1 |
282 |
jeremybenn |
/* Intrinsics for ST Microelectronics Loongson-2E/2F SIMD operations.
|
2 |
|
|
|
3 |
|
|
Copyright (C) 2008, 2009 Free Software Foundation, Inc.
|
4 |
|
|
Contributed by CodeSourcery.
|
5 |
|
|
|
6 |
|
|
This file is part of GCC.
|
7 |
|
|
|
8 |
|
|
GCC is free software; you can redistribute it and/or modify it
|
9 |
|
|
under the terms of the GNU General Public License as published
|
10 |
|
|
by the Free Software Foundation; either version 3, or (at your
|
11 |
|
|
option) any later version.
|
12 |
|
|
|
13 |
|
|
GCC is distributed in the hope that it will be useful, but WITHOUT
|
14 |
|
|
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
|
15 |
|
|
or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
|
16 |
|
|
License for more details.
|
17 |
|
|
|
18 |
|
|
Under Section 7 of GPL version 3, you are granted additional
|
19 |
|
|
permissions described in the GCC Runtime Library Exception, version
|
20 |
|
|
3.1, as published by the Free Software Foundation.
|
21 |
|
|
|
22 |
|
|
You should have received a copy of the GNU General Public License and
|
23 |
|
|
a copy of the GCC Runtime Library Exception along with this program;
|
24 |
|
|
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
25 |
|
|
<http://www.gnu.org/licenses/>. */
|
26 |
|
|
|
27 |
|
|
#ifndef _GCC_LOONGSON_H
|
28 |
|
|
#define _GCC_LOONGSON_H
|
29 |
|
|
|
30 |
|
|
#if !defined(__mips_loongson_vector_rev)
|
31 |
|
|
# error "You must select -march=loongson2e or -march=loongson2f to use loongson.h"
|
32 |
|
|
#endif
|
33 |
|
|
|
34 |
|
|
#ifdef __cplusplus
|
35 |
|
|
extern "C" {
|
36 |
|
|
#endif
|
37 |
|
|
|
38 |
|
|
#include <stdint.h>
|
39 |
|
|
|
40 |
|
|
/* Vectors of unsigned bytes, halfwords and words. */
|
41 |
|
|
typedef uint8_t uint8x8_t __attribute__((vector_size (8)));
|
42 |
|
|
typedef uint16_t uint16x4_t __attribute__((vector_size (8)));
|
43 |
|
|
typedef uint32_t uint32x2_t __attribute__((vector_size (8)));
|
44 |
|
|
|
45 |
|
|
/* Vectors of signed bytes, halfwords and words. */
|
46 |
|
|
typedef int8_t int8x8_t __attribute__((vector_size (8)));
|
47 |
|
|
typedef int16_t int16x4_t __attribute__((vector_size (8)));
|
48 |
|
|
typedef int32_t int32x2_t __attribute__((vector_size (8)));
|
49 |
|
|
|
50 |
|
|
/* SIMD intrinsics.
