1 /*===---- emmintrin.h - SSE2 intrinsics ------------------------------------===
3 * Permission is hereby granted, free of charge, to any person obtaining a copy
4 * of this software and associated documentation files (the "Software"), to deal
5 * in the Software without restriction, including without limitation the rights
6 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7 * copies of the Software, and to permit persons to whom the Software is
8 * furnished to do so, subject to the following conditions:
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
21 *===-----------------------------------------------------------------------===
28 #error "SSE2 instruction set not enabled"
31 #include <xmmintrin.h>
33 typedef double __m128d __attribute__((__vector_size__(16)));
34 typedef long long __m128i __attribute__((__vector_size__(16)));
37 typedef double __v2df __attribute__ ((__vector_size__ (16)));
38 typedef long long __v2di __attribute__ ((__vector_size__ (16)));
39 typedef short __v8hi __attribute__((__vector_size__(16)));
40 typedef char __v16qi __attribute__((__vector_size__(16)));
42 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
43 _mm_add_sd(__m128d a, __m128d b)
49 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
50 _mm_add_pd(__m128d a, __m128d b)
55 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
56 _mm_sub_sd(__m128d a, __m128d b)
62 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
63 _mm_sub_pd(__m128d a, __m128d b)
68 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
69 _mm_mul_sd(__m128d a, __m128d b)
75 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
76 _mm_mul_pd(__m128d a, __m128d b)
81 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
82 _mm_div_sd(__m128d a, __m128d b)
88 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
89 _mm_div_pd(__m128d a, __m128d b)
94 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
95 _mm_sqrt_sd(__m128d a, __m128d b)
97 __m128d c = __builtin_ia32_sqrtsd(b);
98 return (__m128d) { c[0], a[1] };
101 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
102 _mm_sqrt_pd(__m128d a)
104 return __builtin_ia32_sqrtpd(a);
107 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
108 _mm_min_sd(__m128d a, __m128d b)
110 return __builtin_ia32_minsd(a, b);
113 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
114 _mm_min_pd(__m128d a, __m128d b)
116 return __builtin_ia32_minpd(a, b);
119 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
120 _mm_max_sd(__m128d a, __m128d b)
122 return __builtin_ia32_maxsd(a, b);
125 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
126 _mm_max_pd(__m128d a, __m128d b)
128 return __builtin_ia32_maxpd(a, b);
131 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
132 _mm_and_pd(__m128d a, __m128d b)
134 return (__m128d)((__v4si)a & (__v4si)b);
137 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
138 _mm_andnot_pd(__m128d a, __m128d b)
140 return (__m128d)(~(__v4si)a & (__v4si)b);
143 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
144 _mm_or_pd(__m128d a, __m128d b)
146 return (__m128d)((__v4si)a | (__v4si)b);
149 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
150 _mm_xor_pd(__m128d a, __m128d b)
152 return (__m128d)((__v4si)a ^ (__v4si)b);
155 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
156 _mm_cmpeq_pd(__m128d a, __m128d b)
158 return (__m128d)__builtin_ia32_cmppd(a, b, 0);
161 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
162 _mm_cmplt_pd(__m128d a, __m128d b)
164 return (__m128d)__builtin_ia32_cmppd(a, b, 1);
167 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
168 _mm_cmple_pd(__m128d a, __m128d b)
170 return (__m128d)__builtin_ia32_cmppd(a, b, 2);
173 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
174 _mm_cmpgt_pd(__m128d a, __m128d b)
176 return (__m128d)__builtin_ia32_cmppd(b, a, 1);
179 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
180 _mm_cmpge_pd(__m128d a, __m128d b)
182 return (__m128d)__builtin_ia32_cmppd(b, a, 2);
185 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
