Lines Matching refs:__m512h

20 typedef _Float16 __m512h __attribute__((__vector_size__(64), __aligned__(64)));
34 static __inline__ _Float16 __DEFAULT_FN_ATTRS512 _mm512_cvtsh_h(__m512h __a) {
51 static __inline __m512h __DEFAULT_FN_ATTRS512 _mm512_setzero_ph(void) {
52 return (__m512h){0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
61 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_undefined_ph(void) {
62 return (__m512h)__builtin_ia32_undef512();
65 static __inline __m512h __DEFAULT_FN_ATTRS512 _mm512_set1_ph(_Float16 __h) {
66 return (__m512h)(__v32hf){__h, __h, __h, __h, __h, __h, __h, __h,
72 static __inline __m512h __DEFAULT_FN_ATTRS512
81 return (__m512h)(__v32hf){__h32, __h31, __h30, __h29, __h28, __h27, __h26,
96 static __inline __m512h __DEFAULT_FN_ATTRS512
98 return (__m512h)_mm512_set1_ps(__builtin_bit_cast(float, h));
109 static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_castph_ps(__m512h __a) {
121 static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_castph_pd(__m512h __a) {
135 _mm512_castph_si512(__m512h __a) {
147 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_castps_ph(__m512 __a) {
148 return (__m512h)__a;
159 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_castpd_ph(__m512d __a) {
160 return (__m512h)__a;
172 static __inline__ __m512h __DEFAULT_FN_ATTRS512
174 return (__m512h)__a;
183 _mm512_castph512_ph128(__m512h __a) {
188 _mm512_castph512_ph256(__m512h __a) {
199 static __inline__ __m512h __DEFAULT_FN_ATTRS512
206 static __inline__ __m512h __DEFAULT_FN_ATTRS512
245 static __inline__ __m512h __DEFAULT_FN_ATTRS512
265 static __inline__ __m512h __DEFAULT_FN_ATTRS512
351 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_add_ph(__m512h __A,
352 __m512h __B) {
353 return (__m512h)((__v32hf)__A + (__v32hf)__B);
356 static __inline__ __m512h __DEFAULT_FN_ATTRS512
357 _mm512_mask_add_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
358 return (__m512h)__builtin_ia32_selectph_512(
362 static __inline__ __m512h __DEFAULT_FN_ATTRS512
363 _mm512_maskz_add_ph(__mmask32 __U, __m512h __A, __m512h __B) {
364 return (__m512h)__builtin_ia32_selectph_512((__mmask32)__U,
370 ((__m512h)__builtin_ia32_addph512((__v32hf)(__m512h)(A), \
371 (__v32hf)(__m512h)(B), (int)(R)))
374 ((__m512h)__builtin_ia32_selectph_512( \
376 (__v32hf)(__m512h)(W)))
379 ((__m512h)__builtin_ia32_selectph_512( \
383 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_sub_ph(__m512h __A,
384 __m512h __B) {
385 return (__m512h)((__v32hf)__A - (__v32hf)__B);
388 static __inline__ __m512h __DEFAULT_FN_ATTRS512
389 _mm512_mask_sub_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
390 return (__m512h)__builtin_ia32_selectph_512(
394 static __inline__ __m512h __DEFAULT_FN_ATTRS512
395 _mm512_maskz_sub_ph(__mmask32 __U, __m512h __A, __m512h __B) {
396 return (__m512h)__builtin_ia32_selectph_512((__mmask32)__U,
402 ((__m512h)__builtin_ia32_subph512((__v32hf)(__m512h)(A), \
403 (__v32hf)(__m512h)(B), (int)(R)))
406 ((__m512h)__builtin_ia32_selectph_512( \
408 (__v32hf)(__m512h)(W)))
411 ((__m512h)__builtin_ia32_selectph_512( \
415 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_mul_ph(__m512h __A,
416 __m512h __B) {
417 return (__m512h)((__v32hf)__A * (__v32hf)__B);
420 static __inline__ __m512h __DEFAULT_FN_ATTRS512
421 _mm512_mask_mul_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
422 return (__m512h)__builtin_ia32_selectph_512(
