1 | /* Copyright (C) 2016 Povilas Kanapickas <povilas@radix.lt> |
2 | |
3 | Distributed under the Boost Software License, Version 1.0. |
4 | (See accompanying file LICENSE_1_0.txt or copy at |
5 | http://www.boost.org/LICENSE_1_0.txt) |
6 | */ |
7 | |
8 | #ifndef LIBSIMDPP_SIMDPP_DETAIL_INSN_STORE_MASKED_H |
9 | #define LIBSIMDPP_SIMDPP_DETAIL_INSN_STORE_MASKED_H |
10 | |
11 | #ifndef LIBSIMDPP_SIMD_H |
12 | #error "This file must be included through simd.h" |
13 | #endif |
14 | |
15 | #include <simdpp/types.h> |
16 | #include <simdpp/core/load.h> |
17 | #include <simdpp/core/store.h> |
18 | #include <simdpp/detail/null/memory.h> |
19 | #include <simdpp/detail/align.h> |
20 | |
21 | namespace simdpp { |
22 | namespace SIMDPP_ARCH_NAMESPACE { |
23 | namespace detail { |
24 | namespace insn { |
25 | |
26 | static SIMDPP_INL |
27 | void i_store_masked(char* p, const uint32<4>& a, const mask_int32<4>& mask) |
28 | { |
29 | #if SIMDPP_USE_NULL |
30 | null::store_masked(p, a, mask); |
31 | #elif SIMDPP_USE_AVX512VL |
32 | _mm_mask_store_epi32(p, mask.native(), a.native()); |
33 | #elif SIMDPP_USE_AVX2 |
34 | _mm_maskstore_epi32(reinterpret_cast<int*>(p), mask.native(), a.native()); |
35 | #elif SIMDPP_USE_AVX |
36 | _mm_maskstore_ps(reinterpret_cast<float*>(p), mask.native(), |
37 | _mm_castsi128_ps(a.native())); |
38 | #elif SIMDPP_USE_SSE2 || SIMDPP_USE_NEON || SIMDPP_USE_ALTIVEC || SIMDPP_USE_MSA |
39 | uint32<4> b = load(p); |
40 | b = blend(a, b, mask); |
41 | store(p, b); |
42 | #endif |
43 | } |
44 | |
45 | #if SIMDPP_USE_AVX2 |
46 | static SIMDPP_INL |
47 | void i_store_masked(char* p, const uint32<8>& a, const mask_int32<8>& mask) |
48 | { |
49 | #if SIMDPP_USE_AVX512VL |
50 | _mm256_mask_store_epi32(p, mask.native(), a.native()); |
51 | #else |
52 | _mm256_maskstore_epi32(reinterpret_cast<int*>(p), mask.native(), a.native()); |
53 | #endif |
54 | } |
55 | #endif |
56 | |
57 | #if SIMDPP_USE_AVX512F |
58 | static SIMDPP_INL |
59 | void i_store_masked(char* p, const uint32<16>& a, const mask_int32<16>& mask) |
60 | { |
61 | _mm512_mask_store_epi32(reinterpret_cast<int*>(p), mask.native(), a.native()); |
62 | } |
63 | #endif |
64 | |
65 | // ----------------------------------------------------------------------------- |
66 | |
67 | static SIMDPP_INL |
68 | void i_store_masked(char* p, const uint64<2>& a, const mask_int64<2>& mask) |
69 | { |
70 | #if SIMDPP_USE_AVX512VL |
71 | #if __INTEL_COMPILER |
72 | _mm_mask_store_epi64(reinterpret_cast<__int64*>(p), mask.native(), |
73 | a.native()); |
74 | #else |
75 | _mm_mask_store_epi64(reinterpret_cast<long long*>(p), mask.native(), |
76 | a.native()); |
77 | #endif |
78 | #elif SIMDPP_USE_AVX2 |
79 | #if __INTEL_COMPILER |
80 | _mm_maskstore_epi64(reinterpret_cast<__int64*>(p), mask.native(), a.native()); |
81 | #else |
82 | _mm_maskstore_epi64(reinterpret_cast<long long*>(p), mask.native(), a.native()); |
83 | #endif |
84 | #elif SIMDPP_USE_AVX |
85 | _mm_maskstore_pd(reinterpret_cast<double*>(p), mask.native(), _mm_castsi128_pd(a.native())); |
86 | #elif SIMDPP_USE_SSE2 || SIMDPP_USE_NEON || SIMDPP_USE_VSX_207 || SIMDPP_USE_MSA |
87 | uint64<2> b = load(p); |
88 | b = blend(a, b, mask); |
89 | store(p, b); |
90 | #elif SIMDPP_USE_NULL || SIMDPP_USE_ALTIVEC |
91 | null::store_masked(p, a, mask); |
92 | #endif |
93 | } |
94 | |
95 | #if SIMDPP_USE_AVX2 |
96 | static SIMDPP_INL |
97 | void i_store_masked(char* p, const uint64<4>& a, const mask_int64<4>& mask) |
98 | { |
99 | #if SIMDPP_USE_AVX512VL |
100 | #if __INTEL_COMPILER |
101 | _mm256_mask_store_epi64(reinterpret_cast<__int64*>(p), mask.native(), |
102 | a.native()); |
103 | #else |
104 | _mm256_mask_store_epi64(reinterpret_cast<long long*>(p), mask.native(), |
105 | a.native()); |
106 | #endif |
107 | #else |
108 | #if __INTEL_COMPILER |
109 | _mm256_maskstore_epi64(reinterpret_cast<__int64*>(p), mask.native(), a.native()); |
110 | #else |
111 | _mm256_maskstore_epi64(reinterpret_cast<long long*>(p), mask.native(), a.native()); |
112 | #endif |
113 | #endif |
114 | } |
