Lines Matching refs:__m128
31 typedef float __m128 __attribute__((__vector_size__(16))); typedef
60 static __inline__ __m128 __DEFAULT_FN_ATTRS
61 _mm_add_ss(__m128 __a, __m128 __b) in _mm_add_ss()
80 static __inline__ __m128 __DEFAULT_FN_ATTRS
81 _mm_add_ps(__m128 __a, __m128 __b) in _mm_add_ps()
83 return (__m128)((__v4sf)__a + (__v4sf)__b); in _mm_add_ps()
102 static __inline__ __m128 __DEFAULT_FN_ATTRS
103 _mm_sub_ss(__m128 __a, __m128 __b) in _mm_sub_ss()
123 static __inline__ __m128 __DEFAULT_FN_ATTRS
124 _mm_sub_ps(__m128 __a, __m128 __b) in _mm_sub_ps()
126 return (__m128)((__v4sf)__a - (__v4sf)__b); in _mm_sub_ps()
145 static __inline__ __m128 __DEFAULT_FN_ATTRS
146 _mm_mul_ss(__m128 __a, __m128 __b) in _mm_mul_ss()
165 static __inline__ __m128 __DEFAULT_FN_ATTRS
166 _mm_mul_ps(__m128 __a, __m128 __b) in _mm_mul_ps()
168 return (__m128)((__v4sf)__a * (__v4sf)__b); in _mm_mul_ps()
187 static __inline__ __m128 __DEFAULT_FN_ATTRS
188 _mm_div_ss(__m128 __a, __m128 __b) in _mm_div_ss()
206 static __inline__ __m128 __DEFAULT_FN_ATTRS
207 _mm_div_ps(__m128 __a, __m128 __b) in _mm_div_ps()
209 return (__m128)((__v4sf)__a / (__v4sf)__b); in _mm_div_ps()
224 static __inline__ __m128 __DEFAULT_FN_ATTRS
225 _mm_sqrt_ss(__m128 __a) in _mm_sqrt_ss()
227 __m128 __c = __builtin_ia32_sqrtss((__v4sf)__a); in _mm_sqrt_ss()
228 return (__m128) { __c[0], __a[1], __a[2], __a[3] }; in _mm_sqrt_ss()
242 static __inline__ __m128 __DEFAULT_FN_ATTRS
243 _mm_sqrt_ps(__m128 __a) in _mm_sqrt_ps()
260 static __inline__ __m128 __DEFAULT_FN_ATTRS
261 _mm_rcp_ss(__m128 __a) in _mm_rcp_ss()
263 __m128 __c = __builtin_ia32_rcpss((__v4sf)__a); in _mm_rcp_ss()
264 return (__m128) { __c[0], __a[1], __a[2], __a[3] }; in _mm_rcp_ss()
278 static __inline__ __m128 __DEFAULT_FN_ATTRS
279 _mm_rcp_ps(__m128 __a) in _mm_rcp_ps()
297 static __inline__ __m128 __DEFAULT_FN_ATTRS
298 _mm_rsqrt_ss(__m128 __a) in _mm_rsqrt_ss()
300 __m128 __c = __builtin_ia32_rsqrtss((__v4sf)__a); in _mm_rsqrt_ss()
301 return (__m128) { __c[0], __a[1], __a[2], __a[3] }; in _mm_rsqrt_ss()
315 static __inline__ __m128 __DEFAULT_FN_ATTRS
316 _mm_rsqrt_ps(__m128 __a) in _mm_rsqrt_ps()
338 static __inline__ __m128 __DEFAULT_FN_ATTRS
339 _mm_min_ss(__m128 __a, __m128 __b) in _mm_min_ss()
357 static __inline__ __m128 __DEFAULT_FN_ATTRS
358 _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()
399 static __inline__ __m128 __DEFAULT_FN_ATTRS
400 _mm_max_ps(__m128 __a, __m128 __b) in _mm_max_ps()
417 static __inline__ __m128 __DEFAULT_FN_ATTRS
418 _mm_and_ps(__m128 __a, __m128 __b) in _mm_and_ps()
420 return (__m128)((__v4su)__a & (__v4su)__b); in _mm_and_ps()