426 static __inline__ __m512h __DEFAULT_FN_ATTRS512
427 _mm512_maskz_mul_ph(__mmask32 __U, __m512h __A, __m512h __B) {
428 return (__m512h)__builtin_ia32_selectph_512((__mmask32)__U,
434 ((__m512h)__builtin_ia32_mulph512((__v32hf)(__m512h)(A), \
435 (__v32hf)(__m512h)(B), (int)(R)))
438 ((__m512h)__builtin_ia32_selectph_512( \
440 (__v32hf)(__m512h)(W)))
443 ((__m512h)__builtin_ia32_selectph_512( \
447 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_div_ph(__m512h __A,
448 __m512h __B) {
449 return (__m512h)((__v32hf)__A / (__v32hf)__B);
452 static __inline__ __m512h __DEFAULT_FN_ATTRS512
453 _mm512_mask_div_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
454 return (__m512h)__builtin_ia32_selectph_512(
458 static __inline__ __m512h __DEFAULT_FN_ATTRS512
459 _mm512_maskz_div_ph(__mmask32 __U, __m512h __A, __m512h __B) {
460 return (__m512h)__builtin_ia32_selectph_512((__mmask32)__U,
466 ((__m512h)__builtin_ia32_divph512((__v32hf)(__m512h)(A), \
467 (__v32hf)(__m512h)(B), (int)(R)))
470 ((__m512h)__builtin_ia32_selectph_512( \
472 (__v32hf)(__m512h)(W)))
475 ((__m512h)__builtin_ia32_selectph_512( \
479 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_min_ph(__m512h __A,
480 __m512h __B) {
481 return (__m512h)__builtin_ia32_minph512((__v32hf)__A, (__v32hf)__B,
485 static __inline__ __m512h __DEFAULT_FN_ATTRS512
486 _mm512_mask_min_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
487 return (__m512h)__builtin_ia32_selectph_512(
491 static __inline__ __m512h __DEFAULT_FN_ATTRS512
492 _mm512_maskz_min_ph(__mmask32 __U, __m512h __A, __m512h __B) {
493 return (__m512h)__builtin_ia32_selectph_512((__mmask32)__U,
499 ((__m512h)__builtin_ia32_minph512((__v32hf)(__m512h)(A), \
500 (__v32hf)(__m512h)(B), (int)(R)))
503 ((__m512h)__builtin_ia32_selectph_512( \
505 (__v32hf)(__m512h)(W)))
508 ((__m512h)__builtin_ia32_selectph_512( \
512 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_max_ph(__m512h __A,
513 __m512h __B) {
514 return (__m512h)__builtin_ia32_maxph512((__v32hf)__A, (__v32hf)__B,
518 static __inline__ __m512h __DEFAULT_FN_ATTRS512
519 _mm512_mask_max_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
520 return (__m512h)__builtin_ia32_selectph_512(
524 static __inline__ __m512h __DEFAULT_FN_ATTRS512
525 _mm512_maskz_max_ph(__mmask32 __U, __m512h __A, __m512h __B) {
526 return (__m512h)__builtin_ia32_selectph_512((__mmask32)__U,
532 ((__m512h)__builtin_ia32_maxph512((__v32hf)(__m512h)(A), \
533 (__v32hf)(__m512h)(B), (int)(R)))
536 ((__m512h)__builtin_ia32_selectph_512( \
538 (__v32hf)(__m512h)(W)))
541 ((__m512h)__builtin_ia32_selectph_512( \
545 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_abs_ph(__m512h __A) {
546 return (__m512h)_mm512_and_epi32(_mm512_set1_epi32(0x7FFF7FFF), (__m512i)__A);
549 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_conj_pch(__m512h __A) {
550 return (__m512h)_mm512_xor_ps((__m512)__A, _mm512_set1_ps(-0.0f));
553 static __inline__ __m512h __DEFAULT_FN_ATTRS512
554 _mm512_mask_conj_pch(__m512h __W, __mmask16 __U, __m512h __A) {
555 return (__m512h)__builtin_ia32_selectps_512(
559 static __inline__ __m512h __DEFAULT_FN_ATTRS512
560 _mm512_maskz_conj_pch(__mmask16 __U, __m512h __A) {
561 return (__m512h)__builtin_ia32_selectps_512((__mmask16)__U,
789 ((__mmask32)__builtin_ia32_cmpph512_mask((__v32hf)(__m512h)(A), \
790 (__v32hf)(__m512h)(B), (int)(P), \
794 ((__mmask32)__builtin_ia32_cmpph512_mask((__v32hf)(__m512h)(A), \
795 (__v32hf)(__m512h)(B), (int)(P), \
