blob: b0df91f205b73b5d3d898a217ac33bc7c8ac77e9 [file] [log] [blame]
Craig Topperdec792e2011-12-19 05:04:33 +00001/*===---- avx2intrin.h - AVX2 intrinsics -----------------------------------===
2 *
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:
9 *
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
12 *
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
19 * THE SOFTWARE.
20 *
21 *===-----------------------------------------------------------------------===
22 */
23
24#ifndef __IMMINTRIN_H
25#error "Never use <avx2intrin.h> directly; include <immintrin.h> instead."
26#endif
27
Richard Smith49e56442013-07-14 05:41:45 +000028#ifndef __AVX2INTRIN_H
29#define __AVX2INTRIN_H
30
Craig Topperdec792e2011-12-19 05:04:33 +000031/* SSE4 Multiple Packed Sums of Absolute Difference. */
32#define _mm256_mpsadbw_epu8(X, Y, M) __builtin_ia32_mpsadbw256((X), (Y), (M))
33
34static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000035_mm256_abs_epi8(__m256i __a)
Craig Topperdec792e2011-12-19 05:04:33 +000036{
David Blaikie3302f2b2013-01-16 23:08:36 +000037 return (__m256i)__builtin_ia32_pabsb256((__v32qi)__a);
Craig Topperdec792e2011-12-19 05:04:33 +000038}
39
40static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000041_mm256_abs_epi16(__m256i __a)
Craig Topperdec792e2011-12-19 05:04:33 +000042{
David Blaikie3302f2b2013-01-16 23:08:36 +000043 return (__m256i)__builtin_ia32_pabsw256((__v16hi)__a);
Craig Topperdec792e2011-12-19 05:04:33 +000044}
45
46static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000047_mm256_abs_epi32(__m256i __a)
Craig Topperdec792e2011-12-19 05:04:33 +000048{
David Blaikie3302f2b2013-01-16 23:08:36 +000049 return (__m256i)__builtin_ia32_pabsd256((__v8si)__a);
Craig Topperdec792e2011-12-19 05:04:33 +000050}
51
52static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000053_mm256_packs_epi16(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000054{
David Blaikie3302f2b2013-01-16 23:08:36 +000055 return (__m256i)__builtin_ia32_packsswb256((__v16hi)__a, (__v16hi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000056}
57
58static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000059_mm256_packs_epi32(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000060{
David Blaikie3302f2b2013-01-16 23:08:36 +000061 return (__m256i)__builtin_ia32_packssdw256((__v8si)__a, (__v8si)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000062}
63
64static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000065_mm256_packus_epi16(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000066{
David Blaikie3302f2b2013-01-16 23:08:36 +000067 return (__m256i)__builtin_ia32_packuswb256((__v16hi)__a, (__v16hi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000068}
69
70static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
71_mm256_packus_epi32(__m256i __V1, __m256i __V2)
72{
73 return (__m256i) __builtin_ia32_packusdw256((__v8si)__V1, (__v8si)__V2);
74}
75
76static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000077_mm256_add_epi8(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000078{
David Blaikie3302f2b2013-01-16 23:08:36 +000079 return (__m256i)((__v32qi)__a + (__v32qi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000080}
81
82static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000083_mm256_add_epi16(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000084{
David Blaikie3302f2b2013-01-16 23:08:36 +000085 return (__m256i)((__v16hi)__a + (__v16hi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000086}
87
88static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000089_mm256_add_epi32(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000090{
David Blaikie3302f2b2013-01-16 23:08:36 +000091 return (__m256i)((__v8si)__a + (__v8si)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000092}
93
94static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +000095_mm256_add_epi64(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000096{
David Blaikie3302f2b2013-01-16 23:08:36 +000097 return __a + __b;
Craig Topperdec792e2011-12-19 05:04:33 +000098}
99
100static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000101_mm256_adds_epi8(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000102{
David Blaikie3302f2b2013-01-16 23:08:36 +0000103 return (__m256i)__builtin_ia32_paddsb256((__v32qi)__a, (__v32qi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000104}
105
106static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000107_mm256_adds_epi16(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000108{
David Blaikie3302f2b2013-01-16 23:08:36 +0000109 return (__m256i)__builtin_ia32_paddsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000110}
111
112static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000113_mm256_adds_epu8(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000114{
David Blaikie3302f2b2013-01-16 23:08:36 +0000115 return (__m256i)__builtin_ia32_paddusb256((__v32qi)__a, (__v32qi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000116}
117
118static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000119_mm256_adds_epu16(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000120{
David Blaikie3302f2b2013-01-16 23:08:36 +0000121 return (__m256i)__builtin_ia32_paddusw256((__v16hi)__a, (__v16hi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000122}
123
124#define _mm256_alignr_epi8(a, b, n) __extension__ ({ \
125 __m256i __a = (a); \
126 __m256i __b = (b); \
127 (__m256i)__builtin_ia32_palignr256((__v32qi)__a, (__v32qi)__b, (n)); })
128
129static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000130_mm256_and_si256(__m256i __a, __m256i __b)
Craig Toppera557e1c2011-12-19 09:03:48 +0000131{
David Blaikie3302f2b2013-01-16 23:08:36 +0000132 return __a & __b;
Craig Toppera557e1c2011-12-19 09:03:48 +0000133}
134
135static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000136_mm256_andnot_si256(__m256i __a, __m256i __b)
Craig Toppera557e1c2011-12-19 09:03:48 +0000137{
David Blaikie3302f2b2013-01-16 23:08:36 +0000138 return ~__a & __b;
Craig Toppera557e1c2011-12-19 09:03:48 +0000139}
140
141static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000142_mm256_avg_epu8(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000143{
David Blaikie3302f2b2013-01-16 23:08:36 +0000144 return (__m256i)__builtin_ia32_pavgb256((__v32qi)__a, (__v32qi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000145}
146
147static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000148_mm256_avg_epu16(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000149{
David Blaikie3302f2b2013-01-16 23:08:36 +0000150 return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000151}
152
153static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
