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_CMP_NEQ_H |
9 | #define LIBSIMDPP_SIMDPP_DETAIL_INSN_CMP_NEQ_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_not.h> |
17 | #include <simdpp/core/cmp_eq.h> |
18 | #include <simdpp/detail/not_implemented.h> |
19 | #include <simdpp/detail/null/compare.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 | static SIMDPP_INL |
28 | mask_int8x16 i_cmp_neq(const uint8x16& a, const uint8x16& b) |
29 | { |
30 | #if SIMDPP_USE_NULL |
31 | return detail::null::cmp_neq(a, b); |
32 | #elif SIMDPP_USE_AVX512VL |
33 | return _mm_cmpneq_epi8_mask(a.native(), b.native()); |
34 | #elif SIMDPP_USE_XOP && !SIMDPP_WORKAROUND_XOP_COM |
35 | return _mm_comneq_epi8(a.native(), b.native()); |
36 | #else |
37 | return bit_not(cmp_eq(a, b)); |
38 | #endif |
39 | } |
40 | |
41 | #if SIMDPP_USE_AVX512VL |
42 | static SIMDPP_INL |
43 | mask_int8<16> i_cmp_neq(const mask_int8<16>& a, const mask_int8<16>& b) |
44 | { |
45 | return _mm512_kxor(a.native(), b.native()); |
46 | } |
47 | #endif |
48 | |
49 | #if SIMDPP_USE_AVX2 |
50 | static SIMDPP_INL |
51 | mask_int8x32 i_cmp_neq(const uint8x32& a, const uint8x32& b) |
52 | { |
53 | #if SIMDPP_USE_AVX512VL |
54 | return _mm256_cmpneq_epi8_mask(a.native(), b.native()); |
55 | #else |
56 | return bit_not(cmp_eq(a, b)); |
57 | #endif |
58 | } |
59 | #endif |
60 | |
61 | #if SIMDPP_USE_AVX512VL |
62 | static SIMDPP_INL |
63 | mask_int8<32> i_cmp_neq(const mask_int8<32>& a, const mask_int8<32>& b) |
64 | { |
65 | return _mm512_kxor(a.native(), b.native()); |
66 | } |
67 | #endif |
68 | |
69 | #if SIMDPP_USE_AVX512BW |
70 | SIMDPP_INL mask_int8<64> i_cmp_neq(const uint8<64>& a, const uint8<64>& b) |
71 | { |
72 | return _mm512_cmpneq_epi8_mask(a.native(), b.native()); |
73 | } |
74 | |
75 | SIMDPP_INL mask_int8<64> i_cmp_neq(const mask_int8<64>& a, const mask_int8<64>& b) |
76 | { |
77 | return _mm512_kxor(a.native(), b.native()); |
78 | } |
79 | #endif |
80 | |
81 | // ----------------------------------------------------------------------------- |
82 | |
83 | static SIMDPP_INL |
84 | mask_int16x8 i_cmp_neq(const uint16x8& a, const uint16x8& b) |
85 | { |
86 | #if SIMDPP_USE_NULL |
87 | return detail::null::cmp_neq(a, b); |
88 | #elif SIMDPP_USE_XOP && !SIMDPP_WORKAROUND_XOP_COM |
89 | return _mm_comneq_epi16(a.native(), b.native()); |
90 | #else |
91 | return bit_not(cmp_eq(a, b)); |
92 | #endif |
93 | } |
94 | |
95 | #if SIMDPP_USE_AVX512VL |
96 | static SIMDPP_INL |
97 | mask_int16<8> i_cmp_neq(const mask_int16<8>& a, const mask_int16<8>& b) |
98 | { |
99 | return _mm512_kxor(a.native(), b.native()); |
100 | } |
101 | #endif |
102 | |
103 | #if SIMDPP_USE_AVX2 |
104 | static SIMDPP_INL |
105 | mask_int16x16 i_cmp_neq(const uint16x16& a, const uint16x16& b) |
106 | { |
107 | return bit_not(cmp_eq(a, b)); |
108 | } |
109 | #endif |
110 | |
111 | #if SIMDPP_USE_AVX512VL |
112 | static SIMDPP_INL |