846 static __inline__ __m512h __DEFAULT_FN_ATTRS512
848 return *(const __m512h *)__p;
860 static __inline__ __m512h __DEFAULT_FN_ATTRS512
899 __m512h __A) {
900 *(__m512h *)__P = __A;
914 __m512h __A) {
968 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_rcp_ph(__m512h __A) {
969 return (__m512h)__builtin_ia32_rcpph512_mask(
973 static __inline__ __m512h __DEFAULT_FN_ATTRS512
974 _mm512_mask_rcp_ph(__m512h __W, __mmask32 __U, __m512h __A) {
975 return (__m512h)__builtin_ia32_rcpph512_mask((__v32hf)__A, (__v32hf)__W,
979 static __inline__ __m512h __DEFAULT_FN_ATTRS512
980 _mm512_maskz_rcp_ph(__mmask32 __U, __m512h __A) {
981 return (__m512h)__builtin_ia32_rcpph512_mask(
985 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_rsqrt_ph(__m512h __A) {
986 return (__m512h)__builtin_ia32_rsqrtph512_mask(
990 static __inline__ __m512h __DEFAULT_FN_ATTRS512
991 _mm512_mask_rsqrt_ph(__m512h __W, __mmask32 __U, __m512h __A) {
992 return (__m512h)__builtin_ia32_rsqrtph512_mask((__v32hf)__A, (__v32hf)__W,
996 static __inline__ __m512h __DEFAULT_FN_ATTRS512
997 _mm512_maskz_rsqrt_ph(__mmask32 __U, __m512h __A) {
998 return (__m512h)__builtin_ia32_rsqrtph512_mask(
1003 ((__m512h)__builtin_ia32_getmantph512_mask( \
1004 (__v32hf)(__m512h)(A), (int)(((C) << 2) | (B)), \
1009 ((__m512h)__builtin_ia32_getmantph512_mask( \
1010 (__v32hf)(__m512h)(A), (int)(((C) << 2) | (B)), (__v32hf)(__m512h)(W), \
1014 ((__m512h)__builtin_ia32_getmantph512_mask( \
1015 (__v32hf)(__m512h)(A), (int)(((C) << 2) | (B)), \
1019 ((__m512h)__builtin_ia32_getmantph512_mask( \
1020 (__v32hf)(__m512h)(A), (int)(((C) << 2) | (B)), \
1024 ((__m512h)__builtin_ia32_getmantph512_mask( \
1025 (__v32hf)(__m512h)(A), (int)(((C) << 2) | (B)), (__v32hf)(__m512h)(W), \
1029 ((__m512h)__builtin_ia32_getmantph512_mask( \
1030 (__v32hf)(__m512h)(A), (int)(((C) << 2) | (B)), \
1033 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_getexp_ph(__m512h __A) {
1034 return (__m512h)__builtin_ia32_getexpph512_mask(
1039 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1040 _mm512_mask_getexp_ph(__m512h __W, __mmask32 __U, __m512h __A) {
1041 return (__m512h)__builtin_ia32_getexpph512_mask(
1045 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1046 _mm512_maskz_getexp_ph(__mmask32 __U, __m512h __A) {
1047 return (__m512h)__builtin_ia32_getexpph512_mask(
1053 ((__m512h)__builtin_ia32_getexpph512_mask((__v32hf)(__m512h)(A), \
1058 ((__m512h)__builtin_ia32_getexpph512_mask( \
1059 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(W), (__mmask32)(U), (int)(R)))
1062 ((__m512h)__builtin_ia32_getexpph512_mask((__v32hf)(__m512h)(A), \
1066 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_scalef_ph(__m512h __A,
1067 __m512h __B) {
1068 return (__m512h)__builtin_ia32_scalefph512_mask(
1073 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1074 _mm512_mask_scalef_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
1075 return (__m512h)__builtin_ia32_scalefph512_mask((__v32hf)__A, (__v32hf)__B,
1080 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1081 _mm512_maskz_scalef_ph(__mmask32 __U, __m512h __A, __m512h __B) {
1082 return (__m512h)__builtin_ia32_scalefph512_mask(
1088 ((__m512h)__builtin_ia32_scalefph512_mask( \
1089 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), \
1093 ((__m512h)__builtin_ia32_scalefph512_mask( \
1094 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(W), \