|
51 |
|
|
Unless otherwise noted, calls to the functions below will expand into
|
52 |
|
|
precisely one machine instruction, modulo any moves required to
|
53 |
|
|
satisfy register allocation constraints. */
|
54 |
|
|
|
55 |
|
|
/* Pack with signed saturation. */
|
56 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
57 |
|
|
packsswh (int32x2_t s, int32x2_t t)
|
58 |
|
|
{
|
59 |
|
|
return __builtin_loongson_packsswh (s, t);
|
60 |
|
|
}
|
61 |
|
|
|
62 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
63 |
|
|
packsshb (int16x4_t s, int16x4_t t)
|
64 |
|
|
{
|
65 |
|
|
return __builtin_loongson_packsshb (s, t);
|
66 |
|
|
}
|
67 |
|
|
|
68 |
|
|
/* Pack with unsigned saturation. */
|
69 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
70 |
|
|
packushb (uint16x4_t s, uint16x4_t t)
|
71 |
|
|
{
|
72 |
|
|
return __builtin_loongson_packushb (s, t);
|
73 |
|
|
}
|
74 |
|
|
|
75 |
|
|
/* Vector addition, treating overflow by wraparound. */
|
76 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
77 |
|
|
paddw_u (uint32x2_t s, uint32x2_t t)
|
78 |
|
|
{
|
79 |
|
|
return __builtin_loongson_paddw_u (s, t);
|
80 |
|
|
}
|
81 |
|
|
|
82 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
83 |
|
|
paddh_u (uint16x4_t s, uint16x4_t t)
|
84 |
|
|
{
|
85 |
|
|
return __builtin_loongson_paddh_u (s, t);
|
86 |
|
|
}
|
87 |
|
|
|
88 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
89 |
|
|
paddb_u (uint8x8_t s, uint8x8_t t)
|
90 |
|
|
{
|
91 |
|
|
return __builtin_loongson_paddb_u (s, t);
|
92 |
|
|
}
|
93 |
|
|
|
94 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
95 |
|
|
paddw_s (int32x2_t s, int32x2_t t)
|
96 |
|
|
{
|
97 |
|
|
return __builtin_loongson_paddw_s (s, t);
|
98 |
|
|
}
|
99 |
|
|
|
100 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
101 |
|
|
paddh_s (int16x4_t s, int16x4_t t)
|
102 |
|
|
{
|
103 |
|
|
return __builtin_loongson_paddh_s (s, t);
|
104 |
|
|
}
|
105 |
|
|
|
106 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
107 |
|
|
paddb_s (int8x8_t s, int8x8_t t)
|
108 |
|
|
{
|
109 |
|
|
return __builtin_loongson_paddb_s (s, t);
|
110 |
|
|
}
|
111 |
|
|
|
112 |
|
|
/* Addition of doubleword integers, treating overflow by wraparound. */
|
113 |
|
|
__extension__ static __inline uint64_t __attribute__ ((__always_inline__))
|
114 |
|
|
paddd_u (uint64_t s, uint64_t t)
|
115 |
|
|
{
|
116 |
|
|
return __builtin_loongson_paddd_u (s, t);
|
117 |
|
|
}
|
118 |
|
|
|
119 |
|
|
__extension__ static __inline int64_t __attribute__ ((__always_inline__))
|
120 |
|
|
paddd_s (int64_t s, int64_t t)
|
121 |
|
|
{
|
122 |
|
|
return __builtin_loongson_paddd_s (s, t);
|
123 |
|
|
}
|
124 |
|
|
|
125 |
|
|
/* Vector addition, treating overflow by signed saturation. */
|
126 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
127 |
|
|
paddsh (int16x4_t s, int16x4_t t)
|
128 |
|
|
{
|
129 |
|
|
return __builtin_loongson_paddsh (s, t);
|
130 |
|
|
}
|
131 |
|
|
|
132 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
133 |
|
|
paddsb (int8x8_t s, int8x8_t t)
|
134 |
|
|
{
|
135 |
|
|
return __builtin_loongson_paddsb (s, t);
|
136 |
|
|
}
|
137 |
|
|
|
138 |
|
|
/* Vector addition, treating overflow by unsigned saturation. */
|
139 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
140 |
|
|
paddush (uint16x4_t s, uint16x4_t t)
|
141 |
|
|
{
|
142 |
|
|
return __builtin_loongson_paddush (s, t);
|
