blob: e1e639de1ba8f252993e7c552fc82fccfe35310c [file] [log] [blame]
Ying Wang60999142013-01-07 13:59:36 -08001/*===---- 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
Stephen Hines996e4dc2013-08-13 01:04:14 -070028#ifndef __AVX2INTRIN_H
29#define __AVX2INTRIN_H
30
Ying Wang60999142013-01-07 13:59:36 -080031/* 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__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070035_mm256_abs_epi8(__m256i __a)
Ying Wang60999142013-01-07 13:59:36 -080036{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070037 return (__m256i)__builtin_ia32_pabsb256((__v32qi)__a);
Ying Wang60999142013-01-07 13:59:36 -080038}
39
40static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070041_mm256_abs_epi16(__m256i __a)
Ying Wang60999142013-01-07 13:59:36 -080042{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070043 return (__m256i)__builtin_ia32_pabsw256((__v16hi)__a);
Ying Wang60999142013-01-07 13:59:36 -080044}
45
46static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070047_mm256_abs_epi32(__m256i __a)
Ying Wang60999142013-01-07 13:59:36 -080048{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070049 return (__m256i)__builtin_ia32_pabsd256((__v8si)__a);
Ying Wang60999142013-01-07 13:59:36 -080050}
51
52static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070053_mm256_packs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -080054{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070055 return (__m256i)__builtin_ia32_packsswb256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -080056}
57
58static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070059_mm256_packs_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -080060{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070061 return (__m256i)__builtin_ia32_packssdw256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -080062}
63
64static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070065_mm256_packus_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -080066{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070067 return (__m256i)__builtin_ia32_packuswb256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -080068}
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__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070077_mm256_add_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -080078{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070079 return (__m256i)((__v32qi)__a + (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -080080}
81
82static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070083_mm256_add_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -080084{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070085 return (__m256i)((__v16hi)__a + (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -080086}
87
88static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070089_mm256_add_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -080090{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070091 return (__m256i)((__v8si)__a + (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -080092}
93
94static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070095_mm256_add_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -080096{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -070097 return __a + __b;
Ying Wang60999142013-01-07 13:59:36 -080098}
99
100static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700101_mm256_adds_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800102{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700103 return (__m256i)__builtin_ia32_paddsb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800104}
105
106static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700107_mm256_adds_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800108{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700109 return (__m256i)__builtin_ia32_paddsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800110}
111
112static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700113_mm256_adds_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800114{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700115 return (__m256i)__builtin_ia32_paddusb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800116}
117
118static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700119_mm256_adds_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800120{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700121 return (__m256i)__builtin_ia32_paddusw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800122}
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__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700130_mm256_and_si256(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800131{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700132 return __a & __b;
Ying Wang60999142013-01-07 13:59:36 -0800133}
134
135static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700136_mm256_andnot_si256(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800137{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700138 return ~__a & __b;
Ying Wang60999142013-01-07 13:59:36 -0800139}
140
141static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700142_mm256_avg_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800143{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700144 return (__m256i)__builtin_ia32_pavgb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800145}
146
147static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700148_mm256_avg_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800149{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700150 return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800151}
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); \
Pirama Arumuga Nainar37fb08a2015-05-11 14:34:37 -0700163 (__m256i)__builtin_shufflevector((__v16hi)__V1, (__v16hi)__V2, \
Stephen Hinese65db132014-05-30 13:26:31 -0700164 (((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)); })
Ying Wang60999142013-01-07 13:59:36 -0800180
181static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700182_mm256_cmpeq_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800183{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700184 return (__m256i)((__v32qi)__a == (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800185}
186
187static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700188_mm256_cmpeq_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800189{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700190 return (__m256i)((__v16hi)__a == (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800191}
192
193static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700194_mm256_cmpeq_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800195{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700196 return (__m256i)((__v8si)__a == (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800197}
198
199static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700200_mm256_cmpeq_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800201{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700202 return (__m256i)(__a == __b);
Ying Wang60999142013-01-07 13:59:36 -0800203}
204
205static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700206_mm256_cmpgt_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800207{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700208 return (__m256i)((__v32qi)__a > (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800209}
210
211static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700212_mm256_cmpgt_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800213{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700214 return (__m256i)((__v16hi)__a > (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800215}
