11 #error "Never use <avx512vlvnniintrin.h> directly; include <immintrin.h> instead."
14 #ifndef __AVX512VLVNNIINTRIN_H
15 #define __AVX512VLVNNIINTRIN_H
18 #define __DEFAULT_FN_ATTRS128 \
19 __attribute__((__always_inline__, __nodebug__, \
20 __target__("avx512vl,avx512vnni,no-evex512"), \
21 __min_vector_width__(128)))
22 #define __DEFAULT_FN_ATTRS256 \
23 __attribute__((__always_inline__, __nodebug__, \
24 __target__("avx512vl,avx512vnni,no-evex512"), \
25 __min_vector_width__(256)))
44 #define _mm256_dpbusd_epi32(S, A, B) \
45 ((__m256i)__builtin_ia32_vpdpbusd256((__v8si)(S), (__v8si)(A), (__v8si)(B)))
64 #define _mm256_dpbusds_epi32(S, A, B) \
65 ((__m256i)__builtin_ia32_vpdpbusds256((__v8si)(S), (__v8si)(A), (__v8si)(B)))
82 #define _mm256_dpwssd_epi32(S, A, B) \
83 ((__m256i)__builtin_ia32_vpdpwssd256((__v8si)(S), (__v8si)(A), (__v8si)(B)))
100 #define _mm256_dpwssds_epi32(S, A, B) \
101 ((__m256i)__builtin_ia32_vpdpwssds256((__v8si)(S), (__v8si)(A), (__v8si)(B)))
120 #define _mm_dpbusd_epi32(S, A, B) \
121 ((__m128i)__builtin_ia32_vpdpbusd128((__v4si)(S), (__v4si)(A), (__v4si)(B)))
140 #define _mm_dpbusds_epi32(S, A, B) \
141 ((__m128i)__builtin_ia32_vpdpbusds128((__v4si)(S), (__v4si)(A), (__v4si)(B)))
158 #define _mm_dpwssd_epi32(S, A, B) \
159 ((__m128i)__builtin_ia32_vpdpwssd128((__v4si)(S), (__v4si)(A), (__v4si)(B)))
176 #define _mm_dpwssds_epi32(S, A, B) \
177 ((__m128i)__builtin_ia32_vpdpwssds128((__v4si)(S), (__v4si)(A), (__v4si)(B)))
182 return (__m256i)__builtin_ia32_selectd_256(__U,
190 return (__m256i)__builtin_ia32_selectd_256(__U,
198 return (__m256i)__builtin_ia32_selectd_256(__U,
206 return (__m256i)__builtin_ia32_selectd_256(__U,
214 return (__m256i)__builtin_ia32_selectd_256(__U,
222 return (__m256i)__builtin_ia32_selectd_256(__U,
230 return (__m256i)__builtin_ia32_selectd_256(__U,
238 return (__m256i)__builtin_ia32_selectd_256(__U,
246 return (__m128i)__builtin_ia32_selectd_128(__U,
254 return (__m128i)__builtin_ia32_selectd_128(__U,
262 return (__m128i)__builtin_ia32_selectd_128(__U,
270 return (__m128i)__builtin_ia32_selectd_128(__U,
278 return (__m128i)__builtin_ia32_selectd_128(__U,
286 return (__m128i)__builtin_ia32_selectd_128(__U,
294 return (__m128i)__builtin_ia32_selectd_128(__U,
302 return (__m128i)__builtin_ia32_selectd_128(__U,
307 #undef __DEFAULT_FN_ATTRS128
308 #undef __DEFAULT_FN_ATTRS256
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_dpwssds_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B)
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_dpwssds_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_dpbusds_epi32(__m256i __S, __mmask8 __U, __m256i __A, __m256i __B)
#define _mm_dpbusd_epi32(S, A, B)
Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in A with corresponding signed 8-bit i...
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_dpbusds_epi32(__m128i __S, __mmask8 __U, __m128i __A, __m128i __B)
#define _mm256_dpbusds_epi32(S, A, B)
Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in A with corresponding signed 8-bit i...
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_dpwssd_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_dpbusd_epi32(__mmask8 __U, __m256i __S, __m256i __A, __m256i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_dpwssd_epi32(__m256i __S, __mmask8 __U, __m256i __A, __m256i __B)
#define _mm_dpwssds_epi32(S, A, B)
Multiply groups of 2 adjacent pairs of signed 16-bit integers in A with corresponding 16-bit integers...
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_dpbusd_epi32(__m256i __S, __mmask8 __U, __m256i __A, __m256i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_dpbusds_epi32(__mmask8 __U, __m256i __S, __m256i __A, __m256i __B)
#define __DEFAULT_FN_ATTRS256
#define _mm256_dpwssds_epi32(S, A, B)
Multiply groups of 2 adjacent pairs of signed 16-bit integers in A with corresponding 16-bit integers...
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_dpbusds_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B)
#define __DEFAULT_FN_ATTRS128
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_dpwssds_epi32(__m256i __S, __mmask8 __U, __m256i __A, __m256i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_dpwssds_epi32(__mmask8 __U, __m256i __S, __m256i __A, __m256i __B)
#define _mm256_dpwssd_epi32(S, A, B)
Multiply groups of 2 adjacent pairs of signed 16-bit integers in A with corresponding 16-bit integers...
#define _mm_dpbusds_epi32(S, A, B)
Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in A with corresponding signed 8-bit i...
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_dpwssd_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)
#define _mm256_dpbusd_epi32(S, A, B)
Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in A with corresponding signed 8-bit i...
static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_dpwssd_epi32(__mmask8 __U, __m256i __S, __m256i __A, __m256i __B)
#define _mm_dpwssd_epi32(S, A, B)
Multiply groups of 2 adjacent pairs of signed 16-bit integers in A with corresponding 16-bit integers...
static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_dpbusd_epi32(__mmask8 __U, __m128i __S, __m128i __A, __m128i __B)
static __inline __m256i __DEFAULT_FN_ATTRS _mm256_setzero_si256(void)
Constructs a 256-bit integer vector initialized to zero.
static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_setzero_si128(void)
Creates a 128-bit integer vector initialized to zero.