1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #pragma once
7 
8 #include <xnnpack/common.h>
9 
10 
11 #if defined(__SSE2__)
12 #include <emmintrin.h>
13 
14 // GCC any, Clang pre-8, Android NDK Clang pre-8.0.7, Apple Clang pre-11, and ICC pre-16
15 #if (defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER)) || \
16     (defined(__clang__) && !defined(__apple_build_version__) && (__clang_major__ < 8)) || \
17     (defined(__clang__) && defined(__ANDROID__) && (__clang_major__ == 8) && (__clang_minor__ == 0) && (__clang_patchlevel__ < 7)) || \
18     (defined(__clang__) && defined(__apple_build_version__) && (__apple_build_version__ < 11000000)) || \
19     (defined(__INTEL_COMPILER) && (__INTEL_COMPILER < 1600))
20 
21 static XNN_INTRINSIC
_mm_loadu_si32(const void * address)22 __m128i _mm_loadu_si32(const void* address) {
23   return _mm_cvtsi32_si128(*((const int*) address));
24 }
25 
26 static XNN_INTRINSIC
_mm_storeu_si32(const void * address,__m128i v)27 void _mm_storeu_si32(const void* address, __m128i v) {
28   *((int*) address) = _mm_cvtsi128_si32(v);
29 }
30 #endif  // GCC any, Clang pre-8, Android NDK Clang pre-8.0.7, Apple Clang pre-11, and ICC pre-16
31 #endif  // SSE2
32 
33 #ifdef __AVX512F__
34 #include <immintrin.h>
35 
36 // GCC pre-7, Clang pre-8, Android NDK Clang pre-8.0.7, Apple Clang pre-11, and ICC pre-18
37 #if (defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && (__GNUC__ < 7)) || \
38     (defined(__clang__) && !defined(__apple_build_version__) && (__clang_major__ < 8)) || \
39     (defined(__clang__) && defined(__ANDROID__) && (__clang_major__ == 8) && (__clang_minor__ == 0) && (__clang_patchlevel__ < 7)) || \
40     (defined(__clang__) && defined(__apple_build_version__) && (__apple_build_version__ < 11000000)) || \
41     (defined(__INTEL_COMPILER) && (__INTEL_COMPILER < 1800))
42 
43 static XNN_INTRINSIC
_cvtu32_mask16(unsigned int mask)44 __mmask16 _cvtu32_mask16(unsigned int mask) {
45   return (__mmask16) mask;
46 }
47 
48 static XNN_INTRINSIC
_cvtu64_mask64(unsigned long long mask)49 __mmask64 _cvtu64_mask64(unsigned long long mask) {
50   return (__mmask64) mask;
51 }
52 
53 static XNN_INTRINSIC
_kshiftli_mask64(__mmask64 a,unsigned int count)54 __mmask64 _kshiftli_mask64(__mmask64 a, unsigned int count) {
55   return (__mmask64) ((unsigned long long) a << count);
56 }
57 
58 static XNN_INTRINSIC
_kshiftri_mask64(__mmask64 a,unsigned int count)59 __mmask64 _kshiftri_mask64(__mmask64 a, unsigned int count) {
60   return (__mmask64) ((unsigned long long) a >> count);
61 }
62 
63 #endif  // GCC pre-7, Clang pre-8, Android NDK Clang pre-8.0.7, Apple Clang pre-11, and ICC pre-18
64 
65 // GCC pre-7, Clang pre-4, and ICC pre-18
66 #if (defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && (__GNUC__ < 7)) || \
67     (defined(__clang__) && (__clang_major__ < 4)) || \
68     (defined(__INTEL_COMPILER) && (__INTEL_COMPILER < 1800))
69 
70 static XNN_INTRINSIC
_mm512_reduce_add_ps(__m512 v)71 float _mm512_reduce_add_ps(__m512 v) {
72 #if __AVX512DQ__
73   const __m256 sum2 = _mm256_add_ps(_mm512_castps512_ps256(v), _mm512_extractf32x8_ps(v, 1));
