58 _mm_add_ss(__m128 __a, __m128 __b)  in _mm_add_ss()
78 _mm_add_ps(__m128 __a, __m128 __b)  in _mm_add_ps()
100 _mm_sub_ss(__m128 __a, __m128 __b)  in _mm_sub_ss()
121 _mm_sub_ps(__m128 __a, __m128 __b)  in _mm_sub_ps()
143 _mm_mul_ss(__m128 __a, __m128 __b)  in _mm_mul_ss()
163 _mm_mul_ps(__m128 __a, __m128 __b)  in _mm_mul_ps()
185 _mm_div_ss(__m128 __a, __m128 __b)  in _mm_div_ss()
204 _mm_div_ps(__m128 __a, __m128 __b)  in _mm_div_ps()
222 _mm_sqrt_ss(__m128 __a)  in _mm_sqrt_ss()
239 _mm_sqrt_ps(__m128 __a)  in _mm_sqrt_ps()
257 _mm_rcp_ss(__m128 __a)  in _mm_rcp_ss()
274 _mm_rcp_ps(__m128 __a)  in _mm_rcp_ps()
293 _mm_rsqrt_ss(__m128 __a)  in _mm_rsqrt_ss()
310 _mm_rsqrt_ps(__m128 __a)  in _mm_rsqrt_ps()
335 _mm_min_ss(__m128 __a, __m128 __b)  in _mm_min_ss()
356 _mm_min_ps(__m128 __a, __m128 __b)  in _mm_min_ps()
381 _mm_max_ss(__m128 __a, __m128 __b)  in _mm_max_ss()
402 _mm_max_ps(__m128 __a, __m128 __b)  in _mm_max_ps()
420 _mm_and_ps(__m128 __a, __m128 __b)  in _mm_and_ps()
442 _mm_andnot_ps(__m128 __a, __m128 __b)  in _mm_andnot_ps()
460 _mm_or_ps(__m128 __a, __m128 __b)  in _mm_or_ps()
479 _mm_xor_ps(__m128 __a, __m128 __b)  in _mm_xor_ps()
504 _mm_cmpeq_ss(__m128 __a, __m128 __b)  in _mm_cmpeq_ss()
525 _mm_cmpeq_ps(__m128 __a, __m128 __b)  in _mm_cmpeq_ps()
551 _mm_cmplt_ss(__m128 __a, __m128 __b)  in _mm_cmplt_ss()
573 _mm_cmplt_ps(__m128 __a, __m128 __b)  in _mm_cmplt_ps()
599 _mm_cmple_ss(__m128 __a, __m128 __b)  in _mm_cmple_ss()
621 _mm_cmple_ps(__m128 __a, __m128 __b)  in _mm_cmple_ps()
647 _mm_cmpgt_ss(__m128 __a, __m128 __b)  in _mm_cmpgt_ss()
671 _mm_cmpgt_ps(__m128 __a, __m128 __b)  in _mm_cmpgt_ps()
697 _mm_cmpge_ss(__m128 __a, __m128 __b)  in _mm_cmpge_ss()
721 _mm_cmpge_ps(__m128 __a, __m128 __b)  in _mm_cmpge_ps()
747 _mm_cmpneq_ss(__m128 __a, __m128 __b)  in _mm_cmpneq_ss()
769 _mm_cmpneq_ps(__m128 __a, __m128 __b)  in _mm_cmpneq_ps()
796 _mm_cmpnlt_ss(__m128 __a, __m128 __b)  in _mm_cmpnlt_ss()
819 _mm_cmpnlt_ps(__m128 __a, __m128 __b)  in _mm_cmpnlt_ps()
846 _mm_cmpnle_ss(__m128 __a, __m128 __b)  in _mm_cmpnle_ss()
869 _mm_cmpnle_ps(__m128 __a, __m128 __b)  in _mm_cmpnle_ps()
896 _mm_cmpngt_ss(__m128 __a, __m128 __b)  in _mm_cmpngt_ss()
921 _mm_cmpngt_ps(__m128 __a, __m128 __b)  in _mm_cmpngt_ps()
948 _mm_cmpnge_ss(__m128 __a, __m128 __b)  in _mm_cmpnge_ss()
973 _mm_cmpnge_ps(__m128 __a, __m128 __b)  in _mm_cmpnge_ps()
1000 _mm_cmpord_ss(__m128 __a, __m128 __b)  in _mm_cmpord_ss()
1024 _mm_cmpord_ps(__m128 __a, __m128 __b)  in _mm_cmpord_ps()
1051 _mm_cmpunord_ss(__m128 __a, __m128 __b)  in _mm_cmpunord_ss()
1075 _mm_cmpunord_ps(__m128 __a, __m128 __b)  in _mm_cmpunord_ps()
1099 _mm_comieq_ss(__m128 __a, __m128 __b)  in _mm_comieq_ss()
1124 _mm_comilt_ss(__m128 __a, __m128 __b)  in _mm_comilt_ss()
1148 _mm_comile_ss(__m128 __a, __m128 __b)  in _mm_comile_ss()
1172 _mm_comigt_ss(__m128 __a, __m128 __b)  in _mm_comigt_ss()
1196 _mm_comige_ss(__m128 __a, __m128 __b)  in _mm_comige_ss()
1220 _mm_comineq_ss(__m128 __a, __m128 __b)  in _mm_comineq_ss()
1243 _mm_ucomieq_ss(__m128 __a, __m128 __b)  in _mm_ucomieq_ss()
1267 _mm_ucomilt_ss(__m128 __a, __m128 __b)  in _mm_ucomilt_ss()
1291 _mm_ucomile_ss(__m128 __a, __m128 __b)  in _mm_ucomile_ss()