216
217static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700218_mm256_cmpgt_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800219{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700220 return (__m256i)((__v8si)__a > (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800221}
222
223static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700224_mm256_cmpgt_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800225{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700226 return (__m256i)(__a > __b);
Ying Wang60999142013-01-07 13:59:36 -0800227}
228
229static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700230_mm256_hadd_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800231{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700232 return (__m256i)__builtin_ia32_phaddw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800233}
234
235static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700236_mm256_hadd_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800237{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700238 return (__m256i)__builtin_ia32_phaddd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800239}
240
241static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700242_mm256_hadds_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800243{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700244 return (__m256i)__builtin_ia32_phaddsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800245}
246
247static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700248_mm256_hsub_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800249{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700250 return (__m256i)__builtin_ia32_phsubw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800251}
252
253static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700254_mm256_hsub_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800255{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700256 return (__m256i)__builtin_ia32_phsubd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800257}
258
259static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700260_mm256_hsubs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800261{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700262 return (__m256i)__builtin_ia32_phsubsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800263}
264
265static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700266_mm256_maddubs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800267{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700268 return (__m256i)__builtin_ia32_pmaddubsw256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800269}
270
271static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700272_mm256_madd_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800273{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700274 return (__m256i)__builtin_ia32_pmaddwd256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800275}
276
277static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700278_mm256_max_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800279{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700280 return (__m256i)__builtin_ia32_pmaxsb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800281}
282
283static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700284_mm256_max_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800285{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700286 return (__m256i)__builtin_ia32_pmaxsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800287}
288
289static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700290_mm256_max_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800291{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700292 return (__m256i)__builtin_ia32_pmaxsd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800293}
294
295static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700296_mm256_max_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800297{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700298 return (__m256i)__builtin_ia32_pmaxub256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800299}
300
301static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700302_mm256_max_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800303{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700304 return (__m256i)__builtin_ia32_pmaxuw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800305}
306
307static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700308_mm256_max_epu32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800309{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700310 return (__m256i)__builtin_ia32_pmaxud256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800311}
312
313static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700314_mm256_min_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800315{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700316 return (__m256i)__builtin_ia32_pminsb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800317}
318
319static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700320_mm256_min_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800321{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700322 return (__m256i)__builtin_ia32_pminsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800323}
324
325static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700326_mm256_min_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800327{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700328 return (__m256i)__builtin_ia32_pminsd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800329}
330
331static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700332_mm256_min_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800333{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700334 return (__m256i)__builtin_ia32_pminub256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800335}
336
337static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700338_mm256_min_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800339{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700340 return (__m256i)__builtin_ia32_pminuw256 ((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800341}
342
343static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700344_mm256_min_epu32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800345{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700346 return (__m256i)__builtin_ia32_pminud256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800347}
348
349static __inline__ int __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700350_mm256_movemask_epi8(__m256i __a)
Ying Wang60999142013-01-07 13:59:36 -0800351{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700352 return __builtin_ia32_pmovmskb256((__v32qi)__a);
Ying Wang60999142013-01-07 13:59:36 -0800353}
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
427static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700428_mm256_mul_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800429{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700430 return (__m256i)__builtin_ia32_pmuldq256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800431}
432
433static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700434_mm256_mulhrs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800435{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700436 return (__m256i)__builtin_ia32_pmulhrsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800437}
438
