25 #error "Never use <xopintrin.h> directly; include <x86intrin.h> instead."
34 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("xop")))
37 _mm_maccs_epi16(__m128i __A, __m128i __B, __m128i __C)
39 return (__m128i)__builtin_ia32_vpmacssww((__v8hi)__A, (__v8hi)__B, (__v8hi)__C);
43 _mm_macc_epi16(__m128i __A, __m128i __B, __m128i __C)
45 return (__m128i)__builtin_ia32_vpmacsww((__v8hi)__A, (__v8hi)__B, (__v8hi)__C);
49 _mm_maccsd_epi16(__m128i __A, __m128i __B, __m128i __C)
51 return (__m128i)__builtin_ia32_vpmacsswd((__v8hi)__A, (__v8hi)__B, (__v4si)__C);
55 _mm_maccd_epi16(__m128i __A, __m128i __B, __m128i __C)
57 return (__m128i)__builtin_ia32_vpmacswd((__v8hi)__A, (__v8hi)__B, (__v4si)__C);
61 _mm_maccs_epi32(__m128i __A, __m128i __B, __m128i __C)
63 return (__m128i)__builtin_ia32_vpmacssdd((__v4si)__A, (__v4si)__B, (__v4si)__C);
67 _mm_macc_epi32(__m128i __A, __m128i __B, __m128i __C)
69 return (__m128i)__builtin_ia32_vpmacsdd((__v4si)__A, (__v4si)__B, (__v4si)__C);
73 _mm_maccslo_epi32(__m128i __A, __m128i __B, __m128i __C)
75 return (__m128i)__builtin_ia32_vpmacssdql((__v4si)__A, (__v4si)__B, (__v2di)__C);
79 _mm_macclo_epi32(__m128i __A, __m128i __B, __m128i __C)
81 return (__m128i)__builtin_ia32_vpmacsdql((__v4si)__A, (__v4si)__B, (__v2di)__C);
85 _mm_maccshi_epi32(__m128i __A, __m128i __B, __m128i __C)
87 return (__m128i)__builtin_ia32_vpmacssdqh((__v4si)__A, (__v4si)__B, (__v2di)__C);
91 _mm_macchi_epi32(__m128i __A, __m128i __B, __m128i __C)
93 return (__m128i)__builtin_ia32_vpmacsdqh((__v4si)__A, (__v4si)__B, (__v2di)__C);
97 _mm_maddsd_epi16(__m128i __A, __m128i __B, __m128i __C)
99 return (__m128i)__builtin_ia32_vpmadcsswd((__v8hi)__A, (__v8hi)__B, (__v4si)__C);
103 _mm_maddd_epi16(__m128i __A, __m128i __B, __m128i __C)
105 return (__m128i)__builtin_ia32_vpmadcswd((__v8hi)__A, (__v8hi)__B, (__v4si)__C);
109 _mm_haddw_epi8(__m128i __A)
111 return (__m128i)__builtin_ia32_vphaddbw((__v16qi)__A);
115 _mm_haddd_epi8(__m128i __A)
117 return (__m128i)__builtin_ia32_vphaddbd((__v16qi)__A);
121 _mm_haddq_epi8(__m128i __A)
123 return (__m128i)__builtin_ia32_vphaddbq((__v16qi)__A);
127 _mm_haddd_epi16(__m128i __A)
129 return (__m128i)__builtin_ia32_vphaddwd((__v8hi)__A);
133 _mm_haddq_epi16(__m128i __A)
135 return (__m128i)__builtin_ia32_vphaddwq((__v8hi)__A);
139 _mm_haddq_epi32(__m128i __A)
141 return (__m128i)__builtin_ia32_vphadddq((__v4si)__A);
145 _mm_haddw_epu8(__m128i __A)
147 return (__m128i)__builtin_ia32_vphaddubw((__v16qi)__A);
151 _mm_haddd_epu8(__m128i __A)
153 return (__m128i)__builtin_ia32_vphaddubd((__v16qi)__A);
157 _mm_haddq_epu8(__m128i __A)
159 return (__m128i)__builtin_ia32_vphaddubq((__v16qi)__A);
163 _mm_haddd_epu16(__m128i __A)
165 return (__m128i)__builtin_ia32_vphadduwd((__v8hi)__A);
169 _mm_haddq_epu16(__m128i __A)
171 return (__m128i)__builtin_ia32_vphadduwq((__v8hi)__A);
175 _mm_haddq_epu32(__m128i __A)
177 return (__m128i)__builtin_ia32_vphaddudq((__v4si)__A);
181 _mm_hsubw_epi8(__m128i __A)
183 return (__m128i)__builtin_ia32_vphsubbw((__v16qi)__A);