439 static __inline__ __m128 __DEFAULT_FN_ATTRS
440 _mm_andnot_ps(__m128 __a, __m128 __b) in _mm_andnot_ps()
442 return (__m128)(~(__v4su)__a & (__v4su)__b); in _mm_andnot_ps()
457 static __inline__ __m128 __DEFAULT_FN_ATTRS
458 _mm_or_ps(__m128 __a, __m128 __b) in _mm_or_ps()
460 return (__m128)((__v4su)__a | (__v4su)__b); in _mm_or_ps()
476 static __inline__ __m128 __DEFAULT_FN_ATTRS
477 _mm_xor_ps(__m128 __a, __m128 __b) in _mm_xor_ps()
479 return (__m128)((__v4su)__a ^ (__v4su)__b); in _mm_xor_ps()
498 static __inline__ __m128 __DEFAULT_FN_ATTRS
499 _mm_cmpeq_ss(__m128 __a, __m128 __b) in _mm_cmpeq_ss()
501 return (__m128)__builtin_ia32_cmpeqss((__v4sf)__a, (__v4sf)__b); in _mm_cmpeq_ss()
516 static __inline__ __m128 __DEFAULT_FN_ATTRS
517 _mm_cmpeq_ps(__m128 __a, __m128 __b) in _mm_cmpeq_ps()
519 return (__m128)__builtin_ia32_cmpeqps((__v4sf)__a, (__v4sf)__b); in _mm_cmpeq_ps()
539 static __inline__ __m128 __DEFAULT_FN_ATTRS
540 _mm_cmplt_ss(__m128 __a, __m128 __b) in _mm_cmplt_ss()
542 return (__m128)__builtin_ia32_cmpltss((__v4sf)__a, (__v4sf)__b); in _mm_cmplt_ss()
558 static __inline__ __m128 __DEFAULT_FN_ATTRS
559 _mm_cmplt_ps(__m128 __a, __m128 __b) in _mm_cmplt_ps()
561 return (__m128)__builtin_ia32_cmpltps((__v4sf)__a, (__v4sf)__b); in _mm_cmplt_ps()
582 static __inline__ __m128 __DEFAULT_FN_ATTRS
583 _mm_cmple_ss(__m128 __a, __m128 __b) in _mm_cmple_ss()
585 return (__m128)__builtin_ia32_cmpless((__v4sf)__a, (__v4sf)__b); in _mm_cmple_ss()
601 static __inline__ __m128 __DEFAULT_FN_ATTRS
602 _mm_cmple_ps(__m128 __a, __m128 __b) in _mm_cmple_ps()
604 return (__m128)__builtin_ia32_cmpleps((__v4sf)__a, (__v4sf)__b); in _mm_cmple_ps()
624 static __inline__ __m128 __DEFAULT_FN_ATTRS
625 _mm_cmpgt_ss(__m128 __a, __m128 __b) in _mm_cmpgt_ss()
627 return (__m128)__builtin_shufflevector((__v4sf)__a, in _mm_cmpgt_ss()
645 static __inline__ __m128 __DEFAULT_FN_ATTRS
646 _mm_cmpgt_ps(__m128 __a, __m128 __b) in _mm_cmpgt_ps()
648 return (__m128)__builtin_ia32_cmpltps((__v4sf)__b, (__v4sf)__a); in _mm_cmpgt_ps()
669 static __inline__ __m128 __DEFAULT_FN_ATTRS
670 _mm_cmpge_ss(__m128 __a, __m128 __b) in _mm_cmpge_ss()
672 return (__m128)__builtin_shufflevector((__v4sf)__a, in _mm_cmpge_ss()
690 static __inline__ __m128 __DEFAULT_FN_ATTRS
691 _mm_cmpge_ps(__m128 __a, __m128 __b) in _mm_cmpge_ps()
693 return (__m128)__builtin_ia32_cmpleps((__v4sf)__b, (__v4sf)__a); in _mm_cmpge_ps()
712 static __inline__ __m128 __DEFAULT_FN_ATTRS
713 _mm_cmpneq_ss(__m128 __a, __m128 __b) in _mm_cmpneq_ss()
715 return (__m128)__builtin_ia32_cmpneqss((__v4sf)__a, (__v4sf)__b); in _mm_cmpneq_ss()