1098 ((__m512h)__builtin_ia32_scalefph512_mask( \
1099 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), \
1103 ((__m512h)__builtin_ia32_rndscaleph_mask( \
1104 (__v32hf)(__m512h)(A), (int)(B), (__v32hf)(__m512h)(A), (__mmask32)-1, \
1108 ((__m512h)__builtin_ia32_rndscaleph_mask( \
1109 (__v32hf)(__m512h)(C), (int)(imm), (__v32hf)(__m512h)(A), \
1113 ((__m512h)__builtin_ia32_rndscaleph_mask( \
1114 (__v32hf)(__m512h)(B), (int)(imm), (__v32hf)_mm512_setzero_ph(), \
1118 ((__m512h)__builtin_ia32_rndscaleph_mask((__v32hf)(__m512h)(C), (int)(imm), \
1119 (__v32hf)(__m512h)(A), \
1123 ((__m512h)__builtin_ia32_rndscaleph_mask((__v32hf)(__m512h)(B), (int)(imm), \
1128 ((__m512h)__builtin_ia32_rndscaleph_mask((__v32hf)(__m512h)(A), (int)(imm), \
1133 ((__m512h)__builtin_ia32_reduceph512_mask( \
1134 (__v32hf)(__m512h)(A), (int)(imm), (__v32hf)_mm512_undefined_ph(), \
1138 ((__m512h)__builtin_ia32_reduceph512_mask( \
1139 (__v32hf)(__m512h)(A), (int)(imm), (__v32hf)(__m512h)(W), \
1143 ((__m512h)__builtin_ia32_reduceph512_mask( \
1144 (__v32hf)(__m512h)(A), (int)(imm), (__v32hf)_mm512_setzero_ph(), \
1148 ((__m512h)__builtin_ia32_reduceph512_mask((__v32hf)(__m512h)(A), (int)(imm), \
1149 (__v32hf)(__m512h)(W), \
1153 ((__m512h)__builtin_ia32_reduceph512_mask((__v32hf)(__m512h)(A), (int)(imm), \
1158 ((__m512h)__builtin_ia32_reduceph512_mask((__v32hf)(__m512h)(A), (int)(imm), \
1366 ((__m512h)__builtin_ia32_sqrtph512((__v32hf)(__m512h)(A), (int)(R)))
1369 ((__m512h)__builtin_ia32_selectph_512( \
1371 (__v32hf)(__m512h)(W)))
1374 ((__m512h)__builtin_ia32_selectph_512( \
1378 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_sqrt_ph(__m512h __A) {
1379 return (__m512h)__builtin_ia32_sqrtph512((__v32hf)__A,
1383 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1384 _mm512_mask_sqrt_ph(__m512h __W, __mmask32 __U, __m512h __A) {
1385 return (__m512h)__builtin_ia32_selectph_512(
1388 (__v32hf)(__m512h)(__W));
1391 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1392 _mm512_maskz_sqrt_ph(__mmask32 __U, __m512h __A) {
1393 return (__m512h)__builtin_ia32_selectph_512(
1439 ((__mmask32)__builtin_ia32_fpclassph512_mask((__v32hf)(__m512h)(A), \
1443 ((__mmask32)__builtin_ia32_fpclassph512_mask((__v32hf)(__m512h)(A), \
1681 _mm512_cvtph_epi16(__m512h __A) {
1688 _mm512_mask_cvtph_epi16(__m512i __W, __mmask32 __U, __m512h __A) {
1694 _mm512_maskz_cvtph_epi16(__mmask32 __U, __m512h __A) {
1715 _mm512_cvttph_epi16(__m512h __A) {
1722 _mm512_mask_cvttph_epi16(__m512i __W, __mmask32 __U, __m512h __A) {
1728 _mm512_maskz_cvttph_epi16(__mmask32 __U, __m512h __A) {
1735 ((__m512h)__builtin_ia32_vcvtw2ph512_mask((__v32hi)(A), \
1740 ((__m512h)__builtin_ia32_vcvtw2ph512_mask((__v32hi)(A), (__v32hf)(W), \
1744 ((__m512h)__builtin_ia32_vcvtw2ph512_mask( \
1747 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1749 return (__m512h)__builtin_ia32_vcvtw2ph512_mask(
1754 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1755 _mm512_mask_cvtepi16_ph(__m512h __W, __mmask32 __U, __m512i __A) {
1756 return (__m512h)__builtin_ia32_vcvtw2ph512_mask(
1760 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1762 return (__m512h)__builtin_ia32_vcvtw2ph512_mask(
1782 _mm512_cvtph_epu16(__m512h __A) {
1789 _mm512_mask_cvtph_epu16(__m512i __W, __mmask32 __U, __m512h __A) {
1795 _mm512_maskz_cvtph_epu16(__mmask32 __U, __m512h __A) {
1816 _mm512_cvttph_epu16(__m512h __A) {
1823 _mm512_mask_cvttph_epu16(__m512i __W, __mmask32 __U, __m512h __A) {