154_mm256_blendv_epi8(__m256i __V1, __m256i __V2, __m256i __M)
155{
156 return (__m256i)__builtin_ia32_pblendvb256((__v32qi)__V1, (__v32qi)__V2,
157 (__v32qi)__M);
158}
159
160#define _mm256_blend_epi16(V1, V2, M) __extension__ ({ \
161 __m256i __V1 = (V1); \
162 __m256i __V2 = (V2); \
Ekaterina Romanovab929ad72015-04-10 02:39:45 +0000163 (__m256i)__builtin_shufflevector((__v16hi)__V1, (__v16hi)__V2, \
Filipe Cabecinhas5d289b42014-05-13 02:37:02 +0000164 (((M) & 0x01) ? 16 : 0), \
165 (((M) & 0x02) ? 17 : 1), \
166 (((M) & 0x04) ? 18 : 2), \
167 (((M) & 0x08) ? 19 : 3), \
168 (((M) & 0x10) ? 20 : 4), \
169 (((M) & 0x20) ? 21 : 5), \
170 (((M) & 0x40) ? 22 : 6), \
171 (((M) & 0x80) ? 23 : 7), \
172 (((M) & 0x01) ? 24 : 8), \
173 (((M) & 0x02) ? 25 : 9), \
174 (((M) & 0x04) ? 26 : 10), \
175 (((M) & 0x08) ? 27 : 11), \
176 (((M) & 0x10) ? 28 : 12), \
177 (((M) & 0x20) ? 29 : 13), \
178 (((M) & 0x40) ? 30 : 14), \
179 (((M) & 0x80) ? 31 : 15)); })
Craig Toppera89747d2011-12-20 09:55:26 +0000180
181static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000182_mm256_cmpeq_epi8(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000183{
David Blaikie3302f2b2013-01-16 23:08:36 +0000184 return (__m256i)((__v32qi)__a == (__v32qi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000185}
186
187static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000188_mm256_cmpeq_epi16(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000189{
David Blaikie3302f2b2013-01-16 23:08:36 +0000190 return (__m256i)((__v16hi)__a == (__v16hi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000191}
192
193static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000194_mm256_cmpeq_epi32(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000195{
David Blaikie3302f2b2013-01-16 23:08:36 +0000196 return (__m256i)((__v8si)__a == (__v8si)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000197}
198
199static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000200_mm256_cmpeq_epi64(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000201{
David Blaikie3302f2b2013-01-16 23:08:36 +0000202 return (__m256i)(__a == __b);
Craig Toppera89747d2011-12-20 09:55:26 +0000203}
204
205static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000206_mm256_cmpgt_epi8(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000207{
David Blaikie3302f2b2013-01-16 23:08:36 +0000208 return (__m256i)((__v32qi)__a > (__v32qi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000209}
210
211static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000212_mm256_cmpgt_epi16(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000213{
David Blaikie3302f2b2013-01-16 23:08:36 +0000214 return (__m256i)((__v16hi)__a > (__v16hi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000215}
216
217static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000218_mm256_cmpgt_epi32(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000219{
David Blaikie3302f2b2013-01-16 23:08:36 +0000220 return (__m256i)((__v8si)__a > (__v8si)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000221}
222
223static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000224_mm256_cmpgt_epi64(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000225{
David Blaikie3302f2b2013-01-16 23:08:36 +0000226 return (__m256i)(__a > __b);
Craig Toppera89747d2011-12-20 09:55:26 +0000227}
228
229static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000230_mm256_hadd_epi16(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000231{
David Blaikie3302f2b2013-01-16 23:08:36 +0000232 return (__m256i)__builtin_ia32_phaddw256((__v16hi)__a, (__v16hi)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000233}
234
235static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000236_mm256_hadd_epi32(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000237{
David Blaikie3302f2b2013-01-16 23:08:36 +0000238 return (__m256i)__builtin_ia32_phaddd256((__v8si)__a, (__v8si)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000239}
240
241static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000242_mm256_hadds_epi16(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000243{
David Blaikie3302f2b2013-01-16 23:08:36 +0000244 return (__m256i)__builtin_ia32_phaddsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000245}
246
247static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000248_mm256_hsub_epi16(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000249{
David Blaikie3302f2b2013-01-16 23:08:36 +0000250 return (__m256i)__builtin_ia32_phsubw256((__v16hi)__a, (__v16hi)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000251}
252
253static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000254_mm256_hsub_epi32(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000255{
David Blaikie3302f2b2013-01-16 23:08:36 +0000256 return (__m256i)__builtin_ia32_phsubd256((__v8si)__a, (__v8si)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000257}
258
259static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000260_mm256_hsubs_epi16(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000261{
David Blaikie3302f2b2013-01-16 23:08:36 +0000262 return (__m256i)__builtin_ia32_phsubsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000263}
264
265static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000266_mm256_maddubs_epi16(__m256i __a, __m256i __b)
Craig Toppera73baa82011-12-21 08:35:05 +0000267{
David Blaikie3302f2b2013-01-16 23:08:36 +0000268 return (__m256i)__builtin_ia32_pmaddubsw256((__v32qi)__a, (__v32qi)__b);
Craig Toppera73baa82011-12-21 08:35:05 +0000269}
270
271static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000272_mm256_madd_epi16(__m256i __a, __m256i __b)
Craig Toppera73baa82011-12-21 08:35:05 +0000273{
David Blaikie3302f2b2013-01-16 23:08:36 +0000274 return (__m256i)__builtin_ia32_pmaddwd256((__v16hi)__a, (__v16hi)__b);
Craig Toppera73baa82011-12-21 08:35:05 +0000275}
276
277static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000278_mm256_max_epi8(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000279{
David Blaikie3302f2b2013-01-16 23:08:36 +0000280 return (__m256i)__builtin_ia32_pmaxsb256((__v32qi)__a, (__v32qi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000281}
282
283static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000284_mm256_max_epi16(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000285{
David Blaikie3302f2b2013-01-16 23:08:36 +0000286 return (__m256i)__builtin_ia32_pmaxsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000287}
288
289static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000290_mm256_max_epi32(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000291{
David Blaikie3302f2b2013-01-16 23:08:36 +0000292 return (__m256i)__builtin_ia32_pmaxsd256((__v8si)__a, (__v8si)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000293}
