61 _mm_add_ss(__m128 __a, __m128 __b)  in _mm_add_ss()
81 _mm_add_ps(__m128 __a, __m128 __b)  in _mm_add_ps()
103 _mm_sub_ss(__m128 __a, __m128 __b)  in _mm_sub_ss()
124 _mm_sub_ps(__m128 __a, __m128 __b)  in _mm_sub_ps()
146 _mm_mul_ss(__m128 __a, __m128 __b)  in _mm_mul_ss()
166 _mm_mul_ps(__m128 __a, __m128 __b)  in _mm_mul_ps()
188 _mm_div_ss(__m128 __a, __m128 __b)  in _mm_div_ss()
207 _mm_div_ps(__m128 __a, __m128 __b)  in _mm_div_ps()
225 _mm_sqrt_ss(__m128 __a)  in _mm_sqrt_ss()
243 _mm_sqrt_ps(__m128 __a)  in _mm_sqrt_ps()
261 _mm_rcp_ss(__m128 __a)  in _mm_rcp_ss()
279 _mm_rcp_ps(__m128 __a)  in _mm_rcp_ps()
298 _mm_rsqrt_ss(__m128 __a)  in _mm_rsqrt_ss()
316 _mm_rsqrt_ps(__m128 __a)  in _mm_rsqrt_ps()
339 _mm_min_ss(__m128 __a, __m128 __b)  in _mm_min_ss()
358 _mm_min_ps(__m128 __a, __m128 __b)  in _mm_min_ps()
381 _mm_max_ss(__m128 __a, __m128 __b)  in _mm_max_ss()
400 _mm_max_ps(__m128 __a, __m128 __b)  in _mm_max_ps()
418 _mm_and_ps(__m128 __a, __m128 __b)  in _mm_and_ps()
440 _mm_andnot_ps(__m128 __a, __m128 __b)  in _mm_andnot_ps()
458 _mm_or_ps(__m128 __a, __m128 __b)  in _mm_or_ps()
477 _mm_xor_ps(__m128 __a, __m128 __b)  in _mm_xor_ps()
499 _mm_cmpeq_ss(__m128 __a, __m128 __b)  in _mm_cmpeq_ss()
517 _mm_cmpeq_ps(__m128 __a, __m128 __b)  in _mm_cmpeq_ps()
540 _mm_cmplt_ss(__m128 __a, __m128 __b)  in _mm_cmplt_ss()
559 _mm_cmplt_ps(__m128 __a, __m128 __b)  in _mm_cmplt_ps()
583 _mm_cmple_ss(__m128 __a, __m128 __b)  in _mm_cmple_ss()
602 _mm_cmple_ps(__m128 __a, __m128 __b)  in _mm_cmple_ps()
625 _mm_cmpgt_ss(__m128 __a, __m128 __b)  in _mm_cmpgt_ss()
646 _mm_cmpgt_ps(__m128 __a, __m128 __b)  in _mm_cmpgt_ps()
670 _mm_cmpge_ss(__m128 __a, __m128 __b)  in _mm_cmpge_ss()
691 _mm_cmpge_ps(__m128 __a, __m128 __b)  in _mm_cmpge_ps()
713 _mm_cmpneq_ss(__m128 __a, __m128 __b)  in _mm_cmpneq_ss()
731 _mm_cmpneq_ps(__m128 __a, __m128 __b)  in _mm_cmpneq_ps()
754 _mm_cmpnlt_ss(__m128 __a, __m128 __b)  in _mm_cmpnlt_ss()
773 _mm_cmpnlt_ps(__m128 __a, __m128 __b)  in _mm_cmpnlt_ps()
797 _mm_cmpnle_ss(__m128 __a, __m128 __b)  in _mm_cmpnle_ss()
816 _mm_cmpnle_ps(__m128 __a, __m128 __b)  in _mm_cmpnle_ps()
840 _mm_cmpngt_ss(__m128 __a, __m128 __b)  in _mm_cmpngt_ss()
861 _mm_cmpngt_ps(__m128 __a, __m128 __b)  in _mm_cmpngt_ps()
885 _mm_cmpnge_ss(__m128 __a, __m128 __b)  in _mm_cmpnge_ss()