186 _mm_cmpord_pd(__m128d a, __m128d b)
188 return (__m128d)__builtin_ia32_cmppd(a, b, 7);
191 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
192 _mm_cmpunord_pd(__m128d a, __m128d b)
194 return (__m128d)__builtin_ia32_cmppd(a, b, 3);
197 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
198 _mm_cmpneq_pd(__m128d a, __m128d b)
200 return (__m128d)__builtin_ia32_cmppd(a, b, 4);
203 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
204 _mm_cmpnlt_pd(__m128d a, __m128d b)
206 return (__m128d)__builtin_ia32_cmppd(a, b, 5);
209 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
210 _mm_cmpnle_pd(__m128d a, __m128d b)
212 return (__m128d)__builtin_ia32_cmppd(a, b, 6);
215 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
216 _mm_cmpngt_pd(__m128d a, __m128d b)
218 return (__m128d)__builtin_ia32_cmppd(b, a, 5);
221 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
222 _mm_cmpnge_pd(__m128d a, __m128d b)
224 return (__m128d)__builtin_ia32_cmppd(b, a, 6);
227 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
228 _mm_cmpeq_sd(__m128d a, __m128d b)
230 return (__m128d)__builtin_ia32_cmpsd(a, b, 0);
233 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
234 _mm_cmplt_sd(__m128d a, __m128d b)
236 return (__m128d)__builtin_ia32_cmpsd(a, b, 1);
239 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
240 _mm_cmple_sd(__m128d a, __m128d b)
242 return (__m128d)__builtin_ia32_cmpsd(a, b, 2);
245 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
246 _mm_cmpgt_sd(__m128d a, __m128d b)
248 return (__m128d)__builtin_ia32_cmpsd(b, a, 1);
251 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
252 _mm_cmpge_sd(__m128d a, __m128d b)
254 return (__m128d)__builtin_ia32_cmpsd(b, a, 2);
257 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
258 _mm_cmpord_sd(__m128d a, __m128d b)
260 return (__m128d)__builtin_ia32_cmpsd(a, b, 7);
263 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
264 _mm_cmpunord_sd(__m128d a, __m128d b)
266 return (__m128d)__builtin_ia32_cmpsd(a, b, 3);
269 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
270 _mm_cmpneq_sd(__m128d a, __m128d b)
272 return (__m128d)__builtin_ia32_cmpsd(a, b, 4);
275 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
276 _mm_cmpnlt_sd(__m128d a, __m128d b)
278 return (__m128d)__builtin_ia32_cmpsd(a, b, 5);
281 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
282 _mm_cmpnle_sd(__m128d a, __m128d b)
284 return (__m128d)__builtin_ia32_cmpsd(a, b, 6);
287 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
288 _mm_cmpngt_sd(__m128d a, __m128d b)
290 return (__m128d)__builtin_ia32_cmpsd(b, a, 5);
293 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
294 _mm_cmpnge_sd(__m128d a, __m128d b)
296 return (__m128d)__builtin_ia32_cmpsd(b, a, 6);
299 static __inline__ int __attribute__((__always_inline__, __nodebug__))
300 _mm_comieq_sd(__m128d a, __m128d b)
302 return __builtin_ia32_comisdeq(a, b);
305 static __inline__ int __attribute__((__always_inline__, __nodebug__))
306 _mm_comilt_sd(__m128d a, __m128d b)
308 return __builtin_ia32_comisdlt(a, b);
311 static __inline__ int __attribute__((__always_inline__, __nodebug__))
312 _mm_comile_sd(__m128d a, __m128d b)
314 return __builtin_ia32_comisdle(a, b);
317 static __inline__ int __attribute__((__always_inline__, __nodebug__))
318 _mm_comigt_sd(__m128d a, __m128d b)
320 return __builtin_ia32_comisdgt(a, b);
323 static __inline__ int __attribute__((__always_inline__, __nodebug__))
324 _mm_comineq_sd(__m128d a, __m128d b)
326 return __builtin_ia32_comisdneq(a, b);
329 static __inline__ int __attribute__((__always_inline__, __nodebug__))
330 _mm_ucomieq_sd(__m128d a, __m128d b)
332 return __builtin_ia32_ucomisdeq(a, b);
335 static __inline__ int __attribute__((__always_inline__, __nodebug__))
336 _mm_ucomilt_sd(__m128d a, __m128d b)
338 return __builtin_ia32_ucomisdlt(a, b);
341 static __inline__ int __attribute__((__always_inline__, __nodebug__))
342 _mm_ucomile_sd(__m128d a, __m128d b)
344 return __builtin_ia32_ucomisdle(a, b);
