Lines Matching defs:__m128

21 typedef float __m128 __attribute__((__vector_size__(16), __aligned__(16)));
86 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
87 _mm_add_ss(__m128 __a, __m128 __b) {
105 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
106 _mm_add_ps(__m128 __a, __m128 __b) {
107 return (__m128)((__v4sf)__a + (__v4sf)__b);
126 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
127 _mm_sub_ss(__m128 __a, __m128 __b) {
146 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
147 _mm_sub_ps(__m128 __a, __m128 __b) {
148 return (__m128)((__v4sf)__a - (__v4sf)__b);
167 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
168 _mm_mul_ss(__m128 __a, __m128 __b) {
186 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
187 _mm_mul_ps(__m128 __a, __m128 __b) {
188 return (__m128)((__v4sf)__a * (__v4sf)__b);
207 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
208 _mm_div_ss(__m128 __a, __m128 __b) {
225 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
226 _mm_div_ps(__m128 __a, __m128 __b) {
227 return (__m128)((__v4sf)__a / (__v4sf)__b);
242 static __inline__ __m128 __DEFAULT_FN_ATTRS
243 _mm_sqrt_ss(__m128 __a)
245 return (__m128)__builtin_ia32_sqrtss((__v4sf)__a);
259 static __inline__ __m128 __DEFAULT_FN_ATTRS
260 _mm_sqrt_ps(__m128 __a)
277 static __inline__ __m128 __DEFAULT_FN_ATTRS
278 _mm_rcp_ss(__m128 __a)
280 return (__m128)__builtin_ia32_rcpss((__v4sf)__a);
294 static __inline__ __m128 __DEFAULT_FN_ATTRS
295 _mm_rcp_ps(__m128 __a)
297 return (__m128)__builtin_ia32_rcpps((__v4sf)__a);
313 static __inline__ __m128 __DEFAULT_FN_ATTRS
314 _mm_rsqrt_ss(__m128 __a)
330 static __inline__ __m128 __DEFAULT_FN_ATTRS
331 _mm_rsqrt_ps(__m128 __a)
355 static __inline__ __m128 __DEFAULT_FN_ATTRS
356 _mm_min_ss(__m128 __a, __m128 __b)
376 static __inline__ __m128 __DEFAULT_FN_ATTRS
377 _mm_min_ps(__m128 __a, __m128 __b)
401 static __inline__ __m128 __DEFAULT_FN_ATTRS
402 _mm_max_ss(__m128 __a, __m128 __b)
422 static __inline__ __m128 __DEFAULT_FN_ATTRS
423 _mm_max_ps(__m128 __a, __m128 __b)
440 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
441 _mm_and_ps(__m128 __a, __m128 __b) {
442 return (__m128)((__v4su)__a & (__v4su)__b);
461 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
462 _mm_andnot_ps(__m128 __a, __m128 __b) {
463 return (__m128)(~(__v4su)__a & (__v4su)__b);
478 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
479 _mm_or_ps(__m128 __a, __m128 __b) {
480 return (__m128)((__v4su)__a | (__v4su)__b);
496 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
497 _mm_xor_ps(__m128 __a, __m128 __b) {
498 return (__m128)((__v4su)__a ^ (__v4su)__b);
520 static __inline__ __m128 __DEFAULT_FN_ATTRS
521 _mm_cmpeq_ss(__m128 __a, __m128 __b)
523 return (__m128)__builtin_ia32_cmpeqss((__v4sf)__a, (__v4sf)__b);
541 static __inline__ __m128 __DEFAULT_FN_ATTRS
542 _mm_cmpeq_ps(__m128 __a, __m128 __b)
544 return (__m128)__builtin_ia32_cmpeqps((__v4sf)__a, (__v4sf)__b);
567 static __inline__ __m128 __DEFAULT_FN_ATTRS
568 _mm_cmplt_ss(__m128 __a, __m128 __b)
570 return (__m128)__builtin_ia32_cmpltss((__v4sf)__a, (__v4sf)__b);
589 static __inline__ __m128 __DEFAULT_FN_ATTRS
590 _mm_cmplt_ps(__m128 __a, __m128 __b)
592 return (__m128)__builtin_ia32_cmpltps((__v4sf)__a, (__v4sf)__b);
615 static __inline__ __m128 __DEFAULT_FN_ATTRS
616 _mm_cmple_ss(__m128 __a, __m128 __b)
618 return (__m128)__builtin_ia32_cmpless((__v4sf)__a, (__v4sf)__b);
637 static __inline__ __m128 __DEFAULT_FN_ATTRS
638 _mm_cmple_ps(__m128 __a, __m128 __b)
640 return (__m128)__builtin_ia32_cmpleps((__v4sf)__a, (__v4sf)__b);
663 static __inline__ __m128 __DEFAULT_FN_ATTRS
664 _mm_cmpgt_ss(__m128 __a, __m128 __b)
666 return (__m128)__builtin_shufflevector((__v4sf)__a,
