Lines Matching refs:__v4sf
1221 (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \
1222 (__v4sf)(__m128)(b), (int)(p), \
1226 (__mmask8)__builtin_ia32_cmpps128_mask((__v4sf)(__m128)(a), \
1227 (__v4sf)(__m128)(b), (int)(p), \
1387 return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, in _mm_mask_fmadd_ps()
1388 (__v4sf) __B, in _mm_mask_fmadd_ps()
1389 (__v4sf) __C, in _mm_mask_fmadd_ps()
1396 return (__m128) __builtin_ia32_vfmaddps128_mask3 ((__v4sf) __A, in _mm_mask3_fmadd_ps()
1397 (__v4sf) __B, in _mm_mask3_fmadd_ps()
1398 (__v4sf) __C, in _mm_mask3_fmadd_ps()
1405 return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, in _mm_maskz_fmadd_ps()
1406 (__v4sf) __B, in _mm_maskz_fmadd_ps()
1407 (__v4sf) __C, in _mm_maskz_fmadd_ps()
1414 return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, in _mm_mask_fmsub_ps()
1415 (__v4sf) __B, in _mm_mask_fmsub_ps()
1416 -(__v4sf) __C, in _mm_mask_fmsub_ps()
1423 return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, in _mm_maskz_fmsub_ps()
1424 (__v4sf) __B, in _mm_maskz_fmsub_ps()
1425 -(__v4sf) __C, in _mm_maskz_fmsub_ps()
1432 return (__m128) __builtin_ia32_vfmaddps128_mask3 (-(__v4sf) __A, in _mm_mask3_fnmadd_ps()
1433 (__v4sf) __B, in _mm_mask3_fnmadd_ps()
1434 (__v4sf) __C, in _mm_mask3_fnmadd_ps()
1441 return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A, in _mm_maskz_fnmadd_ps()
1442 (__v4sf) __B, in _mm_maskz_fnmadd_ps()
1443 (__v4sf) __C, in _mm_maskz_fnmadd_ps()
1450 return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A, in _mm_maskz_fnmsub_ps()
1451 (__v4sf) __B, in _mm_maskz_fnmsub_ps()
1452 -(__v4sf) __C, in _mm_maskz_fnmsub_ps()
1627 return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, in _mm_mask_fmaddsub_ps()
1628 (__v4sf) __B, in _mm_mask_fmaddsub_ps()
1629 (__v4sf) __C, in _mm_mask_fmaddsub_ps()
1636 return (__m128) __builtin_ia32_vfmaddsubps128_mask3 ((__v4sf) __A, in _mm_mask3_fmaddsub_ps()
1637 (__v4sf) __B, in _mm_mask3_fmaddsub_ps()
1638 (__v4sf) __C, in _mm_mask3_fmaddsub_ps()
1645 return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, in _mm_maskz_fmaddsub_ps()
1646 (__v4sf) __B, in _mm_maskz_fmaddsub_ps()
1647 (__v4sf) __C, in _mm_maskz_fmaddsub_ps()
1654 return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, in _mm_mask_fmsubadd_ps()
1655 (__v4sf) __B, in _mm_mask_fmsubadd_ps()
1656 -(__v4sf) __C, in _mm_mask_fmsubadd_ps()
1663 return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, in _mm_maskz_fmsubadd_ps()
1664 (__v4sf) __B, in _mm_maskz_fmsubadd_ps()
1665 -(__v4sf) __C, in _mm_maskz_fmsubadd_ps()
1736 return (__m128) __builtin_ia32_vfmsubps128_mask3 ((__v4sf) __A, in _mm_mask3_fmsub_ps()
1737 (__v4sf) __B, in _mm_mask3_fmsub_ps()
1738 (__v4sf) __C, in _mm_mask3_fmsub_ps()
1774 return (__m128) __builtin_ia32_vfmsubaddps128_mask3 ((__v4sf) __A, in _mm_mask3_fmsubadd_ps()
1775 (__v4sf) __B, in _mm_mask3_fmsubadd_ps()
1776 (__v4sf) __C, in _mm_mask3_fmsubadd_ps()
1810 return (__m128) __builtin_ia32_vfnmaddps128_mask ((__v4sf) __A, in _mm_mask_fnmadd_ps()
1811 (__v4sf) __B, in _mm_mask_fnmadd_ps()
