Lines Matching defs:__b
43 _mm_add_ss(__m128 __a, __m128 __b)
45 __a[0] += __b[0];
50 _mm_add_ps(__m128 __a, __m128 __b)
52 return __a + __b;
56 _mm_sub_ss(__m128 __a, __m128 __b)
58 __a[0] -= __b[0];
63 _mm_sub_ps(__m128 __a, __m128 __b)
65 return __a - __b;
69 _mm_mul_ss(__m128 __a, __m128 __b)
71 __a[0] *= __b[0];
76 _mm_mul_ps(__m128 __a, __m128 __b)
78 return __a * __b;
82 _mm_div_ss(__m128 __a, __m128 __b)
84 __a[0] /= __b[0];
89 _mm_div_ps(__m128 __a, __m128 __b)
91 return __a / __b;
134 _mm_min_ss(__m128 __a, __m128 __b)
136 return __builtin_ia32_minss(__a, __b);
140 _mm_min_ps(__m128 __a, __m128 __b)
142 return __builtin_ia32_minps(__a, __b);
146 _mm_max_ss(__m128 __a, __m128 __b)
148 return __builtin_ia32_maxss(__a, __b);
152 _mm_max_ps(__m128 __a, __m128 __b)
154 return __builtin_ia32_maxps(__a, __b);
158 _mm_and_ps(__m128 __a, __m128 __b)
160 return (__m128)((__v4si)__a & (__v4si)__b);
164 _mm_andnot_ps(__m128 __a, __m128 __b)
166 return (__m128)(~(__v4si)__a & (__v4si)__b);
170 _mm_or_ps(__m128 __a, __m128 __b)
172 return (__m128)((__v4si)__a | (__v4si)__b);
176 _mm_xor_ps(__m128 __a, __m128 __b)
178 return (__m128)((__v4si)__a ^ (__v4si)__b);
182 _mm_cmpeq_ss(__m128 __a, __m128 __b)
184 return (__m128)__builtin_ia32_cmpeqss(__a, __b);
188 _mm_cmpeq_ps(__m128 __a, __m128 __b)
190 return (__m128)__builtin_ia32_cmpeqps(__a, __b);
194 _mm_cmplt_ss(__m128 __a, __m128 __b)
196 return (__m128)__builtin_ia32_cmpltss(__a, __b);
200 _mm_cmplt_ps(__m128 __a, __m128 __b)
202 return (__m128)__builtin_ia32_cmpltps(__a, __b);
206 _mm_cmple_ss(__m128 __a, __m128 __b)
208 return (__m128)__builtin_ia32_cmpless(__a, __b);
212 _mm_cmple_ps(__m128 __a, __m128 __b)
214 return (__m128)__builtin_ia32_cmpleps(__a, __b);
218 _mm_cmpgt_ss(__m128 __a, __m128 __b)
221 __builtin_ia32_cmpltss(__b, __a),
226 _mm_cmpgt_ps(__m128 __a, __m128 __b)
228 return (__m128)__builtin_ia32_cmpltps(__b, __a);
232 _mm_cmpge_ss(__m128 __a, __m128 __b)
235 __builtin_ia32_cmpless(__b, __a),
240 _mm_cmpge_ps(__m128 __a, __m128 __b)
242 return (__m128)__builtin_ia32_cmpleps(__b, __a);
246 _mm_cmpneq_ss(__m128 __a, __m128 __b)
248 return (__m128)__builtin_ia32_cmpneqss(__a, __b);
252 _mm_cmpneq_ps(__m128 __a, __m128 __b)
254 return (__m128)__builtin_ia32_cmpneqps(__a, __b);
258 _mm_cmpnlt_ss(__m128 __a, __m128 __b)
260 return (__m128)__builtin_ia32_cmpnltss(__a, __b);
264 _mm_cmpnlt_ps(__m128 __a, __m128 __b)
266 return (__m128)__builtin_ia32_cmpnltps(__a, __b);