1315 _mm_ucomigt_ss(__m128 __a, __m128 __b)  in _mm_ucomigt_ss()
1339 _mm_ucomige_ss(__m128 __a, __m128 __b)  in _mm_ucomige_ss()
1362 _mm_ucomineq_ss(__m128 __a, __m128 __b)  in _mm_ucomineq_ss()
1384 _mm_cvtss_si32(__m128 __a)  in _mm_cvtss_si32()
1406 _mm_cvt_ss2si(__m128 __a)  in _mm_cvt_ss2si()
1430 _mm_cvtss_si64(__m128 __a)  in _mm_cvtss_si64()
1452 _mm_cvtps_pi32(__m128 __a)  in _mm_cvtps_pi32()
1472 _mm_cvt_ps2pi(__m128 __a)  in _mm_cvt_ps2pi()
1494 _mm_cvttss_si32(__m128 __a)  in _mm_cvttss_si32()
1516 _mm_cvtt_ss2si(__m128 __a)  in _mm_cvtt_ss2si()
1539 _mm_cvttss_si64(__m128 __a)  in _mm_cvttss_si64()
1562 _mm_cvttps_pi32(__m128 __a)  in _mm_cvttps_pi32()
1583 _mm_cvtt_ps2pi(__m128 __a)  in _mm_cvtt_ps2pi()
1605 _mm_cvtsi32_ss(__m128 __a, int __b)  in _mm_cvtsi32_ss()
1628 _mm_cvt_si2ss(__m128 __a, int __b)  in _mm_cvt_si2ss()
1652 _mm_cvtsi64_ss(__m128 __a, long long __b)  in _mm_cvtsi64_ss()
1678 _mm_cvtpi32_ps(__m128 __a, __m64 __b)  in _mm_cvtpi32_ps()
1701 _mm_cvt_pi2ps(__m128 __a, __m64 __b)  in _mm_cvt_pi2ps()
1718 _mm_cvtss_f32(__m128 __a)  in _mm_cvtss_f32()
1739 _mm_loadh_pi(__m128 __a, const __m64 *__p)  in _mm_loadh_pi()
1766 _mm_loadl_pi(__m128 __a, const __m64 *__p)  in _mm_loadl_pi()
1879   __m128 __a = _mm_load_ps(__p);  in _mm_loadr_ps()  local
2035 _mm_storeh_pi(__m64 *__p, __m128 __a)  in _mm_storeh_pi()
2056 _mm_storel_pi(__m64 *__p, __m128 __a)  in _mm_storel_pi()
2077 _mm_store_ss(float *__p, __m128 __a)  in _mm_store_ss()
2098 _mm_storeu_ps(float *__p, __m128 __a)  in _mm_storeu_ps()
2119 _mm_store_ps(float *__p, __m128 __a)  in _mm_store_ps()
2138 _mm_store1_ps(float *__p, __m128 __a)  in _mm_store1_ps()
2158 _mm_store_ps1(float *__p, __m128 __a)  in _mm_store_ps1()
2177 _mm_storer_ps(float *__p, __m128 __a)  in _mm_storer_ps()
2235 _mm_stream_pi(void *__p, __m64 __a)  in _mm_stream_pi()
2254 _mm_stream_ps(void *__p, __m128 __a)  in _mm_stream_ps()
2346 _mm_max_pi16(__m64 __a, __m64 __b)  in _mm_max_pi16()
2365 _mm_max_pu8(__m64 __a, __m64 __b)  in _mm_max_pu8()
2384 _mm_min_pi16(__m64 __a, __m64 __b)  in _mm_min_pi16()
2403 _mm_min_pu8(__m64 __a, __m64 __b)  in _mm_min_pu8()
2421 _mm_movemask_pi8(__m64 __a)  in _mm_movemask_pi8()
2440 _mm_mulhi_pu16(__m64 __a, __m64 __b)  in _mm_mulhi_pu16()
2525 _mm_avg_pu8(__m64 __a, __m64 __b)  in _mm_avg_pu8()
2544 _mm_avg_pu16(__m64 __a, __m64 __b)  in _mm_avg_pu16()
2566 _mm_sad_pu8(__m64 __a, __m64 __b)  in _mm_sad_pu8()
2745 _mm_unpackhi_ps(__m128 __a, __m128 __b)  in _mm_unpackhi_ps()
2767 _mm_unpacklo_ps(__m128 __a, __m128 __b)  in _mm_unpacklo_ps()
2789 _mm_move_ss(__m128 __a, __m128 __b)  in _mm_move_ss()
2811 _mm_movehl_ps(__m128 __a, __m128 __b)  in _mm_movehl_ps()
2832 _mm_movelh_ps(__m128 __a, __m128 __b)  in _mm_movelh_ps()
2850 _mm_cvtpi16_ps(__m64 __a)  in _mm_cvtpi16_ps()
2880 _mm_cvtpu16_ps(__m64 __a)  in _mm_cvtpu16_ps()
2909 _mm_cvtpi8_ps(__m64 __a)  in _mm_cvtpi8_ps()
2934 _mm_cvtpu8_ps(__m64 __a)  in _mm_cvtpu8_ps()
2961 _mm_cvtpi32x2_ps(__m64 __a, __m64 __b)  in _mm_cvtpi32x2_ps()
2990 _mm_cvtps_pi16(__m128 __a)  in _mm_cvtps_pi16()
3020 _mm_cvtps_pi8(__m128 __a)  in _mm_cvtps_pi8()
3045 _mm_movemask_ps(__m128 __a)  in _mm_movemask_ps()