1 /*===---- xmmintrin.h - SSE 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 "SSE instruction set not enabled"
33 typedef int __v4si __attribute__((__vector_size__(16)));
34 typedef float __v4sf __attribute__((__vector_size__(16)));
35 typedef float __m128 __attribute__((__vector_size__(16)));
37 // This header should only be included in a hosted environment as it depends on
38 // a standard library to provide allocation routines.
40 #include <mm_malloc.h>
43 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
44 _mm_add_ss(__m128 a, __m128 b)
50 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
51 _mm_add_ps(__m128 a, __m128 b)
56 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
57 _mm_sub_ss(__m128 a, __m128 b)
63 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
64 _mm_sub_ps(__m128 a, __m128 b)
69 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
70 _mm_mul_ss(__m128 a, __m128 b)
76 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
77 _mm_mul_ps(__m128 a, __m128 b)
82 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
83 _mm_div_ss(__m128 a, __m128 b)
89 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
90 _mm_div_ps(__m128 a, __m128 b)
95 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
98 return __builtin_ia32_sqrtss(a);
101 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
102 _mm_sqrt_ps(__m128 a)
104 return __builtin_ia32_sqrtps(a);
107 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
110 return __builtin_ia32_rcpss(a);
113 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
116 return __builtin_ia32_rcpps(a);
119 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
120 _mm_rsqrt_ss(__m128 a)
122 return __builtin_ia32_rsqrtss(a);
125 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
126 _mm_rsqrt_ps(__m128 a)
128 return __builtin_ia32_rsqrtps(a);
131 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
132 _mm_min_ss(__m128 a, __m128 b)
134 return __builtin_ia32_minss(a, b);
137 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
138 _mm_min_ps(__m128 a, __m128 b)
140 return __builtin_ia32_minps(a, b);
143 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
144 _mm_max_ss(__m128 a, __m128 b)
146 return __builtin_ia32_maxss(a, b);
149 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
150 _mm_max_ps(__m128 a, __m128 b)
152 return __builtin_ia32_maxps(a, b);
155 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
156 _mm_and_ps(__m128 a, __m128 b)
158 return (__m128)((__v4si)a & (__v4si)b);
161 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
162 _mm_andnot_ps(__m128 a, __m128 b)
164 return (__m128)(~(__v4si)a & (__v4si)b);
167 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
168 _mm_or_ps(__m128 a, __m128 b)
170 return (__m128)((__v4si)a | (__v4si)b);
173 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
174 _mm_xor_ps(__m128 a, __m128 b)
176 return (__m128)((__v4si)a ^ (__v4si)b);
179 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
180 _mm_cmpeq_ss(__m128 a, __m128 b)
182 return (__m128)__builtin_ia32_cmpss(a, b, 0);
185 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
186 _mm_cmpeq_ps(__m128 a, __m128 b)
188 return (__m128)__builtin_ia32_cmpps(a, b, 0);
191 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
192 _mm_cmplt_ss(__m128 a, __m128 b)
194 return (__m128)__builtin_ia32_cmpss(a, b, 1);
197 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
198 _mm_cmplt_ps(__m128 a, __m128 b)
200 return (__m128)__builtin_ia32_cmpps(a, b, 1);
203 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
204 _mm_cmple_ss(__m128 a, __m128 b)
206 return (__m128)__builtin_ia32_cmpss(a, b, 2);
209 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
210 _mm_cmple_ps(__m128 a, __m128 b)
212 return (__m128)__builtin_ia32_cmpps(a, b, 2);
215 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
