54 _mm_add_ss(__m128 __a, __m128 __b)  in _mm_add_ss()
74 _mm_add_ps(__m128 __a, __m128 __b)  in _mm_add_ps()
96 _mm_sub_ss(__m128 __a, __m128 __b)  in _mm_sub_ss()
117 _mm_sub_ps(__m128 __a, __m128 __b)  in _mm_sub_ps()
139 _mm_mul_ss(__m128 __a, __m128 __b)  in _mm_mul_ss()
159 _mm_mul_ps(__m128 __a, __m128 __b)  in _mm_mul_ps()
181 _mm_div_ss(__m128 __a, __m128 __b)  in _mm_div_ss()
200 _mm_div_ps(__m128 __a, __m128 __b)  in _mm_div_ps()
218 _mm_sqrt_ss(__m128 __a)  in _mm_sqrt_ss()
235 _mm_sqrt_ps(__m128 __a)  in _mm_sqrt_ps()
253 _mm_rcp_ss(__m128 __a)  in _mm_rcp_ss()
270 _mm_rcp_ps(__m128 __a)  in _mm_rcp_ps()
289 _mm_rsqrt_ss(__m128 __a)  in _mm_rsqrt_ss()
306 _mm_rsqrt_ps(__m128 __a)  in _mm_rsqrt_ps()
329 _mm_min_ss(__m128 __a, __m128 __b)  in _mm_min_ss()
348 _mm_min_ps(__m128 __a, __m128 __b)  in _mm_min_ps()
371 _mm_max_ss(__m128 __a, __m128 __b)  in _mm_max_ss()
390 _mm_max_ps(__m128 __a, __m128 __b)  in _mm_max_ps()
408 _mm_and_ps(__m128 __a, __m128 __b)  in _mm_and_ps()
430 _mm_andnot_ps(__m128 __a, __m128 __b)  in _mm_andnot_ps()
448 _mm_or_ps(__m128 __a, __m128 __b)  in _mm_or_ps()
467 _mm_xor_ps(__m128 __a, __m128 __b)  in _mm_xor_ps()
489 _mm_cmpeq_ss(__m128 __a, __m128 __b)  in _mm_cmpeq_ss()
507 _mm_cmpeq_ps(__m128 __a, __m128 __b)  in _mm_cmpeq_ps()
530 _mm_cmplt_ss(__m128 __a, __m128 __b)  in _mm_cmplt_ss()
549 _mm_cmplt_ps(__m128 __a, __m128 __b)  in _mm_cmplt_ps()
573 _mm_cmple_ss(__m128 __a, __m128 __b)  in _mm_cmple_ss()
592 _mm_cmple_ps(__m128 __a, __m128 __b)  in _mm_cmple_ps()
615 _mm_cmpgt_ss(__m128 __a, __m128 __b)  in _mm_cmpgt_ss()
636 _mm_cmpgt_ps(__m128 __a, __m128 __b)  in _mm_cmpgt_ps()
660 _mm_cmpge_ss(__m128 __a, __m128 __b)  in _mm_cmpge_ss()
681 _mm_cmpge_ps(__m128 __a, __m128 __b)  in _mm_cmpge_ps()
704 _mm_cmpneq_ss(__m128 __a, __m128 __b)  in _mm_cmpneq_ss()
723 _mm_cmpneq_ps(__m128 __a, __m128 __b)  in _mm_cmpneq_ps()
747 _mm_cmpnlt_ss(__m128 __a, __m128 __b)  in _mm_cmpnlt_ss()
767 _mm_cmpnlt_ps(__m128 __a, __m128 __b)  in _mm_cmpnlt_ps()
792 _mm_cmpnle_ss(__m128 __a, __m128 __b)  in _mm_cmpnle_ss()
812 _mm_cmpnle_ps(__m128 __a, __m128 __b)  in _mm_cmpnle_ps()
837 _mm_cmpngt_ss(__m128 __a, __m128 __b)  in _mm_cmpngt_ss()
859 _mm_cmpngt_ps(__m128 __a, __m128 __b)  in _mm_cmpngt_ps()
884 _mm_cmpnge_ss(__m128 __a, __m128 __b)  in _mm_cmpnge_ss()
906 _mm_cmpnge_ps(__m128 __a, __m128 __b)  in _mm_cmpnge_ps()
931 _mm_cmpord_ss(__m128 __a, __m128 __b)  in _mm_cmpord_ss()
951 _mm_cmpord_ps(__m128 __a, __m128 __b)  in _mm_cmpord_ps()
976 _mm_cmpunord_ss(__m128 __a, __m128 __b)  in _mm_cmpunord_ss()
996 _mm_cmpunord_ps(__m128 __a, __m128 __b)  in _mm_cmpunord_ps()
1020 _mm_comieq_ss(__m128 __a, __m128 __b)  in _mm_comieq_ss()
1045 _mm_comilt_ss(__m128 __a, __m128 __b)  in _mm_comilt_ss()
1069 _mm_comile_ss(__m128 __a, __m128 __b)  in _mm_comile_ss()
1093 _mm_comigt_ss(__m128 __a, __m128 __b)  in _mm_comigt_ss()
1117 _mm_comige_ss(__m128 __a, __m128 __b)  in _mm_comige_ss()
1141 _mm_comineq_ss(__m128 __a, __m128 __b)  in _mm_comineq_ss()
1165 _mm_ucomieq_ss(__m128 __a, __m128 __b)  in _mm_ucomieq_ss()
1189 _mm_ucomilt_ss(__m128 __a, __m128 __b)  in _mm_ucomilt_ss()
1214 _mm_ucomile_ss(__m128 __a, __m128 __b)  in _mm_ucomile_ss()