113 | mask_int16<16> i_cmp_neq(const mask_int16<16>& a, const mask_int16<16>& b) |
114 | { |
115 | return _mm512_kxor(a.native(), b.native()); |
116 | } |
117 | #endif |
118 | |
119 | |
120 | #if SIMDPP_USE_AVX512BW |
121 | SIMDPP_INL mask_int16<32> i_cmp_neq(const uint16<32>& a, const uint16<32>& b) |
122 | { |
123 | return _mm512_cmpneq_epi16_mask(a.native(), b.native()); |
124 | } |
125 | |
126 | SIMDPP_INL mask_int16<32> i_cmp_neq(const mask_int16<32>& a, const mask_int16<32>& b) |
127 | { |
128 | return _mm512_kxor(a.native(), b.native()); |
129 | } |
130 | #endif |
131 | |
132 | // ----------------------------------------------------------------------------- |
133 | |
134 | static SIMDPP_INL |
135 | mask_int32x4 i_cmp_neq(const uint32x4& a, const uint32x4& b) |
136 | { |
137 | #if SIMDPP_USE_NULL |
138 | return detail::null::cmp_neq(a, b); |
139 | #elif SIMDPP_USE_XOP && !SIMDPP_WORKAROUND_XOP_COM |
140 | return _mm_comneq_epi32(a.native(), b.native()); |
141 | #else |
142 | return bit_not(cmp_eq(a, b)); |
143 | #endif |
144 | } |
145 | |
146 | #if SIMDPP_USE_AVX512VL |
147 | static SIMDPP_INL |
148 | mask_int32<4> i_cmp_neq(const mask_int32<4>& a, const mask_int32<4>& b) |
149 | { |
150 | return _mm512_kxor(a.native(), b.native()); |
151 | } |
152 | #endif |
153 | |
154 | #if SIMDPP_USE_AVX2 |
155 | static SIMDPP_INL |
156 | mask_int32x8 i_cmp_neq(const uint32x8& a, const uint32x8& b) |
157 | { |
158 | return bit_not(cmp_eq(a, b)); |
159 | } |
160 | #endif |
161 | |
162 | #if SIMDPP_USE_AVX512VL |
163 | static SIMDPP_INL |
164 | mask_int32<8> i_cmp_neq(const mask_int32<8>& a, const mask_int32<8>& b) |
165 | { |
166 | return _mm512_kxor(a.native(), b.native()); |
167 | } |
168 | #endif |
169 | |
170 | #if SIMDPP_USE_AVX512F |
171 | static SIMDPP_INL |
172 | mask_int32<16> i_cmp_neq(const uint32<16>& a, const uint32<16>& b) |
173 | { |
174 | return _mm512_cmpneq_epu32_mask(a.native(), b.native()); |
175 | } |
176 | |
177 | static SIMDPP_INL |
178 | mask_int32<16> i_cmp_neq(const mask_int32<16>& a, const mask_int32<16>& b) |
179 | { |
180 | return _mm512_kxor(a.native(), b.native()); |
181 | } |
182 | #endif |
183 | |
184 | // ----------------------------------------------------------------------------- |
185 | |
186 | static SIMDPP_INL |
187 | mask_int64x2 i_cmp_neq(const uint64x2& a, const uint64x2& b) |
188 | { |
189 | #if SIMDPP_USE_XOP && !SIMDPP_WORKAROUND_XOP_COM |
190 | return _mm_comneq_epi64(a.native(), b.native()); |
191 | #elif SIMDPP_USE_SSE4_1 || SIMDPP_USE_NEON || SIMDPP_USE_VSX_207 || SIMDPP_USE_MSA |
192 | return bit_not(cmp_eq(a, b)); |
193 | #elif SIMDPP_USE_SSE2 |
194 | uint64x2 r32, r32s; |
195 | r32 = (uint32x4)cmp_eq(uint32x4(a), uint32x4(b)); |
196 | // swap the 32-bit halves |
197 | r32s = bit_or(shift_l<32>(r32), shift_r<32>(r32)); |
198 | // combine the results. Each 32-bit half is ORed with the neighbouring pair |
199 | r32 = bit_or(r32, r32s); |
200 | return r32; |
201 | #elif SIMDPP_USE_NULL || SIMDPP_USE_ALTIVEC |