74 #else
75   const __m256 sum2 = _mm256_add_ps(_mm512_castps512_ps256(v), _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(v), 1)));
76 #endif
77   const __m128 sum4 = _mm_add_ps(_mm256_castps256_ps128(sum2), _mm256_extractf128_ps(sum2, 1));
78   const __m128 sum8 = _mm_add_ps(sum4, _mm_movehl_ps(sum4, sum4));
79   const __m128 sum16 = _mm_add_ss(sum8, _mm_movehdup_ps(sum8));
80   return _mm_cvtss_f32(sum16);
81 }
82 
83 static XNN_INTRINSIC
_mm512_reduce_max_ps(__m512 v)84 float _mm512_reduce_max_ps(__m512 v) {
85 #if __AVX512DQ__
86   const __m256 sum2 = _mm256_max_ps(_mm512_castps512_ps256(v), _mm512_extractf32x8_ps(v, 1));
87 #else
88   const __m256 sum2 = _mm256_max_ps(_mm512_castps512_ps256(v), _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(v), 1)));
89 #endif
90   const __m128 sum4 = _mm_max_ps(_mm256_castps256_ps128(sum2), _mm256_extractf128_ps(sum2, 1));
91   const __m128 sum8 = _mm_max_ps(sum4, _mm_movehl_ps(sum4, sum4));
92   const __m128 sum16 = _mm_max_ss(sum8, _mm_movehdup_ps(sum8));
93   return _mm_cvtss_f32(sum16);
94 }
95 
96 #endif  // GCC pre-7, Clang pre-4, and ICC pre-18
97 
98 #if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && (__GNUC__ < 9)
99 static XNN_INTRINSIC
_mm512_set_epi8(char e63,char e62,char e61,char e60,char e59,char e58,char e57,char e56,char e55,char e54,char e53,char e52,char e51,char e50,char e49,char e48,char e47,char e46,char e45,char e44,char e43,char e42,char e41,char e40,char e39,char e38,char e37,char e36,char e35,char e34,char e33,char e32,char e31,char e30,char e29,char e28,char e27,char e26,char e25,char e24,char e23,char e22,char e21,char e20,char e19,char e18,char e17,char e16,char e15,char e14,char e13,char e12,char e11,char e10,char e09,char e08,char e07,char e06,char e05,char e04,char e03,char e02,char e01,char e00)100 __m512i _mm512_set_epi8(
101   char e63, char e62, char e61, char e60,
102   char e59, char e58, char e57, char e56,
103   char e55, char e54, char e53, char e52,
104   char e51, char e50, char e49, char e48,
105   char e47, char e46, char e45, char e44,
106   char e43, char e42, char e41, char e40,
107   char e39, char e38, char e37, char e36,
108   char e35, char e34, char e33, char e32,
109   char e31, char e30, char e29, char e28,
110   char e27, char e26, char e25, char e24,
111   char e23, char e22, char e21, char e20,
112   char e19, char e18, char e17, char e16,
113   char e15, char e14, char e13, char e12,
114   char e11, char e10, char e09, char e08,
115   char e07, char e06, char e05, char e04,
116   char e03, char e02, char e01, char e00)
117 {
118   return (__m512i) (__v64qi) {
119     e00, e01, e02, e03, e04, e05, e06, e07,
120     e08, e09, e10, e11, e12, e13, e14, e15,
121     e16, e17, e18, e19, e20, e21, e22, e23,
122     e24, e25, e26, e27, e28, e29, e30, e31,
123     e32, e33, e34, e35, e36, e37, e38, e39,
124     e40, e41, e42, e43, e44, e45, e46, e47,
125     e48, e49, e50, e51, e52, e53, e54, e55,
126     e56, e57, e58, e59, e60, e61, e62, e63
127   };
128 }
129 #endif  // GCC pre-9
130 
131 #endif  // __AVX512F__
132