44 _mm_add_ss(__m128 a, __m128 b)  in _mm_add_ss()
51 _mm_add_ps(__m128 a, __m128 b)  in _mm_add_ps()
57 _mm_sub_ss(__m128 a, __m128 b)  in _mm_sub_ss()
64 _mm_sub_ps(__m128 a, __m128 b)  in _mm_sub_ps()
70 _mm_mul_ss(__m128 a, __m128 b)  in _mm_mul_ss()
77 _mm_mul_ps(__m128 a, __m128 b)  in _mm_mul_ps()
83 _mm_div_ss(__m128 a, __m128 b)  in _mm_div_ss()
90 _mm_div_ps(__m128 a, __m128 b)  in _mm_div_ps()
96 _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()
114 _mm_rcp_ps(__m128 a)  in _mm_rcp_ps()
120 _mm_rsqrt_ss(__m128 a)  in _mm_rsqrt_ss()
126 _mm_rsqrt_ps(__m128 a)  in _mm_rsqrt_ps()
132 _mm_min_ss(__m128 a, __m128 b)  in _mm_min_ss()
138 _mm_min_ps(__m128 a, __m128 b)  in _mm_min_ps()
144 _mm_max_ss(__m128 a, __m128 b)  in _mm_max_ss()
150 _mm_max_ps(__m128 a, __m128 b)  in _mm_max_ps()
156 _mm_and_ps(__m128 a, __m128 b)  in _mm_and_ps()
162 _mm_andnot_ps(__m128 a, __m128 b)  in _mm_andnot_ps()
168 _mm_or_ps(__m128 a, __m128 b)  in _mm_or_ps()
174 _mm_xor_ps(__m128 a, __m128 b)  in _mm_xor_ps()
180 _mm_cmpeq_ss(__m128 a, __m128 b)  in _mm_cmpeq_ss()
186 _mm_cmpeq_ps(__m128 a, __m128 b)  in _mm_cmpeq_ps()
192 _mm_cmplt_ss(__m128 a, __m128 b)  in _mm_cmplt_ss()
198 _mm_cmplt_ps(__m128 a, __m128 b)  in _mm_cmplt_ps()
204 _mm_cmple_ss(__m128 a, __m128 b)  in _mm_cmple_ss()
210 _mm_cmple_ps(__m128 a, __m128 b)  in _mm_cmple_ps()
216 _mm_cmpgt_ss(__m128 a, __m128 b)  in _mm_cmpgt_ss()
222 _mm_cmpgt_ps(__m128 a, __m128 b)  in _mm_cmpgt_ps()
228 _mm_cmpge_ss(__m128 a, __m128 b)  in _mm_cmpge_ss()
234 _mm_cmpge_ps(__m128 a, __m128 b)  in _mm_cmpge_ps()
240 _mm_cmpneq_ss(__m128 a, __m128 b)  in _mm_cmpneq_ss()
246 _mm_cmpneq_ps(__m128 a, __m128 b)  in _mm_cmpneq_ps()
252 _mm_cmpnlt_ss(__m128 a, __m128 b)  in _mm_cmpnlt_ss()
258 _mm_cmpnlt_ps(__m128 a, __m128 b)  in _mm_cmpnlt_ps()
264 _mm_cmpnle_ss(__m128 a, __m128 b)  in _mm_cmpnle_ss()
270 _mm_cmpnle_ps(__m128 a, __m128 b)  in _mm_cmpnle_ps()
276 _mm_cmpngt_ss(__m128 a, __m128 b)  in _mm_cmpngt_ss()
282 _mm_cmpngt_ps(__m128 a, __m128 b)  in _mm_cmpngt_ps()
288 _mm_cmpnge_ss(__m128 a, __m128 b)  in _mm_cmpnge_ss()
294 _mm_cmpnge_ps(__m128 a, __m128 b)  in _mm_cmpnge_ps()
300 _mm_cmpord_ss(__m128 a, __m128 b)  in _mm_cmpord_ss()
306 _mm_cmpord_ps(__m128 a, __m128 b)  in _mm_cmpord_ps()
312 _mm_cmpunord_ss(__m128 a, __m128 b)  in _mm_cmpunord_ss()
318 _mm_cmpunord_ps(__m128 a, __m128 b)  in _mm_cmpunord_ps()
324 _mm_comieq_ss(__m128 a, __m128 b)  in _mm_comieq_ss()
330 _mm_comilt_ss(__m128 a, __m128 b)  in _mm_comilt_ss()
336 _mm_comile_ss(__m128 a, __m128 b)  in _mm_comile_ss()
342 _mm_comigt_ss(__m128 a, __m128 b)  in _mm_comigt_ss()
348 _mm_comige_ss(__m128 a, __m128 b)  in _mm_comige_ss()
354 _mm_comineq_ss(__m128 a, __m128 b)  in _mm_comineq_ss()
360 _mm_ucomieq_ss(__m128 a, __m128 b)  in _mm_ucomieq_ss()
366 _mm_ucomilt_ss(__m128 a, __m128 b)  in _mm_ucomilt_ss()
372 _mm_ucomile_ss(__m128 a, __m128 b)  in _mm_ucomile_ss()
378 _mm_ucomigt_ss(__m128 a, __m128 b)  in _mm_ucomigt_ss()
384 _mm_ucomige_ss(__m128 a, __m128 b)  in _mm_ucomige_ss()
390 _mm_ucomineq_ss(__m128 a, __m128 b)  in _mm_ucomineq_ss()