1812 (__v4sf) __C, in _mm_mask_fnmadd_ps()
1864 return (__m128) __builtin_ia32_vfnmsubps128_mask ((__v4sf) __A, in _mm_mask_fnmsub_ps()
1865 (__v4sf) __B, in _mm_mask_fnmsub_ps()
1866 (__v4sf) __C, in _mm_mask_fnmsub_ps()
1873 return (__m128) __builtin_ia32_vfnmsubps128_mask3 ((__v4sf) __A, in _mm_mask3_fnmsub_ps()
1874 (__v4sf) __B, in _mm_mask3_fnmsub_ps()
1875 (__v4sf) __C, in _mm_mask3_fnmsub_ps()
1933 return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A, in _mm_mask_add_ps()
1934 (__v4sf) __B, in _mm_mask_add_ps()
1935 (__v4sf) __W, in _mm_mask_add_ps()
1941 return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A, in _mm_maskz_add_ps()
1942 (__v4sf) __B, in _mm_maskz_add_ps()
1943 (__v4sf) in _mm_maskz_add_ps()
1996 (__v4sf) __W, in _mm_mask_blend_ps()
1997 (__v4sf) __A); in _mm_mask_blend_ps()
2083 return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, in _mm_mask_compress_ps()
2084 (__v4sf) __W, in _mm_mask_compress_ps()
2090 return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, in _mm_maskz_compress_ps()
2091 (__v4sf) in _mm_maskz_compress_ps()
2171 __builtin_ia32_compressstoresf128_mask ((__v4sf *) __P, in _mm_mask_compressstoreu_ps()
2172 (__v4sf) __A, in _mm_mask_compressstoreu_ps()
2230 (__v4sf) __W, in _mm_mask_cvtepi32_ps()
2237 (__v4sf) in _mm_maskz_cvtepi32_ps()
2290 (__v4sf) __W, in _mm_mask_cvtpd_ps()
2297 (__v4sf) in _mm_maskz_cvtpd_ps()
2305 (__v4sf) __W, in _mm256_mask_cvtpd_ps()
2312 (__v4sf) in _mm256_maskz_cvtpd_ps()
2365 return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, in _mm_mask_cvtps_epi32()
2372 return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, in _mm_maskz_cvtps_epi32()
2395 return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, in _mm_mask_cvtps_pd()
2402 return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, in _mm_maskz_cvtps_pd()
2410 return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, in _mm256_mask_cvtps_pd()
2417 return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, in _mm256_maskz_cvtps_pd()
2425 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, in _mm_cvtps_epu32()
2433 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, in _mm_mask_cvtps_epu32()
2440 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, in _mm_maskz_cvtps_epu32()
2547 return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, in _mm_mask_cvttps_epi32()
2554 return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, in _mm_maskz_cvttps_epi32()
2577 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, in _mm_cvttps_epu32()
2585 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, in _mm_mask_cvttps_epu32()
2592 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, in _mm_maskz_cvttps_epu32()
2670 (__v4sf) in _mm_cvtepu32_ps()
2678 (__v4sf) __W, in _mm_mask_cvtepu32_ps()
2685 (__v4sf) in _mm_maskz_cvtepu32_ps()
2750 return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A, in _mm_mask_div_ps()
2751 (__v4sf) __B, in _mm_mask_div_ps()
2752 (__v4sf) __W, in _mm_mask_div_ps()
2758 return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A, in _mm_maskz_div_ps()