115 | #endif |
116 | |
117 | #if SIMDPP_USE_AVX512F |
118 | static SIMDPP_INL |
119 | void i_store_masked(char* p, const uint64<8>& a, const mask_int64<8>& mask) |
120 | { |
121 | #if __INTEL_COMPILER |
122 | _mm512_mask_store_epi64(reinterpret_cast<__int64*>(p), mask.native(), a.native()); |
123 | #else |
124 | _mm512_mask_store_epi64(reinterpret_cast<long long*>(p), mask.native(), a.native()); |
125 | #endif |
126 | } |
127 | #endif |
128 | |
129 | // ----------------------------------------------------------------------------- |
130 | |
131 | static SIMDPP_INL |
132 | void i_store_masked(char* p, const float32<4>& a, const mask_float32<4>& mask) |
133 | { |
134 | #if SIMDPP_USE_NULL |
135 | null::store_masked(p, a, mask); |
136 | #elif SIMDPP_USE_AVX512VL |
137 | _mm_mask_store_ps(reinterpret_cast<float*>(p), mask.native(), a.native()); |
138 | #elif SIMDPP_USE_AVX |
139 | _mm_maskstore_ps(reinterpret_cast<float*>(p), |
140 | _mm_castps_si128(mask.native()), a.native()); |
141 | #elif SIMDPP_USE_SSE2 || SIMDPP_USE_NEON || SIMDPP_USE_ALTIVEC || SIMDPP_USE_MSA |
142 | float32<4> b = load(p); |
143 | b = blend(a, b, mask); |
144 | store(p, b); |
145 | #endif |
146 | } |
147 | |
148 | #if SIMDPP_USE_AVX |
149 | static SIMDPP_INL |
150 | void i_store_masked(char* p, const float32<8>& a, const mask_float32<8>& mask) |
151 | { |
152 | #if SIMDPP_USE_AVX512VL |
153 | _mm256_mask_store_ps(reinterpret_cast<float*>(p), mask.native(), |
154 | a.native()); |
155 | #else |
156 | _mm256_maskstore_ps(reinterpret_cast<float*>(p), |
157 | _mm256_castps_si256(mask.native()), a.native()); |
158 | #endif |
159 | } |
160 | #endif |
161 | |
162 | #if SIMDPP_USE_AVX512F |
163 | static SIMDPP_INL |
164 | void i_store_masked(char* p, const float32<16>& a, const mask_float32<16>& mask) |
165 | { |
166 | _mm512_mask_store_ps(reinterpret_cast<float*>(p), mask.native(), a.native()); |
167 | } |
168 | #endif |
169 | |
170 | // ----------------------------------------------------------------------------- |
171 | |
172 | static SIMDPP_INL |
173 | void i_store_masked(char* p, const float64<2>& a, const mask_float64<2>& mask) |
174 | { |
175 | #if SIMDPP_USE_AVX512VL |
176 | _mm_mask_store_pd(reinterpret_cast<double*>(p), mask.native(), a.native()); |
177 | #elif SIMDPP_USE_AVX |
178 | _mm_maskstore_pd(reinterpret_cast<double*>(p), |
179 | _mm_castpd_si128(mask.native()), a.native()); |
180 | #elif SIMDPP_USE_SSE2 || SIMDPP_USE_NEON64 || SIMDPP_USE_VSX_206 || SIMDPP_USE_MSA |
181 | float64<2> b = load(p); |
182 | b = blend(a, b, mask); |
183 | store(p, b); |
184 | #elif SIMDPP_USE_NULL || SIMDPP_USE_NEON32 || SIMDPP_USE_ALTIVEC |
185 | null::store_masked(p, a, mask); |
186 | #endif |
187 | } |
188 | |
189 | #if SIMDPP_USE_AVX |
190 | static SIMDPP_INL |
191 | void i_store_masked(char* p, const float64<4>& a, const mask_float64<4>& mask) |
192 | { |
193 | #if SIMDPP_USE_AVX512VL |
194 | _mm256_mask_store_pd(reinterpret_cast<double*>(p), mask.native(), |
195 | a.native()); |
196 | #else |
197 | _mm256_maskstore_pd(reinterpret_cast<double*>(p), |
198 | _mm256_castpd_si256(mask.native()), a.native()); |
199 | #endif |
200 | } |
201 | #endif |
202 | |
203 | #if SIMDPP_USE_AVX512F |
204 | static SIMDPP_INL |
205 | void i_store_masked(char* p, const float64<8>& a, const mask_float64<8>& mask) |
206 | { |
207 | _mm512_mask_store_pd(reinterpret_cast<double*>(p), mask.native(), a.native()); |
208 | } |
209 | #endif |
210 | |
211 | // ----------------------------------------------------------------------------- |
212 | |
213 | template<class V, class M> |
214 | void i_store_masked(char* p, const V& a, const M& mask) |
215 | { |
216 | const unsigned veclen = V::base_vector_type::length_bytes; |
217 | |
218 | for (unsigned i = 0; i < a.vec_length; ++i) { |
219 | i_store_masked(p, a.vec(i), mask.vec(i)); |
220 | p += veclen; |
221 | } |
222 | } |
223 | |
224 | } // namespace insn |
225 | } // namespace detail |
226 | } // namespace SIMDPP_ARCH_NAMESPACE |
227 | } // namespace simdpp |
228 | |
229 | #endif |
230 | |
231 | |