730 static __inline__ __m128 __DEFAULT_FN_ATTRS
731 _mm_cmpneq_ps(__m128 __a, __m128 __b) in _mm_cmpneq_ps()
733 return (__m128)__builtin_ia32_cmpneqps((__v4sf)__a, (__v4sf)__b); in _mm_cmpneq_ps()
753 static __inline__ __m128 __DEFAULT_FN_ATTRS
754 _mm_cmpnlt_ss(__m128 __a, __m128 __b) in _mm_cmpnlt_ss()
756 return (__m128)__builtin_ia32_cmpnltss((__v4sf)__a, (__v4sf)__b); in _mm_cmpnlt_ss()
772 static __inline__ __m128 __DEFAULT_FN_ATTRS
773 _mm_cmpnlt_ps(__m128 __a, __m128 __b) in _mm_cmpnlt_ps()
775 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__a, (__v4sf)__b); in _mm_cmpnlt_ps()
796 static __inline__ __m128 __DEFAULT_FN_ATTRS
797 _mm_cmpnle_ss(__m128 __a, __m128 __b) in _mm_cmpnle_ss()
799 return (__m128)__builtin_ia32_cmpnless((__v4sf)__a, (__v4sf)__b); in _mm_cmpnle_ss()
815 static __inline__ __m128 __DEFAULT_FN_ATTRS
816 _mm_cmpnle_ps(__m128 __a, __m128 __b) in _mm_cmpnle_ps()
818 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__a, (__v4sf)__b); in _mm_cmpnle_ps()
839 static __inline__ __m128 __DEFAULT_FN_ATTRS
840 _mm_cmpngt_ss(__m128 __a, __m128 __b) in _mm_cmpngt_ss()
842 return (__m128)__builtin_shufflevector((__v4sf)__a, in _mm_cmpngt_ss()
860 static __inline__ __m128 __DEFAULT_FN_ATTRS
861 _mm_cmpngt_ps(__m128 __a, __m128 __b) in _mm_cmpngt_ps()
863 return (__m128)__builtin_ia32_cmpnltps((__v4sf)__b, (__v4sf)__a); in _mm_cmpngt_ps()
884 static __inline__ __m128 __DEFAULT_FN_ATTRS
885 _mm_cmpnge_ss(__m128 __a, __m128 __b) in _mm_cmpnge_ss()
887 return (__m128)__builtin_shufflevector((__v4sf)__a, in _mm_cmpnge_ss()
905 static __inline__ __m128 __DEFAULT_FN_ATTRS
906 _mm_cmpnge_ps(__m128 __a, __m128 __b) in _mm_cmpnge_ps()
908 return (__m128)__builtin_ia32_cmpnleps((__v4sf)__b, (__v4sf)__a); in _mm_cmpnge_ps()
929 static __inline__ __m128 __DEFAULT_FN_ATTRS
930 _mm_cmpord_ss(__m128 __a, __m128 __b) in _mm_cmpord_ss()
932 return (__m128)__builtin_ia32_cmpordss((__v4sf)__a, (__v4sf)__b); in _mm_cmpord_ss()
948 static __inline__ __m128 __DEFAULT_FN_ATTRS
949 _mm_cmpord_ps(__m128 __a, __m128 __b) in _mm_cmpord_ps()
951 return (__m128)__builtin_ia32_cmpordps((__v4sf)__a, (__v4sf)__b); in _mm_cmpord_ps()
972 static __inline__ __m128 __DEFAULT_FN_ATTRS
973 _mm_cmpunord_ss(__m128 __a, __m128 __b) in _mm_cmpunord_ss()
975 return (__m128)__builtin_ia32_cmpunordss((__v4sf)__a, (__v4sf)__b); in _mm_cmpunord_ss()
991 static __inline__ __m128 __DEFAULT_FN_ATTRS
992 _mm_cmpunord_ps(__m128 __a, __m128 __b) in _mm_cmpunord_ps()
994 return (__m128)__builtin_ia32_cmpunordps((__v4sf)__a, (__v4sf)__b); in _mm_cmpunord_ps()
1012 _mm_comieq_ss(__m128 __a, __m128 __b) in _mm_comieq_ss()