1829 _mm512_maskz_cvttph_epu16(__mmask32 __U, __m512h __A) {
1836 ((__m512h)__builtin_ia32_vcvtuw2ph512_mask((__v32hu)(A), \
1841 ((__m512h)__builtin_ia32_vcvtuw2ph512_mask((__v32hu)(A), (__v32hf)(W), \
1845 ((__m512h)__builtin_ia32_vcvtuw2ph512_mask( \
1848 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1850 return (__m512h)__builtin_ia32_vcvtuw2ph512_mask(
1855 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1856 _mm512_mask_cvtepu16_ph(__m512h __W, __mmask32 __U, __m512i __A) {
1857 return (__m512h)__builtin_ia32_vcvtuw2ph512_mask(
1861 static __inline__ __m512h __DEFAULT_FN_ATTRS512
1863 return (__m512h)__builtin_ia32_vcvtuw2ph512_mask(
2445 ((__m512h)__builtin_ia32_vfmaddph512_mask( \
2446 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2450 ((__m512h)__builtin_ia32_vfmaddph512_mask( \
2451 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2455 ((__m512h)__builtin_ia32_vfmaddph512_mask3( \
2456 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2460 ((__m512h)__builtin_ia32_vfmaddph512_maskz( \
2461 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2465 ((__m512h)__builtin_ia32_vfmaddph512_mask( \
2466 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2470 ((__m512h)__builtin_ia32_vfmaddph512_mask( \
2471 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2475 ((__m512h)__builtin_ia32_vfmaddph512_maskz( \
2476 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2480 ((__m512h)__builtin_ia32_vfmaddph512_mask( \
2481 (__v32hf)(__m512h)(A), -(__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2485 ((__m512h)__builtin_ia32_vfmaddph512_mask3( \
2486 -(__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2490 ((__m512h)__builtin_ia32_vfmaddph512_maskz( \
2491 -(__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2495 ((__m512h)__builtin_ia32_vfmaddph512_mask( \
2496 (__v32hf)(__m512h)(A), -(__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2500 ((__m512h)__builtin_ia32_vfmaddph512_maskz( \
2501 -(__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2504 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_fmadd_ph(__m512h __A,
2505 __m512h __B,
2506 __m512h __C) {
2507 return (__m512h)__builtin_ia32_vfmaddph512_mask((__v32hf)__A, (__v32hf)__B,
2512 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2513 _mm512_mask_fmadd_ph(__m512h __A, __mmask32 __U, __m512h __B, __m512h __C) {
2514 return (__m512h)__builtin_ia32_vfmaddph512_mask((__v32hf)__A, (__v32hf)__B,
2519 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2520 _mm512_mask3_fmadd_ph(__m512h __A, __m512h __B, __m512h __C, __mmask32 __U) {
2521 return (__m512h)__builtin_ia32_vfmaddph512_mask3((__v32hf)__A, (__v32hf)__B,
2526 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2527 _mm512_maskz_fmadd_ph(__mmask32 __U, __m512h __A, __m512h __B, __m512h __C) {
2528 return (__m512h)__builtin_ia32_vfmaddph512_maskz((__v32hf)__A, (__v32hf)__B,
2533 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_fmsub_ph(__m512h __A,
2534 __m512h __B,
2535 __m512h __C) {
2536 return (__m512h)__builtin_ia32_vfmaddph512_mask((__v32hf)__A, (__v32hf)__B,
2541 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2542 _mm512_mask_fmsub_ph(__m512h __A, __mmask32 __U, __m512h __B, __m512h __C) {
2543 return (__m512h)__builtin_ia32_vfmaddph512_mask((__v32hf)__A, (__v32hf)__B,
2548 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2549 _mm512_maskz_fmsub_ph(__mmask32 __U, __m512h __A, __m512h __B, __m512h __C) {
