Lines Matching refs:FORCE_INLINE
62 #define FORCE_INLINE static inline __attribute__((always_inline)) macro
66 #ifdef FORCE_INLINE
67 #undef FORCE_INLINE
69 #define FORCE_INLINE static inline macro
301 FORCE_INLINE uint8x16x4_t vld1q_u8_x4(const uint8_t *p) in vld1q_u8_x4()
406 FORCE_INLINE void _mm_prefetch(const void *p, int i) in _mm_prefetch()
417 FORCE_INLINE float _mm_cvtss_f32(__m128 a) in _mm_cvtss_f32()
424 FORCE_INLINE __m128i _mm_setzero_si128(void) in _mm_setzero_si128()
431 FORCE_INLINE __m128 _mm_setzero_ps(void) in _mm_setzero_ps()
441 FORCE_INLINE __m128 _mm_set1_ps(float _w) in _mm_set1_ps()
448 FORCE_INLINE __m128 _mm_set_ps1(float _w) in _mm_set_ps1()
455 FORCE_INLINE __m128 _mm_set_ps(float w, float z, float y, float x) in _mm_set_ps()
464 FORCE_INLINE __m128 _mm_set_ss(float a) in _mm_set_ss()
473 FORCE_INLINE __m128 _mm_setr_ps(float w, float z, float y, float x) in _mm_setr_ps()
486 FORCE_INLINE __m128i _mm_setr_epi16(short w0, in _mm_setr_epi16()
501 FORCE_INLINE __m128i _mm_setr_epi32(int i3, int i2, int i1, int i0) in _mm_setr_epi32()
509 FORCE_INLINE __m128i _mm_setr_epi64(__m64 e1, __m64 e0) in _mm_setr_epi64()
522 FORCE_INLINE __m128i _mm_set1_epi8(signed char w) in _mm_set1_epi8()
535 FORCE_INLINE __m128i _mm_set1_epi16(short w) in _mm_set1_epi16()
542 FORCE_INLINE __m128i _mm_set_epi8(signed char b15, in _mm_set_epi8()
569 FORCE_INLINE __m128i _mm_set_epi16(short i7, in _mm_set_epi16()
584 FORCE_INLINE __m128i _mm_setr_epi8(signed char b0, in _mm_setr_epi8()
617 FORCE_INLINE __m128i _mm_set1_epi32(int _i) in _mm_set1_epi32()
624 FORCE_INLINE __m128i _mm_set1_epi64(__m64 _i) in _mm_set1_epi64()
631 FORCE_INLINE __m128i _mm_set1_epi64x(int64_t _i) in _mm_set1_epi64x()
638 FORCE_INLINE __m128i _mm_set_epi32(int i3, int i2, int i1, int i0) in _mm_set_epi32()
647 FORCE_INLINE __m128i _mm_set_epi64x(int64_t i1, int64_t i2) in _mm_set_epi64x()
656 FORCE_INLINE __m128i _mm_set_epi64(__m64 i1, __m64 i2) in _mm_set_epi64()
664 FORCE_INLINE __m128d _mm_set_pd(double e1, double e0) in _mm_set_pd()
676 FORCE_INLINE void _mm_store_ps(float *p, __m128 a) in _mm_store_ps()
683 FORCE_INLINE void _mm_storeu_ps(float *p, __m128 a) in _mm_storeu_ps()
690 FORCE_INLINE void _mm_store_si128(__m128i *p, __m128i a) in _mm_store_si128()
697 FORCE_INLINE void _mm_storeu_si128(__m128i *p, __m128i a) in _mm_storeu_si128()
704 FORCE_INLINE void _mm_store_ss(float *p, __m128 a) in _mm_store_ss()
713 FORCE_INLINE void _mm_store_pd(double *mem_addr, __m128d a) in _mm_store_pd()
726 FORCE_INLINE void _mm_storeu_pd(double *mem_addr, __m128d a) in _mm_storeu_pd()
733 FORCE_INLINE void _mm_storel_epi64(__m128i *a, __m128i b) in _mm_storel_epi64()
747 FORCE_INLINE void _mm_storel_pi(__m64 *p, __m128 a) in _mm_storel_pi()
759 FORCE_INLINE void _mm_storeh_pi(__m64 *p, __m128 a) in _mm_storeh_pi()
767 FORCE_INLINE __m128 _mm_load1_ps(const float *p) in _mm_load1_ps()
794 FORCE_INLINE __m128 _mm_loadl_pi(__m128 a, __m64 const *p) in _mm_loadl_pi()
810 FORCE_INLINE __m128 _mm_loadr_ps(const float *p) in _mm_loadr_ps()
826 FORCE_INLINE __m128 _mm_loadh_pi(__m128 a, __m64 const *p) in _mm_loadh_pi()
834 FORCE_INLINE __m128 _mm_load_ps(const float *p) in _mm_load_ps()
841 FORCE_INLINE __m128 _mm_loadu_ps(const float *p) in _mm_loadu_ps()