1033 _mm_comilt_ss(__m128 __a, __m128 __b) in _mm_comilt_ss()
1054 _mm_comile_ss(__m128 __a, __m128 __b) in _mm_comile_ss()
1075 _mm_comigt_ss(__m128 __a, __m128 __b) in _mm_comigt_ss()
1096 _mm_comige_ss(__m128 __a, __m128 __b) in _mm_comige_ss()
1117 _mm_comineq_ss(__m128 __a, __m128 __b) in _mm_comineq_ss()
1138 _mm_ucomieq_ss(__m128 __a, __m128 __b) in _mm_ucomieq_ss()
1159 _mm_ucomilt_ss(__m128 __a, __m128 __b) in _mm_ucomilt_ss()
1181 _mm_ucomile_ss(__m128 __a, __m128 __b) in _mm_ucomile_ss()
1203 _mm_ucomigt_ss(__m128 __a, __m128 __b) in _mm_ucomigt_ss()
1225 _mm_ucomige_ss(__m128 __a, __m128 __b) in _mm_ucomige_ss()
1246 _mm_ucomineq_ss(__m128 __a, __m128 __b) in _mm_ucomineq_ss()
1263 _mm_cvtss_si32(__m128 __a) in _mm_cvtss_si32()
1280 _mm_cvt_ss2si(__m128 __a) in _mm_cvt_ss2si()
1299 _mm_cvtss_si64(__m128 __a) in _mm_cvtss_si64()
1317 _mm_cvtps_pi32(__m128 __a) in _mm_cvtps_pi32()
1333 _mm_cvt_ps2pi(__m128 __a) in _mm_cvt_ps2pi()
1351 _mm_cvttss_si32(__m128 __a) in _mm_cvttss_si32()
1369 _mm_cvtt_ss2si(__m128 __a) in _mm_cvtt_ss2si()
1387 _mm_cvttss_si64(__m128 __a) in _mm_cvttss_si64()
1404 _mm_cvttps_pi32(__m128 __a) in _mm_cvttps_pi32()
1421 _mm_cvtt_ps2pi(__m128 __a) in _mm_cvtt_ps2pi()
1442 static __inline__ __m128 __DEFAULT_FN_ATTRS
1443 _mm_cvtsi32_ss(__m128 __a, int __b) in _mm_cvtsi32_ss()
1465 static __inline__ __m128 __DEFAULT_FN_ATTRS
1466 _mm_cvt_si2ss(__m128 __a, int __b) in _mm_cvt_si2ss()
1489 static __inline__ __m128 __DEFAULT_FN_ATTRS
1490 _mm_cvtsi64_ss(__m128 __a, long long __b) in _mm_cvtsi64_ss()
1515 static __inline__ __m128 __DEFAULT_FN_ATTRS
1516 _mm_cvtpi32_ps(__m128 __a, __m64 __b) in _mm_cvtpi32_ps()
1538 static __inline__ __m128 __DEFAULT_FN_ATTRS
1539 _mm_cvt_pi2ps(__m128 __a, __m64 __b) in _mm_cvt_pi2ps()
1556 _mm_cvtss_f32(__m128 __a) in _mm_cvtss_f32()
1576 static __inline__ __m128 __DEFAULT_FN_ATTRS
1577 _mm_loadh_pi(__m128 __a, const __m64 *__p) in _mm_loadh_pi()
1584 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1); in _mm_loadh_pi()
1603 static __inline__ __m128 __DEFAULT_FN_ATTRS
1604 _mm_loadl_pi(__m128 __a, const __m64 *__p) in _mm_loadl_pi()
1611 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1); in _mm_loadl_pi()
1630 static __inline__ __m128 __DEFAULT_FN_ATTRS
1637 return (__m128){ __u, 0, 0, 0 }; in _mm_load_ss()
1652 static __inline__ __m128 __DEFAULT_FN_ATTRS
1659 return (__m128){ __u, __u, __u, __u }; in _mm_load1_ps()
1675 static __inline__ __m128 __DEFAULT_FN_ATTRS
1678 return *(__m128*)__p; in _mm_load_ps()
1692 static __inline__ __m128 __DEFAULT_FN_ATTRS
1696 __m128 __v; in _mm_loadu_ps()