396 _mm_cvtss_si32(__m128 a)  in _mm_cvtss_si32()
402 _mm_cvt_ss2si(__m128 a)  in _mm_cvt_ss2si()
410 _mm_cvtss_si64(__m128 a)  in _mm_cvtss_si64()
418 _mm_cvtps_pi32(__m128 a)  in _mm_cvtps_pi32()
424 _mm_cvt_ps2pi(__m128 a)  in _mm_cvt_ps2pi()
430 _mm_cvttss_si32(__m128 a)  in _mm_cvttss_si32()
436 _mm_cvtt_ss2si(__m128 a)  in _mm_cvtt_ss2si()
442 _mm_cvttss_si64(__m128 a)  in _mm_cvttss_si64()
448 _mm_cvttps_pi32(__m128 a)  in _mm_cvttps_pi32()
454 _mm_cvtt_ps2pi(__m128 a)  in _mm_cvtt_ps2pi()
460 _mm_cvtsi32_ss(__m128 a, int b)  in _mm_cvtsi32_ss()
467 _mm_cvt_si2ss(__m128 a, int b)  in _mm_cvt_si2ss()
475 _mm_cvtsi64_ss(__m128 a, long long b)  in _mm_cvtsi64_ss()
484 _mm_cvtpi32_ps(__m128 a, __m64 b)  in _mm_cvtpi32_ps()
490 _mm_cvt_pi2ps(__m128 a, __m64 b)  in _mm_cvt_pi2ps()
496 _mm_cvtss_f32(__m128 a)  in _mm_cvtss_f32()
502 _mm_loadh_pi(__m128 a, const __m64 *p)  in _mm_loadh_pi()
514 _mm_loadl_pi(__m128 a, const __m64 *p)  in _mm_loadl_pi()
565   __m128 a = _mm_load_ps(p);  in _mm_loadr_ps()  local
607 _mm_storeh_pi(__m64 *p, __m128 a)  in _mm_storeh_pi()
613 _mm_storel_pi(__m64 *p, __m128 a)  in _mm_storel_pi()
619 _mm_store_ss(float *p, __m128 a)  in _mm_store_ss()
628 _mm_storeu_ps(float *p, __m128 a)  in _mm_storeu_ps()
634 _mm_store1_ps(float *p, __m128 a)  in _mm_store1_ps()
641 _mm_store_ps1(float *p, __m128 a)  in _mm_store_ps1()
647 _mm_store_ps(float *p, __m128 a)  in _mm_store_ps()
653 _mm_storer_ps(float *p, __m128 a)  in _mm_storer_ps()
667 #define _mm_prefetch(a, sel) (__builtin_prefetch((void *)(a), 0, (sel)))  argument
670 _mm_stream_pi(__m64 *p, __m64 a)  in _mm_stream_pi()
676 _mm_stream_ps(float *p, __m128 a)  in _mm_stream_ps()
688 _mm_extract_pi16(__m64 a, int n)  in _mm_extract_pi16()
695 _mm_insert_pi16(__m64 a, int d, int n)  in _mm_insert_pi16()
703 _mm_max_pi16(__m64 a, __m64 b)  in _mm_max_pi16()
709 _mm_max_pu8(__m64 a, __m64 b)  in _mm_max_pu8()
715 _mm_min_pi16(__m64 a, __m64 b)  in _mm_min_pi16()
721 _mm_min_pu8(__m64 a, __m64 b)  in _mm_min_pu8()
727 _mm_movemask_pi8(__m64 a)  in _mm_movemask_pi8()
733 _mm_mulhi_pu16(__m64 a, __m64 b)  in _mm_mulhi_pu16()
738 #define _mm_shuffle_pi16(a, n) __extension__ ({ \  argument
749 _mm_avg_pu8(__m64 a, __m64 b)  in _mm_avg_pu8()
755 _mm_avg_pu16(__m64 a, __m64 b)  in _mm_avg_pu16()
761 _mm_sad_pu8(__m64 a, __m64 b)  in _mm_sad_pu8()
778 #define _mm_shuffle_ps(a, b, mask) __extension__ ({ \  argument
787 _mm_unpackhi_ps(__m128 a, __m128 b)  in _mm_unpackhi_ps()
793 _mm_unpacklo_ps(__m128 a, __m128 b)  in _mm_unpacklo_ps()
799 _mm_move_ss(__m128 a, __m128 b)  in _mm_move_ss()
805 _mm_movehl_ps(__m128 a, __m128 b)  in _mm_movehl_ps()
811 _mm_movelh_ps(__m128 a, __m128 b)  in _mm_movelh_ps()
817 _mm_cvtpi16_ps(__m64 a)  in _mm_cvtpi16_ps()
835 _mm_cvtpu16_ps(__m64 a)  in _mm_cvtpu16_ps()
852 _mm_cvtpi8_ps(__m64 a)  in _mm_cvtpi8_ps()
864 _mm_cvtpu8_ps(__m64 a)  in _mm_cvtpu8_ps()
875 _mm_cvtpi32x2_ps(__m64 a, __m64 b)  in _mm_cvtpi32x2_ps()
887 _mm_cvtps_pi16(__m128 a)  in _mm_cvtps_pi16()
899 _mm_cvtps_pi8(__m128 a)  in _mm_cvtps_pi8()
910 _mm_movemask_ps(__m128 a)  in _mm_movemask_ps()