854 FORCE_INLINE __m128i _mm_loadu_si16(const void *p) in _mm_loadu_si16()
866 FORCE_INLINE __m128i _mm_loadu_si64(const void *p) in _mm_loadu_si64()
880 FORCE_INLINE __m128d _mm_load_sd(const double *p) in _mm_load_sd()
897 FORCE_INLINE __m128d _mm_load_pd(const double *p) in _mm_load_pd()
910 FORCE_INLINE __m128d _mm_loadu_pd(const double *p) in _mm_loadu_pd()
918 FORCE_INLINE __m128 _mm_load_ss(const float *p) in _mm_load_ss()
923 FORCE_INLINE __m128i _mm_loadl_epi64(__m128i const *p) in _mm_loadl_epi64()
940 FORCE_INLINE __m128d _mm_loadl_pd(__m128d a, const double *p) in _mm_loadl_pd()
960 FORCE_INLINE __m128d _mm_loadr_pd(const double *p) in _mm_loadr_pd()
973 FORCE_INLINE __m128 _mm_move_ss(__m128 a, __m128 b) in _mm_move_ss()
987 FORCE_INLINE __m128i _mm_move_epi64(__m128i a) in _mm_move_epi64()
995 FORCE_INLINE __m128 _mm_undefined_ps(void) in _mm_undefined_ps()
1012 FORCE_INLINE __m128 _mm_andnot_ps(__m128 a, __m128 b) in _mm_andnot_ps()
1028 FORCE_INLINE __m128d _mm_andnot_pd(__m128d a, __m128d b) in _mm_andnot_pd()
1041 FORCE_INLINE __m128i _mm_andnot_si128(__m128i a, __m128i b) in _mm_andnot_si128()
1054 FORCE_INLINE __m128i _mm_and_si128(__m128i a, __m128i b) in _mm_and_si128()
1069 FORCE_INLINE __m128 _mm_and_ps(__m128 a, __m128 b) in _mm_and_ps()
1084 FORCE_INLINE __m128d _mm_and_pd(__m128d a, __m128d b) in _mm_and_pd()
1093 FORCE_INLINE __m128 _mm_or_ps(__m128 a, __m128 b) in _mm_or_ps()
1102 FORCE_INLINE __m128 _mm_xor_ps(__m128 a, __m128 b) in _mm_xor_ps()
1117 FORCE_INLINE __m128d _mm_xor_pd(__m128d a, __m128d b) in _mm_xor_pd()
1128 FORCE_INLINE __m128i _mm_or_si128(__m128i a, __m128i b) in _mm_or_si128()
1136 FORCE_INLINE __m128i _mm_xor_si128(__m128i a, __m128i b) in _mm_xor_si128()
1145 FORCE_INLINE __m128 _mm_movehdup_ps(__m128 a) in _mm_movehdup_ps()
1161 FORCE_INLINE __m128 _mm_moveldup_ps(__m128 a) in _mm_moveldup_ps()
1180 FORCE_INLINE __m128 _mm_movehl_ps(__m128 __A, __m128 __B) in _mm_movehl_ps()
1193 FORCE_INLINE __m128 _mm_movelh_ps(__m128 __A, __m128 __B) in _mm_movelh_ps()
1209 FORCE_INLINE __m128i _mm_abs_epi32(__m128i a) in _mm_abs_epi32()
1223 FORCE_INLINE __m128i _mm_abs_epi16(__m128i a) in _mm_abs_epi16()
1237 FORCE_INLINE __m128i _mm_abs_epi8(__m128i a) in _mm_abs_epi8()
1251 FORCE_INLINE __m64 _mm_abs_pi32(__m64 a) in _mm_abs_pi32()
1265 FORCE_INLINE __m64 _mm_abs_pi16(__m64 a) in _mm_abs_pi16()
1279 FORCE_INLINE __m64 _mm_abs_pi8(__m64 a) in _mm_abs_pi8()
1286 FORCE_INLINE __m128 _mm_shuffle_ps_1032(__m128 a, __m128 b) in _mm_shuffle_ps_1032()
1296 FORCE_INLINE __m128 _mm_shuffle_ps_2301(__m128 a, __m128 b) in _mm_shuffle_ps_2301()
1303 FORCE_INLINE __m128 _mm_shuffle_ps_0321(__m128 a, __m128 b) in _mm_shuffle_ps_0321()
1312 FORCE_INLINE __m128 _mm_shuffle_ps_2103(__m128 a, __m128 b) in _mm_shuffle_ps_2103()
1321 FORCE_INLINE __m128 _mm_shuffle_ps_1010(__m128 a, __m128 b) in _mm_shuffle_ps_1010()
1328 FORCE_INLINE __m128 _mm_shuffle_ps_1001(__m128 a, __m128 b) in _mm_shuffle_ps_1001()
1335 FORCE_INLINE __m128 _mm_shuffle_ps_0101(__m128 a, __m128 b) in _mm_shuffle_ps_0101()
1344 FORCE_INLINE __m128 _mm_shuffle_ps_3210(__m128 a, __m128 b) in _mm_shuffle_ps_3210()
1351 FORCE_INLINE __m128 _mm_shuffle_ps_0011(__m128 a, __m128 b) in _mm_shuffle_ps_0011()
