50 _mm_add_ss(__m128 __a, __m128 __b)  in _mm_add_ss()
70 _mm_add_ps(__m128 __a, __m128 __b)  in _mm_add_ps()
92 _mm_sub_ss(__m128 __a, __m128 __b)  in _mm_sub_ss()
113 _mm_sub_ps(__m128 __a, __m128 __b)  in _mm_sub_ps()
135 _mm_mul_ss(__m128 __a, __m128 __b)  in _mm_mul_ss()
155 _mm_mul_ps(__m128 __a, __m128 __b)  in _mm_mul_ps()
177 _mm_div_ss(__m128 __a, __m128 __b)  in _mm_div_ss()
196 _mm_div_ps(__m128 __a, __m128 __b)  in _mm_div_ps()
214 _mm_sqrt_ss(__m128 __a)  in _mm_sqrt_ss()
231 _mm_sqrt_ps(__m128 __a)  in _mm_sqrt_ps()
249 _mm_rcp_ss(__m128 __a)  in _mm_rcp_ss()
266 _mm_rcp_ps(__m128 __a)  in _mm_rcp_ps()
285 _mm_rsqrt_ss(__m128 __a)  in _mm_rsqrt_ss()
302 _mm_rsqrt_ps(__m128 __a)  in _mm_rsqrt_ps()
325 _mm_min_ss(__m128 __a, __m128 __b)  in _mm_min_ss()
344 _mm_min_ps(__m128 __a, __m128 __b)  in _mm_min_ps()
367 _mm_max_ss(__m128 __a, __m128 __b)  in _mm_max_ss()
386 _mm_max_ps(__m128 __a, __m128 __b)  in _mm_max_ps()
404 _mm_and_ps(__m128 __a, __m128 __b)  in _mm_and_ps()
426 _mm_andnot_ps(__m128 __a, __m128 __b)  in _mm_andnot_ps()
444 _mm_or_ps(__m128 __a, __m128 __b)  in _mm_or_ps()
463 _mm_xor_ps(__m128 __a, __m128 __b)  in _mm_xor_ps()
485 _mm_cmpeq_ss(__m128 __a, __m128 __b)  in _mm_cmpeq_ss()
503 _mm_cmpeq_ps(__m128 __a, __m128 __b)  in _mm_cmpeq_ps()
526 _mm_cmplt_ss(__m128 __a, __m128 __b)  in _mm_cmplt_ss()
545 _mm_cmplt_ps(__m128 __a, __m128 __b)  in _mm_cmplt_ps()
569 _mm_cmple_ss(__m128 __a, __m128 __b)  in _mm_cmple_ss()
588 _mm_cmple_ps(__m128 __a, __m128 __b)  in _mm_cmple_ps()
611 _mm_cmpgt_ss(__m128 __a, __m128 __b)  in _mm_cmpgt_ss()
632 _mm_cmpgt_ps(__m128 __a, __m128 __b)  in _mm_cmpgt_ps()
656 _mm_cmpge_ss(__m128 __a, __m128 __b)  in _mm_cmpge_ss()
677 _mm_cmpge_ps(__m128 __a, __m128 __b)  in _mm_cmpge_ps()
700 _mm_cmpneq_ss(__m128 __a, __m128 __b)  in _mm_cmpneq_ss()
719 _mm_cmpneq_ps(__m128 __a, __m128 __b)  in _mm_cmpneq_ps()
743 _mm_cmpnlt_ss(__m128 __a, __m128 __b)  in _mm_cmpnlt_ss()
763 _mm_cmpnlt_ps(__m128 __a, __m128 __b)  in _mm_cmpnlt_ps()
788 _mm_cmpnle_ss(__m128 __a, __m128 __b)  in _mm_cmpnle_ss()
808 _mm_cmpnle_ps(__m128 __a, __m128 __b)  in _mm_cmpnle_ps()
833 _mm_cmpngt_ss(__m128 __a, __m128 __b)  in _mm_cmpngt_ss()
855 _mm_cmpngt_ps(__m128 __a, __m128 __b)  in _mm_cmpngt_ps()
880 _mm_cmpnge_ss(__m128 __a, __m128 __b)  in _mm_cmpnge_ss()
902 _mm_cmpnge_ps(__m128 __a, __m128 __b)  in _mm_cmpnge_ps()
927 _mm_cmpord_ss(__m128 __a, __m128 __b)  in _mm_cmpord_ss()
947 _mm_cmpord_ps(__m128 __a, __m128 __b)  in _mm_cmpord_ps()
972 _mm_cmpunord_ss(__m128 __a, __m128 __b)  in _mm_cmpunord_ss()
992 _mm_cmpunord_ps(__m128 __a, __m128 __b)  in _mm_cmpunord_ps()
1016 _mm_comieq_ss(__m128 __a, __m128 __b)  in _mm_comieq_ss()
1041 _mm_comilt_ss(__m128 __a, __m128 __b)  in _mm_comilt_ss()
1065 _mm_comile_ss(__m128 __a, __m128 __b)  in _mm_comile_ss()
1089 _mm_comigt_ss(__m128 __a, __m128 __b)  in _mm_comigt_ss()
1113 _mm_comige_ss(__m128 __a, __m128 __b)  in _mm_comige_ss()
1137 _mm_comineq_ss(__m128 __a, __m128 __b)  in _mm_comineq_ss()
1161 _mm_ucomieq_ss(__m128 __a, __m128 __b)  in _mm_ucomieq_ss()
1185 _mm_ucomilt_ss(__m128 __a, __m128 __b)  in _mm_ucomilt_ss()
1210 _mm_ucomile_ss(__m128 __a, __m128 __b)  in _mm_ucomile_ss()