2550 return (__m512h)__builtin_ia32_vfmaddph512_maskz(
2555 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_fnmadd_ph(__m512h __A,
2556 __m512h __B,
2557 __m512h __C) {
2558 return (__m512h)__builtin_ia32_vfmaddph512_mask((__v32hf)__A, -(__v32hf)__B,
2563 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2564 _mm512_mask3_fnmadd_ph(__m512h __A, __m512h __B, __m512h __C, __mmask32 __U) {
2565 return (__m512h)__builtin_ia32_vfmaddph512_mask3(-(__v32hf)__A, (__v32hf)__B,
2570 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2571 _mm512_maskz_fnmadd_ph(__mmask32 __U, __m512h __A, __m512h __B, __m512h __C) {
2572 return (__m512h)__builtin_ia32_vfmaddph512_maskz(-(__v32hf)__A, (__v32hf)__B,
2577 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_fnmsub_ph(__m512h __A,
2578 __m512h __B,
2579 __m512h __C) {
2580 return (__m512h)__builtin_ia32_vfmaddph512_mask((__v32hf)__A, -(__v32hf)__B,
2585 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2586 _mm512_maskz_fnmsub_ph(__mmask32 __U, __m512h __A, __m512h __B, __m512h __C) {
2587 return (__m512h)__builtin_ia32_vfmaddph512_maskz(
2593 ((__m512h)__builtin_ia32_vfmaddsubph512_mask( \
2594 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2598 ((__m512h)__builtin_ia32_vfmaddsubph512_mask( \
2599 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2603 ((__m512h)__builtin_ia32_vfmaddsubph512_mask3( \
2604 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2608 ((__m512h)__builtin_ia32_vfmaddsubph512_maskz( \
2609 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2613 ((__m512h)__builtin_ia32_vfmaddsubph512_mask( \
2614 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2618 ((__m512h)__builtin_ia32_vfmaddsubph512_mask( \
2619 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2623 ((__m512h)__builtin_ia32_vfmaddsubph512_maskz( \
2624 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2627 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2628 _mm512_fmaddsub_ph(__m512h __A, __m512h __B, __m512h __C) {
2629 return (__m512h)__builtin_ia32_vfmaddsubph512_mask(
2634 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2635 _mm512_mask_fmaddsub_ph(__m512h __A, __mmask32 __U, __m512h __B, __m512h __C) {
2636 return (__m512h)__builtin_ia32_vfmaddsubph512_mask(
2641 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2642 _mm512_mask3_fmaddsub_ph(__m512h __A, __m512h __B, __m512h __C, __mmask32 __U) {
2643 return (__m512h)__builtin_ia32_vfmaddsubph512_mask3(
2648 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2649 _mm512_maskz_fmaddsub_ph(__mmask32 __U, __m512h __A, __m512h __B, __m512h __C) {
2650 return (__m512h)__builtin_ia32_vfmaddsubph512_maskz(
2655 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2656 _mm512_fmsubadd_ph(__m512h __A, __m512h __B, __m512h __C) {
2657 return (__m512h)__builtin_ia32_vfmaddsubph512_mask(
2662 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2663 _mm512_mask_fmsubadd_ph(__m512h __A, __mmask32 __U, __m512h __B, __m512h __C) {
2664 return (__m512h)__builtin_ia32_vfmaddsubph512_mask(
2669 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2670 _mm512_maskz_fmsubadd_ph(__mmask32 __U, __m512h __A, __m512h __B, __m512h __C) {
2671 return (__m512h)__builtin_ia32_vfmaddsubph512_maskz(
2677 ((__m512h)__builtin_ia32_vfmsubph512_mask3( \
2678 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2681 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2682 _mm512_mask3_fmsub_ph(__m512h __A, __m512h __B, __m512h __C, __mmask32 __U) {