216 _mm_cmpgt_ss(__m128 a, __m128 b)
218 return (__m128)__builtin_ia32_cmpss(b, a, 1);
221 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
222 _mm_cmpgt_ps(__m128 a, __m128 b)
224 return (__m128)__builtin_ia32_cmpps(b, a, 1);
227 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
228 _mm_cmpge_ss(__m128 a, __m128 b)
230 return (__m128)__builtin_ia32_cmpss(b, a, 2);
233 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
234 _mm_cmpge_ps(__m128 a, __m128 b)
236 return (__m128)__builtin_ia32_cmpps(b, a, 2);
239 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
240 _mm_cmpneq_ss(__m128 a, __m128 b)
242 return (__m128)__builtin_ia32_cmpss(a, b, 4);
245 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
246 _mm_cmpneq_ps(__m128 a, __m128 b)
248 return (__m128)__builtin_ia32_cmpps(a, b, 4);
251 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
252 _mm_cmpnlt_ss(__m128 a, __m128 b)
254 return (__m128)__builtin_ia32_cmpss(a, b, 5);
257 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
258 _mm_cmpnlt_ps(__m128 a, __m128 b)
260 return (__m128)__builtin_ia32_cmpps(a, b, 5);
263 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
264 _mm_cmpnle_ss(__m128 a, __m128 b)
266 return (__m128)__builtin_ia32_cmpss(a, b, 6);
269 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
270 _mm_cmpnle_ps(__m128 a, __m128 b)
272 return (__m128)__builtin_ia32_cmpps(a, b, 6);
275 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
276 _mm_cmpngt_ss(__m128 a, __m128 b)
278 return (__m128)__builtin_ia32_cmpss(b, a, 5);
281 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
282 _mm_cmpngt_ps(__m128 a, __m128 b)
284 return (__m128)__builtin_ia32_cmpps(b, a, 5);
287 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
288 _mm_cmpnge_ss(__m128 a, __m128 b)
290 return (__m128)__builtin_ia32_cmpss(b, a, 6);
293 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
294 _mm_cmpnge_ps(__m128 a, __m128 b)
296 return (__m128)__builtin_ia32_cmpps(b, a, 6);
299 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
300 _mm_cmpord_ss(__m128 a, __m128 b)
302 return (__m128)__builtin_ia32_cmpss(a, b, 7);
305 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
306 _mm_cmpord_ps(__m128 a, __m128 b)
308 return (__m128)__builtin_ia32_cmpps(a, b, 7);
311 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
312 _mm_cmpunord_ss(__m128 a, __m128 b)
314 return (__m128)__builtin_ia32_cmpss(a, b, 3);
317 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
318 _mm_cmpunord_ps(__m128 a, __m128 b)
320 return (__m128)__builtin_ia32_cmpps(a, b, 3);
323 static __inline__ int __attribute__((__always_inline__, __nodebug__))
324 _mm_comieq_ss(__m128 a, __m128 b)
326 return __builtin_ia32_comieq(a, b);
329 static __inline__ int __attribute__((__always_inline__, __nodebug__))
330 _mm_comilt_ss(__m128 a, __m128 b)
332 return __builtin_ia32_comilt(a, b);
335 static __inline__ int __attribute__((__always_inline__, __nodebug__))
336 _mm_comile_ss(__m128 a, __m128 b)
338 return __builtin_ia32_comile(a, b);
341 static __inline__ int __attribute__((__always_inline__, __nodebug__))
342 _mm_comigt_ss(__m128 a, __m128 b)
344 return __builtin_ia32_comigt(a, b);
347 static __inline__ int __attribute__((__always_inline__, __nodebug__))
348 _mm_comige_ss(__m128 a, __m128 b)
350 return __builtin_ia32_comige(a, b);
353 static __inline__ int __attribute__((__always_inline__, __nodebug__))
354 _mm_comineq_ss(__m128 a, __m128 b)
356 return __builtin_ia32_comineq(a, b);
359 static __inline__ int __attribute__((__always_inline__, __nodebug__))
360 _mm_ucomieq_ss(__m128 a, __m128 b)
362 return __builtin_ia32_ucomieq(a, b);
365 static __inline__ int __attribute__((__always_inline__, __nodebug__))
366 _mm_ucomilt_ss(__m128 a, __m128 b)
368 return __builtin_ia32_ucomilt(a, b);