143 |
|
|
}
|
144 |
|
|
|
145 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
146 |
|
|
paddusb (uint8x8_t s, uint8x8_t t)
|
147 |
|
|
{
|
148 |
|
|
return __builtin_loongson_paddusb (s, t);
|
149 |
|
|
}
|
150 |
|
|
|
151 |
|
|
/* Logical AND NOT. */
|
152 |
|
|
__extension__ static __inline uint64_t __attribute__ ((__always_inline__))
|
153 |
|
|
pandn_ud (uint64_t s, uint64_t t)
|
154 |
|
|
{
|
155 |
|
|
return __builtin_loongson_pandn_ud (s, t);
|
156 |
|
|
}
|
157 |
|
|
|
158 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
159 |
|
|
pandn_uw (uint32x2_t s, uint32x2_t t)
|
160 |
|
|
{
|
161 |
|
|
return __builtin_loongson_pandn_uw (s, t);
|
162 |
|
|
}
|
163 |
|
|
|
164 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
165 |
|
|
pandn_uh (uint16x4_t s, uint16x4_t t)
|
166 |
|
|
{
|
167 |
|
|
return __builtin_loongson_pandn_uh (s, t);
|
168 |
|
|
}
|
169 |
|
|
|
170 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
171 |
|
|
pandn_ub (uint8x8_t s, uint8x8_t t)
|
172 |
|
|
{
|
173 |
|
|
return __builtin_loongson_pandn_ub (s, t);
|
174 |
|
|
}
|
175 |
|
|
|
176 |
|
|
__extension__ static __inline int64_t __attribute__ ((__always_inline__))
|
177 |
|
|
pandn_sd (int64_t s, int64_t t)
|
178 |
|
|
{
|
179 |
|
|
return __builtin_loongson_pandn_sd (s, t);
|
180 |
|
|
}
|
181 |
|
|
|
182 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
183 |
|
|
pandn_sw (int32x2_t s, int32x2_t t)
|
184 |
|
|
{
|
185 |
|
|
return __builtin_loongson_pandn_sw (s, t);
|
186 |
|
|
}
|
187 |
|
|
|
188 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
189 |
|
|
pandn_sh (int16x4_t s, int16x4_t t)
|
190 |
|
|
{
|
191 |
|
|
return __builtin_loongson_pandn_sh (s, t);
|
192 |
|
|
}
|
193 |
|
|
|
194 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
195 |
|
|
pandn_sb (int8x8_t s, int8x8_t t)
|
196 |
|
|
{
|
197 |
|
|
return __builtin_loongson_pandn_sb (s, t);
|
198 |
|
|
}
|
199 |
|
|
|
200 |
|
|
/* Average. */
|
201 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
202 |
|
|
pavgh (uint16x4_t s, uint16x4_t t)
|
203 |
|
|
{
|
204 |
|
|
return __builtin_loongson_pavgh (s, t);
|
205 |
|
|
}
|
206 |
|
|
|
207 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
208 |
|
|
pavgb (uint8x8_t s, uint8x8_t t)
|
209 |
|
|
{
|
210 |
|
|
return __builtin_loongson_pavgb (s, t);
|
211 |
|
|
}
|
212 |
|
|
|
213 |
|
|
/* Equality test. */
|
214 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
215 |
|
|
pcmpeqw_u (uint32x2_t s, uint32x2_t t)
|
216 |
|
|
{
|
217 |
|
|
return __builtin_loongson_pcmpeqw_u (s, t);
|
218 |
|
|
}
|
219 |
|
|
|
220 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
221 |
|
|
pcmpeqh_u (uint16x4_t s, uint16x4_t t)
|
222 |
|
|
{
|
223 |
|
|
return __builtin_loongson_pcmpeqh_u (s, t);
|
224 |
|
|
}
|
225 |
|
|
|
226 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
227 |
|
|
pcmpeqb_u (uint8x8_t s, uint8x8_t t)
|
228 |
|
|
{
|
229 |
|
|
return __builtin_loongson_pcmpeqb_u (s, t);
|
230 |
|
|
}
|
231 |
|
|
|
232 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
233 |
|
|
pcmpeqw_s (int32x2_t s, int32x2_t t)
|
234 |
|
|
{
|
235 |
|
|
return __builtin_loongson_pcmpeqw_s (s, t);
|
236 |
|
|
}
|
237 |
|
|
|
238 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
239 |
|
|
pcmpeqh_s (int16x4_t s, int16x4_t t)
|
240 |
|
|
{
|
241 |
|
|
return __builtin_loongson_pcmpeqh_s (s, t);
|
242 |
|
|
}
|
243 |
|
|
|