347 static __inline__ int __attribute__((__always_inline__, __nodebug__))
348 _mm_ucomigt_sd(__m128d a, __m128d b)
350 return __builtin_ia32_ucomisdgt(a, b);
353 static __inline__ int __attribute__((__always_inline__, __nodebug__))
354 _mm_ucomineq_sd(__m128d a, __m128d b)
356 return __builtin_ia32_ucomisdneq(a, b);
359 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
360 _mm_cvtpd_ps(__m128d a)
362 return __builtin_ia32_cvtpd2ps(a);
365 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
366 _mm_cvtps_pd(__m128 a)
368 return __builtin_ia32_cvtps2pd(a);
371 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
372 _mm_cvtepi32_pd(__m128i a)
374 return __builtin_ia32_cvtdq2pd((__v4si)a);
377 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
378 _mm_cvtpd_epi32(__m128d a)
380 return __builtin_ia32_cvtpd2dq(a);
383 static __inline__ int __attribute__((__always_inline__, __nodebug__))
384 _mm_cvtsd_si32(__m128d a)
386 return __builtin_ia32_cvtsd2si(a);
389 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
390 _mm_cvtsd_ss(__m128 a, __m128d b)
396 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
397 _mm_cvtsi32_sd(__m128d a, int b)
403 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
404 _mm_cvtss_sd(__m128d a, __m128 b)
410 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
411 _mm_cvttpd_epi32(__m128d a)
413 return (__m128i)__builtin_ia32_cvttpd2dq(a);
416 static __inline__ int __attribute__((__always_inline__, __nodebug__))
417 _mm_cvttsd_si32(__m128d a)
422 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
423 _mm_cvtpd_pi32(__m128d a)
425 return (__m64)__builtin_ia32_cvtpd2pi(a);
428 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
429 _mm_cvttpd_pi32(__m128d a)
431 return (__m64)__builtin_ia32_cvttpd2pi(a);
434 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
435 _mm_cvtpi32_pd(__m64 a)
437 return __builtin_ia32_cvtpi2pd((__v2si)a);
440 static __inline__ double __attribute__((__always_inline__, __nodebug__))
441 _mm_cvtsd_f64(__m128d a)
446 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
447 _mm_load_pd(double const *dp)
449 return *(__m128d*)dp;
452 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
453 _mm_load1_pd(double const *dp)
455 return (__m128d){ dp[0], dp[0] };
458 #define _mm_load_pd1(dp) _mm_load1_pd(dp)
460 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
461 _mm_loadr_pd(double const *dp)
463 return (__m128d){ dp[1], dp[0] };
466 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
467 _mm_loadu_pd(double const *dp)
469 return __builtin_ia32_loadupd(dp);
472 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
473 _mm_load_sd(double const *dp)
475 return (__m128d){ *dp, 0.0 };
478 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
479 _mm_loadh_pd(__m128d a, double const *dp)
481 return __builtin_shufflevector(a, *(__m128d *)dp, 0, 2);
484 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
485 _mm_loadl_pd(__m128d a, double const *dp)
487 return __builtin_shufflevector(a, *(__m128d *)dp, 2, 1);
490 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
493 return (__m128d){ w, 0 };
496 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
497 _mm_set1_pd(double w)
499 return (__m128d){ w, w };
502 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
503 _mm_set_pd(double w, double x)
505 return (__m128d){ x, w };
508 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
509 _mm_setr_pd(double w, double x)
511 return (__m128d){ w, x };
514 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
517 return (__m128d){ 0, 0 };
520 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
521 _mm_move_sd(__m128d a, __m128d b)
523 return (__m128d){ b[0], a[1] };
526 static __inline__ void __attribute__((__always_inline__, __nodebug__))
527 _mm_store_sd(double *dp, __m128d a)
532 static __inline__ void __attribute__((__always_inline__, __nodebug__))
533 _mm_store1_pd(double *dp, __m128d a)
539 static __inline__ void __attribute__((__always_inline__, __nodebug__))
540 _mm_store_pd(double *dp, __m128d a)