906 _mm_cmpnge_ps(__m128 __a, __m128 __b)  in _mm_cmpnge_ps()
930 _mm_cmpord_ss(__m128 __a, __m128 __b)  in _mm_cmpord_ss()
949 _mm_cmpord_ps(__m128 __a, __m128 __b)  in _mm_cmpord_ps()
973 _mm_cmpunord_ss(__m128 __a, __m128 __b)  in _mm_cmpunord_ss()
992 _mm_cmpunord_ps(__m128 __a, __m128 __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()
1443 _mm_cvtsi32_ss(__m128 __a, int __b)  in _mm_cvtsi32_ss()
1466 _mm_cvt_si2ss(__m128 __a, int __b)  in _mm_cvt_si2ss()
1490 _mm_cvtsi64_ss(__m128 __a, long long __b)  in _mm_cvtsi64_ss()
1516 _mm_cvtpi32_ps(__m128 __a, __m64 __b)  in _mm_cvtpi32_ps()
1539 _mm_cvt_pi2ps(__m128 __a, __m64 __b)  in _mm_cvt_pi2ps()
1556 _mm_cvtss_f32(__m128 __a)  in _mm_cvtss_f32()
1577 _mm_loadh_pi(__m128 __a, const __m64 *__p)  in _mm_loadh_pi()
1604 _mm_loadl_pi(__m128 __a, const __m64 *__p)  in _mm_loadl_pi()
1717   __m128 __a = _mm_load_ps(__p);  in _mm_loadr_ps()  local
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()
1951 _mm_store_ps(float *__p, __m128 __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()
2062 _mm_stream_pi(__m64 *__p, __m64 __a)  in _mm_stream_pi()
2081 _mm_stream_ps(float *__p, __m128 __a)  in _mm_stream_ps()
2161 _mm_max_pi16(__m64 __a, __m64 __b)  in _mm_max_pi16()
2180 _mm_max_pu8(__m64 __a, __m64 __b)  in _mm_max_pu8()
2199 _mm_min_pi16(__m64 __a, __m64 __b)  in _mm_min_pi16()
2218 _mm_min_pu8(__m64 __a, __m64 __b)  in _mm_min_pu8()
2236 _mm_movemask_pi8(__m64 __a)  in _mm_movemask_pi8()
2255 _mm_mulhi_pu16(__m64 __a, __m64 __b)  in _mm_mulhi_pu16()
2331 _mm_avg_pu8(__m64 __a, __m64 __b)  in _mm_avg_pu8()
2350 _mm_avg_pu16(__m64 __a, __m64 __b)  in _mm_avg_pu16()
2372 _mm_sad_pu8(__m64 __a, __m64 __b)  in _mm_sad_pu8()
2513 _mm_unpackhi_ps(__m128 __a, __m128 __b)  in _mm_unpackhi_ps()
2536 _mm_unpacklo_ps(__m128 __a, __m128 __b)  in _mm_unpacklo_ps()
2557 _mm_move_ss(__m128 __a, __m128 __b)  in _mm_move_ss()
2578 _mm_movehl_ps(__m128 __a, __m128 __b)  in _mm_movehl_ps()
2599 _mm_movelh_ps(__m128 __a, __m128 __b)  in _mm_movelh_ps()
2617 _mm_cvtpi16_ps(__m64 __a)  in _mm_cvtpi16_ps()
2647 _mm_cvtpu16_ps(__m64 __a)  in _mm_cvtpu16_ps()
2676 _mm_cvtpi8_ps(__m64 __a)  in _mm_cvtpi8_ps()
2701 _mm_cvtpu8_ps(__m64 __a)  in _mm_cvtpu8_ps()
2728 _mm_cvtpi32x2_ps(__m64 __a, __m64 __b)  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()