371 static __inline__ int __attribute__((__always_inline__, __nodebug__))
372 _mm_ucomile_ss(__m128 a, __m128 b)
374 return __builtin_ia32_ucomile(a, b);
377 static __inline__ int __attribute__((__always_inline__, __nodebug__))
378 _mm_ucomigt_ss(__m128 a, __m128 b)
380 return __builtin_ia32_ucomigt(a, b);
383 static __inline__ int __attribute__((__always_inline__, __nodebug__))
384 _mm_ucomige_ss(__m128 a, __m128 b)
386 return __builtin_ia32_ucomige(a, b);
389 static __inline__ int __attribute__((__always_inline__, __nodebug__))
390 _mm_ucomineq_ss(__m128 a, __m128 b)
392 return __builtin_ia32_ucomineq(a, b);
395 static __inline__ int __attribute__((__always_inline__, __nodebug__))
396 _mm_cvtss_si32(__m128 a)
398 return __builtin_ia32_cvtss2si(a);
401 static __inline__ int __attribute__((__always_inline__, __nodebug__))
402 _mm_cvt_ss2si(__m128 a)
404 return _mm_cvtss_si32(a);
409 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
410 _mm_cvtss_si64(__m128 a)
412 return __builtin_ia32_cvtss2si64(a);
417 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
418 _mm_cvtps_pi32(__m128 a)
420 return (__m64)__builtin_ia32_cvtps2pi(a);
423 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
424 _mm_cvt_ps2pi(__m128 a)
426 return _mm_cvtps_pi32(a);
429 static __inline__ int __attribute__((__always_inline__, __nodebug__))
430 _mm_cvttss_si32(__m128 a)
435 static __inline__ int __attribute__((__always_inline__, __nodebug__))
436 _mm_cvtt_ss2si(__m128 a)
438 return _mm_cvttss_si32(a);
441 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
442 _mm_cvttss_si64(__m128 a)
447 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
448 _mm_cvttps_pi32(__m128 a)
450 return (__m64)__builtin_ia32_cvttps2pi(a);
453 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
454 _mm_cvtt_ps2pi(__m128 a)
456 return _mm_cvttps_pi32(a);
459 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
460 _mm_cvtsi32_ss(__m128 a, int b)
466 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
467 _mm_cvt_si2ss(__m128 a, int b)
469 return _mm_cvtsi32_ss(a, b);
474 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
475 _mm_cvtsi64_ss(__m128 a, long long b)
483 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
484 _mm_cvtpi32_ps(__m128 a, __m64 b)
486 return __builtin_ia32_cvtpi2ps(a, (__v2si)b);
489 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
490 _mm_cvt_pi2ps(__m128 a, __m64 b)
492 return _mm_cvtpi32_ps(a, b);
495 static __inline__ float __attribute__((__always_inline__, __nodebug__))
496 _mm_cvtss_f32(__m128 a)
501 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
502 _mm_loadh_pi(__m128 a, const __m64 *p)
506 b[1] = *((float*)p+1);
507 return __builtin_shufflevector(a, b, 0, 1, 4, 5);
510 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
511 _mm_loadl_pi(__m128 a, const __m64 *p)
515 b[1] = *((float*)p+1);
516 return __builtin_shufflevector(a, b, 4, 5, 2, 3);
519 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
520 _mm_load_ss(const float *p)
522 return (__m128){ *p, 0, 0, 0 };
525 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
526 _mm_load1_ps(const float *p)
528 return (__m128){ *p, *p, *p, *p };
531 #define _mm_load_ps1(p) _mm_load1_ps(p)
533 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
534 _mm_load_ps(const float *p)
539 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
540 _mm_loadu_ps(const float *p)
544 } __attribute__((packed, may_alias));
545 return ((struct __loadu_ps*)p)->v;
548 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
549 _mm_loadr_ps(const float *p)
551 __m128 a = _mm_load_ps(p);
552 return __builtin_shufflevector(a, a, 3, 2, 1, 0);
555 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
558 return (__m128){ w, 0, 0, 0 };
561 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
564 return (__m128){ w, w, w, w };
567 // Microsoft specific.