2683 return (__m512h)__builtin_ia32_vfmsubph512_mask3((__v32hf)__A, (__v32hf)__B,
2689 ((__m512h)__builtin_ia32_vfmsubaddph512_mask3( \
2690 (__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2693 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2694 _mm512_mask3_fmsubadd_ph(__m512h __A, __m512h __B, __m512h __C, __mmask32 __U) {
2695 return (__m512h)__builtin_ia32_vfmsubaddph512_mask3(
2701 ((__m512h)__builtin_ia32_vfmaddph512_mask( \
2702 (__v32hf)(__m512h)(A), -(__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2705 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2706 _mm512_mask_fnmadd_ph(__m512h __A, __mmask32 __U, __m512h __B, __m512h __C) {
2707 return (__m512h)__builtin_ia32_vfmaddph512_mask((__v32hf)__A, -(__v32hf)__B,
2713 ((__m512h)__builtin_ia32_vfmaddph512_mask( \
2714 (__v32hf)(__m512h)(A), -(__v32hf)(__m512h)(B), -(__v32hf)(__m512h)(C), \
2718 ((__m512h)__builtin_ia32_vfmsubph512_mask3( \
2719 -(__v32hf)(__m512h)(A), (__v32hf)(__m512h)(B), (__v32hf)(__m512h)(C), \
2722 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2723 _mm512_mask_fnmsub_ph(__m512h __A, __mmask32 __U, __m512h __B, __m512h __C) {
2724 return (__m512h)__builtin_ia32_vfmaddph512_mask((__v32hf)__A, -(__v32hf)__B,
2729 static __inline__ __m512h __DEFAULT_FN_ATTRS512
2730 _mm512_mask3_fnmsub_ph(__m512h __A, __m512h __B, __m512h __C, __mmask32 __U) {
2731 return (__m512h)__builtin_ia32_vfmsubph512_mask3(-(__v32hf)__A, (__v32hf)__B,
3098 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_fcmul_pch(__m512h __A,
3099 __m512h __B) {
3100 return (__m512h)__builtin_ia32_vfcmulcph512_mask(
3105 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3106 _mm512_mask_fcmul_pch(__m512h __W, __mmask16 __U, __m512h __A, __m512h __B) {
3107 return (__m512h)__builtin_ia32_vfcmulcph512_mask((__v16sf)__A, (__v16sf)__B,
3112 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3113 _mm512_maskz_fcmul_pch(__mmask16 __U, __m512h __A, __m512h __B) {
3114 return (__m512h)__builtin_ia32_vfcmulcph512_mask(
3120 ((__m512h)__builtin_ia32_vfcmulcph512_mask( \
3121 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), \
3122 (__v16sf)(__m512h)_mm512_undefined_ph(), (__mmask16)-1, (int)(R)))
3125 ((__m512h)__builtin_ia32_vfcmulcph512_mask( \
3126 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(W), \
3130 ((__m512h)__builtin_ia32_vfcmulcph512_mask( \
3131 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), \
3132 (__v16sf)(__m512h)_mm512_setzero_ph(), (__mmask16)(U), (int)(R)))
3134 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_fmul_pch(__m512h __A,
3135 __m512h __B) {
3136 return (__m512h)__builtin_ia32_vfmulcph512_mask(
3141 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3142 _mm512_mask_fmul_pch(__m512h __W, __mmask16 __U, __m512h __A, __m512h __B) {
3143 return (__m512h)__builtin_ia32_vfmulcph512_mask((__v16sf)__A, (__v16sf)__B,
3148 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3149 _mm512_maskz_fmul_pch(__mmask16 __U, __m512h __A, __m512h __B) {
3150 return (__m512h)__builtin_ia32_vfmulcph512_mask(
3156 ((__m512h)__builtin_ia32_vfmulcph512_mask( \
3157 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), \
3158 (__v16sf)(__m512h)_mm512_undefined_ph(), (__mmask16)-1, (int)(R)))
3161 ((__m512h)__builtin_ia32_vfmulcph512_mask( \
3162 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(W), \
3166 ((__m512h)__builtin_ia32_vfmulcph512_mask( \
