diff options
author | Dimitry Andric <dim@FreeBSD.org> | 2018-07-28 11:06:01 +0000 |
---|---|---|
committer | Dimitry Andric <dim@FreeBSD.org> | 2018-07-28 11:06:01 +0000 |
commit | 486754660bb926339aefcf012a3f848592babb8b (patch) | |
tree | ecdbc446c9876f4f120f701c243373cd3cb43db3 /lib/Headers/avx512vlvnniintrin.h | |
parent | 55e6d896ad333f07bb3b1ba487df214fc268a4ab (diff) | |
download | src-e840ccaca250d542d229ef559f5dbc4689265313.tar.gz src-e840ccaca250d542d229ef559f5dbc4689265313.zip |
Vendor import of clang trunk r338150:vendor/clang/clang-trunk-r338150
Diffstat (limited to 'lib/Headers/avx512vlvnniintrin.h')
-rw-r--r-- | lib/Headers/avx512vlvnniintrin.h | 253 |
1 files changed, 111 insertions, 142 deletions
diff --git a/lib/Headers/avx512vlvnniintrin.h b/lib/Headers/avx512vlvnniintrin.h index 745ae8b7ad3d..62382268ec9f 100644 --- a/lib/Headers/avx512vlvnniintrin.h +++ b/lib/Headers/avx512vlvnniintrin.h @@ -29,226 +29,195 @@ #define __AVX512VLVNNIINTRIN_H /* Define the default attributes for the functions in this file. */ -#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512vnni"))) +#define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512vnni"), __min_vector_width__(128))) +#define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512vnni"), __min_vector_width__(256))) -static __inline__ __m256i __DEFAULT_FN_ATTRS +static __inline__ __m256i __DEFAULT_FN_ATTRS256 +_mm256_dpbusd_epi32(__m256i __S, __m256i __A, __m256i __B) +{ + return (__m256i)__builtin_ia32_vpdpbusd256((__v8si)__S, (__v8si)__A, + (__v8si)__B); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_dpbusd_epi32(__m256i __S, __mmask8 __U, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpbusd256_mask ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectd_256(__U, + (__v8si)_mm256_dpbusd_epi32(__S, __A, __B), + (__v8si)__S); } -static __inline__ __m256i __DEFAULT_FN_ATTRS +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_dpbusd_epi32(__mmask8 __U, __m256i __S, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpbusd256_maskz ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectd_256(__U, + (__v8si)_mm256_dpbusd_epi32(__S, __A, __B), + (__v8si)_mm256_setzero_si256()); } -static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_dpbusd_epi32(__m256i __S, __m256i __A, __m256i __B) +static __inline__ __m256i __DEFAULT_FN_ATTRS256 +_mm256_dpbusds_epi32(__m256i __S, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpbusd256_mask ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) -1); + return (__m256i)__builtin_ia32_vpdpbusds256((__v8si)__S, (__v8si)__A, + (__v8si)__B); } -static __inline__ __m256i __DEFAULT_FN_ATTRS +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_dpbusds_epi32(__m256i __S, __mmask8 __U, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpbusds256_mask ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectd_256(__U, + (__v8si)_mm256_dpbusds_epi32(__S, __A, __B), + (__v8si)__S); } -static __inline__ __m256i __DEFAULT_FN_ATTRS +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_dpbusds_epi32(__mmask8 __U, __m256i __S, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpbusds256_maskz ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectd_256(__U, + (__v8si)_mm256_dpbusds_epi32(__S, __A, __B), + (__v8si)_mm256_setzero_si256()); } -static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_dpbusds_epi32(__m256i __S, __m256i __A, __m256i __B) +static __inline__ __m256i __DEFAULT_FN_ATTRS256 +_mm256_dpwssd_epi32(__m256i __S, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpbusds256_mask ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) -1); + return (__m256i)__builtin_ia32_vpdpwssd256((__v8si)__S, (__v8si)__A, + (__v8si)__B); } -static __inline__ __m256i __DEFAULT_FN_ATTRS +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_dpwssd_epi32(__m256i __S, __mmask8 __U, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpwssd256_mask ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectd_256(__U, + (__v8si)_mm256_dpwssd_epi32(__S, __A, __B), + (__v8si)__S); } -static __inline__ __m256i __DEFAULT_FN_ATTRS +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_dpwssd_epi32(__mmask8 __U, __m256i __S, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpwssd256_maskz ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectd_256(__U, + (__v8si)_mm256_dpwssd_epi32(__S, __A, __B), + (__v8si)_mm256_setzero_si256()); } -static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_dpwssd_epi32(__m256i __S, __m256i __A, __m256i __B) +static __inline__ __m256i __DEFAULT_FN_ATTRS256 +_mm256_dpwssds_epi32(__m256i __S, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpwssd256_mask ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) -1); + return (__m256i)__builtin_ia32_vpdpwssds256((__v8si)__S, (__v8si)__A, + (__v8si)__B); } -static __inline__ __m256i __DEFAULT_FN_ATTRS +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_dpwssds_epi32(__m256i __S, __mmask8 __U, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpwssds256_mask ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectd_256(__U, + (__v8si)_mm256_dpwssds_epi32(__S, __A, __B), + (__v8si)__S); } -static __inline__ __m256i __DEFAULT_FN_ATTRS +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_dpwssds_epi32(__mmask8 __U, __m256i __S, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_vpdpwssds256_maskz ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) __U); + return (__m256i)__builtin_ia32_selectd_256(__U, + (__v8si)_mm256_dpwssds_epi32(__S, __A, __B), + (__v8si)_mm256_setzero_si256()); } -static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_dpwssds_epi32(__m256i __S, __m256i __A, __m256i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_dpbusd_epi32(__m128i __S, __m128i __A, __m128i __B) { - return (__m256i) __builtin_ia32_vpdpwssds256_mask ((__v8si) __S, - (__v8si) __A, - (__v8si) __B, - (__mmask8) -1); + return (__m128i)__builtin_ia32_vpdpbusd128((__v4si)__S, (__v4si)__A, + (__v4si)__B); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_mask_dpbusd_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_mask_dpbusd_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpbusd128_mask ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectd_128(__U, + (__v4si)_mm_dpbusd_epi32(__S, __A, __B), + (__v4si)__S); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_maskz_dpbusd_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_maskz_dpbusd_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpbusd128_maskz ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectd_128(__U, + (__v4si)_mm_dpbusd_epi32(__S, __A, __B), + (__v4si)_mm_setzero_si128()); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_dpbusd_epi32(__m128i __S, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_dpbusds_epi32(__m128i __S, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpbusd128_mask ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) -1); + return (__m128i)__builtin_ia32_vpdpbusds128((__v4si)__S, (__v4si)__A, + (__v4si)__B); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_mask_dpbusds_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_mask_dpbusds_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpbusds128_mask ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectd_128(__U, + (__v4si)_mm_dpbusds_epi32(__S, __A, __B), + (__v4si)__S); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_maskz_dpbusds_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_maskz_dpbusds_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpbusds128_maskz ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectd_128(__U, + (__v4si)_mm_dpbusds_epi32(__S, __A, __B), + (__v4si)_mm_setzero_si128()); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_dpbusds_epi32(__m128i __S, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_dpwssd_epi32(__m128i __S, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpbusds128_mask ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) -1); + return (__m128i)__builtin_ia32_vpdpwssd128((__v4si)__S, (__v4si)__A, + (__v4si)__B); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_mask_dpwssd_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_mask_dpwssd_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpwssd128_mask ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectd_128(__U, + (__v4si)_mm_dpwssd_epi32(__S, __A, __B), + (__v4si)__S); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_maskz_dpwssd_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_maskz_dpwssd_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpwssd128_maskz ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectd_128(__U, + (__v4si)_mm_dpwssd_epi32(__S, __A, __B), + (__v4si)_mm_setzero_si128()); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_dpwssd_epi32(__m128i __S, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_dpwssds_epi32(__m128i __S, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpwssd128_mask ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) -1); + return (__m128i)__builtin_ia32_vpdpwssds128((__v4si)__S, (__v4si)__A, + (__v4si)__B); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_mask_dpwssds_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_mask_dpwssds_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpwssds128_mask ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectd_128(__U, + (__v4si)_mm_dpwssds_epi32(__S, __A, __B), + (__v4si)__S); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_maskz_dpwssds_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_maskz_dpwssds_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_vpdpwssds128_maskz ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) __U); + return (__m128i)__builtin_ia32_selectd_128(__U, + (__v4si)_mm_dpwssds_epi32(__S, __A, __B), + (__v4si)_mm_setzero_si128()); } -static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm128_dpwssds_epi32(__m128i __S, __m128i __A, __m128i __B) -{ - return (__m128i) __builtin_ia32_vpdpwssds128_mask ((__v4si) __S, - (__v4si) __A, - (__v4si) __B, - (__mmask8) -1); -} - - -#undef __DEFAULT_FN_ATTRS +#undef __DEFAULT_FN_ATTRS128 +#undef __DEFAULT_FN_ATTRS256 #endif |