43 _mm_add_ss(__m128 __a, __m128 __b)  in _mm_add_ss()
50 _mm_add_ps(__m128 __a, __m128 __b)  in _mm_add_ps()
56 _mm_sub_ss(__m128 __a, __m128 __b)  in _mm_sub_ss()
63 _mm_sub_ps(__m128 __a, __m128 __b)  in _mm_sub_ps()
69 _mm_mul_ss(__m128 __a, __m128 __b)  in _mm_mul_ss()
76 _mm_mul_ps(__m128 __a, __m128 __b)  in _mm_mul_ps()
82 _mm_div_ss(__m128 __a, __m128 __b)  in _mm_div_ss()
89 _mm_div_ps(__m128 __a, __m128 __b)  in _mm_div_ps()
95 _mm_sqrt_ss(__m128 __a)  in _mm_sqrt_ss()
102 _mm_sqrt_ps(__m128 __a)  in _mm_sqrt_ps()
108 _mm_rcp_ss(__m128 __a)  in _mm_rcp_ss()
115 _mm_rcp_ps(__m128 __a)  in _mm_rcp_ps()
121 _mm_rsqrt_ss(__m128 __a)  in _mm_rsqrt_ss()
128 _mm_rsqrt_ps(__m128 __a)  in _mm_rsqrt_ps()
134 _mm_min_ss(__m128 __a, __m128 __b)  in _mm_min_ss()
140 _mm_min_ps(__m128 __a, __m128 __b)  in _mm_min_ps()
146 _mm_max_ss(__m128 __a, __m128 __b)  in _mm_max_ss()
152 _mm_max_ps(__m128 __a, __m128 __b)  in _mm_max_ps()
158 _mm_and_ps(__m128 __a, __m128 __b)  in _mm_and_ps()
164 _mm_andnot_ps(__m128 __a, __m128 __b)  in _mm_andnot_ps()
170 _mm_or_ps(__m128 __a, __m128 __b)  in _mm_or_ps()
176 _mm_xor_ps(__m128 __a, __m128 __b)  in _mm_xor_ps()
182 _mm_cmpeq_ss(__m128 __a, __m128 __b)  in _mm_cmpeq_ss()
188 _mm_cmpeq_ps(__m128 __a, __m128 __b)  in _mm_cmpeq_ps()
194 _mm_cmplt_ss(__m128 __a, __m128 __b)  in _mm_cmplt_ss()
200 _mm_cmplt_ps(__m128 __a, __m128 __b)  in _mm_cmplt_ps()
206 _mm_cmple_ss(__m128 __a, __m128 __b)  in _mm_cmple_ss()
212 _mm_cmple_ps(__m128 __a, __m128 __b)  in _mm_cmple_ps()
218 _mm_cmpgt_ss(__m128 __a, __m128 __b)  in _mm_cmpgt_ss()
226 _mm_cmpgt_ps(__m128 __a, __m128 __b)  in _mm_cmpgt_ps()
232 _mm_cmpge_ss(__m128 __a, __m128 __b)  in _mm_cmpge_ss()
240 _mm_cmpge_ps(__m128 __a, __m128 __b)  in _mm_cmpge_ps()
246 _mm_cmpneq_ss(__m128 __a, __m128 __b)  in _mm_cmpneq_ss()
252 _mm_cmpneq_ps(__m128 __a, __m128 __b)  in _mm_cmpneq_ps()
258 _mm_cmpnlt_ss(__m128 __a, __m128 __b)  in _mm_cmpnlt_ss()
264 _mm_cmpnlt_ps(__m128 __a, __m128 __b)  in _mm_cmpnlt_ps()
270 _mm_cmpnle_ss(__m128 __a, __m128 __b)  in _mm_cmpnle_ss()
276 _mm_cmpnle_ps(__m128 __a, __m128 __b)  in _mm_cmpnle_ps()
282 _mm_cmpngt_ss(__m128 __a, __m128 __b)  in _mm_cmpngt_ss()
290 _mm_cmpngt_ps(__m128 __a, __m128 __b)  in _mm_cmpngt_ps()
296 _mm_cmpnge_ss(__m128 __a, __m128 __b)  in _mm_cmpnge_ss()
304 _mm_cmpnge_ps(__m128 __a, __m128 __b)  in _mm_cmpnge_ps()
310 _mm_cmpord_ss(__m128 __a, __m128 __b)  in _mm_cmpord_ss()
316 _mm_cmpord_ps(__m128 __a, __m128 __b)  in _mm_cmpord_ps()
322 _mm_cmpunord_ss(__m128 __a, __m128 __b)  in _mm_cmpunord_ss()
328 _mm_cmpunord_ps(__m128 __a, __m128 __b)  in _mm_cmpunord_ps()
334 _mm_comieq_ss(__m128 __a, __m128 __b)  in _mm_comieq_ss()
340 _mm_comilt_ss(__m128 __a, __m128 __b)  in _mm_comilt_ss()
346 _mm_comile_ss(__m128 __a, __m128 __b)  in _mm_comile_ss()
352 _mm_comigt_ss(__m128 __a, __m128 __b)  in _mm_comigt_ss()
358 _mm_comige_ss(__m128 __a, __m128 __b)  in _mm_comige_ss()
364 _mm_comineq_ss(__m128 __a, __m128 __b)  in _mm_comineq_ss()
370 _mm_ucomieq_ss(__m128 __a, __m128 __b)  in _mm_ucomieq_ss()
376 _mm_ucomilt_ss(__m128 __a, __m128 __b)  in _mm_ucomilt_ss()
382 _mm_ucomile_ss(__m128 __a, __m128 __b)  in _mm_ucomile_ss()