270 _mm_cmpnle_ss(__m128 __a, __m128 __b)
272 return (__m128)__builtin_ia32_cmpnless(__a, __b);
276 _mm_cmpnle_ps(__m128 __a, __m128 __b)
278 return (__m128)__builtin_ia32_cmpnleps(__a, __b);
282 _mm_cmpngt_ss(__m128 __a, __m128 __b)
285 __builtin_ia32_cmpnltss(__b, __a),
290 _mm_cmpngt_ps(__m128 __a, __m128 __b)
292 return (__m128)__builtin_ia32_cmpnltps(__b, __a);
296 _mm_cmpnge_ss(__m128 __a, __m128 __b)
299 __builtin_ia32_cmpnless(__b, __a),
304 _mm_cmpnge_ps(__m128 __a, __m128 __b)
306 return (__m128)__builtin_ia32_cmpnleps(__b, __a);
310 _mm_cmpord_ss(__m128 __a, __m128 __b)
312 return (__m128)__builtin_ia32_cmpordss(__a, __b);
316 _mm_cmpord_ps(__m128 __a, __m128 __b)
318 return (__m128)__builtin_ia32_cmpordps(__a, __b);
322 _mm_cmpunord_ss(__m128 __a, __m128 __b)
324 return (__m128)__builtin_ia32_cmpunordss(__a, __b);
328 _mm_cmpunord_ps(__m128 __a, __m128 __b)
330 return (__m128)__builtin_ia32_cmpunordps(__a, __b);
334 _mm_comieq_ss(__m128 __a, __m128 __b)
336 return __builtin_ia32_comieq(__a, __b);
340 _mm_comilt_ss(__m128 __a, __m128 __b)
342 return __builtin_ia32_comilt(__a, __b);
346 _mm_comile_ss(__m128 __a, __m128 __b)
348 return __builtin_ia32_comile(__a, __b);
352 _mm_comigt_ss(__m128 __a, __m128 __b)
354 return __builtin_ia32_comigt(__a, __b);
358 _mm_comige_ss(__m128 __a, __m128 __b)
360 return __builtin_ia32_comige(__a, __b);
364 _mm_comineq_ss(__m128 __a, __m128 __b)
366 return __builtin_ia32_comineq(__a, __b);
370 _mm_ucomieq_ss(__m128 __a, __m128 __b)
372 return __builtin_ia32_ucomieq(__a, __b);
376 _mm_ucomilt_ss(__m128 __a, __m128 __b)
378 return __builtin_ia32_ucomilt(__a, __b);
382 _mm_ucomile_ss(__m128 __a, __m128 __b)
384 return __builtin_ia32_ucomile(__a, __b);
388 _mm_ucomigt_ss(__m128 __a, __m128 __b)
390 return __builtin_ia32_ucomigt(__a, __b);
394 _mm_ucomige_ss(__m128 __a, __m128 __b)
396 return __builtin_ia32_ucomige(__a, __b);
400 _mm_ucomineq_ss(__m128 __a, __m128 __b)
402 return __builtin_ia32_ucomineq(__a, __b);
470 _mm_cvtsi32_ss(__m128 __a, int __b)
472 __a[0] = __b;
477 _mm_cvt_si2ss(__m128 __a, int __b)
479 return _mm_cvtsi32_ss(__a, __b);
485 _mm_cvtsi64_ss(__m128 __a, long long __b)
487 __a[0] = __b;
494 _mm_cvtpi32_ps(__m128 __a, __m64 __b)
496 return __builtin_ia32_cvtpi2ps(__a, (__v2si)__b);
500 _mm_cvt_pi2ps(__m128 __a, __m64 __b)
502 return _mm_cvtpi32_ps(__a, __b);
518 __mm_loadh_pi_v2f32 __b = ((struct __mm_loadh_pi_struct*)__p)->__u;
519 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