2759 (__v4sf) __B, in _mm_maskz_div_ps()
2760 (__v4sf) in _mm_maskz_div_ps()
2913 return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, in _mm_mask_expandloadu_ps()
2914 (__v4sf) __W, in _mm_mask_expandloadu_ps()
2920 return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P, in _mm_maskz_expandloadu_ps()
2921 (__v4sf) in _mm_maskz_expandloadu_ps()
2979 return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, in _mm_mask_expand_ps()
2980 (__v4sf) __W, in _mm_mask_expand_ps()
2986 return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, in _mm_maskz_expand_ps()
2987 (__v4sf) in _mm_maskz_expand_ps()
3085 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, in _mm_getexp_ps()
3086 (__v4sf) in _mm_getexp_ps()
3093 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, in _mm_mask_getexp_ps()
3094 (__v4sf) __W, in _mm_mask_getexp_ps()
3100 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, in _mm_maskz_getexp_ps()
3101 (__v4sf) in _mm_maskz_getexp_ps()
3166 return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, in _mm_mask_max_ps()
3167 (__v4sf) __B, in _mm_mask_max_ps()
3168 (__v4sf) __W, in _mm_mask_max_ps()
3174 return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, in _mm_maskz_max_ps()
3175 (__v4sf) __B, in _mm_maskz_max_ps()
3176 (__v4sf) in _mm_maskz_max_ps()
3235 return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, in _mm_mask_min_ps()
3236 (__v4sf) __B, in _mm_mask_min_ps()
3237 (__v4sf) __W, in _mm_mask_min_ps()
3243 return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, in _mm_maskz_min_ps()
3244 (__v4sf) __B, in _mm_maskz_min_ps()
3245 (__v4sf) in _mm_maskz_min_ps()
3304 return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A, in _mm_mask_mul_ps()
3305 (__v4sf) __B, in _mm_mask_mul_ps()
3306 (__v4sf) __W, in _mm_mask_mul_ps()
3312 return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A, in _mm_maskz_mul_ps()
3313 (__v4sf) __B, in _mm_maskz_mul_ps()
3314 (__v4sf) in _mm_maskz_mul_ps()
3798 (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \
3799 (__v4sf)_mm_setzero_ps(), \
3804 (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \
3805 (__v4sf)(__m128)(W), \
3810 (__m128)__builtin_ia32_rndscaleps_128_mask((__v4sf)(__m128)(A), (int)(imm), \
3811 (__v4sf)_mm_setzero_ps(), \
3886 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, in _mm_scalef_ps()
3887 (__v4sf) __B, in _mm_scalef_ps()
3888 (__v4sf) in _mm_scalef_ps()
3895 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, in _mm_mask_scalef_ps()
3896 (__v4sf) __B, in _mm_mask_scalef_ps()
3897 (__v4sf) __W, in _mm_mask_scalef_ps()
3903 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, in _mm_maskz_scalef_ps()
3904 (__v4sf) __B, in _mm_maskz_scalef_ps()
3905 (__v4sf) in _mm_maskz_scalef_ps()
3979 (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \
3984 (__v2di)(__m128i)(index), (__v4sf)(__m128)(v1), \
3999 (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \
4004 (__v4di)(__m256i)(index), (__v4sf)(__m128)(v1), \
4059 (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \
4064 (__v4si)(__m128i)(index), (__v4sf)(__m128)(v1), \