545 static __inline__ void __attribute__((__always_inline__, __nodebug__))
546 _mm_storeu_pd(double *dp, __m128d a)
548 __builtin_ia32_storeupd(dp, a);
551 static __inline__ void __attribute__((__always_inline__, __nodebug__))
552 _mm_storer_pd(double *dp, __m128d a)
558 static __inline__ void __attribute__((__always_inline__, __nodebug__))
559 _mm_storeh_pd(double *dp, __m128d a)
564 static __inline__ void __attribute__((__always_inline__, __nodebug__))
565 _mm_storel_pd(double *dp, __m128d a)
570 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
571 _mm_add_epi8(__m128i a, __m128i b)
573 return (__m128i)((__v16qi)a + (__v16qi)b);
576 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
577 _mm_add_epi16(__m128i a, __m128i b)
579 return (__m128i)((__v8hi)a + (__v8hi)b);
582 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
583 _mm_add_epi32(__m128i a, __m128i b)
585 return (__m128i)((__v4si)a + (__v4si)b);
588 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
589 _mm_add_si64(__m64 a, __m64 b)
594 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
595 _mm_add_epi64(__m128i a, __m128i b)
600 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
601 _mm_adds_epi8(__m128i a, __m128i b)
603 return (__m128i)__builtin_ia32_paddsb128((__v16qi)a, (__v16qi)b);
606 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
607 _mm_adds_epi16(__m128i a, __m128i b)
609 return (__m128i)__builtin_ia32_paddsw128((__v8hi)a, (__v8hi)b);
612 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
613 _mm_adds_epu8(__m128i a, __m128i b)
615 return (__m128i)__builtin_ia32_paddusb128((__v16qi)a, (__v16qi)b);
618 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
619 _mm_adds_epu16(__m128i a, __m128i b)
621 return (__m128i)__builtin_ia32_paddusw128((__v8hi)a, (__v8hi)b);
624 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
625 _mm_avg_epu8(__m128i a, __m128i b)
627 return (__m128i)__builtin_ia32_pavgb128((__v16qi)a, (__v16qi)b);
630 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
631 _mm_avg_epu16(__m128i a, __m128i b)
633 return (__m128i)__builtin_ia32_pavgw128((__v8hi)a, (__v8hi)b);
636 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
637 _mm_madd_epi16(__m128i a, __m128i b)
639 return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)a, (__v8hi)b);
642 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
643 _mm_max_epi16(__m128i a, __m128i b)
645 return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)a, (__v8hi)b);
648 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
649 _mm_max_epu8(__m128i a, __m128i b)
651 return (__m128i)__builtin_ia32_pmaxub128((__v16qi)a, (__v16qi)b);
654 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
655 _mm_min_epi16(__m128i a, __m128i b)
657 return (__m128i)__builtin_ia32_pminsw128((__v8hi)a, (__v8hi)b);
660 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
661 _mm_min_epu8(__m128i a, __m128i b)
663 return (__m128i)__builtin_ia32_pminub128((__v16qi)a, (__v16qi)b);
666 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
667 _mm_mulhi_epi16(__m128i a, __m128i b)
669 return (__m128i)__builtin_ia32_pmulhw128((__v8hi)a, (__v8hi)b);
672 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
673 _mm_mulhi_epu16(__m128i a, __m128i b)
675 return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)a, (__v8hi)b);
678 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
679 _mm_mullo_epi16(__m128i a, __m128i b)
681 return (__m128i)((__v8hi)a * (__v8hi)b);
684 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
685 _mm_mul_su32(__m64 a, __m64 b)
687 return __builtin_ia32_pmuludq((__v2si)a, (__v2si)b);
690 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
691 _mm_mul_epu32(__m128i a, __m128i b)
693 return __builtin_ia32_pmuludq128((__v4si)a, (__v4si)b);
696 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
697 _mm_sad_epu8(__m128i a, __m128i b)
699 return __builtin_ia32_psadbw128((__v16qi)a, (__v16qi)b);
702 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
703 _mm_sub_epi8(__m128i a, __m128i b)