568 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
571 return _mm_set1_ps(w);
574 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
575 _mm_set_ps(float z, float y, float x, float w)
577 return (__m128){ w, x, y, z };
580 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
581 _mm_setr_ps(float z, float y, float x, float w)
583 return (__m128){ z, y, x, w };
586 static __inline__ __m128 __attribute__((__always_inline__))
589 return (__m128){ 0, 0, 0, 0 };
592 static __inline__ void __attribute__((__always_inline__))
593 _mm_storeh_pi(__m64 *p, __m128 a)
595 __builtin_ia32_storehps((__v2si *)p, a);
598 static __inline__ void __attribute__((__always_inline__))
599 _mm_storel_pi(__m64 *p, __m128 a)
601 __builtin_ia32_storelps((__v2si *)p, a);
604 static __inline__ void __attribute__((__always_inline__))
605 _mm_store_ss(float *p, __m128 a)
610 static __inline__ void __attribute__((__always_inline__, __nodebug__))
611 _mm_storeu_ps(float *p, __m128 a)
613 __builtin_ia32_storeups(p, a);
616 static __inline__ void __attribute__((__always_inline__, __nodebug__))
617 _mm_store1_ps(float *p, __m128 a)
619 a = __builtin_shufflevector(a, a, 0, 0, 0, 0);
623 static __inline__ void __attribute__((__always_inline__, __nodebug__))
624 _mm_store_ps1(float *p, __m128 a)
626 return _mm_store1_ps(p, a);
629 static __inline__ void __attribute__((__always_inline__, __nodebug__))
630 _mm_store_ps(float *p, __m128 a)
635 static __inline__ void __attribute__((__always_inline__, __nodebug__))
636 _mm_storer_ps(float *p, __m128 a)
638 a = __builtin_shufflevector(a, a, 3, 2, 1, 0);
642 #define _MM_HINT_T0 3
643 #define _MM_HINT_T1 2
644 #define _MM_HINT_T2 1
645 #define _MM_HINT_NTA 0
647 /* FIXME: We have to #define this because "sel" must be a constant integer, and
648 Sema doesn't do any form of constant propagation yet. */
650 #define _mm_prefetch(a, sel) (__builtin_prefetch((void *)(a), 0, sel))
652 static __inline__ void __attribute__((__always_inline__, __nodebug__))
653 _mm_stream_pi(__m64 *p, __m64 a)
655 __builtin_ia32_movntq(p, a);
658 static __inline__ void __attribute__((__always_inline__, __nodebug__))
659 _mm_stream_ps(float *p, __m128 a)
661 __builtin_ia32_movntps(p, a);
664 static __inline__ void __attribute__((__always_inline__, __nodebug__))
667 __builtin_ia32_sfence();
670 static __inline__ int __attribute__((__always_inline__, __nodebug__))
671 _mm_extract_pi16(__m64 a, int n)
673 __v4hi b = (__v4hi)a;
674 return (unsigned short)b[n & 3];
677 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
678 _mm_insert_pi16(__m64 a, int d, int n)
680 __v4hi b = (__v4hi)a;
685 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
686 _mm_max_pi16(__m64 a, __m64 b)
688 return (__m64)__builtin_ia32_pmaxsw((__v4hi)a, (__v4hi)b);
691 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
692 _mm_max_pu8(__m64 a, __m64 b)
694 return (__m64)__builtin_ia32_pmaxub((__v8qi)a, (__v8qi)b);
697 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
698 _mm_min_pi16(__m64 a, __m64 b)
700 return (__m64)__builtin_ia32_pminsw((__v4hi)a, (__v4hi)b);
703 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
704 _mm_min_pu8(__m64 a, __m64 b)
706 return (__m64)__builtin_ia32_pminub((__v8qi)a, (__v8qi)b);
709 static __inline__ int __attribute__((__always_inline__, __nodebug__))
