128 __a, __m128 __b) in _mm_cmpunord_ps()
1077 return (__m128)__builtin_ia32_cmpunordps((__v4sf)__a, (__v4sf)__b); in _mm_cmpunord_ps()
1099 _mm_comieq_ss(__m128 __a, __m128 __b) in _mm_comieq_ss()
1124 _mm_comilt_ss(__m128 __a, __m128 __b) in _mm_comilt_ss()
1148 _mm_comile_ss(__m128 __a, __m128 __b) in _mm_comile_ss()
1172 _mm_comigt_ss(__m128 __a, __m128 __b) in _mm_comigt_ss()
1196 _mm_comige_ss(__m128 __a, __m128 __b) in _mm_comige_ss()
1220 _mm_comineq_ss(__m128 __a, __m128 __b) in _mm_comineq_ss()
1243 _mm_ucomieq_ss(__m128 __a, __m128 __b) in _mm_ucomieq_ss()
1267 _mm_ucomilt_ss(__m128 __a, __m128 __b) in _mm_ucomilt_ss()
1291 _mm_ucomile_ss(__m128 __a, __m128 __b) in _mm_ucomile_ss()
1315 _mm_ucomigt_ss(__m128 __a, __m128 __b) in _mm_ucomigt_ss()
1339 _mm_ucomige_ss(__m128 __a, __m128 __b) in _mm_ucomige_ss()
1362 _mm_ucomineq_ss(__m128 __a, __m128 __b) in _mm_ucomineq_ss()
1384 _mm_cvtss_si32(__m128 __a) in _mm_cvtss_si32()
1406 _mm_cvt_ss2si(__m128 __a) in _mm_cvt_ss2si()
1430 _mm_cvtss_si64(__m128 __a) in _mm_cvtss_si64()
1452 _mm_cvtps_pi32(__m128 __a) in _mm_cvtps_pi32()
1472 _mm_cvt_ps2pi(__m128 __a) in _mm_cvt_ps2pi()
1494 _mm_cvttss_si32(__m128 __a) in _mm_cvttss_si32()
1516 _mm_cvtt_ss2si(__m128 __a) in _mm_cvtt_ss2si()
1539 _mm_cvttss_si64(__m128 __a) in _mm_cvttss_si64()
1562 _mm_cvttps_pi32(__m128 __a) in _mm_cvttps_pi32()
1583 _mm_cvtt_ps2pi(__m128 __a) in _mm_cvtt_ps2pi()
1604 static __inline__ __m128 __DEFAULT_FN_ATTRS
1605 _mm_cvtsi32_ss(__m128 __a, int __b) in _mm_cvtsi32_ss()
1627 static __inline__ __m128 __DEFAULT_FN_ATTRS
1628 _mm_cvt_si2ss(__m128 __a, int __b) in _mm_cvt_si2ss()
1651 static __inline__ __m128 __DEFAULT_FN_ATTRS
1652 _mm_cvtsi64_ss(__m128 __a, long long __b) in _mm_cvtsi64_ss()
1677 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
1678 _mm_cvtpi32_ps(__m128 __a, __m64 __b) in _mm_cvtpi32_ps()
1700 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
1701 _mm_cvt_pi2ps(__m128 __a, __m64 __b) in _mm_cvt_pi2ps()
1718 _mm_cvtss_f32(__m128 __a) in _mm_cvtss_f32()
1738 static __inline__ __m128 __DEFAULT_FN_ATTRS
1739 _mm_loadh_pi(__m128 __a, const __m64 *__p) in _mm_loadh_pi()
1746 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1); in _mm_loadh_pi()
1765 static __inline__ __m128 __DEFAULT_FN_ATTRS
1766 _mm_loadl_pi(__m128 __a, const __m64 *__p) in _mm_loadl_pi()
1773 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1); in _mm_loadl_pi()
1792 static __inline__ __m128 __DEFAULT_FN_ATTRS
1799 return __extension__ (__m128){ __u, 0, 0, 0 }; in _mm_load_ss()
1814 static __inline__ __m128 __DEFAULT_FN_ATTRS
1821 return __extension__ (__m128){ __u, __u, __u, __u }; in _mm_load1_ps()
1837 static __inline__ __m128 __DEFAULT_FN_ATTRS
1840 return *(const __m128*)__p; in _mm_load_ps()
1854 static __inline__ __m128 __DEFAULT_FN_ATTRS
1876 static __inline__ __m128 __DEFAULT_FN_ATTRS