705 return (__m128i)((__v16qi)a - (__v16qi)b);
708 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
709 _mm_sub_epi16(__m128i a, __m128i b)
711 return (__m128i)((__v8hi)a - (__v8hi)b);
714 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
715 _mm_sub_epi32(__m128i a, __m128i b)
717 return (__m128i)((__v4si)a - (__v4si)b);
720 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
721 _mm_sub_si64(__m64 a, __m64 b)
726 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
727 _mm_sub_epi64(__m128i a, __m128i b)
732 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
733 _mm_subs_epi8(__m128i a, __m128i b)
735 return (__m128i)__builtin_ia32_psubsb128((__v16qi)a, (__v16qi)b);
738 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
739 _mm_subs_epi16(__m128i a, __m128i b)
741 return (__m128i)__builtin_ia32_psubsw128((__v8hi)a, (__v8hi)b);
744 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
745 _mm_subs_epu8(__m128i a, __m128i b)
747 return (__m128i)__builtin_ia32_psubusb128((__v16qi)a, (__v16qi)b);
750 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
751 _mm_subs_epu16(__m128i a, __m128i b)
753 return (__m128i)__builtin_ia32_psubusw128((__v8hi)a, (__v8hi)b);
756 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
757 _mm_and_si128(__m128i a, __m128i b)
762 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
763 _mm_andnot_si128(__m128i a, __m128i b)
768 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
769 _mm_or_si128(__m128i a, __m128i b)
774 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
775 _mm_xor_si128(__m128i a, __m128i b)
780 #define _mm_slli_si128(VEC, IMM) \
781 ((__m128i)__builtin_ia32_pslldqi128((__m128i)(VEC), (IMM)*8))
783 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
784 _mm_slli_epi16(__m128i a, int count)
786 return (__m128i)__builtin_ia32_psllwi128((__v8hi)a, count);
789 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
790 _mm_sll_epi16(__m128i a, __m128i count)
792 return (__m128i)__builtin_ia32_psllw128((__v8hi)a, (__v8hi)count);
795 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
796 _mm_slli_epi32(__m128i a, int count)
798 return (__m128i)__builtin_ia32_pslldi128((__v4si)a, count);
801 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
802 _mm_sll_epi32(__m128i a, __m128i count)
804 return (__m128i)__builtin_ia32_pslld128((__v4si)a, (__v4si)count);
807 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
808 _mm_slli_epi64(__m128i a, int count)
810 return __builtin_ia32_psllqi128(a, count);
813 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
814 _mm_sll_epi64(__m128i a, __m128i count)
816 return __builtin_ia32_psllq128(a, count);
819 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
820 _mm_srai_epi16(__m128i a, int count)
822 return (__m128i)__builtin_ia32_psrawi128((__v8hi)a, count);
825 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
826 _mm_sra_epi16(__m128i a, __m128i count)
828 return (__m128i)__builtin_ia32_psraw128((__v8hi)a, (__v8hi)count);
831 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
832 _mm_srai_epi32(__m128i a, int count)
834 return (__m128i)__builtin_ia32_psradi128((__v4si)a, count);
837 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
838 _mm_sra_epi32(__m128i a, __m128i count)
840 return (__m128i)__builtin_ia32_psrad128((__v4si)a, (__v4si)count);
844 #define _mm_srli_si128(VEC, IMM) \
845 ((__m128i)__builtin_ia32_psrldqi128((__m128i)(VEC), (IMM)*8))
847 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
848 _mm_srli_epi16(__m128i a, int count)
850 return (__m128i)__builtin_ia32_psrlwi128((__v8hi)a, count);
853 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
854 _mm_srl_epi16(__m128i a, __m128i count)
856 return (__m128i)__builtin_ia32_psrlw128((__v8hi)a, (__v8hi)count);
859 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
860 _mm_srli_epi32(__m128i a, int count)
862 return (__m128i)__builtin_ia32_psrldi128((__v4si)a, count);
865 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
866 _mm_srl_epi32(__m128i a, __m128i count)
