lib/Headers/avx512fp16intrin.h

357 _mm512_mask_add_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
359       (__mmask32)__U, (__v32hf)_mm512_add_ph(__A, __B), (__v32hf)__W);
389 _mm512_mask_sub_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
391       (__mmask32)__U, (__v32hf)_mm512_sub_ph(__A, __B), (__v32hf)__W);
421 _mm512_mask_mul_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
423       (__mmask32)__U, (__v32hf)_mm512_mul_ph(__A, __B), (__v32hf)__W);
453 _mm512_mask_div_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
455       (__mmask32)__U, (__v32hf)_mm512_div_ph(__A, __B), (__v32hf)__W);
486 _mm512_mask_min_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
488       (__mmask32)__U, (__v32hf)_mm512_min_ph(__A, __B), (__v32hf)__W);
519 _mm512_mask_max_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
521       (__mmask32)__U, (__v32hf)_mm512_max_ph(__A, __B), (__v32hf)__W);
554 _mm512_mask_conj_pch(__m512h __W, __mmask16 __U, __m512h __A) {
556       (__mmask16)__U, (__v16sf)_mm512_conj_pch(__A), (__v16sf)__W);
572 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_add_sh(__m128h __W,
577   return __builtin_ia32_selectsh_128(__U, __A, __W);
608 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_sub_sh(__m128h __W,
613   return __builtin_ia32_selectsh_128(__U, __A, __W);
644 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_mul_sh(__m128h __W,
649   return __builtin_ia32_selectsh_128(__U, __A, __W);
680 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_div_sh(__m128h __W,
685   return __builtin_ia32_selectsh_128(__U, __A, __W);
717 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_min_sh(__m128h __W,
722                                                   (__v8hf)__W, (__mmask8)__U,
756 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_max_sh(__m128h __W,
761                                                   (__v8hf)__W, (__mmask8)__U,
833 _mm_mask_load_sh(__m128h __W, __mmask8 __U, const void *__A) {
835       (__v8hf)__W, (__v8hf)_mm_setzero_ph(), 0, 8, 8, 8, 8, 8, 8, 8);
892 static __inline__ void __DEFAULT_FN_ATTRS128 _mm_mask_store_sh(void *__W,
895   __builtin_ia32_storesh128_mask((__v8hf *)__W, __A, __U & 1);
944 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_move_sh(__m128h __W,
948   return __builtin_ia32_selectsh_128(__U, _mm_move_sh(__A, __B), __W);
974 _mm512_mask_rcp_ph(__m512h __W, __mmask32 __U, __m512h __A) {
975   return (__m512h)__builtin_ia32_rcpph512_mask((__v32hf)__A, (__v32hf)__W,
991 _mm512_mask_rsqrt_ph(__m512h __W, __mmask32 __U, __m512h __A) {
992   return (__m512h)__builtin_ia32_rsqrtph512_mask((__v32hf)__A, (__v32hf)__W,
1040 _mm512_mask_getexp_ph(__m512h __W, __mmask32 __U, __m512h __A) {
1042       (__v32hf)__A, (__v32hf)__W, (__mmask32)__U, _MM_FROUND_CUR_DIRECTION);
1074 _mm512_mask_scalef_ph(__m512h __W, __mmask32 __U, __m512h __A, __m512h __B) {
1076                                                   (__v32hf)__W, (__mmask32)__U,
1168 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_rcp_sh(__m128h __W,
1173                                             (__v8hf)__W, (__mmask8)__U);
1189 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_rsqrt_sh(__m128h __W,
1194                                               (__v8hf)__W, (__mmask8)__U);
1246 _mm_mask_getexp_sh(__m128h __W, __mmask8 __U, __m128h __A, __m128h __B) {
1248       (__v8hf)__A, (__v8hf)__B, (__v8hf)__W, (__mmask8)__U,
1282 _mm_mask_scalef_sh(__m128h __W, __mmask8 __U, __m128h __A, __m128h __B) {
1284                                                      (__v8hf)__W, (__mmask8)__U,