1358 FORCE_INLINE __m128 _mm_shuffle_ps_0022(__m128 a, __m128 b) in _mm_shuffle_ps_0022()
1366 FORCE_INLINE __m128 _mm_shuffle_ps_2200(__m128 a, __m128 b) in _mm_shuffle_ps_2200()
1374 FORCE_INLINE __m128 _mm_shuffle_ps_3202(__m128 a, __m128 b) in _mm_shuffle_ps_3202()
1384 FORCE_INLINE __m128 _mm_shuffle_ps_1133(__m128 a, __m128 b) in _mm_shuffle_ps_1133()
1392 FORCE_INLINE __m128 _mm_shuffle_ps_2010(__m128 a, __m128 b) in _mm_shuffle_ps_2010()
1401 FORCE_INLINE __m128 _mm_shuffle_ps_2001(__m128 a, __m128 b) in _mm_shuffle_ps_2001()
1410 FORCE_INLINE __m128 _mm_shuffle_ps_2032(__m128 a, __m128 b) in _mm_shuffle_ps_2032()
1528 FORCE_INLINE __m128i _mm_shuffle_epi_1032(__m128i a) in _mm_shuffle_epi_1032()
1538 FORCE_INLINE __m128i _mm_shuffle_epi_2301(__m128i a) in _mm_shuffle_epi_2301()
1547 FORCE_INLINE __m128i _mm_shuffle_epi_0321(__m128i a) in _mm_shuffle_epi_0321()
1555 FORCE_INLINE __m128i _mm_shuffle_epi_2103(__m128i a) in _mm_shuffle_epi_2103()
1563 FORCE_INLINE __m128i _mm_shuffle_epi_1010(__m128i a) in _mm_shuffle_epi_1010()
1571 FORCE_INLINE __m128i _mm_shuffle_epi_1001(__m128i a) in _mm_shuffle_epi_1001()
1581 FORCE_INLINE __m128i _mm_shuffle_epi_0101(__m128i a) in _mm_shuffle_epi_0101()
1587 FORCE_INLINE __m128i _mm_shuffle_epi_2211(__m128i a) in _mm_shuffle_epi_2211()
1594 FORCE_INLINE __m128i _mm_shuffle_epi_0122(__m128i a) in _mm_shuffle_epi_0122()
1601 FORCE_INLINE __m128i _mm_shuffle_epi_3332(__m128i a) in _mm_shuffle_epi_3332()
1611 FORCE_INLINE __m128i _mm_shuffle_epi8(__m128i a, __m128i b) in _mm_shuffle_epi8()
1858 FORCE_INLINE __m128i _mm_blendv_epi8(__m128i _a, __m128i _b, __m128i _mask) in _mm_blendv_epi8()
1874 FORCE_INLINE __m128i _mm_srai_epi16(__m128i a, int imm) in _mm_srai_epi16()
1907 FORCE_INLINE __m128i _mm_slli_epi32(__m128i a, int imm) in _mm_slli_epi32()
1919 FORCE_INLINE __m128i _mm_slli_epi64(__m128i a, int imm) in _mm_slli_epi64()
2091 FORCE_INLINE __m128i _mm_sll_epi16(__m128i a, __m128i count) in _mm_sll_epi16()
2110 FORCE_INLINE __m128i _mm_sll_epi32(__m128i a, __m128i count) in _mm_sll_epi32()
2127 FORCE_INLINE __m128i _mm_sll_epi64(__m128i a, __m128i count) in _mm_sll_epi64()
2146 FORCE_INLINE __m128i _mm_srl_epi16(__m128i a, __m128i count) in _mm_srl_epi16()
2165 FORCE_INLINE __m128i _mm_srl_epi32(__m128i a, __m128i count) in _mm_srl_epi32()
2182 FORCE_INLINE __m128i _mm_srl_epi64(__m128i a, __m128i count) in _mm_srl_epi64()
2196 FORCE_INLINE int _mm_movemask_epi8(__m128i a) in _mm_movemask_epi8()
2295 FORCE_INLINE __m64 _mm_movepi64_pi64(__m128i a) in _mm_movepi64_pi64()
2307 FORCE_INLINE __m128i _mm_movpi64_epi64(__m64 a) in _mm_movpi64_epi64()
2317 FORCE_INLINE int _mm_movemask_ps(__m128 a) in _mm_movemask_ps()
2340 FORCE_INLINE int _mm_test_all_ones(__m128i a) in _mm_test_all_ones()
2349 FORCE_INLINE int _mm_test_all_zeros(__m128i a, __m128i mask) in _mm_test_all_zeros()
2367 FORCE_INLINE __m128 _mm_sub_ps(__m128 a, __m128 b) in _mm_sub_ps()
2382 FORCE_INLINE __m128 _mm_sub_ss(__m128 a, __m128 b) in _mm_sub_ss()
2391 FORCE_INLINE __m128i _mm_sub_epi64(__m128i a, __m128i b) in _mm_sub_epi64()
2406 FORCE_INLINE __m128i _mm_sub_epi32(__m128i a, __m128i b) in _mm_sub_epi32()
2412 FORCE_INLINE __m128i _mm_sub_epi16(__m128i a, __m128i b) in _mm_sub_epi16()