294
295static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000296_mm256_max_epu8(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000297{
David Blaikie3302f2b2013-01-16 23:08:36 +0000298 return (__m256i)__builtin_ia32_pmaxub256((__v32qi)__a, (__v32qi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000299}
300
301static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000302_mm256_max_epu16(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000303{
David Blaikie3302f2b2013-01-16 23:08:36 +0000304 return (__m256i)__builtin_ia32_pmaxuw256((__v16hi)__a, (__v16hi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000305}
306
307static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000308_mm256_max_epu32(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000309{
David Blaikie3302f2b2013-01-16 23:08:36 +0000310 return (__m256i)__builtin_ia32_pmaxud256((__v8si)__a, (__v8si)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000311}
312
313static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000314_mm256_min_epi8(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000315{
David Blaikie3302f2b2013-01-16 23:08:36 +0000316 return (__m256i)__builtin_ia32_pminsb256((__v32qi)__a, (__v32qi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000317}
318
319static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000320_mm256_min_epi16(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000321{
David Blaikie3302f2b2013-01-16 23:08:36 +0000322 return (__m256i)__builtin_ia32_pminsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000323}
324
325static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000326_mm256_min_epi32(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000327{
David Blaikie3302f2b2013-01-16 23:08:36 +0000328 return (__m256i)__builtin_ia32_pminsd256((__v8si)__a, (__v8si)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000329}
330
331static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000332_mm256_min_epu8(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000333{
David Blaikie3302f2b2013-01-16 23:08:36 +0000334 return (__m256i)__builtin_ia32_pminub256((__v32qi)__a, (__v32qi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000335}
336
337static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000338_mm256_min_epu16(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000339{
David Blaikie3302f2b2013-01-16 23:08:36 +0000340 return (__m256i)__builtin_ia32_pminuw256 ((__v16hi)__a, (__v16hi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000341}
342
343static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000344_mm256_min_epu32(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000345{
David Blaikie3302f2b2013-01-16 23:08:36 +0000346 return (__m256i)__builtin_ia32_pminud256((__v8si)__a, (__v8si)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000347}
348
349static __inline__ int __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000350_mm256_movemask_epi8(__m256i __a)
Craig Topper1f2460a2011-12-22 09:18:58 +0000351{
David Blaikie3302f2b2013-01-16 23:08:36 +0000352 return __builtin_ia32_pmovmskb256((__v32qi)__a);
Craig Topper1f2460a2011-12-22 09:18:58 +0000353}
354
355static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
356_mm256_cvtepi8_epi16(__m128i __V)
357{
358 return (__m256i)__builtin_ia32_pmovsxbw256((__v16qi)__V);
359}
360
361static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
362_mm256_cvtepi8_epi32(__m128i __V)
363{
364 return (__m256i)__builtin_ia32_pmovsxbd256((__v16qi)__V);
365}
366
367static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
368_mm256_cvtepi8_epi64(__m128i __V)
369{
370 return (__m256i)__builtin_ia32_pmovsxbq256((__v16qi)__V);
371}
372
373static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
374_mm256_cvtepi16_epi32(__m128i __V)
375{
376 return (__m256i)__builtin_ia32_pmovsxwd256((__v8hi)__V);
377}
378
379static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
380_mm256_cvtepi16_epi64(__m128i __V)
381{
382 return (__m256i)__builtin_ia32_pmovsxwq256((__v8hi)__V);
383}
384
385static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
386_mm256_cvtepi32_epi64(__m128i __V)
387{
388 return (__m256i)__builtin_ia32_pmovsxdq256((__v4si)__V);
389}
390
391static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
392_mm256_cvtepu8_epi16(__m128i __V)
393{
394 return (__m256i)__builtin_ia32_pmovzxbw256((__v16qi)__V);
395}
396
397static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
398_mm256_cvtepu8_epi32(__m128i __V)
399{
400 return (__m256i)__builtin_ia32_pmovzxbd256((__v16qi)__V);
401}
402
403static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
404_mm256_cvtepu8_epi64(__m128i __V)
405{
406 return (__m256i)__builtin_ia32_pmovzxbq256((__v16qi)__V);
407}
408
409static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
410_mm256_cvtepu16_epi32(__m128i __V)
411{
412 return (__m256i)__builtin_ia32_pmovzxwd256((__v8hi)__V);
413}
414
415static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
416_mm256_cvtepu16_epi64(__m128i __V)
417{
418 return (__m256i)__builtin_ia32_pmovzxwq256((__v8hi)__V);
419}
420
421static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
422_mm256_cvtepu32_epi64(__m128i __V)
423{
424 return (__m256i)__builtin_ia32_pmovzxdq256((__v4si)__V);
425}
426
Craig Topper235a3652011-12-23 08:31:16 +0000427static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000428_mm256_mul_epi32(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000429{
David Blaikie3302f2b2013-01-16 23:08:36 +0000430 return (__m256i)__builtin_ia32_pmuldq256((__v8si)__a, (__v8si)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000431}
432
433static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000434_mm256_mulhrs_epi16(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000435{
David Blaikie3302f2b2013-01-16 23:08:36 +0000436 return (__m256i)__builtin_ia32_pmulhrsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000437}
438
439static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000440_mm256_mulhi_epu16(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000441{
David Blaikie3302f2b2013-01-16 23:08:36 +0000442 return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000443}
444
445static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000446_mm256_mulhi_epi16(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000447{
David Blaikie3302f2b2013-01-16 23:08:36 +0000448 return (__m256i)__builtin_ia32_pmulhw256((__v16hi)__a, (__v16hi)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000449}
450
451static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000452_mm256_mullo_epi16(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000453{