3167 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), \
3168 (__v16sf)(__m512h)_mm512_setzero_ph(), (__mmask16)(U), (int)(R)))
3170 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_fcmadd_pch(__m512h __A,
3171 __m512h __B,
3172 __m512h __C) {
3173 return (__m512h)__builtin_ia32_vfcmaddcph512_mask3(
3178 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3179 _mm512_mask_fcmadd_pch(__m512h __A, __mmask16 __U, __m512h __B, __m512h __C) {
3180 return (__m512h)__builtin_ia32_vfcmaddcph512_mask(
3185 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3186 _mm512_mask3_fcmadd_pch(__m512h __A, __m512h __B, __m512h __C, __mmask16 __U) {
3187 return (__m512h)__builtin_ia32_vfcmaddcph512_mask3(
3192 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3193 _mm512_maskz_fcmadd_pch(__mmask16 __U, __m512h __A, __m512h __B, __m512h __C) {
3194 return (__m512h)__builtin_ia32_vfcmaddcph512_maskz(
3200 ((__m512h)__builtin_ia32_vfcmaddcph512_mask3( \
3201 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(C), \
3205 ((__m512h)__builtin_ia32_vfcmaddcph512_mask( \
3206 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(C), \
3210 ((__m512h)__builtin_ia32_vfcmaddcph512_mask3( \
3211 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(C), \
3215 ((__m512h)__builtin_ia32_vfcmaddcph512_maskz( \
3216 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(C), \
3219 static __inline__ __m512h __DEFAULT_FN_ATTRS512 _mm512_fmadd_pch(__m512h __A,
3220 __m512h __B,
3221 __m512h __C) {
3222 return (__m512h)__builtin_ia32_vfmaddcph512_mask3((__v16sf)__A, (__v16sf)__B,
3227 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3228 _mm512_mask_fmadd_pch(__m512h __A, __mmask16 __U, __m512h __B, __m512h __C) {
3229 return (__m512h)__builtin_ia32_vfmaddcph512_mask((__v16sf)__A, (__v16sf)__B,
3234 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3235 _mm512_mask3_fmadd_pch(__m512h __A, __m512h __B, __m512h __C, __mmask16 __U) {
3236 return (__m512h)__builtin_ia32_vfmaddcph512_mask3(
3241 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3242 _mm512_maskz_fmadd_pch(__mmask16 __U, __m512h __A, __m512h __B, __m512h __C) {
3243 return (__m512h)__builtin_ia32_vfmaddcph512_maskz(
3249 ((__m512h)__builtin_ia32_vfmaddcph512_mask3( \
3250 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(C), \
3254 ((__m512h)__builtin_ia32_vfmaddcph512_mask( \
3255 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(C), \
3259 ((__m512h)__builtin_ia32_vfmaddcph512_mask3( \
3260 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(C), \
3264 ((__m512h)__builtin_ia32_vfmaddcph512_maskz( \
3265 (__v16sf)(__m512h)(A), (__v16sf)(__m512h)(B), (__v16sf)(__m512h)(C), \
3269 _mm512_reduce_add_ph(__m512h __W) {
3274 _mm512_reduce_mul_ph(__m512h __W) {
3279 _mm512_reduce_max_ph(__m512h __V) {
3284 _mm512_reduce_min_ph(__m512h __V) {
3288 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3289 _mm512_mask_blend_ph(__mmask32 __U, __m512h __A, __m512h __W) {
3290 return (__m512h)__builtin_ia32_selectph_512((__mmask32)__U, (__v32hf)__W,
3294 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3295 _mm512_permutex2var_ph(__m512h __A, __m512i __I, __m512h __B) {
3296 return (__m512h)__builtin_ia32_vpermi2varhi512((__v32hi)__A, (__v32hi)__I,
3300 static __inline__ __m512h __DEFAULT_FN_ATTRS512
3301 _mm512_permutexvar_ph(__m512i __A, __m512h __B) {
3302 return (__m512h)__builtin_ia32_permvarhi512((__v32hi)__B, (__v32hi)__A);