868 return (__m128i)__builtin_ia32_psrld128((__v4si)a, (__v4si)count);
871 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
872 _mm_srli_epi64(__m128i a, int count)
874 return __builtin_ia32_psrlqi128(a, count);
877 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
878 _mm_srl_epi64(__m128i a, __m128i count)
880 return __builtin_ia32_psrlq128(a, count);
883 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
884 _mm_cmpeq_epi8(__m128i a, __m128i b)
886 return (__m128i)((__v16qi)a == (__v16qi)b);
889 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
890 _mm_cmpeq_epi16(__m128i a, __m128i b)
892 return (__m128i)((__v8hi)a == (__v8hi)b);
895 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
896 _mm_cmpeq_epi32(__m128i a, __m128i b)
898 return (__m128i)((__v4si)a == (__v4si)b);
901 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
902 _mm_cmpgt_epi8(__m128i a, __m128i b)
904 return (__m128i)((__v16qi)a > (__v16qi)b);
907 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
908 _mm_cmpgt_epi16(__m128i a, __m128i b)
910 return (__m128i)((__v8hi)a > (__v8hi)b);
913 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
914 _mm_cmpgt_epi32(__m128i a, __m128i b)
916 return (__m128i)((__v4si)a > (__v4si)b);
919 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
920 _mm_cmplt_epi8(__m128i a, __m128i b)
922 return _mm_cmpgt_epi8(b,a);
925 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
926 _mm_cmplt_epi16(__m128i a, __m128i b)
928 return _mm_cmpgt_epi16(b,a);
931 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
932 _mm_cmplt_epi32(__m128i a, __m128i b)
934 return _mm_cmpgt_epi32(b,a);
938 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
939 _mm_cvtsi64_sd(__m128d a, long long b)
945 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
946 _mm_cvtsd_si64(__m128d a)
948 return __builtin_ia32_cvtsd2si64(a);
951 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
952 _mm_cvttsd_si64(__m128d a)
958 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
959 _mm_cvtepi32_ps(__m128i a)
961 return __builtin_ia32_cvtdq2ps((__v4si)a);
964 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
965 _mm_cvtps_epi32(__m128 a)
967 return (__m128i)__builtin_ia32_cvtps2dq(a);
970 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
971 _mm_cvttps_epi32(__m128 a)
973 return (__m128i)__builtin_ia32_cvttps2dq(a);
976 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
977 _mm_cvtsi32_si128(int a)
979 return (__m128i)(__v4si){ a, 0, 0, 0 };
983 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
984 _mm_cvtsi64_si128(long long a)
986 return (__m128i){ a, 0 };
990 static __inline__ int __attribute__((__always_inline__, __nodebug__))
991 _mm_cvtsi128_si32(__m128i a)
993 __v4si b = (__v4si)a;
998 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
999 _mm_cvtsi128_si64(__m128i a)
1005 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1006 _mm_load_si128(__m128i const *p)
1011 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1012 _mm_loadu_si128(__m128i const *p)
1014 return (__m128i)__builtin_ia32_loaddqu((char const *)p);
1017 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1018 _mm_loadl_epi64(__m128i const *p)
1020 return (__m128i) { *(long long*)p, 0};
1023 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1024 _mm_set_epi64x(long long q1, long long q0)
1026 return (__m128i){ q0, q1 };
1029 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1030 _mm_set_epi64(__m64 q1, __m64 q0)
1032 return (__m128i){ (long long)q0, (long long)q1 };
1035 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1036 _mm_set_epi32(int i3, int i2, int i1, int i0)
1038 return (__m128i)(__v4si){ i0, i1, i2, i3};
1041 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1042 _mm_set_epi16(short w7, short w6, short w5, short w4, short w3, short w2, short w1, short w0)
1044 return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 };
1047 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1048 _mm_set_epi8(char b15, char b14, char b13, char b12, char b11, char b10, char b9, char b8, char b7, char b6, char b5, char b4, char b3, char b2, char b1, char b0)