187 _mm_hsubd_epi16(__m128i __A)
189 return (__m128i)__builtin_ia32_vphsubwd((__v8hi)__A);
193 _mm_hsubq_epi32(__m128i __A)
195 return (__m128i)__builtin_ia32_vphsubdq((__v4si)__A);
199 _mm_cmov_si128(__m128i __A, __m128i __B, __m128i __C)
201 return (__m128i)__builtin_ia32_vpcmov(__A, __B, __C);
205 _mm256_cmov_si256(__m256i __A, __m256i __B, __m256i __C)
207 return (__m256i)__builtin_ia32_vpcmov_256(__A, __B, __C);
211 _mm_perm_epi8(__m128i __A, __m128i __B, __m128i __C)
213 return (__m128i)__builtin_ia32_vpperm((__v16qi)__A, (__v16qi)__B, (__v16qi)__C);
217 _mm_rot_epi8(__m128i __A, __m128i __B)
219 return (__m128i)__builtin_ia32_vprotb((__v16qi)__A, (__v16qi)__B);
223 _mm_rot_epi16(__m128i __A, __m128i __B)
225 return (__m128i)__builtin_ia32_vprotw((__v8hi)__A, (__v8hi)__B);
229 _mm_rot_epi32(__m128i __A, __m128i __B)
231 return (__m128i)__builtin_ia32_vprotd((__v4si)__A, (__v4si)__B);
235 _mm_rot_epi64(__m128i __A, __m128i __B)
237 return (__m128i)__builtin_ia32_vprotq((__v2di)__A, (__v2di)__B);
240 #define _mm_roti_epi8(A, N) __extension__ ({ \
241 (__m128i)__builtin_ia32_vprotbi((__v16qi)(__m128i)(A), (N)); })
243 #define _mm_roti_epi16(A, N) __extension__ ({ \
244 (__m128i)__builtin_ia32_vprotwi((__v8hi)(__m128i)(A), (N)); })
246 #define _mm_roti_epi32(A, N) __extension__ ({ \
247 (__m128i)__builtin_ia32_vprotdi((__v4si)(__m128i)(A), (N)); })
249 #define _mm_roti_epi64(A, N) __extension__ ({ \
250 (__m128i)__builtin_ia32_vprotqi((__v2di)(__m128i)(A), (N)); })
253 _mm_shl_epi8(__m128i __A, __m128i __B)
255 return (__m128i)__builtin_ia32_vpshlb((__v16qi)__A, (__v16qi)__B);
259 _mm_shl_epi16(__m128i __A, __m128i __B)
261 return (__m128i)__builtin_ia32_vpshlw((__v8hi)__A, (__v8hi)__B);
265 _mm_shl_epi32(__m128i __A, __m128i __B)
267 return (__m128i)__builtin_ia32_vpshld((__v4si)__A, (__v4si)__B);
271 _mm_shl_epi64(__m128i __A, __m128i __B)
273 return (__m128i)__builtin_ia32_vpshlq((__v2di)__A, (__v2di)__B);
277 _mm_sha_epi8(__m128i __A, __m128i __B)
279 return (__m128i)__builtin_ia32_vpshab((__v16qi)__A, (__v16qi)__B);
283 _mm_sha_epi16(__m128i __A, __m128i __B)
285 return (__m128i)__builtin_ia32_vpshaw((__v8hi)__A, (__v8hi)__B);
289 _mm_sha_epi32(__m128i __A, __m128i __B)
291 return (__m128i)__builtin_ia32_vpshad((__v4si)__A, (__v4si)__B);
295 _mm_sha_epi64(__m128i __A, __m128i __B)
297 return (__m128i)__builtin_ia32_vpshaq((__v2di)__A, (__v2di)__B);
300 #define _mm_com_epu8(A, B, N) __extension__ ({ \
301 (__m128i)__builtin_ia32_vpcomub((__v16qi)(__m128i)(A), \
302 (__v16qi)(__m128i)(B), (N)); })
304 #define _mm_com_epu16(A, B, N) __extension__ ({ \
305 (__m128i)__builtin_ia32_vpcomuw((__v8hi)(__m128i)(A), \
306 (__v8hi)(__m128i)(B), (N)); })
308 #define _mm_com_epu32(A, B, N) __extension__ ({ \
309 (__m128i)__builtin_ia32_vpcomud((__v4si)(__m128i)(A), \
310 (__v4si)(__m128i)(B), (N)); })
312 #define _mm_com_epu64(A, B, N) __extension__ ({ \
313 (__m128i)__builtin_ia32_vpcomuq((__v2di)(__m128i)(A), \
314 (__v2di)(__m128i)(B), (N)); })