David Blaikie3302f2b2013-01-16 23:08:36 +0000454 return (__m256i)((__v16hi)__a * (__v16hi)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000455}
456
457static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000458_mm256_mullo_epi32 (__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000459{
David Blaikie3302f2b2013-01-16 23:08:36 +0000460 return (__m256i)((__v8si)__a * (__v8si)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000461}
462
463static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000464_mm256_mul_epu32(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000465{
David Blaikie3302f2b2013-01-16 23:08:36 +0000466 return __builtin_ia32_pmuludq256((__v8si)__a, (__v8si)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000467}
468
Craig Topper1f2460a2011-12-22 09:18:58 +0000469static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000470_mm256_or_si256(__m256i __a, __m256i __b)
Craig Toppera557e1c2011-12-19 09:03:48 +0000471{
David Blaikie3302f2b2013-01-16 23:08:36 +0000472 return __a | __b;
Craig Toppera557e1c2011-12-19 09:03:48 +0000473}
474
475static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000476_mm256_sad_epu8(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000477{
David Blaikie3302f2b2013-01-16 23:08:36 +0000478 return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000479}
480
481static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000482_mm256_shuffle_epi8(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000483{
David Blaikie3302f2b2013-01-16 23:08:36 +0000484 return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000485}
486
487#define _mm256_shuffle_epi32(a, imm) __extension__ ({ \
488 __m256i __a = (a); \
489 (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)_mm256_set1_epi32(0), \
490 (imm) & 0x3, ((imm) & 0xc) >> 2, \
491 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
492 4 + (((imm) & 0x03) >> 0), \
493 4 + (((imm) & 0x0c) >> 2), \
494 4 + (((imm) & 0x30) >> 4), \
495 4 + (((imm) & 0xc0) >> 6)); })
496
497#define _mm256_shufflehi_epi16(a, imm) __extension__ ({ \
498 __m256i __a = (a); \
499 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
500 0, 1, 2, 3, \
501 4 + (((imm) & 0x03) >> 0), \
502 4 + (((imm) & 0x0c) >> 2), \
503 4 + (((imm) & 0x30) >> 4), \
504 4 + (((imm) & 0xc0) >> 6), \
505 8, 9, 10, 11, \
506 12 + (((imm) & 0x03) >> 0), \
507 12 + (((imm) & 0x0c) >> 2), \
508 12 + (((imm) & 0x30) >> 4), \
509 12 + (((imm) & 0xc0) >> 6)); })
510
511#define _mm256_shufflelo_epi16(a, imm) __extension__ ({ \
512 __m256i __a = (a); \
513 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
514 (imm) & 0x3,((imm) & 0xc) >> 2, \
515 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
516 4, 5, 6, 7, \
517 8 + (((imm) & 0x03) >> 0), \
518 8 + (((imm) & 0x0c) >> 2), \
519 8 + (((imm) & 0x30) >> 4), \
520 8 + (((imm) & 0xc0) >> 6), \
521 12, 13, 14, 15); })
522
523static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000524_mm256_sign_epi8(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000525{
David Blaikie3302f2b2013-01-16 23:08:36 +0000526 return (__m256i)__builtin_ia32_psignb256((__v32qi)__a, (__v32qi)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000527}
528
529static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000530_mm256_sign_epi16(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000531{
David Blaikie3302f2b2013-01-16 23:08:36 +0000532 return (__m256i)__builtin_ia32_psignw256((__v16hi)__a, (__v16hi)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000533}
534
535static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000536_mm256_sign_epi32(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000537{
David Blaikie3302f2b2013-01-16 23:08:36 +0000538 return (__m256i)__builtin_ia32_psignd256((__v8si)__a, (__v8si)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000539}
540
541#define _mm256_slli_si256(a, count) __extension__ ({ \
542 __m256i __a = (a); \
543 (__m256i)__builtin_ia32_pslldqi256(__a, (count)*8); })
544
Michael Kuperstein61681832015-05-19 13:05:46 +0000545#define _mm256_bslli_epi128(a, count) _mm256_slli_si256((a), (count))
546
Craig Topperf4bb9522011-12-24 03:28:57 +0000547static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000548_mm256_slli_epi16(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000549{
David Blaikie3302f2b2013-01-16 23:08:36 +0000550 return (__m256i)__builtin_ia32_psllwi256((__v16hi)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000551}
552
553static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000554_mm256_sll_epi16(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000555{
David Blaikie3302f2b2013-01-16 23:08:36 +0000556 return (__m256i)__builtin_ia32_psllw256((__v16hi)__a, (__v8hi)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000557}
558
559static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000560_mm256_slli_epi32(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000561{
David Blaikie3302f2b2013-01-16 23:08:36 +0000562 return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000563}
564
565static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000566_mm256_sll_epi32(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000567{
David Blaikie3302f2b2013-01-16 23:08:36 +0000568 return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000569}
570
571static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000572_mm256_slli_epi64(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000573{
David Blaikie3302f2b2013-01-16 23:08:36 +0000574 return __builtin_ia32_psllqi256(__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000575}
576
577static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000578_mm256_sll_epi64(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000579{
David Blaikie3302f2b2013-01-16 23:08:36 +0000580 return __builtin_ia32_psllq256(__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000581}
582
583static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000584_mm256_srai_epi16(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000585{
David Blaikie3302f2b2013-01-16 23:08:36 +0000586 return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000587}
588
589static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000590_mm256_sra_epi16(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000591{
David Blaikie3302f2b2013-01-16 23:08:36 +0000592 return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000593}
594
595static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000596_mm256_srai_epi32(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000597{
David Blaikie3302f2b2013-01-16 23:08:36 +0000598 return (__m256i)__builtin_ia32_psradi256((__v8si)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000599}
600