2418 FORCE_INLINE __m128i _mm_sub_epi8(__m128i a, __m128i b) in _mm_sub_epi8()
2429 FORCE_INLINE __m64 _mm_sub_si64(__m64 a, __m64 b) in _mm_sub_si64()
2438 FORCE_INLINE __m128i _mm_subs_epu16(__m128i a, __m128i b) in _mm_subs_epu16()
2453 FORCE_INLINE __m128i _mm_subs_epu8(__m128i a, __m128i b) in _mm_subs_epu8()
2468 FORCE_INLINE __m128i _mm_subs_epi8(__m128i a, __m128i b) in _mm_subs_epi8()
2483 FORCE_INLINE __m128i _mm_subs_epi16(__m128i a, __m128i b) in _mm_subs_epi16()
2489 FORCE_INLINE __m128i _mm_adds_epu16(__m128i a, __m128i b) in _mm_adds_epu16()
2509 FORCE_INLINE __m128i _mm_sign_epi8(__m128i _a, __m128i _b) in _mm_sign_epi8()
2548 FORCE_INLINE __m128i _mm_sign_epi16(__m128i _a, __m128i _b) in _mm_sign_epi16()
2585 FORCE_INLINE __m128i _mm_sign_epi32(__m128i _a, __m128i _b) in _mm_sign_epi32()
2625 FORCE_INLINE __m64 _mm_sign_pi16(__m64 _a, __m64 _b) in _mm_sign_pi16()
2666 FORCE_INLINE __m64 _mm_sign_pi32(__m64 _a, __m64 _b) in _mm_sign_pi32()
2707 FORCE_INLINE __m64 _mm_sign_pi8(__m64 _a, __m64 _b) in _mm_sign_pi8()
2741 FORCE_INLINE __m64 _mm_avg_pu16(__m64 a, __m64 b) in _mm_avg_pu16()
2756 FORCE_INLINE __m64 _mm_avg_pu8(__m64 a, __m64 b) in _mm_avg_pu8()
2793 FORCE_INLINE __m128i _mm_avg_epu8(__m128i a, __m128i b) in _mm_avg_epu8()
2808 FORCE_INLINE __m128i _mm_avg_epu16(__m128i a, __m128i b) in _mm_avg_epu16()
2822 FORCE_INLINE __m128 _mm_add_ps(__m128 a, __m128 b) in _mm_add_ps()
2831 FORCE_INLINE __m128d _mm_add_pd(__m128d a, __m128d b) in _mm_add_pd()
2851 FORCE_INLINE __m64 _mm_add_si64(__m64 a, __m64 b) in _mm_add_si64()
2859 FORCE_INLINE __m128 _mm_add_ss(__m128 a, __m128 b) in _mm_add_ss()
2870 FORCE_INLINE __m128i _mm_add_epi64(__m128i a, __m128i b) in _mm_add_epi64()
2885 FORCE_INLINE __m128i _mm_add_epi32(__m128i a, __m128i b) in _mm_add_epi32()
2894 FORCE_INLINE __m128i _mm_add_epi16(__m128i a, __m128i b) in _mm_add_epi16()
2903 FORCE_INLINE __m128i _mm_add_epi8(__m128i a, __m128i b) in _mm_add_epi8()
2918 FORCE_INLINE __m128i _mm_adds_epi16(__m128i a, __m128i b) in _mm_adds_epi16()
2933 FORCE_INLINE __m128i _mm_adds_epi8(__m128i a, __m128i b) in _mm_adds_epi8()
2942 FORCE_INLINE __m128i _mm_adds_epu8(__m128i a, __m128i b) in _mm_adds_epu8()
2957 FORCE_INLINE __m128i _mm_mullo_epi16(__m128i a, __m128i b) in _mm_mullo_epi16()
2966 FORCE_INLINE __m128i _mm_mullo_epi32(__m128i a, __m128i b) in _mm_mullo_epi32()
2993 FORCE_INLINE __m128 _mm_mul_ps(__m128 a, __m128 b) in _mm_mul_ps()
3007 FORCE_INLINE __m128 _mm_mul_ss(__m128 a, __m128 b) in _mm_mul_ss()
3017 FORCE_INLINE __m128i _mm_mul_epu32(__m128i a, __m128i b) in _mm_mul_epu32()
3031 FORCE_INLINE __m64 _mm_mul_su32(__m64 a, __m64 b) in _mm_mul_su32()
3042 FORCE_INLINE __m128i _mm_mul_epi32(__m128i a, __m128i b) in _mm_mul_epi32()
3058 FORCE_INLINE __m128i _mm_madd_epi16(__m128i a, __m128i b) in _mm_madd_epi16()
3080 FORCE_INLINE __m128i _mm_mulhrs_epi16(__m128i a, __m128i b) in _mm_mulhrs_epi16()
3110 FORCE_INLINE __m128i _mm_maddubs_epi16(__m128i _a, __m128i _b) in _mm_maddubs_epi16()
3149 FORCE_INLINE __m128 _mm_fmadd_ps(__m128 a, __m128 b, __m128 c) in _mm_fmadd_ps()
3165 FORCE_INLINE __m128 _mm_addsub_ps(__m128 a, __m128 b) in _mm_addsub_ps()
3176 FORCE_INLINE __m128i _mm_sad_epu8(__m128i a, __m128i b) in _mm_sad_epu8()