1239 _mm_ucomigt_ss(__m128 __a, __m128 __b)  in _mm_ucomigt_ss()
1264 _mm_ucomige_ss(__m128 __a, __m128 __b)  in _mm_ucomige_ss()
1288 _mm_ucomineq_ss(__m128 __a, __m128 __b)  in _mm_ucomineq_ss()
1306 _mm_cvtss_si32(__m128 __a)  in _mm_cvtss_si32()
1324 _mm_cvt_ss2si(__m128 __a)  in _mm_cvt_ss2si()
1344 _mm_cvtss_si64(__m128 __a)  in _mm_cvtss_si64()
1362 _mm_cvtps_pi32(__m128 __a)  in _mm_cvtps_pi32()
1378 _mm_cvt_ps2pi(__m128 __a)  in _mm_cvt_ps2pi()
1397 _mm_cvttss_si32(__m128 __a)  in _mm_cvttss_si32()
1416 _mm_cvtt_ss2si(__m128 __a)  in _mm_cvtt_ss2si()
1436 _mm_cvttss_si64(__m128 __a)  in _mm_cvttss_si64()
1455 _mm_cvttps_pi32(__m128 __a)  in _mm_cvttps_pi32()
1472 _mm_cvtt_ps2pi(__m128 __a)  in _mm_cvtt_ps2pi()
1494 _mm_cvtsi32_ss(__m128 __a, int __b)  in _mm_cvtsi32_ss()
1517 _mm_cvt_si2ss(__m128 __a, int __b)  in _mm_cvt_si2ss()
1541 _mm_cvtsi64_ss(__m128 __a, long long __b)  in _mm_cvtsi64_ss()
1567 _mm_cvtpi32_ps(__m128 __a, __m64 __b)  in _mm_cvtpi32_ps()
1590 _mm_cvt_pi2ps(__m128 __a, __m64 __b)  in _mm_cvt_pi2ps()
1607 _mm_cvtss_f32(__m128 __a)  in _mm_cvtss_f32()
1628 _mm_loadh_pi(__m128 __a, const __m64 *__p)  in _mm_loadh_pi()
1655 _mm_loadl_pi(__m128 __a, const __m64 *__p)  in _mm_loadl_pi()
1768   __m128 __a = _mm_load_ps(__p);  in _mm_loadr_ps()  local
1924 _mm_storeh_pi(__m64 *__p, __m128 __a)  in _mm_storeh_pi()
1945 _mm_storel_pi(__m64 *__p, __m128 __a)  in _mm_storel_pi()
1966 _mm_store_ss(float *__p, __m128 __a)  in _mm_store_ss()
1987 _mm_storeu_ps(float *__p, __m128 __a)  in _mm_storeu_ps()
2008 _mm_store_ps(float *__p, __m128 __a)  in _mm_store_ps()
2027 _mm_store1_ps(float *__p, __m128 __a)  in _mm_store1_ps()
2047 _mm_store_ps1(float *__p, __m128 __a)  in _mm_store_ps1()
2066 _mm_storer_ps(float *__p, __m128 __a)  in _mm_storer_ps()
2124 _mm_stream_pi(__m64 *__p, __m64 __a)  in _mm_stream_pi()
2143 _mm_stream_ps(float *__p, __m128 __a)  in _mm_stream_ps()
2235 _mm_max_pi16(__m64 __a, __m64 __b)  in _mm_max_pi16()
2254 _mm_max_pu8(__m64 __a, __m64 __b)  in _mm_max_pu8()
2273 _mm_min_pi16(__m64 __a, __m64 __b)  in _mm_min_pi16()
2292 _mm_min_pu8(__m64 __a, __m64 __b)  in _mm_min_pu8()
2310 _mm_movemask_pi8(__m64 __a)  in _mm_movemask_pi8()
2329 _mm_mulhi_pu16(__m64 __a, __m64 __b)  in _mm_mulhi_pu16()
2414 _mm_avg_pu8(__m64 __a, __m64 __b)  in _mm_avg_pu8()
2433 _mm_avg_pu16(__m64 __a, __m64 __b)  in _mm_avg_pu16()
2455 _mm_sad_pu8(__m64 __a, __m64 __b)  in _mm_sad_pu8()
2634 _mm_unpackhi_ps(__m128 __a, __m128 __b)  in _mm_unpackhi_ps()
2656 _mm_unpacklo_ps(__m128 __a, __m128 __b)  in _mm_unpacklo_ps()
2678 _mm_move_ss(__m128 __a, __m128 __b)  in _mm_move_ss()
2700 _mm_movehl_ps(__m128 __a, __m128 __b)  in _mm_movehl_ps()
2721 _mm_movelh_ps(__m128 __a, __m128 __b)  in _mm_movelh_ps()
2739 _mm_cvtpi16_ps(__m64 __a)  in _mm_cvtpi16_ps()
2769 _mm_cvtpu16_ps(__m64 __a)  in _mm_cvtpu16_ps()
2798 _mm_cvtpi8_ps(__m64 __a)  in _mm_cvtpi8_ps()
2823 _mm_cvtpu8_ps(__m64 __a)  in _mm_cvtpu8_ps()
2850 _mm_cvtpi32x2_ps(__m64 __a, __m64 __b)  in _mm_cvtpi32x2_ps()
2879 _mm_cvtps_pi16(__m128 __a)  in _mm_cvtps_pi16()
2909 _mm_cvtps_pi8(__m128 __a)  in _mm_cvtps_pi8()
2934 _mm_movemask_ps(__m128 __a)  in _mm_movemask_ps()