1879 __m128 __a = _mm_load_ps(__p); in _mm_loadr_ps()
1890 static __inline__ __m128 __DEFAULT_FN_ATTRS
1893 return (__m128)__builtin_ia32_undef128(); in _mm_undefined_ps()
1910 static __inline__ __m128 __DEFAULT_FN_ATTRS
1913 return __extension__ (__m128){ __w, 0.0f, 0.0f, 0.0f }; in _mm_set_ss()
1928 static __inline__ __m128 __DEFAULT_FN_ATTRS
1931 return __extension__ (__m128){ __w, __w, __w, __w }; in _mm_set1_ps()
1947 static __inline__ __m128 __DEFAULT_FN_ATTRS
1974 static __inline__ __m128 __DEFAULT_FN_ATTRS
1977 return __extension__ (__m128){ __w, __x, __y, __z }; in _mm_set_ps()
2002 static __inline__ __m128 __DEFAULT_FN_ATTRS
2005 return __extension__ (__m128){ __z, __y, __x, __w }; in _mm_setr_ps()
2017 static __inline__ __m128 __DEFAULT_FN_ATTRS
2020 return __extension__ (__m128){ 0.0f, 0.0f, 0.0f, 0.0f }; in _mm_setzero_ps()
2035 _mm_storeh_pi(__m64 *__p, __m128 __a) in _mm_storeh_pi()
2056 _mm_storel_pi(__m64 *__p, __m128 __a) in _mm_storel_pi()
2077 _mm_store_ss(float *__p, __m128 __a) in _mm_store_ss()
2098 _mm_storeu_ps(float *__p, __m128 __a) in _mm_storeu_ps()
2119 _mm_store_ps(float *__p, __m128 __a) in _mm_store_ps()
2121 *(__m128*)__p = __a; in _mm_store_ps()
2138 _mm_store1_ps(float *__p, __m128 __a) in _mm_store1_ps()
2158 _mm_store_ps1(float *__p, __m128 __a) in _mm_store_ps1()
2177 _mm_storer_ps(float *__p, __m128 __a) in _mm_storer_ps()
2254 _mm_stream_ps(void *__p, __m128 __a) in _mm_stream_ps()
2725 ((__m128)__builtin_ia32_shufps((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), \
2744 static __inline__ __m128 __DEFAULT_FN_ATTRS
2745 _mm_unpackhi_ps(__m128 __a, __m128 __b) in _mm_unpackhi_ps()
2766 static __inline__ __m128 __DEFAULT_FN_ATTRS
2767 _mm_unpacklo_ps(__m128 __a, __m128 __b) in _mm_unpacklo_ps()
2788 static __inline__ __m128 __DEFAULT_FN_ATTRS
2789 _mm_move_ss(__m128 __a, __m128 __b) in _mm_move_ss()
2810 static __inline__ __m128 __DEFAULT_FN_ATTRS
2811 _mm_movehl_ps(__m128 __a, __m128 __b) in _mm_movehl_ps()
2831 static __inline__ __m128 __DEFAULT_FN_ATTRS
2832 _mm_movelh_ps(__m128 __a, __m128 __b) in _mm_movelh_ps()
2849 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2853 __m128 __r; in _mm_cvtpi16_ps()
2879 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2883 __m128 __r; in _mm_cvtpu16_ps()
2908 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2933 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2960 static __inline__ __m128 __DEFAULT_FN_ATTRS_MMX
2963 __m128 __c; in _mm_cvtpi32x2_ps()
2990 _mm_cvtps_pi16(__m128 __a) in _mm_cvtps_pi16()
3020 _mm_cvtps_pi8(__m128 __a) in _mm_cvtps_pi8()
3045 _mm_movemask_ps(__m128 __a) in _mm_movemask_ps()
3093 ((__m128)__builtin_ia32_cmpps((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), (c)))
3128 ((__m128)__builtin_ia32_cmpss((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), (c)))
3172 __m128 tmp3, tmp2, tmp1, tmp0; \