244 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
245 |
|
|
pcmpeqb_s (int8x8_t s, int8x8_t t)
|
246 |
|
|
{
|
247 |
|
|
return __builtin_loongson_pcmpeqb_s (s, t);
|
248 |
|
|
}
|
249 |
|
|
|
250 |
|
|
/* Greater-than test. */
|
251 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
252 |
|
|
pcmpgtw_u (uint32x2_t s, uint32x2_t t)
|
253 |
|
|
{
|
254 |
|
|
return __builtin_loongson_pcmpgtw_u (s, t);
|
255 |
|
|
}
|
256 |
|
|
|
257 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
258 |
|
|
pcmpgth_u (uint16x4_t s, uint16x4_t t)
|
259 |
|
|
{
|
260 |
|
|
return __builtin_loongson_pcmpgth_u (s, t);
|
261 |
|
|
}
|
262 |
|
|
|
263 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
264 |
|
|
pcmpgtb_u (uint8x8_t s, uint8x8_t t)
|
265 |
|
|
{
|
266 |
|
|
return __builtin_loongson_pcmpgtb_u (s, t);
|
267 |
|
|
}
|
268 |
|
|
|
269 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
270 |
|
|
pcmpgtw_s (int32x2_t s, int32x2_t t)
|
271 |
|
|
{
|
272 |
|
|
return __builtin_loongson_pcmpgtw_s (s, t);
|
273 |
|
|
}
|
274 |
|
|
|
275 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
276 |
|
|
pcmpgth_s (int16x4_t s, int16x4_t t)
|
277 |
|
|
{
|
278 |
|
|
return __builtin_loongson_pcmpgth_s (s, t);
|
279 |
|
|
}
|
280 |
|
|
|
281 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
282 |
|
|
pcmpgtb_s (int8x8_t s, int8x8_t t)
|
283 |
|
|
{
|
284 |
|
|
return __builtin_loongson_pcmpgtb_s (s, t);
|
285 |
|
|
}
|
286 |
|
|
|
287 |
|
|
/* Extract halfword. */
|
288 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
289 |
|
|
pextrh_u (uint16x4_t s, int field /* 0--3 */)
|
290 |
|
|
{
|
291 |
|
|
return __builtin_loongson_pextrh_u (s, field);
|
292 |
|
|
}
|
293 |
|
|
|
294 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
295 |
|
|
pextrh_s (int16x4_t s, int field /* 0--3 */)
|
296 |
|
|
{
|
297 |
|
|
return __builtin_loongson_pextrh_s (s, field);
|
298 |
|
|
}
|
299 |
|
|
|
300 |
|
|
/* Insert halfword. */
|
301 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
302 |
|
|
pinsrh_0_u (uint16x4_t s, uint16x4_t t)
|
303 |
|
|
{
|
304 |
|
|
return __builtin_loongson_pinsrh_0_u (s, t);
|
305 |
|
|
}
|
306 |
|
|
|
307 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
308 |
|
|
pinsrh_1_u (uint16x4_t s, uint16x4_t t)
|
309 |
|
|
{
|
310 |
|
|
return __builtin_loongson_pinsrh_1_u (s, t);
|
311 |
|
|
}
|
312 |
|
|
|
313 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
314 |
|
|
pinsrh_2_u (uint16x4_t s, uint16x4_t t)
|
315 |
|
|
{
|
316 |
|
|
return __builtin_loongson_pinsrh_2_u (s, t);
|
317 |
|
|
}
|
318 |
|
|
|
319 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
320 |
|
|
pinsrh_3_u (uint16x4_t s, uint16x4_t t)
|
321 |
|
|
{
|
322 |
|
|
return __builtin_loongson_pinsrh_3_u (s, t);
|
323 |
|
|
}
|
324 |
|
|
|
325 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
326 |
|
|
pinsrh_0_s (int16x4_t s, int16x4_t t)
|
327 |
|
|
{
|
328 |
|
|
return __builtin_loongson_pinsrh_0_s (s, t);
|
329 |
|
|
}
|
330 |
|
|
|
331 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
332 |
|
|
pinsrh_1_s (int16x4_t s, int16x4_t t)
|
333 |
|
|
{
|
334 |
|
|
return __builtin_loongson_pinsrh_1_s (s, t);
|
335 |
|
|
}
|
336 |
|
|
|
337 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
338 |
|
|
pinsrh_2_s (int16x4_t s, int16x4_t t)
|
339 |
|
|
{
|
340 |
|
|
return __builtin_loongson_pinsrh_2_s (s, t);
|
341 |
|
|
}
|
342 |
|
|
|