316 #define _mm_com_epi8(A, B, N) __extension__ ({ \
317 (__m128i)__builtin_ia32_vpcomb((__v16qi)(__m128i)(A), \
318 (__v16qi)(__m128i)(B), (N)); })
320 #define _mm_com_epi16(A, B, N) __extension__ ({ \
321 (__m128i)__builtin_ia32_vpcomw((__v8hi)(__m128i)(A), \
322 (__v8hi)(__m128i)(B), (N)); })
324 #define _mm_com_epi32(A, B, N) __extension__ ({ \
325 (__m128i)__builtin_ia32_vpcomd((__v4si)(__m128i)(A), \
326 (__v4si)(__m128i)(B), (N)); })
328 #define _mm_com_epi64(A, B, N) __extension__ ({ \
329 (__m128i)__builtin_ia32_vpcomq((__v2di)(__m128i)(A), \
330 (__v2di)(__m128i)(B), (N)); })
332 #define _MM_PCOMCTRL_LT 0
333 #define _MM_PCOMCTRL_LE 1
334 #define _MM_PCOMCTRL_GT 2
335 #define _MM_PCOMCTRL_GE 3
336 #define _MM_PCOMCTRL_EQ 4
337 #define _MM_PCOMCTRL_NEQ 5
338 #define _MM_PCOMCTRL_FALSE 6
339 #define _MM_PCOMCTRL_TRUE 7
342 _mm_comlt_epu8(__m128i __A, __m128i __B)
348 _mm_comle_epu8(__m128i __A, __m128i __B)
354 _mm_comgt_epu8(__m128i __A, __m128i __B)
360 _mm_comge_epu8(__m128i __A, __m128i __B)
366 _mm_comeq_epu8(__m128i __A, __m128i __B)
372 _mm_comneq_epu8(__m128i __A, __m128i __B)
378 _mm_comfalse_epu8(__m128i __A, __m128i __B)
384 _mm_comtrue_epu8(__m128i __A, __m128i __B)
390 _mm_comlt_epu16(__m128i __A, __m128i __B)
396 _mm_comle_epu16(__m128i __A, __m128i __B)
402 _mm_comgt_epu16(__m128i __A, __m128i __B)
408 _mm_comge_epu16(__m128i __A, __m128i __B)
414 _mm_comeq_epu16(__m128i __A, __m128i __B)
420 _mm_comneq_epu16(__m128i __A, __m128i __B)
426 _mm_comfalse_epu16(__m128i __A, __m128i __B)
432 _mm_comtrue_epu16(__m128i __A, __m128i __B)
438 _mm_comlt_epu32(__m128i __A, __m128i __B)
444 _mm_comle_epu32(__m128i __A, __m128i __B)
450 _mm_comgt_epu32(__m128i __A, __m128i __B)
456 _mm_comge_epu32(__m128i __A, __m128i __B)
462 _mm_comeq_epu32(__m128i __A, __m128i __B)
468 _mm_comneq_epu32(__m128i __A, __m128i __B)
474 _mm_comfalse_epu32(__m128i __A, __m128i __B)
480 _mm_comtrue_epu32(__m128i __A, __m128i __B)
486 _mm_comlt_epu64(__m128i __A, __m128i __B)
492 _mm_comle_epu64(__m128i __A, __m128i __B)
498 _mm_comgt_epu64(__m128i __A, __m128i __B)
504 _mm_comge_epu64(__m128i __A, __m128i __B)
510 _mm_comeq_epu64(__m128i __A, __m128i __B)
516 _mm_comneq_epu64(__m128i __A, __m128i __B)
522 _mm_comfalse_epu64(__m128i __A, __m128i __B)
528 _mm_comtrue_epu64(__m128i __A, __m128i __B)
534 _mm_comlt_epi8(__m128i __A, __m128i __B)
540 _mm_comle_epi8(__m128i __A, __m128i __B)
546 _mm_comgt_epi8(__m128i __A, __m128i __B)
552 _mm_comge_epi8(__m128i __A, __m128i __B)
558 _mm_comeq_epi8(__m128i __A, __m128i __B)
564 _mm_comneq_epi8(__m128i __A, __m128i __B)
570 _mm_comfalse_epi8(__m128i __A, __m128i __B)
576 _mm_comtrue_epi8(__m128i __A, __m128i __B)
582 _mm_comlt_epi16(__m128i __A, __m128i __B)
588 _mm_comle_epi16(__m128i __A, __m128i __B)
594 _mm_comgt_epi16(__m128i __A, __m128i __B)
600 _mm_comge_epi16(__m128i __A, __m128i __B)
606 _mm_comeq_epi16(__m128i __A, __m128i __B)
612 _mm_comneq_epi16(__m128i __A, __m128i __B)
618 _mm_comfalse_epi16(__m128i __A, __m128i __B)