601static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000602_mm256_sra_epi32(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000603{
David Blaikie3302f2b2013-01-16 23:08:36 +0000604 return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000605}
606
607#define _mm256_srli_si256(a, count) __extension__ ({ \
608 __m256i __a = (a); \
609 (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); })
610
Michael Kuperstein61681832015-05-19 13:05:46 +0000611#define _mm256_bsrli_epi128(a, count) _mm256_srli_si256((a), (count))
612
Craig Topperf4bb9522011-12-24 03:28:57 +0000613static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000614_mm256_srli_epi16(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000615{
David Blaikie3302f2b2013-01-16 23:08:36 +0000616 return (__m256i)__builtin_ia32_psrlwi256((__v16hi)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000617}
618
619static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000620_mm256_srl_epi16(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000621{
David Blaikie3302f2b2013-01-16 23:08:36 +0000622 return (__m256i)__builtin_ia32_psrlw256((__v16hi)__a, (__v8hi)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000623}
624
625static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000626_mm256_srli_epi32(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000627{
David Blaikie3302f2b2013-01-16 23:08:36 +0000628 return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000629}
630
631static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000632_mm256_srl_epi32(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000633{
David Blaikie3302f2b2013-01-16 23:08:36 +0000634 return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000635}
636
637static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000638_mm256_srli_epi64(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000639{
David Blaikie3302f2b2013-01-16 23:08:36 +0000640 return __builtin_ia32_psrlqi256(__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000641}
642
643static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000644_mm256_srl_epi64(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000645{
David Blaikie3302f2b2013-01-16 23:08:36 +0000646 return __builtin_ia32_psrlq256(__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000647}
648
649static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000650_mm256_sub_epi8(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +0000651{
David Blaikie3302f2b2013-01-16 23:08:36 +0000652 return (__m256i)((__v32qi)__a - (__v32qi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +0000653}
654
655static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000656_mm256_sub_epi16(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +0000657{
David Blaikie3302f2b2013-01-16 23:08:36 +0000658 return (__m256i)((__v16hi)__a - (__v16hi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +0000659}
660
661static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000662_mm256_sub_epi32(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +0000663{
David Blaikie3302f2b2013-01-16 23:08:36 +0000664 return (__m256i)((__v8si)__a - (__v8si)__b);
Craig Topperdec792e2011-12-19 05:04:33 +0000665}
666
667static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000668_mm256_sub_epi64(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +0000669{
David Blaikie3302f2b2013-01-16 23:08:36 +0000670 return __a - __b;
Craig Topperdec792e2011-12-19 05:04:33 +0000671}
Craig Topper94aba2c2011-12-19 07:03:25 +0000672
673static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000674_mm256_subs_epi8(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000675{
David Blaikie3302f2b2013-01-16 23:08:36 +0000676 return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000677}
678
679static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000680_mm256_subs_epi16(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000681{
David Blaikie3302f2b2013-01-16 23:08:36 +0000682 return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000683}
684
685static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000686_mm256_subs_epu8(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000687{
David Blaikie3302f2b2013-01-16 23:08:36 +0000688 return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000689}
690
691static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000692_mm256_subs_epu16(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000693{
David Blaikie3302f2b2013-01-16 23:08:36 +0000694 return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000695}
696
Craig Toppera557e1c2011-12-19 09:03:48 +0000697static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000698_mm256_unpackhi_epi8(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000699{
David Blaikie3302f2b2013-01-16 23:08:36 +0000700 return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 8, 32+8, 9, 32+9, 10, 32+10, 11, 32+11, 12, 32+12, 13, 32+13, 14, 32+14, 15, 32+15, 24, 32+24, 25, 32+25, 26, 32+26, 27, 32+27, 28, 32+28, 29, 32+29, 30, 32+30, 31, 32+31);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000701}
702
703static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000704_mm256_unpackhi_epi16(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000705{
David Blaikie3302f2b2013-01-16 23:08:36 +0000706 return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000707}
708
709static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000710_mm256_unpackhi_epi32(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000711{
David Blaikie3302f2b2013-01-16 23:08:36 +0000712 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 2, 8+2, 3, 8+3, 6, 8+6, 7, 8+7);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000713}
714
715static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000716_mm256_unpackhi_epi64(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000717{
David Blaikie3302f2b2013-01-16 23:08:36 +0000718 return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000719}
720
721static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000722_mm256_unpacklo_epi8(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000723{
David Blaikie3302f2b2013-01-16 23:08:36 +0000724 return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 0, 32+0, 1, 32+1, 2, 32+2, 3, 32+3, 4, 32+4, 5, 32+5, 6, 32+6, 7, 32+7, 16, 32+16, 17, 32+17, 18, 32+18, 19, 32+19, 20, 32+20, 21, 32+21, 22, 32+22, 23, 32+23);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000725}
726
727static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000728_mm256_unpacklo_epi16(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000729{
David Blaikie3302f2b2013-01-16 23:08:36 +0000730 return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000731}
732