388 _mm_ucomigt_ss(__m128 __a, __m128 __b)  in _mm_ucomigt_ss()
394 _mm_ucomige_ss(__m128 __a, __m128 __b)  in _mm_ucomige_ss()
400 _mm_ucomineq_ss(__m128 __a, __m128 __b)  in _mm_ucomineq_ss()
406 _mm_cvtss_si32(__m128 __a)  in _mm_cvtss_si32()
412 _mm_cvt_ss2si(__m128 __a)  in _mm_cvt_ss2si()
420 _mm_cvtss_si64(__m128 __a)  in _mm_cvtss_si64()
428 _mm_cvtps_pi32(__m128 __a)  in _mm_cvtps_pi32()
434 _mm_cvt_ps2pi(__m128 __a)  in _mm_cvt_ps2pi()
440 _mm_cvttss_si32(__m128 __a)  in _mm_cvttss_si32()
446 _mm_cvtt_ss2si(__m128 __a)  in _mm_cvtt_ss2si()
452 _mm_cvttss_si64(__m128 __a)  in _mm_cvttss_si64()
458 _mm_cvttps_pi32(__m128 __a)  in _mm_cvttps_pi32()
464 _mm_cvtt_ps2pi(__m128 __a)  in _mm_cvtt_ps2pi()
470 _mm_cvtsi32_ss(__m128 __a, int __b)  in _mm_cvtsi32_ss()
477 _mm_cvt_si2ss(__m128 __a, int __b)  in _mm_cvt_si2ss()
485 _mm_cvtsi64_ss(__m128 __a, long long __b)  in _mm_cvtsi64_ss()
494 _mm_cvtpi32_ps(__m128 __a, __m64 __b)  in _mm_cvtpi32_ps()
500 _mm_cvt_pi2ps(__m128 __a, __m64 __b)  in _mm_cvt_pi2ps()
506 _mm_cvtss_f32(__m128 __a)  in _mm_cvtss_f32()
512 _mm_loadh_pi(__m128 __a, const __m64 *__p)  in _mm_loadh_pi()
524 _mm_loadl_pi(__m128 __a, const __m64 *__p)  in _mm_loadl_pi()
575   __m128 __a = _mm_load_ps(__p);  in _mm_loadr_ps()  local
623 _mm_storeh_pi(__m64 *__p, __m128 __a)  in _mm_storeh_pi()
629 _mm_storel_pi(__m64 *__p, __m128 __a)  in _mm_storel_pi()
635 _mm_store_ss(float *__p, __m128 __a)  in _mm_store_ss()
644 _mm_storeu_ps(float *__p, __m128 __a)  in _mm_storeu_ps()
650 _mm_store1_ps(float *__p, __m128 __a)  in _mm_store1_ps()
657 _mm_store_ps1(float *__p, __m128 __a)  in _mm_store_ps1()
663 _mm_store_ps(float *__p, __m128 __a)  in _mm_store_ps()
669 _mm_storer_ps(float *__p, __m128 __a)  in _mm_storer_ps()
688 _mm_stream_pi(__m64 *__p, __m64 __a)  in _mm_stream_pi()
694 _mm_stream_ps(float *__p, __m128 __a)  in _mm_stream_ps()
706 _mm_extract_pi16(__m64 __a, int __n)  in _mm_extract_pi16()
713 _mm_insert_pi16(__m64 __a, int __d, int __n)  in _mm_insert_pi16()
721 _mm_max_pi16(__m64 __a, __m64 __b)  in _mm_max_pi16()
727 _mm_max_pu8(__m64 __a, __m64 __b)  in _mm_max_pu8()
733 _mm_min_pi16(__m64 __a, __m64 __b)  in _mm_min_pi16()
739 _mm_min_pu8(__m64 __a, __m64 __b)  in _mm_min_pu8()
745 _mm_movemask_pi8(__m64 __a)  in _mm_movemask_pi8()
751 _mm_mulhi_pu16(__m64 __a, __m64 __b)  in _mm_mulhi_pu16()
766 _mm_avg_pu8(__m64 __a, __m64 __b)  in _mm_avg_pu8()
772 _mm_avg_pu16(__m64 __a, __m64 __b)  in _mm_avg_pu16()
778 _mm_sad_pu8(__m64 __a, __m64 __b)  in _mm_sad_pu8()
802 _mm_unpackhi_ps(__m128 __a, __m128 __b)  in _mm_unpackhi_ps()
808 _mm_unpacklo_ps(__m128 __a, __m128 __b)  in _mm_unpacklo_ps()
814 _mm_move_ss(__m128 __a, __m128 __b)  in _mm_move_ss()
820 _mm_movehl_ps(__m128 __a, __m128 __b)  in _mm_movehl_ps()
826 _mm_movelh_ps(__m128 __a, __m128 __b)  in _mm_movelh_ps()
832 _mm_cvtpi16_ps(__m64 __a)  in _mm_cvtpi16_ps()
850 _mm_cvtpu16_ps(__m64 __a)  in _mm_cvtpu16_ps()
867 _mm_cvtpi8_ps(__m64 __a)  in _mm_cvtpi8_ps()
879 _mm_cvtpu8_ps(__m64 __a)  in _mm_cvtpu8_ps()
890 _mm_cvtpi32x2_ps(__m64 __a, __m64 __b)  in _mm_cvtpi32x2_ps()
902 _mm_cvtps_pi16(__m128 __a)  in _mm_cvtps_pi16()
914 _mm_cvtps_pi8(__m128 __a)  in _mm_cvtps_pi8()
925 _mm_movemask_ps(__m128 __a)  in _mm_movemask_ps()