Lines Matching refs:__m128
35 typedef float __m128 __attribute__((__vector_size__(16)));
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__))
96 _mm_sqrt_ss(__m128 a)
101 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
102 _mm_sqrt_ps(__m128 a)
107 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
108 _mm_rcp_ss(__m128 a)
113 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
114 _mm_rcp_ps(__m128 a)
119 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
120 _mm_rsqrt_ss(__m128 a)
125 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
126 _mm_rsqrt_ps(__m128 a)
131 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
132 _mm_min_ss(__m128 a, __m128 b)
137 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
138 _mm_min_ps(__m128 a, __m128 b)
143 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
144 _mm_max_ss(__m128 a, __m128 b)
149 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
150 _mm_max_ps(__m128 a, __m128 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);
324 _mm_comieq_ss(__m128 a, __m128 b)
330 _mm_comilt_ss(__m128 a, __m128 b)
336 _mm_comile_ss(__m128 a, __m128 b)
342 _mm_comigt_ss(__m128 a, __m128 b)
348 _mm_comige_ss(__m128 a, __m128 b)
354 _mm_comineq_ss(__m128 a, __m128 b)
360 _mm_ucomieq_ss(__m128 a, __m128 b)
366 _mm_ucomilt_ss(__m128 a, __m128 b)
372 _mm_ucomile_ss(__m128 a, __m128 b)
378 _mm_ucomigt_ss(__m128 a, __m128 b)
384 _mm_ucomige_ss(__m128 a, __m128 b)
390 _mm_ucomineq_ss(__m128 a, __m128 b)
396 _mm_cvtss_si32(__m128 a)
402 _mm_cvt_ss2si(__m128 a)
410 _mm_cvtss_si64(__m128 a)
418 _mm_cvtps_pi32(__m128 a)
424 _mm_cvt_ps2pi(__m128 a)
430 _mm_cvttss_si32(__m128 a)
436 _mm_cvtt_ss2si(__m128 a)
442 _mm_cvttss_si64(__m128 a)
448 _mm_cvttps_pi32(__m128 a)
454 _mm_cvtt_ps2pi(__m128 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)
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)
489 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
490 _mm_cvt_pi2ps(__m128 a, __m64 b)
496 _mm_cvtss_f32(__m128 a)
501 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
502 _mm_loadh_pi(__m128 a, const __m64 *p)
504 __m128 b;
510 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
511 _mm_loadl_pi(__m128 a, const __m64 *p)
513 __m128 b;
519 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
522 return (__m128){ *p, 0, 0, 0 };
525 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
528 return (__m128){ *p, *p, *p, *p };
533 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
536 return *(__m128*)p;
539 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
543 __m128 v;
548 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
551 __m128 a = _mm_load_ps(p);
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 };
568 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
574 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
577 return (__m128){ w, x, y, z };
580 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
583 return (__m128){ z, y, x, w };
586 static __inline__ __m128 __attribute__((__always_inline__))
589 return (__m128){ 0, 0, 0, 0 };
593 _mm_storeh_pi(__m64 *p, __m128 a)
599 _mm_storel_pi(__m64 *p, __m128 a)
605 _mm_store_ss(float *p, __m128 a)
611 _mm_storeu_ps(float *p, __m128 a)
617 _mm_store1_ps(float *p, __m128 a)
624 _mm_store_ps1(float *p, __m128 a)
630 _mm_store_ps(float *p, __m128 a)
632 *(__m128 *)p = a;
636 _mm_storer_ps(float *p, __m128 a)
659 _mm_stream_ps(float *p, __m128 a)
766 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
767 _mm_unpackhi_ps(__m128 a, __m128 b)
772 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
773 _mm_unpacklo_ps(__m128 a, __m128 b)
778 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
779 _mm_move_ss(__m128 a, __m128 b)
784 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
785 _mm_movehl_ps(__m128 a, __m128 b)
790 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
791 _mm_movelh_ps(__m128 a, __m128 b)
796 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
800 __m128 r;
814 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
818 __m128 r;
831 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
843 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
854 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
857 __m128 c;
867 _mm_cvtps_pi16(__m128 a)
879 _mm_cvtps_pi8(__m128 a)
890 _mm_movemask_ps(__m128 a)
935 __m128 tmp3, tmp2, tmp1, tmp0; \