3190 FORCE_INLINE __m64 _mm_sad_pu8(__m64 a, __m64 b) in _mm_sad_pu8()
3221 FORCE_INLINE __m128 _mm_div_ps(__m128 a, __m128 b) in _mm_div_ps()
3236 FORCE_INLINE __m128 _mm_div_ss(__m128 a, __m128 b) in _mm_div_ss()
3247 FORCE_INLINE __m128 _mm_rcp_ps(__m128 in) in _mm_rcp_ps()
3263 FORCE_INLINE __m128 _mm_rcp_ss(__m128 a) in _mm_rcp_ss()
3278 FORCE_INLINE __m128 _mm_sqrt_ps(__m128 in) in _mm_sqrt_ps()
3293 FORCE_INLINE __m128 _mm_sqrt_ss(__m128 in) in _mm_sqrt_ss()
3304 FORCE_INLINE __m128 _mm_rsqrt_ps(__m128 in) in _mm_rsqrt_ps()
3314 FORCE_INLINE __m128 _mm_rsqrt_ss(__m128 in) in _mm_rsqrt_ss()
3328 FORCE_INLINE __m64 _mm_max_pi16(__m64 a, __m64 b) in _mm_max_pi16()
3348 FORCE_INLINE __m128 _mm_max_ps(__m128 a, __m128 b) in _mm_max_ps()
3369 FORCE_INLINE __m64 _mm_max_pu8(__m64 a, __m64 b) in _mm_max_pu8()
3395 FORCE_INLINE __m64 _mm_min_pi16(__m64 a, __m64 b) in _mm_min_pi16()
3415 FORCE_INLINE __m128 _mm_min_ps(__m128 a, __m128 b) in _mm_min_ps()
3436 FORCE_INLINE __m64 _mm_min_pu8(__m64 a, __m64 b) in _mm_min_pu8()
3456 FORCE_INLINE __m128 _mm_max_ss(__m128 a, __m128 b) in _mm_max_ss()
3466 FORCE_INLINE __m128 _mm_min_ss(__m128 a, __m128 b) in _mm_min_ss()
3476 FORCE_INLINE __m128i _mm_max_epu8(__m128i a, __m128i b) in _mm_max_epu8()
3485 FORCE_INLINE __m128i _mm_min_epu8(__m128i a, __m128i b) in _mm_min_epu8()
3494 FORCE_INLINE __m128i _mm_min_epi16(__m128i a, __m128i b) in _mm_min_epi16()
3503 FORCE_INLINE __m128i _mm_max_epi8(__m128i a, __m128i b) in _mm_max_epi8()
3512 FORCE_INLINE __m128i _mm_max_epi16(__m128i a, __m128i b) in _mm_max_epi16()
3529 FORCE_INLINE __m128i _mm_max_epi32(__m128i a, __m128i b) in _mm_max_epi32()
3545 FORCE_INLINE __m128i _mm_min_epi32(__m128i a, __m128i b) in _mm_min_epi32()
3554 FORCE_INLINE __m128i _mm_max_epu32(__m128i a, __m128i b) in _mm_max_epu32()
3563 FORCE_INLINE __m128i _mm_min_epu32(__m128i a, __m128i b) in _mm_min_epu32()
3573 FORCE_INLINE __m64 _mm_mulhi_pu16(__m64 a, __m64 b) in _mm_mulhi_pu16()
3588 FORCE_INLINE __m128i _mm_mulhi_epi16(__m128i a, __m128i b) in _mm_mulhi_epi16()
3608 FORCE_INLINE __m128 _mm_hadd_ps(__m128 a, __m128 b) in _mm_hadd_ps()
3625 FORCE_INLINE __m128i _mm_hadd_epi16(__m128i _a, __m128i _b) in _mm_hadd_epi16()
3641 FORCE_INLINE __m128 _mm_hsub_ps(__m128 _a, __m128 _b) in _mm_hsub_ps()
3657 FORCE_INLINE __m64 _mm_hadd_pi16(__m64 a, __m64 b) in _mm_hadd_pi16()
3666 FORCE_INLINE __m64 _mm_hadd_pi32(__m64 a, __m64 b) in _mm_hadd_pi32()
3674 FORCE_INLINE __m128i _mm_hsub_epi16(__m128i _a, __m128i _b) in _mm_hsub_epi16()
3689 FORCE_INLINE __m128i _mm_hadds_epi16(__m128i _a, __m128i _b) in _mm_hadds_epi16()
3712 FORCE_INLINE __m128i _mm_hsubs_epi16(__m128i _a, __m128i _b) in _mm_hsubs_epi16()
3734 FORCE_INLINE __m128i _mm_hadd_epi32(__m128i _a, __m128i _b) in _mm_hadd_epi32()
3745 FORCE_INLINE __m128i _mm_hsub_epi32(__m128i _a, __m128i _b) in _mm_hsub_epi32()
3760 FORCE_INLINE void sse2neon_kadd_f32(float *sum, float *c, float y) in sse2neon_kadd_f32()
3772 FORCE_INLINE __m128 _mm_dp_ps(__m128 a, __m128 b, const int imm) in _mm_dp_ps()
3816 FORCE_INLINE __m128 _mm_cmplt_ps(__m128 a, __m128 b) in _mm_cmplt_ps()
3824 FORCE_INLINE __m128 _mm_cmplt_ss(__m128 a, __m128 b) in _mm_cmplt_ss()
3837 FORCE_INLINE __m128 _mm_cmpgt_ps(__m128 a, __m128 b) in _mm_cmpgt_ps()