1384 _mm512_mask_sqrt_ph(__m512h __W, __mmask32 __U, __m512h __A) {
1388       (__v32hf)(__m512h)(__W));
1421 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_sqrt_sh(__m128h __W,
1426       (__v8hf)(__m128h)(__A), (__v8hf)(__m128h)(__B), (__v8hf)(__m128h)(__W),
1473 _mm512_mask_cvtpd_ph(__m128h __W, __mmask8 __U, __m512d __A) {
1475       (__v8df)__A, (__v8hf)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);
1504 _mm512_mask_cvtph_pd(__m512d __W, __mmask8 __U, __m128h __A) {
1506       (__v8hf)__A, (__v8df)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);
1537 static __inline__ __m128 __DEFAULT_FN_ATTRS128 _mm_mask_cvtsh_ss(__m128 __W,
1542                                                      (__v4sf)__W, (__mmask8)__U,
1575 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_cvtss_sh(__m128h __W,
1580       (__v8hf)__A, (__v4sf)__B, (__v8hf)__W, (__mmask8)__U,
1613 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_cvtsd_sh(__m128h __W,
1618       (__v8hf)__A, (__v2df)__B, (__v8hf)__W, (__mmask8)__U,
1650 static __inline__ __m128d __DEFAULT_FN_ATTRS128 _mm_mask_cvtsh_sd(__m128d __W,
1655       (__v2df)__A, (__v8hf)__B, (__v2df)__W, (__mmask8)__U,
1688 _mm512_mask_cvtph_epi16(__m512i __W, __mmask32 __U, __m512h __A) {
1690       (__v32hf)__A, (__v32hi)__W, (__mmask32)__U, _MM_FROUND_CUR_DIRECTION);
1722 _mm512_mask_cvttph_epi16(__m512i __W, __mmask32 __U, __m512h __A) {
1724       (__v32hf)__A, (__v32hi)__W, (__mmask32)__U, _MM_FROUND_CUR_DIRECTION);
1755 _mm512_mask_cvtepi16_ph(__m512h __W, __mmask32 __U, __m512i __A) {
1757       (__v32hi)__A, (__v32hf)__W, (__mmask32)__U, _MM_FROUND_CUR_DIRECTION);
1789 _mm512_mask_cvtph_epu16(__m512i __W, __mmask32 __U, __m512h __A) {
1791       (__v32hf)__A, (__v32hu)__W, (__mmask32)__U, _MM_FROUND_CUR_DIRECTION);
1823 _mm512_mask_cvttph_epu16(__m512i __W, __mmask32 __U, __m512h __A) {
1825       (__v32hf)__A, (__v32hu)__W, (__mmask32)__U, _MM_FROUND_CUR_DIRECTION);
1856 _mm512_mask_cvtepu16_ph(__m512h __W, __mmask32 __U, __m512i __A) {
1858       (__v32hu)__A, (__v32hf)__W, (__mmask32)__U, _MM_FROUND_CUR_DIRECTION);
1890 _mm512_mask_cvtph_epi32(__m512i __W, __mmask16 __U, __m256h __A) {
1892       (__v16hf)__A, (__v16si)__W, (__mmask16)__U, _MM_FROUND_CUR_DIRECTION);
1924 _mm512_mask_cvtph_epu32(__m512i __W, __mmask16 __U, __m256h __A) {
1926       (__v16hf)__A, (__v16su)__W, (__mmask16)__U, _MM_FROUND_CUR_DIRECTION);
1957 _mm512_mask_cvtepi32_ph(__m256h __W, __mmask16 __U, __m512i __A) {
1959       (__v16si)__A, (__v16hf)__W, (__mmask16)__U, _MM_FROUND_CUR_DIRECTION);
1990 _mm512_mask_cvtepu32_ph(__m256h __W, __mmask16 __U, __m512i __A) {
1992       (__v16su)__A, (__v16hf)__W, (__mmask16)__U, _MM_FROUND_CUR_DIRECTION);
2024 _mm512_mask_cvttph_epi32(__m512i __W, __mmask16 __U, __m256h __A) {
2026       (__v16hf)__A, (__v16si)__W, (__mmask16)__U, _MM_FROUND_CUR_DIRECTION);
2058 _mm512_mask_cvttph_epu32(__m512i __W, __mmask16 __U, __m256h __A) {
2060       (__v16hf)__A, (__v16su)__W, (__mmask16)__U, _MM_FROUND_CUR_DIRECTION);
2090 _mm512_mask_cvtepi64_ph(__m128h __W, __mmask8 __U, __m512i __A) {
2092       (__v8di)__A, (__v8hf)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);
2123 _mm512_mask_cvtph_epi64(__m512i __W, __mmask8 __U, __m128h __A) {
2125       (__v8hf)__A, (__v8di)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);
2155 _mm512_mask_cvtepu64_ph(__m128h __W, __mmask8 __U, __m512i __A) {
2157       (__v8du)__A, (__v8hf)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);
2188 _mm512_mask_cvtph_epu64(__m512i __W, __mmask8 __U, __m128h __A) {
2190       (__v8hf)__A, (__v8du)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);