343 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
344 |
|
|
pinsrh_3_s (int16x4_t s, int16x4_t t)
|
345 |
|
|
{
|
346 |
|
|
return __builtin_loongson_pinsrh_3_s (s, t);
|
347 |
|
|
}
|
348 |
|
|
|
349 |
|
|
/* Multiply and add. */
|
350 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
351 |
|
|
pmaddhw (int16x4_t s, int16x4_t t)
|
352 |
|
|
{
|
353 |
|
|
return __builtin_loongson_pmaddhw (s, t);
|
354 |
|
|
}
|
355 |
|
|
|
356 |
|
|
/* Maximum of signed halfwords. */
|
357 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
358 |
|
|
pmaxsh (int16x4_t s, int16x4_t t)
|
359 |
|
|
{
|
360 |
|
|
return __builtin_loongson_pmaxsh (s, t);
|
361 |
|
|
}
|
362 |
|
|
|
363 |
|
|
/* Maximum of unsigned bytes. */
|
364 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
365 |
|
|
pmaxub (uint8x8_t s, uint8x8_t t)
|
366 |
|
|
{
|
367 |
|
|
return __builtin_loongson_pmaxub (s, t);
|
368 |
|
|
}
|
369 |
|
|
|
370 |
|
|
/* Minimum of signed halfwords. */
|
371 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
372 |
|
|
pminsh (int16x4_t s, int16x4_t t)
|
373 |
|
|
{
|
374 |
|
|
return __builtin_loongson_pminsh (s, t);
|
375 |
|
|
}
|
376 |
|
|
|
377 |
|
|
/* Minimum of unsigned bytes. */
|
378 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
379 |
|
|
pminub (uint8x8_t s, uint8x8_t t)
|
380 |
|
|
{
|
381 |
|
|
return __builtin_loongson_pminub (s, t);
|
382 |
|
|
}
|
383 |
|
|
|
384 |
|
|
/* Move byte mask. */
|
385 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
386 |
|
|
pmovmskb_u (uint8x8_t s)
|
387 |
|
|
{
|
388 |
|
|
return __builtin_loongson_pmovmskb_u (s);
|
389 |
|
|
}
|
390 |
|
|
|
391 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
392 |
|
|
pmovmskb_s (int8x8_t s)
|
393 |
|
|
{
|
394 |
|
|
return __builtin_loongson_pmovmskb_s (s);
|
395 |
|
|
}
|
396 |
|
|
|
397 |
|
|
/* Multiply unsigned integers and store high result. */
|
398 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
399 |
|
|
pmulhuh (uint16x4_t s, uint16x4_t t)
|
400 |
|
|
{
|
401 |
|
|
return __builtin_loongson_pmulhuh (s, t);
|
402 |
|
|
}
|
403 |
|
|
|
404 |
|
|
/* Multiply signed integers and store high result. */
|
405 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
406 |
|
|
pmulhh (int16x4_t s, int16x4_t t)
|
407 |
|
|
{
|
408 |
|
|
return __builtin_loongson_pmulhh (s, t);
|
409 |
|
|
}
|
410 |
|
|
|
411 |
|
|
/* Multiply signed integers and store low result. */
|
412 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
413 |
|
|
pmullh (int16x4_t s, int16x4_t t)
|
414 |
|
|
{
|
415 |
|
|
return __builtin_loongson_pmullh (s, t);
|
416 |
|
|
}
|
417 |
|
|
|
418 |
|
|
/* Multiply unsigned word integers. */
|
419 |
|
|
__extension__ static __inline int64_t __attribute__ ((__always_inline__))
|
420 |
|
|
pmuluw (uint32x2_t s, uint32x2_t t)
|
421 |
|
|
{
|
422 |
|
|
return __builtin_loongson_pmuluw (s, t);
|
423 |
|
|
}
|
424 |
|
|
|
425 |
|
|
/* Absolute difference. */
|
426 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
427 |
|
|
pasubub (uint8x8_t s, uint8x8_t t)
|
428 |
|
|
{
|
429 |
|
|
return __builtin_loongson_pasubub (s, t);
|
430 |
|
|
}
|
431 |
|
|
|
432 |
|
|
/* Sum of unsigned byte integers. */
|
433 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
434 |
|
|
biadd (uint8x8_t s)
|
435 |
|
|
{
|
436 |
|
|
return __builtin_loongson_biadd (s);
|
437 |
|
|
}
|
438 |
|
|
|
439 |
|
|
/* Sum of absolute differences.