3845 FORCE_INLINE __m128 _mm_cmpgt_ss(__m128 a, __m128 b) in _mm_cmpgt_ss()
3852 FORCE_INLINE __m128 _mm_cmpge_ps(__m128 a, __m128 b) in _mm_cmpge_ps()
3860 FORCE_INLINE __m128 _mm_cmpge_ss(__m128 a, __m128 b) in _mm_cmpge_ss()
3873 FORCE_INLINE __m128 _mm_cmple_ps(__m128 a, __m128 b) in _mm_cmple_ps()
3881 FORCE_INLINE __m128 _mm_cmple_ss(__m128 a, __m128 b) in _mm_cmple_ss()
3888 FORCE_INLINE __m128 _mm_cmpeq_ps(__m128 a, __m128 b) in _mm_cmpeq_ps()
3896 FORCE_INLINE __m128 _mm_cmpeq_ss(__m128 a, __m128 b) in _mm_cmpeq_ss()
3903 FORCE_INLINE __m128 _mm_cmpneq_ps(__m128 a, __m128 b) in _mm_cmpneq_ps()
3911 FORCE_INLINE __m128 _mm_cmpneq_ss(__m128 a, __m128 b) in _mm_cmpneq_ss()
3918 FORCE_INLINE __m128 _mm_cmpnge_ps(__m128 a, __m128 b) in _mm_cmpnge_ps()
3925 FORCE_INLINE __m128 _mm_cmpnge_ss(__m128 a, __m128 b) in _mm_cmpnge_ss()
3932 FORCE_INLINE __m128 _mm_cmpngt_ps(__m128 a, __m128 b) in _mm_cmpngt_ps()
3939 FORCE_INLINE __m128 _mm_cmpngt_ss(__m128 a, __m128 b) in _mm_cmpngt_ss()
3946 FORCE_INLINE __m128 _mm_cmpnle_ps(__m128 a, __m128 b) in _mm_cmpnle_ps()
3953 FORCE_INLINE __m128 _mm_cmpnle_ss(__m128 a, __m128 b) in _mm_cmpnle_ss()
3960 FORCE_INLINE __m128 _mm_cmpnlt_ps(__m128 a, __m128 b) in _mm_cmpnlt_ps()
3967 FORCE_INLINE __m128 _mm_cmpnlt_ss(__m128 a, __m128 b) in _mm_cmpnlt_ss()
3975 FORCE_INLINE __m128i _mm_cmpeq_epi8(__m128i a, __m128i b) in _mm_cmpeq_epi8()
3984 FORCE_INLINE __m128i _mm_cmpeq_epi16(__m128i a, __m128i b) in _mm_cmpeq_epi16()
3992 FORCE_INLINE __m128i _mm_cmpeq_epi32(__m128i a, __m128i b) in _mm_cmpeq_epi32()
4000 FORCE_INLINE __m128i _mm_cmpeq_epi64(__m128i a, __m128i b) in _mm_cmpeq_epi64()
4018 FORCE_INLINE __m128i _mm_cmplt_epi8(__m128i a, __m128i b) in _mm_cmplt_epi8()
4033 FORCE_INLINE __m128i _mm_cmpgt_epi8(__m128i a, __m128i b) in _mm_cmpgt_epi8()
4048 FORCE_INLINE __m128i _mm_cmplt_epi16(__m128i a, __m128i b) in _mm_cmplt_epi16()
4063 FORCE_INLINE __m128i _mm_cmpgt_epi16(__m128i a, __m128i b) in _mm_cmpgt_epi16()
4073 FORCE_INLINE __m128i _mm_cmplt_epi32(__m128i a, __m128i b) in _mm_cmplt_epi32()
4082 FORCE_INLINE __m128i _mm_cmpgt_epi32(__m128i a, __m128i b) in _mm_cmpgt_epi32()
4090 FORCE_INLINE __m128i _mm_cmpgt_epi64(__m128i a, __m128i b) in _mm_cmpgt_epi64()
4131 FORCE_INLINE __m128 _mm_cmpord_ps(__m128 a, __m128 b) in _mm_cmpord_ps()
4145 FORCE_INLINE __m128 _mm_cmpord_ss(__m128 a, __m128 b) in _mm_cmpord_ss()
4152 FORCE_INLINE __m128 _mm_cmpunord_ps(__m128 a, __m128 b) in _mm_cmpunord_ps()
4163 FORCE_INLINE __m128 _mm_cmpunord_ss(__m128 a, __m128 b) in _mm_cmpunord_ss()
4173 FORCE_INLINE int _mm_comilt_ss(__m128 a, __m128 b) in _mm_comilt_ss()
4188 FORCE_INLINE int _mm_comigt_ss(__m128 a, __m128 b) in _mm_comigt_ss()
4205 FORCE_INLINE int _mm_comile_ss(__m128 a, __m128 b) in _mm_comile_ss()
4222 FORCE_INLINE int _mm_comige_ss(__m128 a, __m128 b) in _mm_comige_ss()
4239 FORCE_INLINE int _mm_comieq_ss(__m128 a, __m128 b) in _mm_comieq_ss()
4256 FORCE_INLINE int _mm_comineq_ss(__m128 a, __m128 b) in _mm_comineq_ss()
4292 FORCE_INLINE __m128 _mm_cvt_pi2ps(__m128 a, __m64 b) in _mm_cvt_pi2ps()
4307 FORCE_INLINE __m128 _mm_cvt_si2ss(__m128 a, int b) in _mm_cvt_si2ss()
4317 FORCE_INLINE int _mm_cvt_ss2si(__m128 a) in _mm_cvt_ss2si()
4345 FORCE_INLINE __m128 _mm_cvtpi16_ps(__m64 a) in _mm_cvtpi16_ps()