687 static __inline__ __m128 __DEFAULT_FN_ATTRS
688 _mm_cmpgt_ps(__m128 __a, __m128 __b)
690 return (__m128)__builtin_ia32_cmpltps((__v4sf)__b, (__v4sf)__a);
713 static __inline__ __m128 __DEFAULT_FN_ATTRS
714 _mm_cmpge_ss(__m128 __a, __m128 __b)
716 return (__m128)__builtin_shufflevector((__v4sf)__a,
737 static __inline__ __m128 __DEFAULT_FN_ATTRS
738 _mm_cmpge_ps(__m128 __a, __m128 __b)
740 return (__m128)__builtin_ia32_cmpleps((__v4sf)__b, (__v4sf)__a);
763 static __inline__ __m128 __DEFAULT_FN_ATTRS
764 _mm_cmpneq_ss(__m128 __a, __m128 __b)
766 return (__m128)__builtin_ia32_cmpneqss((__v4sf)__a, (__v4sf)__b);
785 static __inline__ __m128 __DEFAULT_FN_ATTRS
786 _mm_cmpneq_ps(__m128 __a, __m128 __b)
788 return (__m128)__builtin_ia32_cmpneqps((__v4sf)__a, (__v4sf)__b);
812 static __inline__ __m128 __DEFAULT_FN_ATTRS
813 _mm_cmpnlt_ss(__m128 __a, __m128 __b)
815 return (__m128)__builtin_ia32_cmpnltss((__v4sf)__a, (__v4sf)__b);
835 static __inline__ __m128 __DEFAULT_FN_ATTRS
836 _mm_cmpnlt_ps(__m128 __a, __m128 __b)
838 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__a, (__v4sf)__b);
862 static __inline__ __m128 __DEFAULT_FN_ATTRS
863 _mm_cmpnle_ss(__m128 __a, __m128 __b)
865 return (__m128)__builtin_ia32_cmpnless((__v4sf)__a, (__v4sf)__b);
885 static __inline__ __m128 __DEFAULT_FN_ATTRS
886 _mm_cmpnle_ps(__m128 __a, __m128 __b)
888 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__a, (__v4sf)__b);
912 static __inline__ __m128 __DEFAULT_FN_ATTRS
913 _mm_cmpngt_ss(__m128 __a, __m128 __b)
915 return (__m128)__builtin_shufflevector((__v4sf)__a,
937 static __inline__ __m128 __DEFAULT_FN_ATTRS
938 _mm_cmpngt_ps(__m128 __a, __m128 __b)
940 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__b, (__v4sf)__a);
964 static __inline__ __m128 __DEFAULT_FN_ATTRS
965 _mm_cmpnge_ss(__m128 __a, __m128 __b)
967 return (__m128)__builtin_shufflevector((__v4sf)__a,
989 static __inline__ __m128 __DEFAULT_FN_ATTRS
990 _mm_cmpnge_ps(__m128 __a, __m128 __b)
992 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__b, (__v4sf)__a);
1016 static __inline__ __m128 __DEFAULT_FN_ATTRS
1017 _mm_cmpord_ss(__m128 __a, __m128 __b)
1019 return (__m128)__builtin_ia32_cmpordss((__v4sf)__a, (__v4sf)__b);
1040 static __inline__ __m128 __DEFAULT_FN_ATTRS
1041 _mm_cmpord_ps(__m128 __a, __m128 __b)
1043 return (__m128)__builtin_ia32_cmpordps((__v4sf)__a, (__v4sf)__b);
1067 static __inline__ __m128 __DEFAULT_FN_ATTRS
1068 _mm_cmpunord_ss(__m128 __a, __m128 __b)
1070 return (__m128)__builtin_ia32_cmpunordss((__v4sf)__a, (__v4sf)__b);
1091 static __inline__ __m128 __DEFAULT_FN_ATTRS
1092 _mm_cmpunord_ps(__m128 __a, __m128 __b)
1094 return (__m128)__builtin_ia32_cmpunordps((__v4sf)__a, (__v4sf)__b);
1116 _mm_comieq_ss(__m128 __a, __m128 __b)
1141 _mm_comilt_ss(__m128 __a, __m128 __b)
1165 _mm_comile_ss(__m128 __a, __m128 __b)
1189 _mm_comigt_ss(__m128 __a, __m128 __b)
1213 _mm_comige_ss(__m128 __a, __m128 __b)
1237 _mm_comineq_ss(__m128 __a, __m128 __b)
1260 _mm_ucomieq_ss(__m128 __a, __m128 __b)
1284 _mm_ucomilt_ss(__m128 __a, __m128 __b)
1308 _mm_ucomile_ss(__m128 __a, __m128 __b)
1332 _mm_ucomigt_ss(__m128 __a, __m128 __b)
1356 _mm_ucomige_ss(__m128 __a, __m128 __b)
1379 _mm_ucomineq_ss(__m128 __a, __m128 __b)
1401 _mm_cvtss_si32(__m128 __a)
1423 _mm_cvt_ss2si(__m128 __a)
1447 _mm_cvtss_si64(__m128 __a)
1469 _mm_cvtps_pi32(__m128 __a)
1489 _mm_cvt_ps2pi(__m128 __a)
1511 _mm_cvttss_si32(__m128 __a)
1533 _mm_cvtt_ss2si(__m128 __a)
1556 _mm_cvttss_si64(__m128 __a)
1579 _mm_cvttps_pi32(__m128 __a)
1600 _mm_cvtt_ps2pi(__m128 __a)
1621 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR _mm_cvtsi32_ss(__m128 __a,
1643 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR _mm_cvt_si2ss(__m128 __a,