733static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000734_mm256_unpacklo_epi32(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000735{
David Blaikie3302f2b2013-01-16 23:08:36 +0000736 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 0, 8+0, 1, 8+1, 4, 8+4, 5, 8+5);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000737}
738
739static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000740_mm256_unpacklo_epi64(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000741{
David Blaikie3302f2b2013-01-16 23:08:36 +0000742 return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000743}
744
745static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000746_mm256_xor_si256(__m256i __a, __m256i __b)
Craig Toppera557e1c2011-12-19 09:03:48 +0000747{
David Blaikie3302f2b2013-01-16 23:08:36 +0000748 return __a ^ __b;
Craig Toppera557e1c2011-12-19 09:03:48 +0000749}
Craig Topper94798952011-12-24 05:19:29 +0000750
751static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
752_mm256_stream_load_si256(__m256i *__V)
753{
754 return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V);
755}
756
757static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
758_mm_broadcastss_ps(__m128 __X)
759{
760 return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X);
761}
762
763static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
764_mm256_broadcastss_ps(__m128 __X)
765{
766 return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X);
767}
768
769static __inline__ __m256d __attribute__((__always_inline__, __nodebug__))
770_mm256_broadcastsd_pd(__m128d __X)
771{
772 return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X);
773}
774
775static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Juergen Ributzka2c2dbf42013-08-17 16:40:09 +0000776_mm256_broadcastsi128_si256(__m128i __X)
Craig Topper94798952011-12-24 05:19:29 +0000777{
Juergen Ributzka9baa03f2015-03-03 17:22:53 +0000778 return (__m256i)__builtin_shufflevector(__X, __X, 0, 1, 0, 1);
Craig Topper94798952011-12-24 05:19:29 +0000779}
780
781#define _mm_blend_epi32(V1, V2, M) __extension__ ({ \
782 __m128i __V1 = (V1); \
783 __m128i __V2 = (V2); \
Filipe Cabecinhas5d289b42014-05-13 02:37:02 +0000784 (__m128i)__builtin_shufflevector((__v4si)__V1, (__v4si)__V2, \
785 (((M) & 0x01) ? 4 : 0), \
786 (((M) & 0x02) ? 5 : 1), \
787 (((M) & 0x04) ? 6 : 2), \
788 (((M) & 0x08) ? 7 : 3)); })
Craig Topper94798952011-12-24 05:19:29 +0000789
790#define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \
791 __m256i __V1 = (V1); \
792 __m256i __V2 = (V2); \
Filipe Cabecinhas5d289b42014-05-13 02:37:02 +0000793 (__m256i)__builtin_shufflevector((__v8si)__V1, (__v8si)__V2, \
794 (((M) & 0x01) ? 8 : 0), \
795 (((M) & 0x02) ? 9 : 1), \
796 (((M) & 0x04) ? 10 : 2), \
797 (((M) & 0x08) ? 11 : 3), \
798 (((M) & 0x10) ? 12 : 4), \
799 (((M) & 0x20) ? 13 : 5), \
800 (((M) & 0x40) ? 14 : 6), \
801 (((M) & 0x80) ? 15 : 7)); })
Craig Topper9f009482011-12-24 07:55:14 +0000802
803static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
804_mm256_broadcastb_epi8(__m128i __X)
805{
806 return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X);
807}
808
809static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
810_mm256_broadcastw_epi16(__m128i __X)
811{
812 return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X);
813}
814
815static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
816_mm256_broadcastd_epi32(__m128i __X)
817{
818 return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X);
819}
820
821static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
822_mm256_broadcastq_epi64(__m128i __X)
823{
824 return (__m256i)__builtin_ia32_pbroadcastq256(__X);
825}
826
827static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
828_mm_broadcastb_epi8(__m128i __X)
829{
830 return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X);
831}
832
833static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
834_mm_broadcastw_epi16(__m128i __X)
835{
836 return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X);
837}
838
839
840static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
841_mm_broadcastd_epi32(__m128i __X)
842{
843 return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X);
844}
845
846static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
847_mm_broadcastq_epi64(__m128i __X)
848{
849 return (__m128i)__builtin_ia32_pbroadcastq128(__X);
850}
851
852static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000853_mm256_permutevar8x32_epi32(__m256i __a, __m256i __b)
Craig Topper9f009482011-12-24 07:55:14 +0000854{
David Blaikie3302f2b2013-01-16 23:08:36 +0000855 return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b);
Craig Topper9f009482011-12-24 07:55:14 +0000856}
857
858#define _mm256_permute4x64_pd(V, M) __extension__ ({ \
859 __m256d __V = (V); \
Craig Topper8e578552012-04-15 22:18:10 +0000860 (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \
861 (M) & 0x3, ((M) & 0xc) >> 2, \
862 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
Craig Topper9f009482011-12-24 07:55:14 +0000863
864static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
David Blaikie3302f2b2013-01-16 23:08:36 +0000865_mm256_permutevar8x32_ps(__m256 __a, __m256 __b)
Craig Topper9f009482011-12-24 07:55:14 +0000866{
David Blaikie3302f2b2013-01-16 23:08:36 +0000867 return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b);
Craig Topper9f009482011-12-24 07:55:14 +0000868}
869
870#define _mm256_permute4x64_epi64(V, M) __extension__ ({ \
871 __m256i __V = (V); \
Craig Topper8e578552012-04-15 22:18:10 +0000872 (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \
873 (M) & 0x3, ((M) & 0xc) >> 2, \
874 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
Craig Topper9f009482011-12-24 07:55:14 +0000875
876#define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \
877 __m256i __V1 = (V1); \
878 __m256i __V2 = (V2); \
Craig Topper26e74e52012-04-17 05:16:56 +0000879 (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); })
Craig Topper175543a2011-12-24 17:20:15 +0000880
Sanjay Patel0a6da5d2015-03-12 21:54:24 +0000881#define _mm256_extracti128_si256(V, M) __extension__ ({ \
882 (__m128i)__builtin_shufflevector( \
883 (__v4di)(V), \
884 (__v4di)(_mm256_setzero_si256()), \
885 (((M) & 1) ? 2 : 0), \
886 (((M) & 1) ? 3 : 1) );})
Craig Topper175543a2011-12-24 17:20:15 +0000887
Sanjay Patel0a6da5d2015-03-12 21:54:24 +0000888#define _mm256_inserti128_si256(V1, V2, M) __extension__ ({ \
889 (__m256i)__builtin_shufflevector( \
890 (__v4di)(V1), \
891 (__v4di)_mm256_castsi128_si256((__m128i)(V2)), \
892 (((M) & 1) ? 0 : 4), \
893 (((M) & 1) ? 1 : 5), \
894 (((M) & 1) ? 4 : 2), \
895 (((M) & 1) ? 5 : 3) );})
Craig Topper175543a2011-12-24 17:20:15 +0000896
897static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