202 | return detail::null::cmp_neq(a, b); |
203 | #endif |
204 | } |
205 | |
206 | #if SIMDPP_USE_AVX512VL |
207 | static SIMDPP_INL |
208 | mask_int64<2> i_cmp_neq(const mask_int64<2>& a, const mask_int64<2>& b) |
209 | { |
210 | return _mm512_kxor(a.native(), b.native()); |
211 | } |
212 | #endif |
213 | |
214 | #if SIMDPP_USE_AVX2 |
215 | static SIMDPP_INL |
216 | mask_int64x4 i_cmp_neq(const uint64x4& a, const uint64x4& b) |
217 | { |
218 | return bit_not(cmp_eq(a, b)); |
219 | } |
220 | #endif |
221 | |
222 | #if SIMDPP_USE_AVX512VL |
223 | static SIMDPP_INL |
224 | mask_int64<4> i_cmp_neq(const mask_int64<4>& a, const mask_int64<4>& b) |
225 | { |
226 | return _mm512_kxor(a.native(), b.native()); |
227 | } |
228 | #endif |
229 | |
230 | #if SIMDPP_USE_AVX512F |
231 | static SIMDPP_INL |
232 | mask_int64<8> i_cmp_neq(const uint64<8>& a, const uint64<8>& b) |
233 | { |
234 | return _mm512_cmpneq_epi64_mask(a.native(), b.native()); |
235 | } |
236 | |
237 | static SIMDPP_INL |
238 | mask_int64<8> i_cmp_neq(const mask_int64<8>& a, const mask_int64<8>& b) |
239 | { |
240 | return _mm512_kxor(a.native(), b.native()); |
241 | } |
242 | #endif |
243 | |
244 | // ----------------------------------------------------------------------------- |
245 | |
246 | static SIMDPP_INL |
247 | mask_float32x4 i_cmp_neq(const float32x4& a, const float32x4& b) |
248 | { |
249 | #if SIMDPP_USE_NULL |
250 | return detail::null::cmp_neq(a, b); |
251 | #elif SIMDPP_USE_AVX512VL |
252 | return _mm_cmp_ps_mask(a.native(), b.native(), _CMP_NEQ_UQ); |
253 | #elif SIMDPP_USE_AVX |
254 | return _mm_cmp_ps(a.native(), b.native(), _CMP_NEQ_UQ); |
255 | #elif SIMDPP_USE_SSE2 |
256 | return _mm_cmpneq_ps(a.native(), b.native()); |
257 | #elif SIMDPP_USE_NEON || SIMDPP_USE_ALTIVEC |
258 | return bit_not(cmp_eq(a, b)); |
259 | #elif SIMDPP_USE_MSA |
260 | return (v4f32) __msa_fcune_w(a.native(), b.native()); |
261 | #endif |
262 | } |
263 | |
264 | #if SIMDPP_USE_AVX512VL |
265 | static SIMDPP_INL |
266 | mask_float32<4> i_cmp_neq(const mask_float32<4>& a, const mask_float32<4>& b) |
267 | { |
268 | return _mm512_kxor(a.native(), b.native()); |
269 | } |
270 | #endif |
271 | |
272 | #if SIMDPP_USE_AVX |
273 | static SIMDPP_INL |
274 | mask_float32x8 i_cmp_neq(const float32x8& a, const float32x8& b) |
275 | { |
276 | #if SIMDPP_USE_AVX512VL |
277 | return _mm256_cmp_ps_mask(a.native(), b.native(), _CMP_NEQ_UQ); |
278 | #else |
279 | return _mm256_cmp_ps(a.native(), b.native(), _CMP_NEQ_UQ); |
280 | #endif |
281 | } |
282 | #endif |
283 | |
284 | #if SIMDPP_USE_AVX512VL |
285 | static SIMDPP_INL |
286 | mask_float32<8> i_cmp_neq(const mask_float32<8>& a, const mask_float32<8>& b) |
287 | { |
288 | return _mm512_kxor(a.native(), b.native()); |
289 | } |
290 | #endif |
291 | |
292 | #if SIMDPP_USE_AVX512F |
293 | static SIMDPP_INL |
294 | mask_float32<16> i_cmp_neq(const float32<16>& a, const float32<16>& b) |
295 | { |