530 __mm_loadl_pi_v2f32 __b = ((struct __mm_loadl_pi_struct*)__p)->__u;
531 __m128 __bb = __builtin_shufflevector(__b, __b, 0, 1, 0, 1);
708 __v4hi __b = (__v4hi)__a;
709 return (unsigned short)__b[__n & 3];
715 __v4hi __b = (__v4hi)__a;
716 __b[__n & 3] = __d;
717 return (__m64)__b;
721 _mm_max_pi16(__m64 __a, __m64 __b)
723 return (__m64)__builtin_ia32_pmaxsw((__v4hi)__a, (__v4hi)__b);
727 _mm_max_pu8(__m64 __a, __m64 __b)
729 return (__m64)__builtin_ia32_pmaxub((__v8qi)__a, (__v8qi)__b);
733 _mm_min_pi16(__m64 __a, __m64 __b)
735 return (__m64)__builtin_ia32_pminsw((__v4hi)__a, (__v4hi)__b);
739 _mm_min_pu8(__m64 __a, __m64 __b)
741 return (__m64)__builtin_ia32_pminub((__v8qi)__a, (__v8qi)__b);
751 _mm_mulhi_pu16(__m64 __a, __m64 __b)
753 return (__m64)__builtin_ia32_pmulhuw((__v4hi)__a, (__v4hi)__b);
766 _mm_avg_pu8(__m64 __a, __m64 __b)
768 return (__m64)__builtin_ia32_pavgb((__v8qi)__a, (__v8qi)__b);
772 _mm_avg_pu16(__m64 __a, __m64 __b)
774 return (__m64)__builtin_ia32_pavgw((__v4hi)__a, (__v4hi)__b);
778 _mm_sad_pu8(__m64 __a, __m64 __b)
780 return (__m64)__builtin_ia32_psadbw((__v8qi)__a, (__v8qi)__b);
802 _mm_unpackhi_ps(__m128 __a, __m128 __b)
804 return __builtin_shufflevector(__a, __b, 2, 6, 3, 7);
808 _mm_unpacklo_ps(__m128 __a, __m128 __b)
810 return __builtin_shufflevector(__a, __b, 0, 4, 1, 5);
814 _mm_move_ss(__m128 __a, __m128 __b)
816 return __builtin_shufflevector(__a, __b, 4, 1, 2, 3);
820 _mm_movehl_ps(__m128 __a, __m128 __b)
822 return __builtin_shufflevector(__a, __b, 6, 7, 2, 3);
826 _mm_movelh_ps(__m128 __a, __m128 __b)
828 return __builtin_shufflevector(__a, __b, 0, 1, 4, 5);
834 __m64 __b, __c;
837 __b = _mm_setzero_si64();
838 __b = _mm_cmpgt_pi16(__b, __a);
839 __c = _mm_unpackhi_pi16(__a, __b);
843 __c = _mm_unpacklo_pi16(__a, __b);
852 __m64 __b, __c;
855 __b = _mm_setzero_si64();
856 __c = _mm_unpackhi_pi16(__a, __b);
860 __c = _mm_unpacklo_pi16(__a, __b);
869 __m64 __b;
871 __b = _mm_setzero_si64();
872 __b = _mm_cmpgt_pi8(__b, __a);
873 __b = _mm_unpacklo_pi8(__a, __b);
875 return _mm_cvtpi16_ps(__b);
881 __m64 __b;
883 __b = _mm_setzero_si64();
884 __b = _mm_unpacklo_pi8(__a, __b);
886 return _mm_cvtpi16_ps(__b);
890 _mm_cvtpi32x2_ps(__m64 __a, __m64 __b)
895 __c = _mm_cvtpi32_ps(__c, __b);
904 __m64 __b, __c;
906 __b = _mm_cvtps_pi32(__a);
910 return _mm_packs_pi32(__b, __c);
916 __m64 __b, __c;
918 __b = _mm_cvtps_pi16(__a);
921 return _mm_packs_pi16(__b, __c);