898_mm256_maskload_epi32(int const *__X, __m256i __M)
899{
900 return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M);
901}
902
903static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
904_mm256_maskload_epi64(long long const *__X, __m256i __M)
905{
906 return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M);
907}
908
909static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
910_mm_maskload_epi32(int const *__X, __m128i __M)
911{
912 return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M);
913}
914
915static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
916_mm_maskload_epi64(long long const *__X, __m128i __M)
917{
918 return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M);
919}
920
921static __inline__ void __attribute__((__always_inline__, __nodebug__))
922_mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y)
923{
924 __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y);
925}
926
927static __inline__ void __attribute__((__always_inline__, __nodebug__))
928_mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y)
929{
930 __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y);
931}
932
933static __inline__ void __attribute__((__always_inline__, __nodebug__))
934_mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y)
935{
936 __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y);
937}
938
939static __inline__ void __attribute__((__always_inline__, __nodebug__))
940_mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y)
941{
942 __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y);
943}
944
945static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
946_mm256_sllv_epi32(__m256i __X, __m256i __Y)
947{
948 return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y);
949}
950
951static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
952_mm_sllv_epi32(__m128i __X, __m128i __Y)
953{
954 return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y);
955}
956
957static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
958_mm256_sllv_epi64(__m256i __X, __m256i __Y)
959{
960 return (__m256i)__builtin_ia32_psllv4di(__X, __Y);
961}
962
963static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
964_mm_sllv_epi64(__m128i __X, __m128i __Y)
965{
966 return (__m128i)__builtin_ia32_psllv2di(__X, __Y);
967}
968
969static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
970_mm256_srav_epi32(__m256i __X, __m256i __Y)
971{
972 return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y);
973}
974
975static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
976_mm_srav_epi32(__m128i __X, __m128i __Y)
977{
978 return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y);
979}
980
981static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
982_mm256_srlv_epi32(__m256i __X, __m256i __Y)
983{
984 return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y);
985}
986
987static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
988_mm_srlv_epi32(__m128i __X, __m128i __Y)
989{
990 return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y);
991}
992
993static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
994_mm256_srlv_epi64(__m256i __X, __m256i __Y)
995{
996 return (__m256i)__builtin_ia32_psrlv4di(__X, __Y);
997}
998
999static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1000_mm_srlv_epi64(__m128i __X, __m128i __Y)
1001{
1002 return (__m128i)__builtin_ia32_psrlv2di(__X, __Y);
1003}
Manman Renadd5e9e2012-06-26 19:55:09 +00001004
1005#define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
1006 __m128d __a = (a); \
1007 double const *__m = (m); \
1008 __m128i __i = (i); \
1009 __m128d __mask = (mask); \
1010 (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \
1011 (__v4si)__i, (__v2df)__mask, (s)); })
1012
1013#define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
1014 __m256d __a = (a); \
1015 double const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001016 __m128i __i = (i); \
Manman Renadd5e9e2012-06-26 19:55:09 +00001017 __m256d __mask = (mask); \
1018 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \
Manman Ren86c32502012-06-29 00:54:35 +00001019 (__v4si)__i, (__v4df)__mask, (s)); })
Manman Renadd5e9e2012-06-26 19:55:09 +00001020
1021#define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
1022 __m128d __a = (a); \
1023 double const *__m = (m); \
1024 __m128i __i = (i); \
1025 __m128d __mask = (mask); \
1026 (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \
1027 (__v2di)__i, (__v2df)__mask, (s)); })
1028
1029#define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
1030 __m256d __a = (a); \
1031 double const *__m = (m); \
1032 __m256i __i = (i); \
1033 __m256d __mask = (mask); \
1034 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \
1035 (__v4di)__i, (__v4df)__mask, (s)); })
1036
1037#define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1038 __m128 __a = (a); \
1039 float const *__m = (m); \
1040 __m128i __i = (i); \
1041 __m128 __mask = (mask); \
1042 (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \
1043 (__v4si)__i, (__v4sf)__mask, (s)); })
1044
1045#define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1046 __m256 __a = (a); \
1047 float const *__m = (m); \
1048 __m256i __i = (i); \
1049 __m256 __mask = (mask); \
1050 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \
1051 (__v8si)__i, (__v8sf)__mask, (s)); })
1052
1053#define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1054 __m128 __a = (a); \
1055 float const *__m = (m); \
1056 __m128i __i = (i); \
1057 __m128 __mask = (mask); \
1058 (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \
1059 (__v2di)__i, (__v4sf)__mask, (s)); })
1060
1061#define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
Manman Ren86c32502012-06-29 00:54:35 +00001062 __m128 __a = (a); \
Manman Renadd5e9e2012-06-26 19:55:09 +00001063 float const *__m = (m); \
1064 __m256i __i = (i); \
Manman Ren86c32502012-06-29 00:54:35 +00001065 __m128 __mask = (mask); \
1066 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \
1067 (__v4di)__i, (__v4sf)__mask, (s)); })
1068
1069#define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1070 __m128i __a = (a); \
1071 int const *__m = (m); \
1072 __m128i __i = (i); \
1073 __m128i __mask = (mask); \
1074 (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \
1075 (__v4si)__i, (__v4si)__mask, (s)); })
1076
1077#define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1078 __m256i __a = (a); \
1079 int const *__m = (m); \
1080 __m256i __i = (i); \
1081 __m256i __mask = (mask); \
1082 (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \
1083 (__v8si)__i, (__v8si)__mask, (s)); })
1084
1085#define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1086 __m128i __a = (a); \
1087 int const *__m = (m); \
1088 __m128i __i = (i); \