1235 _mm_ucomigt_ss(__m128 __a, __m128 __b)  in _mm_ucomigt_ss()
1260 _mm_ucomige_ss(__m128 __a, __m128 __b)  in _mm_ucomige_ss()
1284 _mm_ucomineq_ss(__m128 __a, __m128 __b)  in _mm_ucomineq_ss()
1302 _mm_cvtss_si32(__m128 __a)  in _mm_cvtss_si32()
1320 _mm_cvt_ss2si(__m128 __a)  in _mm_cvt_ss2si()
1340 _mm_cvtss_si64(__m128 __a)  in _mm_cvtss_si64()
1358 _mm_cvtps_pi32(__m128 __a)  in _mm_cvtps_pi32()
1374 _mm_cvt_ps2pi(__m128 __a)  in _mm_cvt_ps2pi()
1393 _mm_cvttss_si32(__m128 __a)  in _mm_cvttss_si32()
1412 _mm_cvtt_ss2si(__m128 __a)  in _mm_cvtt_ss2si()
1432 _mm_cvttss_si64(__m128 __a)  in _mm_cvttss_si64()
1451 _mm_cvttps_pi32(__m128 __a)  in _mm_cvttps_pi32()
1468 _mm_cvtt_ps2pi(__m128 __a)  in _mm_cvtt_ps2pi()
1490 _mm_cvtsi32_ss(__m128 __a, int __b)  in _mm_cvtsi32_ss()
1513 _mm_cvt_si2ss(__m128 __a, int __b)  in _mm_cvt_si2ss()
1537 _mm_cvtsi64_ss(__m128 __a, long long __b)  in _mm_cvtsi64_ss()
1563 _mm_cvtpi32_ps(__m128 __a, __m64 __b)  in _mm_cvtpi32_ps()
1586 _mm_cvt_pi2ps(__m128 __a, __m64 __b)  in _mm_cvt_pi2ps()
1603 _mm_cvtss_f32(__m128 __a)  in _mm_cvtss_f32()
1624 _mm_loadh_pi(__m128 __a, const __m64 *__p)  in _mm_loadh_pi()
1651 _mm_loadl_pi(__m128 __a, const __m64 *__p)  in _mm_loadl_pi()
1764   __m128 __a = _mm_load_ps(__p);  in _mm_loadr_ps()  local
1920 _mm_storeh_pi(__m64 *__p, __m128 __a)  in _mm_storeh_pi()
1941 _mm_storel_pi(__m64 *__p, __m128 __a)  in _mm_storel_pi()
1962 _mm_store_ss(float *__p, __m128 __a)  in _mm_store_ss()
1983 _mm_storeu_ps(float *__p, __m128 __a)  in _mm_storeu_ps()
2004 _mm_store_ps(float *__p, __m128 __a)  in _mm_store_ps()
2023 _mm_store1_ps(float *__p, __m128 __a)  in _mm_store1_ps()
2043 _mm_store_ps1(float *__p, __m128 __a)  in _mm_store_ps1()
2062 _mm_storer_ps(float *__p, __m128 __a)  in _mm_storer_ps()
2120 _mm_stream_pi(__m64 *__p, __m64 __a)  in _mm_stream_pi()
2139 _mm_stream_ps(float *__p, __m128 __a)  in _mm_stream_ps()
2231 _mm_max_pi16(__m64 __a, __m64 __b)  in _mm_max_pi16()
2250 _mm_max_pu8(__m64 __a, __m64 __b)  in _mm_max_pu8()
2269 _mm_min_pi16(__m64 __a, __m64 __b)  in _mm_min_pi16()
2288 _mm_min_pu8(__m64 __a, __m64 __b)  in _mm_min_pu8()
2306 _mm_movemask_pi8(__m64 __a)  in _mm_movemask_pi8()
2325 _mm_mulhi_pu16(__m64 __a, __m64 __b)  in _mm_mulhi_pu16()
2407 _mm_avg_pu8(__m64 __a, __m64 __b)  in _mm_avg_pu8()
2426 _mm_avg_pu16(__m64 __a, __m64 __b)  in _mm_avg_pu16()
2448 _mm_sad_pu8(__m64 __a, __m64 __b)  in _mm_sad_pu8()
2624 _mm_unpackhi_ps(__m128 __a, __m128 __b)  in _mm_unpackhi_ps()
2646 _mm_unpacklo_ps(__m128 __a, __m128 __b)  in _mm_unpacklo_ps()
2668 _mm_move_ss(__m128 __a, __m128 __b)  in _mm_move_ss()
2690 _mm_movehl_ps(__m128 __a, __m128 __b)  in _mm_movehl_ps()
2711 _mm_movelh_ps(__m128 __a, __m128 __b)  in _mm_movelh_ps()
2729 _mm_cvtpi16_ps(__m64 __a)  in _mm_cvtpi16_ps()
2759 _mm_cvtpu16_ps(__m64 __a)  in _mm_cvtpu16_ps()
2788 _mm_cvtpi8_ps(__m64 __a)  in _mm_cvtpi8_ps()
2813 _mm_cvtpu8_ps(__m64 __a)  in _mm_cvtpu8_ps()
2840 _mm_cvtpi32x2_ps(__m64 __a, __m64 __b)  in _mm_cvtpi32x2_ps()
2869 _mm_cvtps_pi16(__m128 __a)  in _mm_cvtps_pi16()
2899 _mm_cvtps_pi8(__m128 __a)  in _mm_cvtps_pi8()
2924 _mm_movemask_ps(__m128 __a)  in _mm_movemask_ps()