lib/Headers/avxvnniintrin.h

32 /// \fn __m256i _mm256_dpbusd_epi32(__m256i __S, __m256i __A, __m256i __B)
33 /// \fn __m256i _mm256_dpbusds_epi32(__m256i __S, __m256i __A, __m256i __B)
34 /// \fn __m256i _mm256_dpwssd_epi32(__m256i __S, __m256i __A, __m256i __B)
35 /// \fn __m256i _mm256_dpwssds_epi32(__m256i __S, __m256i __A, __m256i __B)
36 /// \fn __m128i _mm_dpbusd_epi32(__m128i __S, __m128i __A, __m128i __B)
37 /// \fn __m128i _mm_dpbusds_epi32(__m128i __S, __m128i __A, __m128i __B)
38 /// \fn __m128i _mm_dpwssd_epi32(__m128i __S, __m128i __A, __m128i __B)
39 /// \fn __m128i _mm_dpwssds_epi32(__m128i __S, __m128i __A, __m128i __B)
47 /// corresponding signed 8-bit integers in \a __B, producing 4 intermediate signed
55 ///      tmp1.word := Signed(ZeroExtend16(__A.byte[4*j]) * SignExtend16(__B.byte[4*j]))
56 ///      tmp2.word := Signed(ZeroExtend16(__A.byte[4*j+1]) * SignExtend16(__B.byte[4*j+1]))
57 ///      tmp3.word := Signed(ZeroExtend16(__A.byte[4*j+2]) * SignExtend16(__B.byte[4*j+2]))
58 ///      tmp4.word := Signed(ZeroExtend16(__A.byte[4*j+3]) * SignExtend16(__B.byte[4*j+3]))
64 _mm256_dpbusd_avx_epi32(__m256i __S, __m256i __A, __m256i __B)
66   return (__m256i)__builtin_ia32_vpdpbusd256((__v8si)__S, (__v8si)__A, (__v8si)__B);
70 /// corresponding signed 8-bit integers in \a __B, producing 4 intermediate signed
78 ///      tmp1.word := Signed(ZeroExtend16(__A.byte[4*j]) * SignExtend16(__B.byte[4*j]))
79 ///      tmp2.word := Signed(ZeroExtend16(__A.byte[4*j+1]) * SignExtend16(__B.byte[4*j+1]))
80 ///      tmp3.word := Signed(ZeroExtend16(__A.byte[4*j+2]) * SignExtend16(__B.byte[4*j+2]))
81 ///      tmp4.word := Signed(ZeroExtend16(__A.byte[4*j+3]) * SignExtend16(__B.byte[4*j+3]))
87 _mm256_dpbusds_avx_epi32(__m256i __S, __m256i __A, __m256i __B)
89   return (__m256i)__builtin_ia32_vpdpbusds256((__v8si)__S, (__v8si)__A, (__v8si)__B);
93 /// corresponding 16-bit integers in \a __B, producing 2 intermediate signed 32-bit
101 ///      tmp1.dword := SignExtend32(__A.word[2*j]) * SignExtend32(__B.word[2*j])
102 ///      tmp2.dword := SignExtend32(__A.word[2*j+1]) * SignExtend32(__B.word[2*j+1])
108 _mm256_dpwssd_avx_epi32(__m256i __S, __m256i __A, __m256i __B)
110   return (__m256i)__builtin_ia32_vpdpwssd256((__v8si)__S, (__v8si)__A, (__v8si)__B);
114 /// corresponding 16-bit integers in \a __B, producing 2 intermediate signed 32-bit
122 ///      tmp1.dword := SignExtend32(__A.word[2*j]) * SignExtend32(__B.word[2*j])
123 ///      tmp2.dword := SignExtend32(__A.word[2*j+1]) * SignExtend32(__B.word[2*j+1])
129 _mm256_dpwssds_avx_epi32(__m256i __S, __m256i __A, __m256i __B)
131   return (__m256i)__builtin_ia32_vpdpwssds256((__v8si)__S, (__v8si)__A, (__v8si)__B);
135 /// corresponding signed 8-bit integers in \a __B, producing 4 intermediate signed
143 ///      tmp1.word := Signed(ZeroExtend16(__A.byte[4*j]) * SignExtend16(__B.byte[4*j]))
144 ///      tmp2.word := Signed(ZeroExtend16(__A.byte[4*j+1]) * SignExtend16(__B.byte[4*j+1]))
145 ///      tmp3.word := Signed(ZeroExtend16(__A.byte[4*j+2]) * SignExtend16(__B.byte[4*j+2]))
146 ///      tmp4.word := Signed(ZeroExtend16(__A.byte[4*j+3]) * SignExtend16(__B.byte[4*j+3]))
152 _mm_dpbusd_avx_epi32(__m128i __S, __m128i __A, __m128i __B)
154   return (__m128i)__builtin_ia32_vpdpbusd128((__v4si)__S, (__v4si)__A, (__v4si)__B);
158 /// corresponding signed 8-bit integers in \a __B, producing 4 intermediate signed
166 ///      tmp1.word := Signed(ZeroExtend16(__A.byte[4*j]) * SignExtend16(__B.byte[4*j]))
167 ///      tmp2.word := Signed(ZeroExtend16(__A.byte[4*j+1]) * SignExtend16(__B.byte[4*j+1]))
168 ///      tmp3.word := Signed(ZeroExtend16(__A.byte[4*j+2]) * SignExtend16(__B.byte[4*j+2]))
169 ///      tmp4.word := Signed(ZeroExtend16(__A.byte[4*j+3]) * SignExtend16(__B.byte[4*j+3]))
175 _mm_dpbusds_avx_epi32(__m128i __S, __m128i __A, __m128i __B)
177   return (__m128i)__builtin_ia32_vpdpbusds128((__v4si)__S, (__v4si)__A, (__v4si)__B);
181 /// corresponding 16-bit integers in \a __B, producing 2 intermediate signed 32-bit
189 ///      tmp1.dword := SignExtend32(__A.word[2*j]) * SignExtend32(__B.word[2*j])
190 ///      tmp2.dword := SignExtend32(__A.word[2*j+1]) * SignExtend32(__B.word[2*j+1])
196 _mm_dpwssd_avx_epi32(__m128i __S, __m128i __A, __m128i __B)
198   return (__m128i)__builtin_ia32_vpdpwssd128((__v4si)__S, (__v4si)__A, (__v4si)__B);
202 /// corresponding 16-bit integers in \a __B, producing 2 intermediate signed 32-bit
210 ///      tmp1.dword := SignExtend32(__A.word[2*j]) * SignExtend32(__B.word[2*j])
211 ///      tmp2.dword := SignExtend32(__A.word[2*j+1]) * SignExtend32(__B.word[2*j+1])
217 _mm_dpwssds_avx_epi32(__m128i __S, __m128i __A, __m128i __B)
219   return (__m128i)__builtin_ia32_vpdpwssds128((__v4si)__S, (__v4si)__A, (__v4si)__B);