|
440 |
|
|
Note that this intrinsic expands into two machine instructions:
|
441 |
|
|
PASUBUB followed by BIADD. */
|
442 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
443 |
|
|
psadbh (uint8x8_t s, uint8x8_t t)
|
444 |
|
|
{
|
445 |
|
|
return __builtin_loongson_psadbh (s, t);
|
446 |
|
|
}
|
447 |
|
|
|
448 |
|
|
/* Shuffle halfwords. */
|
449 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
450 |
|
|
pshufh_u (uint16x4_t dest, uint16x4_t s, uint8_t order)
|
451 |
|
|
{
|
452 |
|
|
return __builtin_loongson_pshufh_u (dest, s, order);
|
453 |
|
|
}
|
454 |
|
|
|
455 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
456 |
|
|
pshufh_s (int16x4_t dest, int16x4_t s, uint8_t order)
|
457 |
|
|
{
|
458 |
|
|
return __builtin_loongson_pshufh_s (dest, s, order);
|
459 |
|
|
}
|
460 |
|
|
|
461 |
|
|
/* Shift left logical. */
|
462 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
463 |
|
|
psllh_u (uint16x4_t s, uint8_t amount)
|
464 |
|
|
{
|
465 |
|
|
return __builtin_loongson_psllh_u (s, amount);
|
466 |
|
|
}
|
467 |
|
|
|
468 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
469 |
|
|
psllh_s (int16x4_t s, uint8_t amount)
|
470 |
|
|
{
|
471 |
|
|
return __builtin_loongson_psllh_s (s, amount);
|
472 |
|
|
}
|
473 |
|
|
|
474 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
475 |
|
|
psllw_u (uint32x2_t s, uint8_t amount)
|
476 |
|
|
{
|
477 |
|
|
return __builtin_loongson_psllw_u (s, amount);
|
478 |
|
|
}
|
479 |
|
|
|
480 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
481 |
|
|
psllw_s (int32x2_t s, uint8_t amount)
|
482 |
|
|
{
|
483 |
|
|
return __builtin_loongson_psllw_s (s, amount);
|
484 |
|
|
}
|
485 |
|
|
|
486 |
|
|
/* Shift right logical. */
|
487 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
488 |
|
|
psrlh_u (uint16x4_t s, uint8_t amount)
|
489 |
|
|
{
|
490 |
|
|
return __builtin_loongson_psrlh_u (s, amount);
|
491 |
|
|
}
|
492 |
|
|
|
493 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
494 |
|
|
psrlh_s (int16x4_t s, uint8_t amount)
|
495 |
|
|
{
|
496 |
|
|
return __builtin_loongson_psrlh_s (s, amount);
|
497 |
|
|
}
|
498 |
|
|
|
499 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
500 |
|
|
psrlw_u (uint32x2_t s, uint8_t amount)
|
501 |
|
|
{
|
502 |
|
|
return __builtin_loongson_psrlw_u (s, amount);
|
503 |
|
|
}
|
504 |
|
|
|
505 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
506 |
|
|
psrlw_s (int32x2_t s, uint8_t amount)
|
507 |
|
|
{
|
508 |
|
|
return __builtin_loongson_psrlw_s (s, amount);
|
509 |
|
|
}
|
510 |
|
|
|
511 |
|
|
/* Shift right arithmetic. */
|
512 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
513 |
|
|
psrah_u (uint16x4_t s, uint8_t amount)
|
514 |
|
|
{
|
515 |
|
|
return __builtin_loongson_psrah_u (s, amount);
|
516 |
|
|
}
|
517 |
|
|
|
518 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
519 |
|
|
psrah_s (int16x4_t s, uint8_t amount)
|
520 |
|
|
{
|
521 |
|
|
return __builtin_loongson_psrah_s (s, amount);
|
522 |
|
|
}
|