1089 __m128i __mask = (mask); \
1090 (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \
1091 (__v2di)__i, (__v4si)__mask, (s)); })
1092
1093#define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1094 __m128i __a = (a); \
1095 int const *__m = (m); \
1096 __m256i __i = (i); \
1097 __m128i __mask = (mask); \
1098 (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \
1099 (__v4di)__i, (__v4si)__mask, (s)); })
1100
1101#define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1102 __m128i __a = (a); \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001103 long long const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001104 __m128i __i = (i); \
1105 __m128i __mask = (mask); \
1106 (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \
1107 (__v4si)__i, (__v2di)__mask, (s)); })
1108
1109#define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1110 __m256i __a = (a); \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001111 long long const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001112 __m128i __i = (i); \
1113 __m256i __mask = (mask); \
1114 (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \
1115 (__v4si)__i, (__v4di)__mask, (s)); })
1116
1117#define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1118 __m128i __a = (a); \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001119 long long const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001120 __m128i __i = (i); \
1121 __m128i __mask = (mask); \
1122 (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \
1123 (__v2di)__i, (__v2di)__mask, (s)); })
1124
1125#define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1126 __m256i __a = (a); \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001127 long long const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001128 __m256i __i = (i); \
1129 __m256i __mask = (mask); \
1130 (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \
1131 (__v4di)__i, (__v4di)__mask, (s)); })
Manman Renf865ba02012-06-29 05:19:13 +00001132
1133#define _mm_i32gather_pd(m, i, s) __extension__ ({ \
1134 double const *__m = (m); \
1135 __m128i __i = (i); \
1136 (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \
1137 (const __v2df *)__m, (__v4si)__i, \
1138 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1139
1140#define _mm256_i32gather_pd(m, i, s) __extension__ ({ \
1141 double const *__m = (m); \
1142 __m128i __i = (i); \
1143 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \
1144 (const __v4df *)__m, (__v4si)__i, \
1145 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1146
1147#define _mm_i64gather_pd(m, i, s) __extension__ ({ \
1148 double const *__m = (m); \
1149 __m128i __i = (i); \
1150 (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \
1151 (const __v2df *)__m, (__v2di)__i, \
1152 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1153
1154#define _mm256_i64gather_pd(m, i, s) __extension__ ({ \
1155 double const *__m = (m); \
1156 __m256i __i = (i); \
1157 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \
1158 (const __v4df *)__m, (__v4di)__i, \
1159 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1160
1161#define _mm_i32gather_ps(m, i, s) __extension__ ({ \
1162 float const *__m = (m); \
1163 __m128i __i = (i); \
1164 (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \
1165 (const __v4sf *)__m, (__v4si)__i, \
1166 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1167
1168#define _mm256_i32gather_ps(m, i, s) __extension__ ({ \
1169 float const *__m = (m); \
1170 __m256i __i = (i); \
1171 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \
1172 (const __v8sf *)__m, (__v8si)__i, \
1173 (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); })
1174
1175#define _mm_i64gather_ps(m, i, s) __extension__ ({ \
1176 float const *__m = (m); \
1177 __m128i __i = (i); \
1178 (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \
1179 (const __v4sf *)__m, (__v2di)__i, \
1180 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1181
1182#define _mm256_i64gather_ps(m, i, s) __extension__ ({ \
1183 float const *__m = (m); \
1184 __m256i __i = (i); \
1185 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \
1186 (const __v4sf *)__m, (__v4di)__i, \
1187 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1188
1189#define _mm_i32gather_epi32(m, i, s) __extension__ ({ \
1190 int const *__m = (m); \
1191 __m128i __i = (i); \
1192 (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \
1193 (const __v4si *)__m, (__v4si)__i, \
1194 (__v4si)_mm_set1_epi32(-1), (s)); })
1195
1196#define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \
1197 int const *__m = (m); \
1198 __m256i __i = (i); \
1199 (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \
1200 (const __v8si *)__m, (__v8si)__i, \
1201 (__v8si)_mm256_set1_epi32(-1), (s)); })
1202
1203#define _mm_i64gather_epi32(m, i, s) __extension__ ({ \
1204 int const *__m = (m); \
1205 __m128i __i = (i); \
1206 (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \
1207 (const __v4si *)__m, (__v2di)__i, \
1208 (__v4si)_mm_set1_epi32(-1), (s)); })
1209
1210#define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \
1211 int const *__m = (m); \
1212 __m256i __i = (i); \
1213 (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \
1214 (const __v4si *)__m, (__v4di)__i, \
1215 (__v4si)_mm_set1_epi32(-1), (s)); })
1216
1217#define _mm_i32gather_epi64(m, i, s) __extension__ ({ \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001218 long long const *__m = (m); \
Manman Renf865ba02012-06-29 05:19:13 +00001219 __m128i __i = (i); \
1220 (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \
1221 (const __v2di *)__m, (__v4si)__i, \
1222 (__v2di)_mm_set1_epi64x(-1), (s)); })
1223
1224#define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001225 long long const *__m = (m); \
Manman Renf865ba02012-06-29 05:19:13 +00001226 __m128i __i = (i); \
1227 (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \
1228 (const __v4di *)__m, (__v4si)__i, \
1229 (__v4di)_mm256_set1_epi64x(-1), (s)); })
1230
1231#define _mm_i64gather_epi64(m, i, s) __extension__ ({ \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001232 long long const *__m = (m); \
Manman Renf865ba02012-06-29 05:19:13 +00001233 __m128i __i = (i); \
1234 (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \
1235 (const __v2di *)__m, (__v2di)__i, \
1236 (__v2di)_mm_set1_epi64x(-1), (s)); })
1237
1238#define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001239 long long const *__m = (m); \
Manman Renf865ba02012-06-29 05:19:13 +00001240 __m256i __i = (i); \
1241 (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \
1242 (const __v4di *)__m, (__v4di)__i, \
1243 (__v4di)_mm256_set1_epi64x(-1), (s)); })
Richard Smith49e56442013-07-14 05:41:45 +00001244
1245#endif /* __AVX2INTRIN_H */