4129 return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, in _mm_mask_sqrt_ps()
4130 (__v4sf) __W, in _mm_mask_sqrt_ps()
4136 return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, in _mm_maskz_sqrt_ps()
4137 (__v4sf) in _mm_maskz_sqrt_ps()
4194 return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A, in _mm_mask_sub_ps()
4195 (__v4sf) __B, in _mm_mask_sub_ps()
4196 (__v4sf) __W, in _mm_mask_sub_ps()
4202 return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A, in _mm_maskz_sub_ps()
4203 (__v4sf) __B, in _mm_maskz_sub_ps()
4204 (__v4sf) in _mm_maskz_sub_ps()
4271 return (__m128) __builtin_ia32_vpermi2varps128_mask ((__v4sf) __A, in _mm_mask2_permutex2var_ps()
4274 (__v4sf) __B, in _mm_mask2_permutex2var_ps()
4436 (__v4sf) __A, in _mm_permutex2var_ps()
4437 (__v4sf) __B, in _mm_permutex2var_ps()
4446 (__v4sf) __A, in _mm_mask_permutex2var_ps()
4447 (__v4sf) __B, in _mm_mask_permutex2var_ps()
4456 (__v4sf) __A, in _mm_maskz_permutex2var_ps()
4457 (__v4sf) __B, in _mm_maskz_permutex2var_ps()
6047 (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \
6048 (__v4sf)(__m128)(B), \
6053 (__m128)__builtin_ia32_fixupimmps128_mask((__v4sf)(__m128)(A), \
6054 (__v4sf)(__m128)(B), \
6059 (__m128)__builtin_ia32_fixupimmps128_maskz((__v4sf)(__m128)(A), \
6060 (__v4sf)(__m128)(B), \
6119 return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P, in _mm_mask_load_ps()
6120 (__v4sf) __W, in _mm_mask_load_ps()
6127 return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P, in _mm_maskz_load_ps()
6128 (__v4sf) in _mm_maskz_load_ps()
6255 return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P, in _mm_mask_loadu_ps()
6256 (__v4sf) __W, in _mm_mask_loadu_ps()
6263 return (__m128) __builtin_ia32_loadups128_mask ((__v4sf *) __P, in _mm_maskz_loadu_ps()
6264 (__v4sf) in _mm_maskz_loadu_ps()
6305 __builtin_ia32_storeaps128_mask ((__v4sf *) __P, in _mm_mask_store_ps()
6306 (__v4sf) __A, in _mm_mask_store_ps()
6369 __builtin_ia32_storeups128_mask ((__v4sf *) __P, in _mm_mask_storeu_ps()
6370 (__v4sf) __A, in _mm_mask_storeu_ps()
6419 (__v4sf)_mm_unpackhi_ps(__A, __B), in _mm_mask_unpackhi_ps()
6420 (__v4sf)__W); in _mm_mask_unpackhi_ps()
6427 (__v4sf)_mm_unpackhi_ps(__A, __B), in _mm_maskz_unpackhi_ps()
6428 (__v4sf)_mm_setzero_ps()); in _mm_maskz_unpackhi_ps()
6483 (__v4sf)_mm_unpacklo_ps(__A, __B), in _mm_mask_unpacklo_ps()
6484 (__v4sf)__W); in _mm_mask_unpacklo_ps()
6491 (__v4sf)_mm_unpacklo_ps(__A, __B), in _mm_maskz_unpacklo_ps()
6492 (__v4sf)_mm_setzero_ps()); in _mm_maskz_unpacklo_ps()
6566 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, in _mm_rcp14_ps()
6567 (__v4sf) in _mm_rcp14_ps()
6575 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, in _mm_mask_rcp14_ps()
6576 (__v4sf) __W, in _mm_mask_rcp14_ps()
6583 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, in _mm_maskz_rcp14_ps()
6584 (__v4sf) in _mm_maskz_rcp14_ps()
6637 (__v4sf)_mm_permute_ps((X), (C)), \
6638 (__v4sf)(__m128)(W)); })
6642 (__v4sf)_mm_permute_ps((X), (C)), \
6643 (__v4sf)_mm_setzero_ps()); })
