1 | /* Copyright (C) 2011-2014 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_I_AVG_H |
9 | #define LIBSIMDPP_SIMDPP_DETAIL_INSN_I_AVG_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/bit_xor.h> |
17 | #include <simdpp/core/bit_and.h> |
18 | #include <simdpp/core/i_add.h> |
19 | #include <simdpp/core/i_shift_r.h> |
20 | #include <simdpp/detail/vector_array_macros.h> |
21 | |
22 | namespace simdpp { |
23 | namespace SIMDPP_ARCH_NAMESPACE { |
24 | namespace detail { |
25 | namespace insn { |
26 | |
27 | template<class V> SIMDPP_INL |
28 | V v_emul_avg_u32(const V& a, const V& b); |
29 | template<class V> SIMDPP_INL |
30 | V v_emul_avg_i32(const V& a, const V& b); |
31 | |
32 | static SIMDPP_INL |
33 | uint8x16 i_avg(const uint8x16& a, const uint8x16& b) |
34 | { |
35 | #if SIMDPP_USE_NULL |
36 | uint8x16 r; |
37 | for (unsigned i = 0; i < a.length; i++) { |
38 | r.el(i) = (uint16_t(a.el(i)) + b.el(i) + 1) >> 1; |
39 | } |
40 | return r; |
41 | #elif SIMDPP_USE_SSE2 |
42 | return _mm_avg_epu8(a.native(), b.native()); |
43 | #elif SIMDPP_USE_NEON |
44 | return vrhaddq_u8(a.native(), b.native()); |
45 | #elif SIMDPP_USE_ALTIVEC |
46 | return vec_avg(a.native(), b.native()); |
47 | #elif SIMDPP_USE_MSA |
48 | return __msa_aver_u_b(a.native(), b.native()); |
49 | #endif |
50 | } |
51 | |
52 | #if SIMDPP_USE_AVX2 |
53 | static SIMDPP_INL |
54 | uint8x32 i_avg(const uint8x32& a, const uint8x32& b) |
55 | { |
56 | return _mm256_avg_epu8(a.native(), b.native()); |
57 | } |
58 | #endif |
59 | |
60 | #if SIMDPP_USE_AVX512BW |
61 | SIMDPP_INL uint8<64> i_avg(const uint8<64>& a, const uint8<64>& b) |
62 | { |
63 | return _mm512_avg_epu8(a.native(), b.native()); |
64 | } |
65 | #endif |
66 | |
67 | // ----------------------------------------------------------------------------- |
68 | |
69 | static SIMDPP_INL |
70 | int8x16 i_avg(const int8x16& a, const int8x16& b) |
71 | { |
72 | #if SIMDPP_USE_NULL |
73 | int8x16 r; |
74 | for (unsigned i = 0; i < a.length; i++) { |
75 | r.el(i) = (int16_t(a.el(i)) + b.el(i) + 1) >> 1; |
76 | } |
77 | return r; |
78 | #elif SIMDPP_USE_SSE2 |
79 | uint8x16 a2, b2, bias, r; |
80 | bias = make_uint(0x80); |
81 | a2 = bit_xor(a, bias); // add |
82 | b2 = bit_xor(b, bias); // add |
83 | r = i_avg(a2, b2); // unsigned |
84 | r = bit_xor(r, bias); // sub |
85 | return r; |
86 | #elif SIMDPP_USE_NEON |
87 | return vrhaddq_s8(a.native(), b.native()); |
88 | #elif SIMDPP_USE_ALTIVEC |
89 | return vec_avg(a.native(), b.native()); |
90 | #elif SIMDPP_USE_MSA |
91 | return __msa_aver_s_b(a.native(), b.native()); |
92 | #endif |
93 | } |
94 | |
95 | #if SIMDPP_USE_AVX2 |
96 | static SIMDPP_INL |
97 | int8x32 i_avg(const int8x32& a, const int8x32& b) |
98 | { |
99 | uint8x32 a2, b2, bias, r; |
100 | bias = make_uint(0x80); |
101 | a2 = bit_xor(a, bias); // add |
102 | b2 = bit_xor(b, bias); // add |
103 | r = i_avg(a2, b2); // unsigned |
104 | r = bit_xor(r, bias); // sub |
105 | return r; |
106 | } |
107 | #endif |
108 | |
109 | #if SIMDPP_USE_AVX512BW |
110 | SIMDPP_INL int8<64>i_avg(const int8<64>& a, const int8<64>& b) |
111 | { |
112 | uint8<64> a2, b2, bias, r; |
113 | bias = make_uint(0x80); |
114 | a2 = bit_xor(a, bias); // add |