1714 static __inline__ __m128 __DEFAULT_FN_ATTRS
1717 __m128 __a = _mm_load_ps(__p); in _mm_loadr_ps()
1729 static __inline__ __m128 __DEFAULT_FN_ATTRS
1732 return (__m128)__builtin_ia32_undef128(); in _mm_undefined_ps()
1749 static __inline__ __m128 __DEFAULT_FN_ATTRS
1752 return (__m128){ __w, 0, 0, 0 }; in _mm_set_ss()
1767 static __inline__ __m128 __DEFAULT_FN_ATTRS
1770 return (__m128){ __w, __w, __w, __w }; in _mm_set1_ps()
1786 static __inline__ __m128 __DEFAULT_FN_ATTRS
1813 static __inline__ __m128 __DEFAULT_FN_ATTRS
1816 return (__m128){ __w, __x, __y, __z }; in _mm_set_ps()
1841 static __inline__ __m128 __DEFAULT_FN_ATTRS
1844 return (__m128){ __z, __y, __x, __w }; in _mm_setr_ps()
1856 static __inline__ __m128 __DEFAULT_FN_ATTRS
1859 return (__m128){ 0, 0, 0, 0 }; in _mm_setzero_ps()
1874 _mm_storeh_pi(__m64 *__p, __m128 __a) in _mm_storeh_pi()
1891 _mm_storel_pi(__m64 *__p, __m128 __a) in _mm_storel_pi()
1908 _mm_store_ss(float *__p, __m128 __a) in _mm_store_ss()
1929 _mm_storeu_ps(float *__p, __m128 __a) in _mm_storeu_ps()
1932 __m128 __v; in _mm_storeu_ps()
1951 _mm_store_ps(float *__p, __m128 __a) in _mm_store_ps()
1953 *(__m128*)__p = __a; in _mm_store_ps()
1970 _mm_store1_ps(float *__p, __m128 __a) in _mm_store1_ps()
1989 _mm_store_ps1(float *__p, __m128 __a) in _mm_store_ps1()
2008 _mm_storer_ps(float *__p, __m128 __a) in _mm_storer_ps()
2081 _mm_stream_ps(float *__p, __m128 __a) in _mm_stream_ps()
2489 (__m128)__builtin_shufflevector((__v4sf)(__m128)(a), (__v4sf)(__m128)(b), \
2512 static __inline__ __m128 __DEFAULT_FN_ATTRS
2513 _mm_unpackhi_ps(__m128 __a, __m128 __b) in _mm_unpackhi_ps()
2535 static __inline__ __m128 __DEFAULT_FN_ATTRS
2536 _mm_unpacklo_ps(__m128 __a, __m128 __b) in _mm_unpacklo_ps()
2556 static __inline__ __m128 __DEFAULT_FN_ATTRS
2557 _mm_move_ss(__m128 __a, __m128 __b) in _mm_move_ss()
2577 static __inline__ __m128 __DEFAULT_FN_ATTRS
2578 _mm_movehl_ps(__m128 __a, __m128 __b) in _mm_movehl_ps()
2598 static __inline__ __m128 __DEFAULT_FN_ATTRS
2599 _mm_movelh_ps(__m128 __a, __m128 __b) in _mm_movelh_ps()
2616 static __inline__ __m128 __DEFAULT_FN_ATTRS
2620 __m128 __r; in _mm_cvtpi16_ps()
2646 static __inline__ __m128 __DEFAULT_FN_ATTRS
2650 __m128 __r; in _mm_cvtpu16_ps()
2675 static __inline__ __m128 __DEFAULT_FN_ATTRS
2700 static __inline__ __m128 __DEFAULT_FN_ATTRS
2727 static __inline__ __m128 __DEFAULT_FN_ATTRS
2730 __m128 __c; in _mm_cvtpi32x2_ps()
2756 _mm_cvtps_pi16(__m128 __a) in _mm_cvtps_pi16()
2785 _mm_cvtps_pi8(__m128 __a) in _mm_cvtps_pi8()
2810 _mm_movemask_ps(__m128 __a) in _mm_movemask_ps()
2858 __m128 tmp3, tmp2, tmp1, tmp0; \