2221 _mm512_mask_cvttph_epi64(__m512i __W, __mmask8 __U, __m128h __A) {
2223       (__v8hf)__A, (__v8di)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);
2254 _mm512_mask_cvttph_epu64(__m512i __W, __mmask8 __U, __m128h __A) {
2256       (__v8hf)__A, (__v8du)__W, (__mmask8)__U, _MM_FROUND_CUR_DIRECTION);
2400 _mm512_mask_cvtxph_ps(__m512 __W, __mmask16 __U, __m256h __A) {
2402       (__v16hf)__A, (__v16sf)__W, (__mmask16)__U, _MM_FROUND_CUR_DIRECTION);
2432 _mm512_mask_cvtxps_ph(__m256h __W, __mmask16 __U, __m512 __A) {
2434       (__v16sf)__A, (__v16hf)__W, (__mmask16)__U, _MM_FROUND_CUR_DIRECTION);
2736 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_fmadd_sh(__m128h __W,
2739   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, (__v8hf)__A, (__v8hf)__B,
2743 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_fmadd_sh(__m128h __W,
2747   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, (__v8hf)__A, (__v8hf)__B,
2774 _mm_mask3_fmadd_sh(__m128h __W, __m128h __X, __m128h __Y, __mmask8 __U) {
2775   return __builtin_ia32_vfmaddsh3_mask3((__v8hf)__W, (__v8hf)__X, (__v8hf)__Y,
2785 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_fmsub_sh(__m128h __W,
2788   return (__m128h)__builtin_ia32_vfmaddsh3_mask((__v8hf)__W, (__v8hf)__A,
2793 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_fmsub_sh(__m128h __W,
2797   return (__m128h)__builtin_ia32_vfmaddsh3_mask((__v8hf)__W, (__v8hf)__A,
2825 _mm_mask3_fmsub_sh(__m128h __W, __m128h __X, __m128h __Y, __mmask8 __U) {
2826   return __builtin_ia32_vfmsubsh3_mask3((__v8hf)__W, (__v8hf)__X, (__v8hf)__Y,
2836 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_fnmadd_sh(__m128h __W,
2839   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, -(__v8hf)__A, (__v8hf)__B,
2844 _mm_mask_fnmadd_sh(__m128h __W, __mmask8 __U, __m128h __A, __m128h __B) {
2845   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, -(__v8hf)__A, (__v8hf)__B,
2872 _mm_mask3_fnmadd_sh(__m128h __W, __m128h __X, __m128h __Y, __mmask8 __U) {
2873   return __builtin_ia32_vfmaddsh3_mask3((__v8hf)__W, -(__v8hf)__X, (__v8hf)__Y,
2883 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_fnmsub_sh(__m128h __W,
2886   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, -(__v8hf)__A, -(__v8hf)__B,
2891 _mm_mask_fnmsub_sh(__m128h __W, __mmask8 __U, __m128h __A, __m128h __B) {
2892   return __builtin_ia32_vfmaddsh3_mask((__v8hf)__W, -(__v8hf)__A, -(__v8hf)__B,
2919 _mm_mask3_fnmsub_sh(__m128h __W, __m128h __X, __m128h __Y, __mmask8 __U) {
2920   return __builtin_ia32_vfmsubsh3_mask3((__v8hf)__W, -(__v8hf)__X, (__v8hf)__Y,
3032 _mm_mask_fcmul_sch(__m128h __W, __mmask8 __U, __m128h __A, __m128h __B) {
3034                                                 (__v4sf)__W, (__mmask8)__U,
3067 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_fmul_sch(__m128h __W,
3072                                                (__v4sf)__W, (__mmask8)__U,
3106 _mm512_mask_fcmul_pch(__m512h __W, __mmask16 __U, __m512h __A, __m512h __B) {
3108                                                    (__v16sf)__W, (__mmask16)__U,
3142 _mm512_mask_fmul_pch(__m512h __W, __mmask16 __U, __m512h __A, __m512h __B) {
3144                                                   (__v16sf)__W, (__mmask16)__U,
3269 _mm512_reduce_add_ph(__m512h __W) {
3270   return __builtin_ia32_reduce_fadd_ph512(-0.0f16, __W);
3274 _mm512_reduce_mul_ph(__m512h __W) {
3275   return __builtin_ia32_reduce_fmul_ph512(1.0f16, __W);
3289 _mm512_mask_blend_ph(__mmask32 __U, __m512h __A, __m512h __W) {
3290   return (__m512h)__builtin_ia32_selectph_512((__mmask32)__U, (__v32hf)__W,