1050 return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 };
1053 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1054 _mm_set1_epi64x(long long q)
1056 return (__m128i){ q, q };
1059 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1060 _mm_set1_epi64(__m64 q)
1062 return (__m128i){ (long long)q, (long long)q };
1065 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1066 _mm_set1_epi32(int i)
1068 return (__m128i)(__v4si){ i, i, i, i };
1071 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1072 _mm_set1_epi16(short w)
1074 return (__m128i)(__v8hi){ w, w, w, w, w, w, w, w };
1077 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1078 _mm_set1_epi8(char b)
1080 return (__m128i)(__v16qi){ b, b, b, b, b, b, b, b, b, b, b, b, b, b, b, b };
1083 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1084 _mm_setr_epi64(__m64 q0, __m64 q1)
1086 return (__m128i){ (long long)q0, (long long)q1 };
1089 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1090 _mm_setr_epi32(int i0, int i1, int i2, int i3)
1092 return (__m128i)(__v4si){ i0, i1, i2, i3};
1095 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1096 _mm_setr_epi16(short w0, short w1, short w2, short w3, short w4, short w5, short w6, short w7)
1098 return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 };
1101 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1102 _mm_setr_epi8(char b0, char b1, char b2, char b3, char b4, char b5, char b6, char b7, char b8, char b9, char b10, char b11, char b12, char b13, char b14, char b15)
1104 return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 };
1107 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1108 _mm_setzero_si128(void)
1110 return (__m128i){ 0LL, 0LL };
1113 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1114 _mm_store_si128(__m128i *p, __m128i b)
1119 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1120 _mm_storeu_si128(__m128i *p, __m128i b)
1122 __builtin_ia32_storedqu((char *)p, (__v16qi)b);
1125 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1126 _mm_maskmoveu_si128(__m128i d, __m128i n, char *p)
1128 __builtin_ia32_maskmovdqu((__v16qi)d, (__v16qi)n, p);
1131 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1132 _mm_storel_epi64(__m128i *p, __m128i a)
1134 __builtin_ia32_storelv4si((__v2si *)p, a);
1137 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1138 _mm_stream_pd(double *p, __m128d a)
1140 __builtin_ia32_movntpd(p, a);
1143 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1144 _mm_stream_si128(__m128i *p, __m128i a)
1146 __builtin_ia32_movntdq(p, a);
1149 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1150 _mm_stream_si32(int *p, int a)
1152 __builtin_ia32_movnti(p, a);
1155 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1156 _mm_clflush(void const *p)
1158 __builtin_ia32_clflush(p);
1161 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1164 __builtin_ia32_lfence();
1167 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1170 __builtin_ia32_mfence();
1173 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1174 _mm_packs_epi16(__m128i a, __m128i b)
1176 return (__m128i)__builtin_ia32_packsswb128((__v8hi)a, (__v8hi)b);
1179 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1180 _mm_packs_epi32(__m128i a, __m128i b)
1182 return (__m128i)__builtin_ia32_packssdw128((__v4si)a, (__v4si)b);
1185 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1186 _mm_packus_epi16(__m128i a, __m128i b)
1188 return (__m128i)__builtin_ia32_packuswb128((__v8hi)a, (__v8hi)b);
1191 static __inline__ int __attribute__((__always_inline__, __nodebug__))
1192 _mm_extract_epi16(__m128i a, int imm)
1194 __v8hi b = (__v8hi)a;
1195 return (unsigned short)b[imm];
1198 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1199 _mm_insert_epi16(__m128i a, int b, int imm)
1201 __v8hi c = (__v8hi)a;
1206 static __inline__ int __attribute__((__always_inline__, __nodebug__))
1207 _mm_movemask_epi8(__m128i a)