1666 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
1667 _mm_cvtsi64_ss(__m128 __a, long long __b) {
1691 static __inline__ __m128 __DEFAULT_FN_ATTRS_SSE2
1692 _mm_cvtpi32_ps(__m128 __a, __m64 __b)
1694 return (__m128)__builtin_shufflevector(
1717 static __inline__ __m128 __DEFAULT_FN_ATTRS_SSE2
1718 _mm_cvt_pi2ps(__m128 __a, __m64 __b)
1735 _mm_cvtss_f32(__m128 __a) {
1754 static __inline__ __m128 __DEFAULT_FN_ATTRS
1755 _mm_loadh_pi(__m128 __a, const __m64 *__p)
1762 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
1781 static __inline__ __m128 __DEFAULT_FN_ATTRS
1782 _mm_loadl_pi(__m128 __a, const __m64 *__p)
1789 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
1808 static __inline__ __m128 __DEFAULT_FN_ATTRS
1815 return __extension__ (__m128){ __u, 0, 0, 0 };
1830 static __inline__ __m128 __DEFAULT_FN_ATTRS
1837 return __extension__ (__m128){ __u, __u, __u, __u };
1853 static __inline__ __m128 __DEFAULT_FN_ATTRS
1856 return *(const __m128*)__p;
1870 static __inline__ __m128 __DEFAULT_FN_ATTRS
1892 static __inline__ __m128 __DEFAULT_FN_ATTRS
1895 __m128 __a = _mm_load_ps(__p);
1906 static __inline__ __m128 __DEFAULT_FN_ATTRS
1909 return (__m128)__builtin_ia32_undef128();
1926 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
1928 return __extension__ (__m128){ __w, 0.0f, 0.0f, 0.0f };
1943 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
1945 return __extension__ (__m128){ __w, __w, __w, __w };
1961 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
1987 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
1989 return __extension__ (__m128){ __w, __x, __y, __z };
2014 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
2016 return __extension__ (__m128){ __z, __y, __x, __w };
2028 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
2030 return __extension__ (__m128){ 0.0f, 0.0f, 0.0f, 0.0f };
2045 _mm_storeh_pi(__m64 *__p, __m128 __a)
2066 _mm_storel_pi(__m64 *__p, __m128 __a)
2087 _mm_store_ss(float *__p, __m128 __a)
2108 _mm_storeu_ps(float *__p, __m128 __a)
2129 _mm_store_ps(float *__p, __m128 __a)
2131 *(__m128*)__p = __a;
2148 _mm_store1_ps(float *__p, __m128 __a)
2168 _mm_store_ps1(float *__p, __m128 __a)
2187 _mm_storer_ps(float *__p, __m128 __a)
2264 _mm_stream_ps(void *__p, __m128 __a)
2722 /// __m128 _mm_shuffle_ps(__m128 a, __m128 b, const int mask);
2756 ((__m128)__builtin_ia32_shufps((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), \
2775 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
2776 _mm_unpackhi_ps(__m128 __a, __m128 __b) {
2796 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
2797 _mm_unpacklo_ps(__m128 __a, __m128 __b) {
2817 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
2818 _mm_move_ss(__m128 __a, __m128 __b) {
2838 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
2839 _mm_movehl_ps(__m128 __a, __m128 __b) {
2858 static __inline__ __m128 __DEFAULT_FN_ATTRS_CONSTEXPR
2859 _mm_movelh_ps(__m128 __a, __m128 __b) {
2875 static __inline__ __m128 __DEFAULT_FN_ATTRS_SSE2
2893 static __inline__ __m128 __DEFAULT_FN_ATTRS_SSE2
2911 static __inline__ __m128 __DEFAULT_FN_ATTRS_SSE2
2932 static __inline__ __m128 __DEFAULT_FN_ATTRS_SSE2
2956 static __inline__ __m128 __DEFAULT_FN_ATTRS_SSE2
2982 _mm_cvtps_pi16(__m128 __a)
3007 _mm_cvtps_pi8(__m128 __a)
3032 _mm_movemask_ps(__m128 __a)
3058 /// __m128 _mm_cmp_ps(__m128 a, __m128 b, const int c);
3080 ((__m128)__builtin_ia32_cmpps((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), (c)))
3093 /// __m128 _mm_cmp_ss(__m128 a, __m128 b, const int c);
3115 ((__m128)__builtin_ia32_cmpss((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), (c)))
3159 __m128 tmp3, tmp2, tmp1, tmp0; \