1*bed243d3SAndroid Build Coastguard Worker /*===------------- avx512vnniintrin.h - VNNI intrinsics ------------------===
2*bed243d3SAndroid Build Coastguard Worker *
3*bed243d3SAndroid Build Coastguard Worker *
4*bed243d3SAndroid Build Coastguard Worker * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
5*bed243d3SAndroid Build Coastguard Worker * See https://llvm.org/LICENSE.txt for license information.
6*bed243d3SAndroid Build Coastguard Worker * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7*bed243d3SAndroid Build Coastguard Worker *
8*bed243d3SAndroid Build Coastguard Worker *===-----------------------------------------------------------------------===
9*bed243d3SAndroid Build Coastguard Worker */
10*bed243d3SAndroid Build Coastguard Worker #ifndef __IMMINTRIN_H
11*bed243d3SAndroid Build Coastguard Worker #error "Never use <avx512vnniintrin.h> directly; include <immintrin.h> instead."
12*bed243d3SAndroid Build Coastguard Worker #endif
13*bed243d3SAndroid Build Coastguard Worker
14*bed243d3SAndroid Build Coastguard Worker #ifndef __AVX512VNNIINTRIN_H
15*bed243d3SAndroid Build Coastguard Worker #define __AVX512VNNIINTRIN_H
16*bed243d3SAndroid Build Coastguard Worker
17*bed243d3SAndroid Build Coastguard Worker /* Define the default attributes for the functions in this file. */
18*bed243d3SAndroid Build Coastguard Worker #define __DEFAULT_FN_ATTRS \
19*bed243d3SAndroid Build Coastguard Worker __attribute__((__always_inline__, __nodebug__, \
20*bed243d3SAndroid Build Coastguard Worker __target__("avx512vnni,evex512"), __min_vector_width__(512)))
21*bed243d3SAndroid Build Coastguard Worker
22*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_dpbusd_epi32(__m512i __S,__m512i __A,__m512i __B)23*bed243d3SAndroid Build Coastguard Worker _mm512_dpbusd_epi32(__m512i __S, __m512i __A, __m512i __B)
24*bed243d3SAndroid Build Coastguard Worker {
25*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_vpdpbusd512((__v16si)__S, (__v16si)__A,
26*bed243d3SAndroid Build Coastguard Worker (__v16si)__B);
27*bed243d3SAndroid Build Coastguard Worker }
28*bed243d3SAndroid Build Coastguard Worker
29*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_mask_dpbusd_epi32(__m512i __S,__mmask16 __U,__m512i __A,__m512i __B)30*bed243d3SAndroid Build Coastguard Worker _mm512_mask_dpbusd_epi32(__m512i __S, __mmask16 __U, __m512i __A, __m512i __B)
31*bed243d3SAndroid Build Coastguard Worker {
32*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_selectd_512(__U,
33*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_dpbusd_epi32(__S, __A, __B),
34*bed243d3SAndroid Build Coastguard Worker (__v16si)__S);
35*bed243d3SAndroid Build Coastguard Worker }
36*bed243d3SAndroid Build Coastguard Worker
37*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_maskz_dpbusd_epi32(__mmask16 __U,__m512i __S,__m512i __A,__m512i __B)38*bed243d3SAndroid Build Coastguard Worker _mm512_maskz_dpbusd_epi32(__mmask16 __U, __m512i __S, __m512i __A, __m512i __B)
39*bed243d3SAndroid Build Coastguard Worker {
40*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_selectd_512(__U,
41*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_dpbusd_epi32(__S, __A, __B),
42*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_setzero_si512());
43*bed243d3SAndroid Build Coastguard Worker }
44*bed243d3SAndroid Build Coastguard Worker
45*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_dpbusds_epi32(__m512i __S,__m512i __A,__m512i __B)46*bed243d3SAndroid Build Coastguard Worker _mm512_dpbusds_epi32(__m512i __S, __m512i __A, __m512i __B)
47*bed243d3SAndroid Build Coastguard Worker {
48*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_vpdpbusds512((__v16si)__S, (__v16si)__A,
49*bed243d3SAndroid Build Coastguard Worker (__v16si)__B);
50*bed243d3SAndroid Build Coastguard Worker }
51*bed243d3SAndroid Build Coastguard Worker
52*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_mask_dpbusds_epi32(__m512i __S,__mmask16 __U,__m512i __A,__m512i __B)53*bed243d3SAndroid Build Coastguard Worker _mm512_mask_dpbusds_epi32(__m512i __S, __mmask16 __U, __m512i __A, __m512i __B)
54*bed243d3SAndroid Build Coastguard Worker {
55*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_selectd_512(__U,
56*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_dpbusds_epi32(__S, __A, __B),
57*bed243d3SAndroid Build Coastguard Worker (__v16si)__S);
58*bed243d3SAndroid Build Coastguard Worker }