710 _mm_movemask_pi8(__m64 a)
712 return __builtin_ia32_pmovmskb((__v8qi)a);
715 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
716 _mm_mulhi_pu16(__m64 a, __m64 b)
718 return (__m64)__builtin_ia32_pmulhuw((__v4hi)a, (__v4hi)b);
721 #define _mm_shuffle_pi16(a, n) \
722 ((__m64)__builtin_ia32_pshufw(a, n))
724 static __inline__ void __attribute__((__always_inline__, __nodebug__))
725 _mm_maskmove_si64(__m64 d, __m64 n, char *p)
727 __builtin_ia32_maskmovq((__v8qi)d, (__v8qi)n, p);
730 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
731 _mm_avg_pu8(__m64 a, __m64 b)
733 return (__m64)__builtin_ia32_pavgb((__v8qi)a, (__v8qi)b);
736 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
737 _mm_avg_pu16(__m64 a, __m64 b)
739 return (__m64)__builtin_ia32_pavgw((__v4hi)a, (__v4hi)b);
742 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
743 _mm_sad_pu8(__m64 a, __m64 b)
745 return (__m64)__builtin_ia32_psadbw((__v8qi)a, (__v8qi)b);
748 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__))
751 return __builtin_ia32_stmxcsr();
754 static __inline__ void __attribute__((__always_inline__, __nodebug__))
755 _mm_setcsr(unsigned int i)
757 __builtin_ia32_ldmxcsr(i);
760 #define _mm_shuffle_ps(a, b, mask) \
761 (__builtin_shufflevector((__v4sf)(a), (__v4sf)(b), \
762 (mask) & 0x3, ((mask) & 0xc) >> 2, \
763 (((mask) & 0x30) >> 4) + 4, \
764 (((mask) & 0xc0) >> 6) + 4))
766 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
767 _mm_unpackhi_ps(__m128 a, __m128 b)
769 return __builtin_shufflevector(a, b, 2, 6, 3, 7);
772 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
773 _mm_unpacklo_ps(__m128 a, __m128 b)
775 return __builtin_shufflevector(a, b, 0, 4, 1, 5);
778 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
779 _mm_move_ss(__m128 a, __m128 b)
781 return __builtin_shufflevector(a, b, 4, 1, 2, 3);
784 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
785 _mm_movehl_ps(__m128 a, __m128 b)
787 return __builtin_shufflevector(a, b, 6, 7, 2, 3);
790 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
791 _mm_movelh_ps(__m128 a, __m128 b)
793 return __builtin_shufflevector(a, b, 0, 1, 4, 5);
796 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
797 _mm_cvtpi16_ps(__m64 a)
802 b = _mm_setzero_si64();
803 b = _mm_cmpgt_pi16(b, a);
804 c = _mm_unpackhi_pi16(a, b);
805 r = _mm_setzero_ps();
806 r = _mm_cvtpi32_ps(r, c);
807 r = _mm_movelh_ps(r, r);
808 c = _mm_unpacklo_pi16(a, b);
809 r = _mm_cvtpi32_ps(r, c);
814 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
815 _mm_cvtpu16_ps(__m64 a)
820 b = _mm_setzero_si64();
821 c = _mm_unpackhi_pi16(a, b);
822 r = _mm_setzero_ps();
823 r = _mm_cvtpi32_ps(r, c);
824 r = _mm_movelh_ps(r, r);
825 c = _mm_unpacklo_pi16(a, b);
826 r = _mm_cvtpi32_ps(r, c);
831 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
832 _mm_cvtpi8_ps(__m64 a)
836 b = _mm_setzero_si64();
837 b = _mm_cmpgt_pi8(b, a);
838 b = _mm_unpacklo_pi8(a, b);
840 return _mm_cvtpi16_ps(b);
843 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
844 _mm_cvtpu8_ps(__m64 a)
848 b = _mm_setzero_si64();
849 b = _mm_unpacklo_pi8(a, b);
851 return _mm_cvtpi16_ps(b);
854 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
855 _mm_cvtpi32x2_ps(__m64 a, __m64 b)
859 c = _mm_setzero_ps();
860 c = _mm_cvtpi32_ps(c, b);