624 _mm_comtrue_epi16(__m128i __A, __m128i __B)
630 _mm_comlt_epi32(__m128i __A, __m128i __B)
636 _mm_comle_epi32(__m128i __A, __m128i __B)
642 _mm_comgt_epi32(__m128i __A, __m128i __B)
648 _mm_comge_epi32(__m128i __A, __m128i __B)
654 _mm_comeq_epi32(__m128i __A, __m128i __B)
660 _mm_comneq_epi32(__m128i __A, __m128i __B)
666 _mm_comfalse_epi32(__m128i __A, __m128i __B)
672 _mm_comtrue_epi32(__m128i __A, __m128i __B)
678 _mm_comlt_epi64(__m128i __A, __m128i __B)
684 _mm_comle_epi64(__m128i __A, __m128i __B)
690 _mm_comgt_epi64(__m128i __A, __m128i __B)
696 _mm_comge_epi64(__m128i __A, __m128i __B)
702 _mm_comeq_epi64(__m128i __A, __m128i __B)
708 _mm_comneq_epi64(__m128i __A, __m128i __B)
714 _mm_comfalse_epi64(__m128i __A, __m128i __B)
720 _mm_comtrue_epi64(__m128i __A, __m128i __B)
725 #define _mm_permute2_pd(X, Y, C, I) __extension__ ({ \
726 (__m128d)__builtin_ia32_vpermil2pd((__v2df)(__m128d)(X), \
727 (__v2df)(__m128d)(Y), \
728 (__v2di)(__m128i)(C), (I)); })
730 #define _mm256_permute2_pd(X, Y, C, I) __extension__ ({ \
731 (__m256d)__builtin_ia32_vpermil2pd256((__v4df)(__m256d)(X), \
732 (__v4df)(__m256d)(Y), \
733 (__v4di)(__m256i)(C), (I)); })
735 #define _mm_permute2_ps(X, Y, C, I) __extension__ ({ \
736 (__m128)__builtin_ia32_vpermil2ps((__v4sf)(__m128)(X), (__v4sf)(__m128)(Y), \
737 (__v4si)(__m128i)(C), (I)); })
739 #define _mm256_permute2_ps(X, Y, C, I) __extension__ ({ \
740 (__m256)__builtin_ia32_vpermil2ps256((__v8sf)(__m256)(X), \
741 (__v8sf)(__m256)(Y), \
742 (__v8si)(__m256i)(C), (I)); })
745 _mm_frcz_ss(__m128 __A)
747 return (__m128)__builtin_ia32_vfrczss((__v4sf)__A);
751 _mm_frcz_sd(__m128d __A)
753 return (__m128d)__builtin_ia32_vfrczsd((__v2df)__A);
757 _mm_frcz_ps(__m128 __A)
759 return (__m128)__builtin_ia32_vfrczps((__v4sf)__A);
763 _mm_frcz_pd(__m128d __A)
765 return (__m128d)__builtin_ia32_vfrczpd((__v2df)__A);
769 _mm256_frcz_ps(__m256 __A)
771 return (__m256)__builtin_ia32_vfrczps256((__v8sf)__A);
775 _mm256_frcz_pd(__m256d __A)
777 return (__m256d)__builtin_ia32_vfrczpd256((__v4df)__A);
780 #undef __DEFAULT_FN_ATTRS
#define _mm_com_epi16(A, B, N)
Definition: xopintrin.h:320
#define _mm_com_epi32(A, B, N)
Definition: xopintrin.h:324
#define _MM_PCOMCTRL_GT
Definition: xopintrin.h:334
#define __DEFAULT_FN_ATTRS
Definition: xopintrin.h:34
#define _MM_PCOMCTRL_GE
Definition: xopintrin.h:335
#define _mm_com_epu16(A, B, N)
Definition: xopintrin.h:304
#define _mm_com_epu8(A, B, N)
Definition: xopintrin.h:300
#define _MM_PCOMCTRL_NEQ
Definition: xopintrin.h:337
#define _MM_PCOMCTRL_FALSE
Definition: xopintrin.h:338
#define _mm_com_epu32(A, B, N)
Definition: xopintrin.h:308
#define _mm_com_epi64(A, B, N)
Definition: xopintrin.h:328
#define _MM_PCOMCTRL_TRUE
Definition: xopintrin.h:339
#define _MM_PCOMCTRL_LT
Definition: xopintrin.h:332
#define _mm_com_epu64(A, B, N)
Definition: xopintrin.h:312
#define _MM_PCOMCTRL_EQ
Definition: xopintrin.h:336
#define _MM_PCOMCTRL_LE
Definition: xopintrin.h:333
#define _mm_com_epi8(A, B, N)
Definition: xopintrin.h:316