4361 FORCE_INLINE __m128 _mm_cvtpi32_ps(__m128 a, __m64 b) in _mm_cvtpi32_ps()
4380 FORCE_INLINE __m128 _mm_cvtpi32x2_ps(__m64 a, __m64 b) in _mm_cvtpi32x2_ps()
4396 FORCE_INLINE __m128 _mm_cvtpi8_ps(__m64 a) in _mm_cvtpi8_ps()
4412 FORCE_INLINE __m128 _mm_cvtpu16_ps(__m64 a) in _mm_cvtpu16_ps()
4429 FORCE_INLINE __m128 _mm_cvtpu8_ps(__m64 a) in _mm_cvtpu8_ps()
4438 FORCE_INLINE __m128i _mm_cvttps_epi32(__m128 a) in _mm_cvttps_epi32()
4446 FORCE_INLINE __m128 _mm_cvtepi32_ps(__m128i a) in _mm_cvtepi32_ps()
4453 FORCE_INLINE __m128i _mm_cvtepu8_epi16(__m128i a) in _mm_cvtepu8_epi16()
4463 FORCE_INLINE __m128i _mm_cvtepu8_epi32(__m128i a) in _mm_cvtepu8_epi32()
4473 FORCE_INLINE __m128i _mm_cvtepu8_epi64(__m128i a) in _mm_cvtepu8_epi64()
4484 FORCE_INLINE __m128i _mm_cvtepi8_epi16(__m128i a) in _mm_cvtepi8_epi16()
4493 FORCE_INLINE __m128i _mm_cvtepi8_epi32(__m128i a) in _mm_cvtepi8_epi32()
4503 FORCE_INLINE __m128i _mm_cvtepi8_epi64(__m128i a) in _mm_cvtepi8_epi64()
4514 FORCE_INLINE __m128i _mm_cvtepi16_epi32(__m128i a) in _mm_cvtepi16_epi32()
4522 FORCE_INLINE __m128i _mm_cvtepi16_epi64(__m128i a) in _mm_cvtepi16_epi64()
4532 FORCE_INLINE __m128i _mm_cvtepu16_epi32(__m128i a) in _mm_cvtepu16_epi32()
4540 FORCE_INLINE __m128i _mm_cvtepu16_epi64(__m128i a) in _mm_cvtepu16_epi64()
4550 FORCE_INLINE __m128i _mm_cvtepu32_epi64(__m128i a) in _mm_cvtepu32_epi64()
4558 FORCE_INLINE __m128i _mm_cvtepi32_epi64(__m128i a) in _mm_cvtepi32_epi64()
4575 FORCE_INLINE __m128i _mm_cvtps_epi32(__m128 a) in _mm_cvtps_epi32()
4604 FORCE_INLINE int _mm_cvtsi128_si32(__m128i a) in _mm_cvtsi128_si32()
4614 FORCE_INLINE int64_t _mm_cvtsi128_si64(__m128i a) in _mm_cvtsi128_si64()
4635 FORCE_INLINE __m128i _mm_cvtsi32_si128(int a) in _mm_cvtsi32_si128()
4645 FORCE_INLINE __m128i _mm_cvtsi64_si128(int64_t a) in _mm_cvtsi64_si128()
4653 FORCE_INLINE __m128d _mm_castps_pd(__m128 a) in _mm_castps_pd()
4661 FORCE_INLINE __m128i _mm_castps_si128(__m128 a) in _mm_castps_si128()
4669 FORCE_INLINE __m128 _mm_castsi128_ps(__m128i a) in _mm_castsi128_ps()
4676 FORCE_INLINE __m128i _mm_load_si128(const __m128i *p) in _mm_load_si128()
4688 FORCE_INLINE __m128d _mm_load1_pd(const double *p) in _mm_load1_pd()
4705 FORCE_INLINE __m128d _mm_loadh_pd(__m128d a, const double *p) in _mm_loadh_pd()
4736 FORCE_INLINE __m128i _mm_loadu_si128(const __m128i *p) in _mm_loadu_si128()
4747 FORCE_INLINE __m128i _mm_loadu_si32(const void *p) in _mm_loadu_si32()
4765 FORCE_INLINE __m128 _mm_cvtpd_ps(__m128d a) in _mm_cvtpd_ps()
4782 FORCE_INLINE double _mm_cvtsd_f64(__m128d a) in _mm_cvtsd_f64()
4802 FORCE_INLINE __m128d _mm_cvtps_pd(__m128 a) in _mm_cvtps_pd()
4817 FORCE_INLINE __m128i _mm_castpd_si128(__m128d a) in _mm_castpd_si128()
4825 FORCE_INLINE __m128 _mm_blendv_ps(__m128 a, __m128 b, __m128 mask) in _mm_blendv_ps()
4836 FORCE_INLINE __m128 _mm_round_ps(__m128 a, int rounding) in _mm_round_ps()
4882 FORCE_INLINE __m128 _mm_ceil_ps(__m128 a) in _mm_ceil_ps()
4891 FORCE_INLINE __m128 _mm_floor_ps(__m128 a) in _mm_floor_ps()
4917 FORCE_INLINE __m128i _mm_sra_epi16(__m128i a, __m128i count) in _mm_sra_epi16()
4934 FORCE_INLINE __m128i _mm_sra_epi32(__m128i a, __m128i count) in _mm_sra_epi32()
4945 FORCE_INLINE __m128i _mm_packs_epi16(__m128i a, __m128i b) in _mm_packs_epi16()