439static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700440_mm256_mulhi_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800441{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700442 return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800443}
444
445static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700446_mm256_mulhi_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800447{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700448 return (__m256i)__builtin_ia32_pmulhw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800449}
450
451static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700452_mm256_mullo_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800453{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700454 return (__m256i)((__v16hi)__a * (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800455}
456
457static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700458_mm256_mullo_epi32 (__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800459{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700460 return (__m256i)((__v8si)__a * (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800461}
462
463static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700464_mm256_mul_epu32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800465{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700466 return __builtin_ia32_pmuludq256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800467}
468
469static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700470_mm256_or_si256(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800471{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700472 return __a | __b;
Ying Wang60999142013-01-07 13:59:36 -0800473}
474
475static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700476_mm256_sad_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800477{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700478 return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800479}
480
481static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700482_mm256_shuffle_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800483{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700484 return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800485}
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__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700524_mm256_sign_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800525{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700526 return (__m256i)__builtin_ia32_psignb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800527}
528
529static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700530_mm256_sign_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800531{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700532 return (__m256i)__builtin_ia32_psignw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800533}
534
535static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700536_mm256_sign_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800537{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700538 return (__m256i)__builtin_ia32_psignd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800539}
540
541#define _mm256_slli_si256(a, count) __extension__ ({ \
542 __m256i __a = (a); \
543 (__m256i)__builtin_ia32_pslldqi256(__a, (count)*8); })
544
Stephen Hines28c5e1e2015-08-13 18:18:46 -0700545#define _mm256_bslli_epi128(a, count) _mm256_slli_si256((a), (count))
546
Ying Wang60999142013-01-07 13:59:36 -0800547static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700548_mm256_slli_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800549{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700550 return (__m256i)__builtin_ia32_psllwi256((__v16hi)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800551}
552
553static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700554_mm256_sll_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800555{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700556 return (__m256i)__builtin_ia32_psllw256((__v16hi)__a, (__v8hi)__count);
Ying Wang60999142013-01-07 13:59:36 -0800557}
558
559static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700560_mm256_slli_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800561{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700562 return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800563}
564
565static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700566_mm256_sll_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800567{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700568 return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__count);
Ying Wang60999142013-01-07 13:59:36 -0800569}
570
571static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700572_mm256_slli_epi64(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800573{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700574 return __builtin_ia32_psllqi256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800575}
576
577static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700578_mm256_sll_epi64(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800579{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700580 return __builtin_ia32_psllq256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800581}
582
583static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700584_mm256_srai_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800585{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700586 return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800587}
588
589static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700590_mm256_sra_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800591{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700592 return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__count);
Ying Wang60999142013-01-07 13:59:36 -0800593}
594
595static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700596_mm256_srai_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800597{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700598 return (__m256i)__builtin_ia32_psradi256((__v8si)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800599}
600
601static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700602_mm256_sra_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800603{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700604 return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count);
Ying Wang60999142013-01-07 13:59:36 -0800605}
606
607#define _mm256_srli_si256(a, count) __extension__ ({ \
608 __m256i __a = (a); \
609 (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); })
610
Stephen Hines28c5e1e2015-08-13 18:18:46 -0700611#define _mm256_bsrli_epi128(a, count) _mm256_srli_si256((a), (count))
612
Ying Wang60999142013-01-07 13:59:36 -0800613static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700614_mm256_srli_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800615{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700616 return (__m256i)__builtin_ia32_psrlwi256((__v16hi)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800617}
618
619static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700620_mm256_srl_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800621{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700622 return (__m256i)__builtin_ia32_psrlw256((__v16hi)__a, (__v8hi)__count);
Ying Wang60999142013-01-07 13:59:36 -0800623}
624
625static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700626_mm256_srli_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800627{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700628 return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800629}
630