1209 return __builtin_ia32_pmovmskb128((__v16qi)a);
1212 #define _mm_shuffle_epi32(a, imm) \
1213 ((__m128i)__builtin_shufflevector((__v4si)(a), (__v4si) {0}, \
1214 (imm) & 0x3, ((imm) & 0xc) >> 2, \
1215 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6))
1216 #define _mm_shufflelo_epi16(a, imm) \
1217 ((__m128i)__builtin_shufflevector((__v8hi)(a), (__v8hi) {0}, \
1218 (imm) & 0x3, ((imm) & 0xc) >> 2, \
1219 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
1221 #define _mm_shufflehi_epi16(a, imm) \
1222 ((__m128i)__builtin_shufflevector((__v8hi)(a), (__v8hi) {0}, 0, 1, 2, 3, \
1223 4 + (((imm) & 0x03) >> 0), \
1224 4 + (((imm) & 0x0c) >> 2), \
1225 4 + (((imm) & 0x30) >> 4), \
1226 4 + (((imm) & 0xc0) >> 6)))
1228 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1229 _mm_unpackhi_epi8(__m128i a, __m128i b)
1231 return (__m128i)__builtin_shufflevector((__v16qi)a, (__v16qi)b, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15);
1234 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1235 _mm_unpackhi_epi16(__m128i a, __m128i b)
1237 return (__m128i)__builtin_shufflevector((__v8hi)a, (__v8hi)b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7);
1240 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1241 _mm_unpackhi_epi32(__m128i a, __m128i b)
1243 return (__m128i)__builtin_shufflevector((__v4si)a, (__v4si)b, 2, 4+2, 3, 4+3);
1246 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1247 _mm_unpackhi_epi64(__m128i a, __m128i b)
1249 return (__m128i)__builtin_shufflevector(a, b, 1, 2+1);
1252 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1253 _mm_unpacklo_epi8(__m128i a, __m128i b)
1255 return (__m128i)__builtin_shufflevector((__v16qi)a, (__v16qi)b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7);
1258 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1259 _mm_unpacklo_epi16(__m128i a, __m128i b)
1261 return (__m128i)__builtin_shufflevector((__v8hi)a, (__v8hi)b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3);
1264 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1265 _mm_unpacklo_epi32(__m128i a, __m128i b)
1267 return (__m128i)__builtin_shufflevector((__v4si)a, (__v4si)b, 0, 4+0, 1, 4+1);
1270 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1271 _mm_unpacklo_epi64(__m128i a, __m128i b)
1273 return (__m128i)__builtin_shufflevector(a, b, 0, 2+0);
1276 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
1277 _mm_movepi64_pi64(__m128i a)
1282 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1283 _mm_movpi64_pi64(__m64 a)
1285 return (__m128i){ (long long)a, 0 };
1288 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1289 _mm_move_epi64(__m128i a)
1291 return __builtin_shufflevector(a, (__m128i){ 0 }, 0, 2);
1294 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1295 _mm_unpackhi_pd(__m128d a, __m128d b)
1297 return __builtin_shufflevector(a, b, 1, 2+1);
1300 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1301 _mm_unpacklo_pd(__m128d a, __m128d b)
1303 return __builtin_shufflevector(a, b, 0, 2+0);
1306 static __inline__ int __attribute__((__always_inline__, __nodebug__))
1307 _mm_movemask_pd(__m128d a)
1309 return __builtin_ia32_movmskpd(a);
1312 #define _mm_shuffle_pd(a, b, i) \
1313 (__builtin_shufflevector((__m128d)(a), (__m128d)(b), (i) & 1, \
1314 (((i) & 2) >> 1) + 2))
1316 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
1317 _mm_castpd_ps(__m128d in)
1322 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1323 _mm_castpd_si128(__m128d in)
1328 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1329 _mm_castps_pd(__m128 in)
1334 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1335 _mm_castps_si128(__m128 in)
1340 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
1341 _mm_castsi128_ps(__m128i in)
1346 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1347 _mm_castsi128_pd(__m128i in)
1352 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1355 __asm__ volatile ("pause");
1358 #define _MM_SHUFFLE2(x, y) (((x) << 1) | (y))
1360 #endif /* __SSE2__ */
1362 #endif /* __EMMINTRIN_H */