296 | return _mm512_cmp_ps_mask(a.native(), b.native(), _CMP_NEQ_UQ); |
297 | } |
298 | |
299 | static SIMDPP_INL |
300 | mask_float32<16> i_cmp_neq(const mask_float32<16>& a, const mask_float32<16>& b) |
301 | { |
302 | return _mm512_kxor(a.native(), b.native()); |
303 | } |
304 | #endif |
305 | |
306 | // ----------------------------------------------------------------------------- |
307 | |
308 | static SIMDPP_INL |
309 | mask_float64x2 i_cmp_neq(const float64x2& a, const float64x2& b) |
310 | { |
311 | #if SIMDPP_USE_AVX512VL |
312 | return _mm_cmp_pd_mask(a.native(), b.native(), _CMP_NEQ_UQ); |
313 | #elif SIMDPP_USE_AVX |
314 | return _mm_cmp_pd(a.native(), b.native(), _CMP_NEQ_UQ); |
315 | #elif SIMDPP_USE_SSE2 |
316 | return _mm_cmpneq_pd(a.native(), b.native()); |
317 | #elif SIMDPP_USE_NEON64 || SIMDPP_USE_VSX_206 |
318 | return bit_not(cmp_eq(a, b)); |
319 | #elif SIMDPP_USE_MSA |
320 | return (v2f64) __msa_fcune_d(a.native(), b.native()); |
321 | #elif SIMDPP_USE_NULL || SIMDPP_USE_NEON32 || SIMDPP_USE_ALTIVEC |
322 | return detail::null::cmp_neq(a, b); |
323 | #else |
324 | return SIMDPP_NOT_IMPLEMENTED2(a, b); |
325 | #endif |
326 | } |
327 | |
328 | #if SIMDPP_USE_AVX512VL |
329 | static SIMDPP_INL |
330 | mask_float64<2> i_cmp_neq(const mask_float64<2>& a, const mask_float64<2>& b) |
331 | { |
332 | return _mm512_kxor(a.native(), b.native()); |
333 | } |
334 | #endif |
335 | |
336 | #if SIMDPP_USE_AVX |
337 | static SIMDPP_INL |
338 | mask_float64x4 i_cmp_neq(const float64x4& a, const float64x4& b) |
339 | { |
340 | #if SIMDPP_USE_AVX512VL |
341 | return _mm256_cmp_pd_mask(a.native(), b.native(), _CMP_NEQ_UQ); |
342 | #else |
343 | return _mm256_cmp_pd(a.native(), b.native(), _CMP_NEQ_UQ); |
344 | #endif |
345 | } |
346 | #endif |
347 | |
348 | #if SIMDPP_USE_AVX512VL |
349 | static SIMDPP_INL |
350 | mask_float64<4> i_cmp_neq(const mask_float64<4>& a, const mask_float64<4>& b) |
351 | { |
352 | return _mm512_kxor(a.native(), b.native()); |
353 | } |
354 | #endif |
355 | |
356 | #if SIMDPP_USE_AVX512F |
357 | static SIMDPP_INL |
358 | mask_float64<8> i_cmp_neq(const float64<8>& a, const float64<8>& b) |
359 | { |
360 | return _mm512_cmp_pd_mask(a.native(), b.native(), _CMP_NEQ_UQ); |
361 | } |
362 | |
363 | static SIMDPP_INL |
364 | mask_float64<8> i_cmp_neq(const mask_float64<8>& a, const mask_float64<8>& b) |
365 | { |
366 | return _mm512_kxor(a.native(), b.native()); |
367 | } |
368 | #endif |
369 | |
370 | // ----------------------------------------------------------------------------- |
371 | |
372 | template<class V> SIMDPP_INL |
373 | typename V::mask_vector_type i_cmp_neq(const V& a, const V& b) |
374 | { |
375 | SIMDPP_VEC_ARRAY_IMPL2(typename V::mask_vector_type, i_cmp_neq, a, b); |
376 | } |
377 | |
378 | } // namespace insn |
379 | } // namespace detail |
380 | } // namespace SIMDPP_ARCH_NAMESPACE |
381 | } // namespace simdpp |
382 | |
383 | #endif |
384 | |
385 | |