115 | b2 = bit_xor(b, bias); // add |
116 | r = i_avg(a2, b2); // unsigned |
117 | r = bit_xor(r, bias); // sub |
118 | return r; |
119 | } |
120 | #endif |
121 | |
122 | // ----------------------------------------------------------------------------- |
123 | |
124 | static SIMDPP_INL |
125 | uint16x8 i_avg(const uint16x8& a, const uint16x8& b) |
126 | { |
127 | #if SIMDPP_USE_NULL |
128 | uint16x8 r; |
129 | for (unsigned i = 0; i < a.length; i++) { |
130 | r.el(i) = (uint32_t(a.el(i)) + b.el(i) + 1) >> 1; |
131 | } |
132 | return r; |
133 | #elif SIMDPP_USE_SSE2 |
134 | return _mm_avg_epu16(a.native(), b.native()); |
135 | #elif SIMDPP_USE_NEON |
136 | return vrhaddq_u16(a.native(), b.native()); |
137 | #elif SIMDPP_USE_ALTIVEC |
138 | return vec_avg(a.native(), b.native()); |
139 | #elif SIMDPP_USE_MSA |
140 | return __msa_aver_u_h(a.native(), b.native()); |
141 | #endif |
142 | } |
143 | |
144 | #if SIMDPP_USE_AVX2 |
145 | static SIMDPP_INL |
146 | uint16x16 i_avg(const uint16x16& a, const uint16x16& b) |
147 | { |
148 | return _mm256_avg_epu16(a.native(), b.native()); |
149 | } |
150 | #endif |
151 | |
152 | #if SIMDPP_USE_AVX512BW |
153 | SIMDPP_INL uint16<32> i_avg(const uint16<32>& a, const uint16<32>& b) |
154 | { |
155 | return _mm512_avg_epu16(a.native(), b.native()); |
156 | } |
157 | #endif |
158 | |
159 | // ----------------------------------------------------------------------------- |
160 | |
161 | static SIMDPP_INL |
162 | int16x8 i_avg(const int16x8& a, const int16x8& b) |
163 | { |
164 | #if SIMDPP_USE_NULL |
165 | int16x8 r; |
166 | for (unsigned i = 0; i < a.length; i++) { |
167 | r.el(i) = (int32_t(a.el(i)) + b.el(i) + 1) >> 1; |
168 | } |
169 | return r; |
170 | #elif SIMDPP_USE_SSE2 |
171 | uint16x8 a2, b2, r; |
172 | a2 = bit_xor(a, 0x8000); // add |
173 | b2 = bit_xor(b, 0x8000); // add |
174 | r = i_avg(a2, b2); // unsigned |
175 | r = bit_xor(r, 0x8000); // sub |
176 | return r; |
177 | #elif SIMDPP_USE_NEON |
178 | return vrhaddq_s16(a.native(), b.native()); |
179 | #elif SIMDPP_USE_ALTIVEC |
180 | return vec_avg(a.native(), b.native()); |
181 | #elif SIMDPP_USE_MSA |
182 | return __msa_aver_s_h(a.native(), b.native()); |
183 | #endif |
184 | } |
185 | |
186 | #if SIMDPP_USE_AVX2 |
187 | static SIMDPP_INL |
188 | int16x16 i_avg(const int16x16& a, const int16x16& b) |
189 | { |
190 | uint16x16 a2, b2, r; |
191 | a2 = bit_xor(a, 0x8000); // add |
192 | b2 = bit_xor(b, 0x8000); // add |
193 | r = i_avg(a2, b2); // unsigned |
194 | r = bit_xor(r, 0x8000); // sub |
195 | return r; |
196 | } |
197 | #endif |
198 | |
199 | #if SIMDPP_USE_AVX512BW |
200 | SIMDPP_INL int16<32> i_avg(const int16<32>& a, const int16<32>& b) |
201 | { |
202 | uint16<32> a2, b2, r; |
203 | a2 = bit_xor(a, 0x8000); // add |
204 | b2 = bit_xor(b, 0x8000); // add |
205 | r = i_avg(a2, b2); // unsigned |
206 | r = bit_xor(r, 0x8000); // sub |
207 | return r; |
208 | } |
209 | #endif |
210 | |
211 | // ----------------------------------------------------------------------------- |
212 | |
213 | static SIMDPP_INL |
214 | uint32x4 i_avg(const uint32x4& a, const uint32x4& b) |
215 | { |
216 | #if SIMDPP_USE_NULL |
217 | uint32x4 r; |