523 |
|
|
|
524 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
525 |
|
|
psraw_u (uint32x2_t s, uint8_t amount)
|
526 |
|
|
{
|
527 |
|
|
return __builtin_loongson_psraw_u (s, amount);
|
528 |
|
|
}
|
529 |
|
|
|
530 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
531 |
|
|
psraw_s (int32x2_t s, uint8_t amount)
|
532 |
|
|
{
|
533 |
|
|
return __builtin_loongson_psraw_s (s, amount);
|
534 |
|
|
}
|
535 |
|
|
|
536 |
|
|
/* Vector subtraction, treating overflow by wraparound. */
|
537 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
538 |
|
|
psubw_u (uint32x2_t s, uint32x2_t t)
|
539 |
|
|
{
|
540 |
|
|
return __builtin_loongson_psubw_u (s, t);
|
541 |
|
|
}
|
542 |
|
|
|
543 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
544 |
|
|
psubh_u (uint16x4_t s, uint16x4_t t)
|
545 |
|
|
{
|
546 |
|
|
return __builtin_loongson_psubh_u (s, t);
|
547 |
|
|
}
|
548 |
|
|
|
549 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
550 |
|
|
psubb_u (uint8x8_t s, uint8x8_t t)
|
551 |
|
|
{
|
552 |
|
|
return __builtin_loongson_psubb_u (s, t);
|
553 |
|
|
}
|
554 |
|
|
|
555 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
556 |
|
|
psubw_s (int32x2_t s, int32x2_t t)
|
557 |
|
|
{
|
558 |
|
|
return __builtin_loongson_psubw_s (s, t);
|
559 |
|
|
}
|
560 |
|
|
|
561 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
562 |
|
|
psubh_s (int16x4_t s, int16x4_t t)
|
563 |
|
|
{
|
564 |
|
|
return __builtin_loongson_psubh_s (s, t);
|
565 |
|
|
}
|
566 |
|
|
|
567 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
568 |
|
|
psubb_s (int8x8_t s, int8x8_t t)
|
569 |
|
|
{
|
570 |
|
|
return __builtin_loongson_psubb_s (s, t);
|
571 |
|
|
}
|
572 |
|
|
|
573 |
|
|
/* Subtraction of doubleword integers, treating overflow by wraparound. */
|
574 |
|
|
__extension__ static __inline uint64_t __attribute__ ((__always_inline__))
|
575 |
|
|
psubd_u (uint64_t s, uint64_t t)
|
576 |
|
|
{
|
577 |
|
|
return __builtin_loongson_psubd_u (s, t);
|
578 |
|
|
}
|
579 |
|
|
|
580 |
|
|
__extension__ static __inline int64_t __attribute__ ((__always_inline__))
|
581 |
|
|
psubd_s (int64_t s, int64_t t)
|
582 |
|
|
{
|
583 |
|
|
return __builtin_loongson_psubd_s (s, t);
|
584 |
|
|
}
|
585 |
|
|
|
586 |
|
|
/* Vector subtraction, treating overflow by signed saturation. */
|
587 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
588 |
|
|
psubsh (int16x4_t s, int16x4_t t)
|
589 |
|
|
{
|
590 |
|
|
return __builtin_loongson_psubsh (s, t);
|
591 |
|
|
}
|
592 |
|
|
|
593 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
594 |
|
|
psubsb (int8x8_t s, int8x8_t t)
|
595 |
|
|
{
|
596 |
|
|
return __builtin_loongson_psubsb (s, t);
|
597 |
|
|
}
|
598 |
|
|
|
599 |
|
|
/* Vector subtraction, treating overflow by unsigned saturation. */
|
600 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
601 |
|
|
psubush (uint16x4_t s, uint16x4_t t)
|
602 |
|
|
{
|
603 |
|
|
return __builtin_loongson_psubush (s, t);
|
604 |
|
|
}
|
605 |
|
|
|