6701 return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A, in _mm_mask_permutevar_ps()
6703 (__v4sf) __W, in _mm_mask_permutevar_ps()
6710 return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A, in _mm_maskz_permutevar_ps()
6712 (__v4sf) in _mm_maskz_permutevar_ps()
7314 (__v4sf)_mm_shuffle_ps((A), (B), (M)), \
7315 (__v4sf)(__m128)(W)); })
7319 (__v4sf)_mm_shuffle_ps((A), (B), (M)), \
7320 (__v4sf)_mm_setzero_ps()); })
7387 return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, in _mm_rsqrt14_ps()
7388 (__v4sf) in _mm_rsqrt14_ps()
7396 return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, in _mm_mask_rsqrt14_ps()
7397 (__v4sf) __W, in _mm_mask_rsqrt14_ps()
7404 return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, in _mm_maskz_rsqrt14_ps()
7405 (__v4sf) in _mm_maskz_rsqrt14_ps()
7439 return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, in _mm256_broadcast_f32x4()
7447 return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, in _mm256_mask_broadcast_f32x4()
7455 return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, in _mm256_maskz_broadcast_f32x4()
7505 (__v4sf) _mm_broadcastss_ps(__A), in _mm_mask_broadcastss_ps()
7506 (__v4sf) __O); in _mm_mask_broadcastss_ps()
7513 (__v4sf) _mm_broadcastss_ps(__A), in _mm_maskz_broadcastss_ps()
7514 (__v4sf) _mm_setzero_ps()); in _mm_maskz_broadcastss_ps()
8478 (__v4sf)_mm_setzero_ps(), \
8484 (__v4sf)(__m128)(W), \
8490 (__v4sf)_mm_setzero_ps(), \
8513 (__v4sf)(__m128)(B), (int)(imm), \
8519 (__v4sf)(__m128)(B), (int)(imm), \
8525 (__v4sf)(__m128)(B), (int)(imm), \
8587 (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \
8589 (__v4sf)_mm_setzero_ps(), \
8593 (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \
8595 (__v4sf)(__m128)(W), \
8599 (__m128)__builtin_ia32_getmantps128_mask((__v4sf)(__m128)(A), \
8601 (__v4sf)_mm_setzero_ps(), \
8647 (__m128)__builtin_ia32_gather3div4sf((__v4sf)(__m128)(v1_old), \
8659 (__m128)__builtin_ia32_gather3div8sf((__v4sf)(__m128)(v1_old), \
8695 (__m128)__builtin_ia32_gather3siv4sf((__v4sf)(__m128)(v1_old), \
8938 (__v4sf)_mm_movehdup_ps(__A), in _mm_mask_movehdup_ps()
8939 (__v4sf)__W); in _mm_mask_movehdup_ps()
8946 (__v4sf)_mm_movehdup_ps(__A), in _mm_maskz_movehdup_ps()
8947 (__v4sf)_mm_setzero_ps()); in _mm_maskz_movehdup_ps()
8970 (__v4sf)_mm_moveldup_ps(__A), in _mm_mask_moveldup_ps()
8971 (__v4sf)__W); in _mm_mask_moveldup_ps()
8978 (__v4sf)_mm_moveldup_ps(__A), in _mm_maskz_moveldup_ps()
8979 (__v4sf)_mm_setzero_ps()); in _mm_maskz_moveldup_ps()
9054 (__v4sf) __A, in _mm_mask_mov_ps()
9055 (__v4sf) __W); in _mm_mask_mov_ps()
9062 (__v4sf) __A, in _mm_maskz_mov_ps()
9063 (__v4sf) _mm_setzero_ps ()); in _mm_maskz_mov_ps()
9086 (__v4sf) __W, in _mm_mask_cvtph_ps()
9094 (__v4sf) in _mm_maskz_cvtph_ps()
9119 return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION, in _mm_mask_cvtps_ph()
9127 return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, _MM_FROUND_CUR_DIRECTION, in _mm_maskz_cvtps_ph()
9133 (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \
9138 (__m128i)__builtin_ia32_vcvtps2ph_mask((__v4sf)(__m128)(A), (int)(I), \