59*bed243d3SAndroid Build Coastguard Worker
60*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_maskz_dpbusds_epi32(__mmask16 __U,__m512i __S,__m512i __A,__m512i __B)61*bed243d3SAndroid Build Coastguard Worker _mm512_maskz_dpbusds_epi32(__mmask16 __U, __m512i __S, __m512i __A, __m512i __B)
62*bed243d3SAndroid Build Coastguard Worker {
63*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_selectd_512(__U,
64*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_dpbusds_epi32(__S, __A, __B),
65*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_setzero_si512());
66*bed243d3SAndroid Build Coastguard Worker }
67*bed243d3SAndroid Build Coastguard Worker
68*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_dpwssd_epi32(__m512i __S,__m512i __A,__m512i __B)69*bed243d3SAndroid Build Coastguard Worker _mm512_dpwssd_epi32(__m512i __S, __m512i __A, __m512i __B)
70*bed243d3SAndroid Build Coastguard Worker {
71*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_vpdpwssd512((__v16si)__S, (__v16si)__A,
72*bed243d3SAndroid Build Coastguard Worker (__v16si)__B);
73*bed243d3SAndroid Build Coastguard Worker }
74*bed243d3SAndroid Build Coastguard Worker
75*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_mask_dpwssd_epi32(__m512i __S,__mmask16 __U,__m512i __A,__m512i __B)76*bed243d3SAndroid Build Coastguard Worker _mm512_mask_dpwssd_epi32(__m512i __S, __mmask16 __U, __m512i __A, __m512i __B)
77*bed243d3SAndroid Build Coastguard Worker {
78*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_selectd_512(__U,
79*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_dpwssd_epi32(__S, __A, __B),
80*bed243d3SAndroid Build Coastguard Worker (__v16si)__S);
81*bed243d3SAndroid Build Coastguard Worker }
82*bed243d3SAndroid Build Coastguard Worker
83*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_maskz_dpwssd_epi32(__mmask16 __U,__m512i __S,__m512i __A,__m512i __B)84*bed243d3SAndroid Build Coastguard Worker _mm512_maskz_dpwssd_epi32(__mmask16 __U, __m512i __S, __m512i __A, __m512i __B)
85*bed243d3SAndroid Build Coastguard Worker {
86*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_selectd_512(__U,
87*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_dpwssd_epi32(__S, __A, __B),
88*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_setzero_si512());
89*bed243d3SAndroid Build Coastguard Worker }
90*bed243d3SAndroid Build Coastguard Worker
91*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_dpwssds_epi32(__m512i __S,__m512i __A,__m512i __B)92*bed243d3SAndroid Build Coastguard Worker _mm512_dpwssds_epi32(__m512i __S, __m512i __A, __m512i __B)
93*bed243d3SAndroid Build Coastguard Worker {
94*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_vpdpwssds512((__v16si)__S, (__v16si)__A,
95*bed243d3SAndroid Build Coastguard Worker (__v16si)__B);
96*bed243d3SAndroid Build Coastguard Worker }
97*bed243d3SAndroid Build Coastguard Worker
98*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_mask_dpwssds_epi32(__m512i __S,__mmask16 __U,__m512i __A,__m512i __B)99*bed243d3SAndroid Build Coastguard Worker _mm512_mask_dpwssds_epi32(__m512i __S, __mmask16 __U, __m512i __A, __m512i __B)
100*bed243d3SAndroid Build Coastguard Worker {
101*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_selectd_512(__U,
102*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_dpwssds_epi32(__S, __A, __B),
103*bed243d3SAndroid Build Coastguard Worker (__v16si)__S);
104*bed243d3SAndroid Build Coastguard Worker }
105*bed243d3SAndroid Build Coastguard Worker
106*bed243d3SAndroid Build Coastguard Worker static __inline__ __m512i __DEFAULT_FN_ATTRS
_mm512_maskz_dpwssds_epi32(__mmask16 __U,__m512i __S,__m512i __A,__m512i __B)107*bed243d3SAndroid Build Coastguard Worker _mm512_maskz_dpwssds_epi32(__mmask16 __U, __m512i __S, __m512i __A, __m512i __B)
108*bed243d3SAndroid Build Coastguard Worker {
109*bed243d3SAndroid Build Coastguard Worker return (__m512i)__builtin_ia32_selectd_512(__U,
110*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_dpwssds_epi32(__S, __A, __B),
111*bed243d3SAndroid Build Coastguard Worker (__v16si)_mm512_setzero_si512());
112*bed243d3SAndroid Build Coastguard Worker }
113*bed243d3SAndroid Build Coastguard Worker
114*bed243d3SAndroid Build Coastguard Worker #undef __DEFAULT_FN_ATTRS
115*bed243d3SAndroid Build Coastguard Worker
116*bed243d3SAndroid Build Coastguard Worker #endif
117