631static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700632_mm256_srl_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800633{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700634 return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count);
Ying Wang60999142013-01-07 13:59:36 -0800635}
636
637static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700638_mm256_srli_epi64(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800639{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700640 return __builtin_ia32_psrlqi256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800641}
642
643static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700644_mm256_srl_epi64(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800645{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700646 return __builtin_ia32_psrlq256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800647}
648
649static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700650_mm256_sub_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800651{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700652 return (__m256i)((__v32qi)__a - (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800653}
654
655static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700656_mm256_sub_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800657{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700658 return (__m256i)((__v16hi)__a - (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800659}
660
661static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700662_mm256_sub_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800663{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700664 return (__m256i)((__v8si)__a - (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800665}
666
667static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700668_mm256_sub_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800669{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700670 return __a - __b;
Ying Wang60999142013-01-07 13:59:36 -0800671}
672
673static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700674_mm256_subs_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800675{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700676 return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800677}
678
679static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700680_mm256_subs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800681{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700682 return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800683}
684
685static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700686_mm256_subs_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800687{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700688 return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800689}
690
691static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700692_mm256_subs_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800693{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700694 return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800695}
696
697static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700698_mm256_unpackhi_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800699{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700700 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);
Ying Wang60999142013-01-07 13:59:36 -0800701}
702
703static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700704_mm256_unpackhi_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800705{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700706 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);
Ying Wang60999142013-01-07 13:59:36 -0800707}
708
709static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700710_mm256_unpackhi_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800711{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700712 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 2, 8+2, 3, 8+3, 6, 8+6, 7, 8+7);
Ying Wang60999142013-01-07 13:59:36 -0800713}
714
715static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700716_mm256_unpackhi_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800717{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700718 return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3);
Ying Wang60999142013-01-07 13:59:36 -0800719}
720
721static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700722_mm256_unpacklo_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800723{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700724 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);
Ying Wang60999142013-01-07 13:59:36 -0800725}
726
727static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700728_mm256_unpacklo_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800729{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700730 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);
Ying Wang60999142013-01-07 13:59:36 -0800731}
732
733static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700734_mm256_unpacklo_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800735{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700736 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 0, 8+0, 1, 8+1, 4, 8+4, 5, 8+5);
Ying Wang60999142013-01-07 13:59:36 -0800737}
738
739static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700740_mm256_unpacklo_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800741{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700742 return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2);
Ying Wang60999142013-01-07 13:59:36 -0800743}
744
745static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700746_mm256_xor_si256(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800747{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700748 return __a ^ __b;
Ying Wang60999142013-01-07 13:59:36 -0800749}
750
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
Stephen Hines28c5e1e2015-08-13 18:18:46 -0700763static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
764_mm_broadcastsd_pd(__m128d __a)
765{
766 return __builtin_shufflevector(__a, __a, 0, 0);
767}
768
Ying Wang60999142013-01-07 13:59:36 -0800769static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
770_mm256_broadcastss_ps(__m128 __X)
771{
772 return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X);
773}
774
775static __inline__ __m256d __attribute__((__always_inline__, __nodebug__))
776_mm256_broadcastsd_pd(__m128d __X)
777{
778 return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X);
779}
780
781static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hines803c4682014-02-19 18:30:17 -0800782_mm256_broadcastsi128_si256(__m128i __X)
Ying Wang60999142013-01-07 13:59:36 -0800783{
Stephen Hines3f868232015-04-10 09:22:19 -0700784 return (__m256i)__builtin_shufflevector(__X, __X, 0, 1, 0, 1);
Ying Wang60999142013-01-07 13:59:36 -0800785}
786
787#define _mm_blend_epi32(V1, V2, M) __extension__ ({ \
788 __m128i __V1 = (V1); \
789 __m128i __V2 = (V2); \
Stephen Hinese65db132014-05-30 13:26:31 -0700790 (__m128i)__builtin_shufflevector((__v4si)__V1, (__v4si)__V2, \
791 (((M) & 0x01) ? 4 : 0), \
792 (((M) & 0x02) ? 5 : 1), \
793 (((M) & 0x04) ? 6 : 2), \
794 (((M) & 0x08) ? 7 : 3)); })
Ying Wang60999142013-01-07 13:59:36 -0800795
796#define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \
797 __m256i __V1 = (V1); \
798 __m256i __V2 = (V2); \
Stephen Hinese65db132014-05-30 13:26:31 -0700799 (__m256i)__builtin_shufflevector((__v8si)__V1, (__v8si)__V2, \
800 (((M) & 0x01) ? 8 : 0), \
801 (((M) & 0x02) ? 9 : 1), \
802 (((M) & 0x04) ? 10 : 2), \
803 (((M) & 0x08) ? 11 : 3), \
804 (((M) & 0x10) ? 12 : 4), \