861 c = _mm_movelh_ps(c, c);
863 return _mm_cvtpi32_ps(c, a);
866 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
867 _mm_cvtps_pi16(__m128 a)
871 b = _mm_cvtps_pi32(a);
872 a = _mm_movehl_ps(a, a);
873 c = _mm_cvtps_pi32(a);
875 return _mm_packs_pi16(b, c);
878 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
879 _mm_cvtps_pi8(__m128 a)
883 b = _mm_cvtps_pi16(a);
884 c = _mm_setzero_si64();
886 return _mm_packs_pi16(b, c);
889 static __inline__ int __attribute__((__always_inline__, __nodebug__))
890 _mm_movemask_ps(__m128 a)
892 return __builtin_ia32_movmskps(a);
895 #define _MM_SHUFFLE(z, y, x, w) (((z) << 6) | ((y) << 4) | ((x) << 2) | (w))
897 #define _MM_EXCEPT_INVALID (0x0001)
898 #define _MM_EXCEPT_DENORM (0x0002)
899 #define _MM_EXCEPT_DIV_ZERO (0x0004)
900 #define _MM_EXCEPT_OVERFLOW (0x0008)
901 #define _MM_EXCEPT_UNDERFLOW (0x0010)
902 #define _MM_EXCEPT_INEXACT (0x0020)
903 #define _MM_EXCEPT_MASK (0x003f)
905 #define _MM_MASK_INVALID (0x0080)
906 #define _MM_MASK_DENORM (0x0100)
907 #define _MM_MASK_DIV_ZERO (0x0200)
908 #define _MM_MASK_OVERFLOW (0x0400)
909 #define _MM_MASK_UNDERFLOW (0x0800)
910 #define _MM_MASK_INEXACT (0x1000)
911 #define _MM_MASK_MASK (0x1f80)
913 #define _MM_ROUND_NEAREST (0x0000)
914 #define _MM_ROUND_DOWN (0x2000)
915 #define _MM_ROUND_UP (0x4000)
916 #define _MM_ROUND_TOWARD_ZERO (0x6000)
917 #define _MM_ROUND_MASK (0x6000)
919 #define _MM_FLUSH_ZERO_MASK (0x8000)
920 #define _MM_FLUSH_ZERO_ON (0x8000)
921 #define _MM_FLUSH_ZERO_OFF (0x8000)
923 #define _MM_GET_EXCEPTION_MASK() (_mm_getcsr() & _MM_MASK_MASK)
924 #define _MM_GET_EXCEPTION_STATE() (_mm_getcsr() & _MM_EXCEPT_MASK)
925 #define _MM_GET_FLUSH_ZERO_MODE() (_mm_getcsr() & _MM_FLUSH_ZERO_MASK)
926 #define _MM_GET_ROUNDING_MODE() (_mm_getcsr() & _MM_ROUND_MASK)
928 #define _MM_SET_EXCEPTION_MASK(x) (_mm_setcsr((_mm_getcsr() & ~_MM_MASK_MASK) | (x)))
929 #define _MM_SET_EXCEPTION_STATE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_EXCEPT_MASK) | (x)))
930 #define _MM_SET_FLUSH_ZERO_MODE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_FLUSH_ZERO_MASK) | (x)))
931 #define _MM_SET_ROUNDING_MODE(x) (_mm_setcsr((_mm_getcsr() & ~_MM_ROUND_MASK) | (x)))
933 #define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \
935 __m128 tmp3, tmp2, tmp1, tmp0; \
936 tmp0 = _mm_unpacklo_ps((row0), (row1)); \
937 tmp2 = _mm_unpacklo_ps((row2), (row3)); \
938 tmp1 = _mm_unpackhi_ps((row0), (row1)); \
939 tmp3 = _mm_unpackhi_ps((row2), (row3)); \
940 (row0) = _mm_movelh_ps(tmp0, tmp2); \
941 (row1) = _mm_movehl_ps(tmp2, tmp0); \
942 (row2) = _mm_movelh_ps(tmp1, tmp3); \
943 (row3) = _mm_movehl_ps(tmp3, tmp1); \
946 /* Aliases for compatibility. */
947 #define _m_pextrw _mm_extract_pi16
948 #define _m_pinsrw _mm_insert_pi16
949 #define _m_pmaxsw _mm_max_pi16
950 #define _m_pmaxub _mm_max_pu8
951 #define _m_pminsw _mm_min_pi16
952 #define _m_pminub _mm_min_pu8
953 #define _m_pmovmskb _mm_movemask_pi8
954 #define _m_pmulhuw _mm_mulhi_pu16
955 #define _m_pshufw _mm_shuffle_pi16
956 #define _m_maskmovq _mm_maskmove_si64
957 #define _m_pavgb _mm_avg_pu8
958 #define _m_pavgw _mm_avg_pu16
959 #define _m_psadbw _mm_sad_pu8
963 /* Ugly hack for backwards-compatibility (compatible with gcc) */
965 #include <emmintrin.h>
970 #endif /* __XMMINTRIN_H */