606 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
607 |
|
|
psubusb (uint8x8_t s, uint8x8_t t)
|
608 |
|
|
{
|
609 |
|
|
return __builtin_loongson_psubusb (s, t);
|
610 |
|
|
}
|
611 |
|
|
|
612 |
|
|
/* Unpack high data. */
|
613 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
614 |
|
|
punpckhwd_u (uint32x2_t s, uint32x2_t t)
|
615 |
|
|
{
|
616 |
|
|
return __builtin_loongson_punpckhwd_u (s, t);
|
617 |
|
|
}
|
618 |
|
|
|
619 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
620 |
|
|
punpckhhw_u (uint16x4_t s, uint16x4_t t)
|
621 |
|
|
{
|
622 |
|
|
return __builtin_loongson_punpckhhw_u (s, t);
|
623 |
|
|
}
|
624 |
|
|
|
625 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
626 |
|
|
punpckhbh_u (uint8x8_t s, uint8x8_t t)
|
627 |
|
|
{
|
628 |
|
|
return __builtin_loongson_punpckhbh_u (s, t);
|
629 |
|
|
}
|
630 |
|
|
|
631 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
632 |
|
|
punpckhwd_s (int32x2_t s, int32x2_t t)
|
633 |
|
|
{
|
634 |
|
|
return __builtin_loongson_punpckhwd_s (s, t);
|
635 |
|
|
}
|
636 |
|
|
|
637 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
638 |
|
|
punpckhhw_s (int16x4_t s, int16x4_t t)
|
639 |
|
|
{
|
640 |
|
|
return __builtin_loongson_punpckhhw_s (s, t);
|
641 |
|
|
}
|
642 |
|
|
|
643 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
644 |
|
|
punpckhbh_s (int8x8_t s, int8x8_t t)
|
645 |
|
|
{
|
646 |
|
|
return __builtin_loongson_punpckhbh_s (s, t);
|
647 |
|
|
}
|
648 |
|
|
|
649 |
|
|
/* Unpack low data. */
|
650 |
|
|
__extension__ static __inline uint32x2_t __attribute__ ((__always_inline__))
|
651 |
|
|
punpcklwd_u (uint32x2_t s, uint32x2_t t)
|
652 |
|
|
{
|
653 |
|
|
return __builtin_loongson_punpcklwd_u (s, t);
|
654 |
|
|
}
|
655 |
|
|
|
656 |
|
|
__extension__ static __inline uint16x4_t __attribute__ ((__always_inline__))
|
657 |
|
|
punpcklhw_u (uint16x4_t s, uint16x4_t t)
|
658 |
|
|
{
|
659 |
|
|
return __builtin_loongson_punpcklhw_u (s, t);
|
660 |
|
|
}
|
661 |
|
|
|
662 |
|
|
__extension__ static __inline uint8x8_t __attribute__ ((__always_inline__))
|
663 |
|
|
punpcklbh_u (uint8x8_t s, uint8x8_t t)
|
664 |
|
|
{
|
665 |
|
|
return __builtin_loongson_punpcklbh_u (s, t);
|
666 |
|
|
}
|
667 |
|
|
|
668 |
|
|
__extension__ static __inline int32x2_t __attribute__ ((__always_inline__))
|
669 |
|
|
punpcklwd_s (int32x2_t s, int32x2_t t)
|
670 |
|
|
{
|
671 |
|
|
return __builtin_loongson_punpcklwd_s (s, t);
|
672 |
|
|
}
|
673 |
|
|
|
674 |
|
|
__extension__ static __inline int16x4_t __attribute__ ((__always_inline__))
|
675 |
|
|
punpcklhw_s (int16x4_t s, int16x4_t t)
|
676 |
|
|
{
|
677 |
|
|
return __builtin_loongson_punpcklhw_s (s, t);
|
678 |
|
|
}
|
679 |
|
|
|
680 |
|
|
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
|
681 |
|
|
punpcklbh_s (int8x8_t s, int8x8_t t)
|
682 |
|
|
{
|
683 |
|
|
return __builtin_loongson_punpcklbh_s (s, t);
|
684 |
|
|
}
|
685 |
|
|
|
686 |
|
|
#ifdef __cplusplus
|
687 |
|
|
}
|
688 |
|
|
#endif
|
689 |
|
|
|
690 |
|
|
#endif
|