218 | for (unsigned i = 0; i < a.length; i++) { |
219 | r.el(i) = (uint64_t(a.el(i)) + b.el(i) + 1) >> 1; |
220 | } |
221 | return r; |
222 | #elif SIMDPP_USE_SSE2 |
223 | return v_emul_avg_u32(a, b); |
224 | #elif SIMDPP_USE_NEON |
225 | return vrhaddq_u32(a.native(), b.native()); |
226 | #elif SIMDPP_USE_ALTIVEC |
227 | return vec_avg(a.native(), b.native()); |
228 | #elif SIMDPP_USE_MSA |
229 | return __msa_aver_u_w(a.native(), b.native()); |
230 | #endif |
231 | } |
232 | |
233 | #if SIMDPP_USE_AVX2 |
234 | static SIMDPP_INL |
235 | uint32x8 i_avg(const uint32x8& a, const uint32x8& b) |
236 | { |
237 | return v_emul_avg_u32(a, b); |
238 | } |
239 | #endif |
240 | |
241 | #if SIMDPP_USE_AVX512F |
242 | static SIMDPP_INL |
243 | uint32<16> i_avg(const uint32<16>& a, const uint32<16>& b) |
244 | { |
245 | return v_emul_avg_u32(a, b); |
246 | } |
247 | #endif |
248 | |
249 | // ----------------------------------------------------------------------------- |
250 | |
251 | static SIMDPP_INL |
252 | int32x4 i_avg(const int32x4& a, const int32x4& b) |
253 | { |
254 | #if SIMDPP_USE_NULL |
255 | int32x4 r; |
256 | for (unsigned i = 0; i < a.length; i++) { |
257 | r.el(i) = (int64_t(a.el(i)) + b.el(i) + 1) >> 1; |
258 | } |
259 | return r; |
260 | #elif SIMDPP_USE_SSE2 |
261 | return v_emul_avg_i32(a, b); |
262 | |
263 | #elif SIMDPP_USE_NEON |
264 | return vrhaddq_s32(a.native(), b.native()); |
265 | #elif SIMDPP_USE_ALTIVEC |
266 | return vec_avg(a.native(), b.native()); |
267 | #elif SIMDPP_USE_MSA |
268 | return __msa_aver_s_w(a.native(), b.native()); |
269 | #endif |
270 | } |
271 | |
272 | #if SIMDPP_USE_AVX2 |
273 | static SIMDPP_INL |
274 | int32x8 i_avg(const int32x8& a, const int32x8& b) |
275 | { |
276 | return v_emul_avg_i32(a, b); |
277 | } |
278 | #endif |
279 | |
280 | #if SIMDPP_USE_AVX512F |
281 | static SIMDPP_INL |
282 | int32<16> i_avg(const int32<16>& a, const int32<16>& b) |
283 | { |
284 | return v_emul_avg_i32(a, b); |
285 | } |
286 | #endif |
287 | |
288 | // ----------------------------------------------------------------------------- |
289 | |
290 | template<class V> SIMDPP_INL |
291 | V i_avg(const V& a, const V& b) |
292 | { |
293 | SIMDPP_VEC_ARRAY_IMPL2(V, i_avg, a, b); |
294 | } |
295 | |
296 | // generic implementations |
297 | |
298 | template<class V> SIMDPP_INL |
299 | V v_emul_avg_u32(const V& a, const V& b) |
300 | { |
301 | // (x & y) + ((x ^ y) >> 1) + (x ^ y) & 1 |
302 | V x1, x2, round; |
303 | x1 = bit_and(a, b); |
304 | x2 = bit_xor(a, b); |
305 | round = bit_and(x2, 1); |
306 | x1 = add(x1, shift_r<1>(x2)); |
307 | x1 = add(x1, round); |
308 | return x1; |
309 | } |
310 | |
311 | template<class V> SIMDPP_INL |
312 | V v_emul_avg_i32(const V& a, const V& b) |
313 | { |
314 | using VI = typename V::uint_vector_type; |
315 | VI a2, b2, r; |
316 | a2 = bit_xor(a, 0x80000000); // add |
317 | b2 = bit_xor(b, 0x80000000); // add |
318 | r = v_emul_avg_u32(a2, b2); // unsigned |
319 | r = bit_xor(r, 0x80000000); // sub |
320 | return r; |
321 | } |
322 | |
323 | } // namespace insn |
324 | } // namespace detail |
325 | } // namespace SIMDPP_ARCH_NAMESPACE |
326 | } // namespace simdpp |
327 | |
328 | #endif |
329 | |
330 | |