4965 FORCE_INLINE __m128i _mm_packus_epi16(const __m128i a, const __m128i b) in _mm_packus_epi16()
4985 FORCE_INLINE __m128i _mm_packs_epi32(__m128i a, __m128i b) in _mm_packs_epi32()
5003 FORCE_INLINE __m128i _mm_packus_epi32(__m128i a, __m128i b) in _mm_packus_epi32()
5022 FORCE_INLINE __m128i _mm_unpacklo_epi8(__m128i a, __m128i b) in _mm_unpacklo_epi8()
5048 FORCE_INLINE __m128i _mm_unpacklo_epi16(__m128i a, __m128i b) in _mm_unpacklo_epi16()
5070 FORCE_INLINE __m128i _mm_unpacklo_epi32(__m128i a, __m128i b) in _mm_unpacklo_epi32()
5083 FORCE_INLINE __m128i _mm_unpacklo_epi64(__m128i a, __m128i b) in _mm_unpacklo_epi64()
5099 FORCE_INLINE __m128 _mm_unpacklo_ps(__m128 a, __m128 b) in _mm_unpacklo_ps()
5121 FORCE_INLINE __m128 _mm_unpackhi_ps(__m128 a, __m128 b) in _mm_unpackhi_ps()
5146 FORCE_INLINE __m128i _mm_unpackhi_epi8(__m128i a, __m128i b) in _mm_unpackhi_epi8()
5174 FORCE_INLINE __m128i _mm_unpackhi_epi16(__m128i a, __m128i b) in _mm_unpackhi_epi16()
5190 FORCE_INLINE __m128i _mm_unpackhi_epi32(__m128i a, __m128i b) in _mm_unpackhi_epi32()
5208 FORCE_INLINE __m128i _mm_unpackhi_epi64(__m128i a, __m128i b) in _mm_unpackhi_epi64()
5232 FORCE_INLINE __m128i _mm_minpos_epu16(__m128i a) in _mm_minpos_epu16()
5281 FORCE_INLINE int _mm_testc_si128(__m128i a, __m128i b) in _mm_testc_si128()
5294 FORCE_INLINE int _mm_testz_si128(__m128i a, __m128i b) in _mm_testz_si128()
5373 FORCE_INLINE int _mm_popcnt_u32(unsigned int a) in _mm_popcnt_u32()
5400 FORCE_INLINE int64_t _mm_popcnt_u64(uint64_t a) in _mm_popcnt_u64()
5447 FORCE_INLINE uint64x2_t _sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b) in _sse2neon_vmull_p64()
5560 FORCE_INLINE __m128i _mm_clmulepi64_si128(__m128i _a, __m128i _b, const int imm) in _mm_clmulepi64_si128()
5637 FORCE_INLINE __m128i _mm_aesenc_si128(__m128i EncBlock, __m128i RoundKey) in _mm_aesenc_si128()
5713 FORCE_INLINE __m128i _mm_aesenclast_si128(__m128i a, __m128i RoundKey) in _mm_aesenclast_si128()
5746 FORCE_INLINE __m128i _mm_aeskeygenassist_si128(__m128i key, const int rcon) in _mm_aeskeygenassist_si128()
5766 FORCE_INLINE __m128i _mm_aesenc_si128(__m128i a, __m128i b) in _mm_aesenc_si128()
5774 FORCE_INLINE __m128i _mm_aesenclast_si128(__m128i a, __m128i RoundKey) in _mm_aesenclast_si128()
5781 FORCE_INLINE __m128i _mm_aeskeygenassist_si128(__m128i a, const int rcon) in _mm_aeskeygenassist_si128()
5803 FORCE_INLINE void _mm_sfence(void) in _mm_sfence()
5811 FORCE_INLINE void _mm_stream_ps(float *p, __m128 a) in _mm_stream_ps()
5824 FORCE_INLINE void _mm_stream_si128(__m128i *p, __m128i a) in _mm_stream_si128()
5840 FORCE_INLINE __m128i _mm_stream_load_si128(__m128i *p) in _mm_stream_load_si128()
5852 FORCE_INLINE void _mm_clflush(void const *p) in _mm_clflush()
5861 FORCE_INLINE void *_mm_malloc(size_t size, size_t align) in _mm_malloc()
5873 FORCE_INLINE void _mm_free(void *addr) in _mm_free()
5881 FORCE_INLINE uint32_t _mm_crc32_u8(uint32_t crc, uint8_t v) in _mm_crc32_u8()
5902 FORCE_INLINE uint32_t _mm_crc32_u16(uint32_t crc, uint16_t v) in _mm_crc32_u16()
5918 FORCE_INLINE uint32_t _mm_crc32_u32(uint32_t crc, uint32_t v) in _mm_crc32_u32()
5934 FORCE_INLINE uint64_t _mm_crc32_u64(uint64_t crc, uint64_t v) in _mm_crc32_u64()