805 (((M) & 0x20) ? 13 : 5), \
806 (((M) & 0x40) ? 14 : 6), \
807 (((M) & 0x80) ? 15 : 7)); })
Ying Wang60999142013-01-07 13:59:36 -0800808
809static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
810_mm256_broadcastb_epi8(__m128i __X)
811{
812 return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X);
813}
814
815static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
816_mm256_broadcastw_epi16(__m128i __X)
817{
818 return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X);
819}
820
821static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
822_mm256_broadcastd_epi32(__m128i __X)
823{
824 return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X);
825}
826
827static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
828_mm256_broadcastq_epi64(__m128i __X)
829{
830 return (__m256i)__builtin_ia32_pbroadcastq256(__X);
831}
832
833static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
834_mm_broadcastb_epi8(__m128i __X)
835{
836 return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X);
837}
838
839static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
840_mm_broadcastw_epi16(__m128i __X)
841{
842 return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X);
843}
844
845
846static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
847_mm_broadcastd_epi32(__m128i __X)
848{
849 return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X);
850}
851
852static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
853_mm_broadcastq_epi64(__m128i __X)
854{
855 return (__m128i)__builtin_ia32_pbroadcastq128(__X);
856}
857
858static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700859_mm256_permutevar8x32_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800860{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700861 return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800862}
863
864#define _mm256_permute4x64_pd(V, M) __extension__ ({ \
865 __m256d __V = (V); \
866 (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \
867 (M) & 0x3, ((M) & 0xc) >> 2, \
868 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
869
870static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700871_mm256_permutevar8x32_ps(__m256 __a, __m256 __b)
Ying Wang60999142013-01-07 13:59:36 -0800872{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700873 return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b);
Ying Wang60999142013-01-07 13:59:36 -0800874}
875
876#define _mm256_permute4x64_epi64(V, M) __extension__ ({ \
877 __m256i __V = (V); \
878 (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \
879 (M) & 0x3, ((M) & 0xc) >> 2, \
880 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
881
882#define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \
883 __m256i __V1 = (V1); \
884 __m256i __V2 = (V2); \
885 (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); })
886
Stephen Hines3f868232015-04-10 09:22:19 -0700887#define _mm256_extracti128_si256(V, M) __extension__ ({ \
888 (__m128i)__builtin_shufflevector( \
889 (__v4di)(V), \
890 (__v4di)(_mm256_setzero_si256()), \
891 (((M) & 1) ? 2 : 0), \
892 (((M) & 1) ? 3 : 1) );})
Ying Wang60999142013-01-07 13:59:36 -0800893
Stephen Hines3f868232015-04-10 09:22:19 -0700894#define _mm256_inserti128_si256(V1, V2, M) __extension__ ({ \
895 (__m256i)__builtin_shufflevector( \
896 (__v4di)(V1), \
897 (__v4di)_mm256_castsi128_si256((__m128i)(V2)), \
898 (((M) & 1) ? 0 : 4), \
899 (((M) & 1) ? 1 : 5), \
900 (((M) & 1) ? 4 : 2), \
901 (((M) & 1) ? 5 : 3) );})
Ying Wang60999142013-01-07 13:59:36 -0800902
903static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
904_mm256_maskload_epi32(int const *__X, __m256i __M)
905{
906 return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M);
907}
908
909static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
910_mm256_maskload_epi64(long long const *__X, __m256i __M)
911{
912 return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M);
913}
914
915static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
916_mm_maskload_epi32(int const *__X, __m128i __M)
917{
918 return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M);
919}
920
921static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
922_mm_maskload_epi64(long long const *__X, __m128i __M)
923{
924 return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M);
925}
926
927static __inline__ void __attribute__((__always_inline__, __nodebug__))
928_mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y)
929{
930 __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y);
931}
932
933static __inline__ void __attribute__((__always_inline__, __nodebug__))
934_mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y)
935{
936 __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y);
937}
938
939static __inline__ void __attribute__((__always_inline__, __nodebug__))
940_mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y)
941{
942 __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y);
943}
944
945static __inline__ void __attribute__((__always_inline__, __nodebug__))
946_mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y)
947{
948 __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y);
949}
950
951static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
952_mm256_sllv_epi32(__m256i __X, __m256i __Y)
953{
954 return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y);
955}
956
957static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
958_mm_sllv_epi32(__m128i __X, __m128i __Y)
959{
960 return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y);
961}
962
963static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
964_mm256_sllv_epi64(__m256i __X, __m256i __Y)
965{
966 return (__m256i)__builtin_ia32_psllv4di(__X, __Y);
967}
968
969static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
970_mm_sllv_epi64(__m128i __X, __m128i __Y)
971{
972 return (__m128i)__builtin_ia32_psllv2di(__X, __Y);
973}
974
975static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
976_mm256_srav_epi32(__m256i __X, __m256i __Y)
977{
978 return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y);
979}
980
981static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
982_mm_srav_epi32(__m128i __X, __m128i __Y)
983{
984 return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y);
985}
986
987static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
988_mm256_srlv_epi32(__m256i __X, __m256i __Y)
989{
990 return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y);
991}
992
993static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
994_mm_srlv_epi32(__m128i __X, __m128i __Y)
995{
996 return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y);
997}
998
999static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
1000_mm256_srlv_epi64(__m256i __X, __m256i __Y)
1001{
1002 return (__m256i)__builtin_ia32_psrlv4di(__X, __Y);
1003}
1004
1005static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1006_mm_srlv_epi64(__m128i __X, __m128i __Y)
1007{
1008 return (__m128i)__builtin_ia32_psrlv2di(__X, __Y);
1009}
1010
1011#define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
1012 __m128d __a = (a); \
1013 double const *__m = (m); \
1014 __m128i __i = (i); \
1015 __m128d __mask = (mask); \
1016 (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \
1017 (__v4si)__i, (__v2df)__mask, (s)); })
1018
1019#define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
1020 __m256d __a = (a); \
1021 double const *__m = (m); \
1022 __m128i __i = (i); \
1023 __m256d __mask = (mask); \
1024 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \
1025 (__v4si)__i, (__v4df)__mask, (s)); })
1026
1027#define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
1028 __m128d __a = (a); \
1029 double const *__m = (m); \
1030 __m128i __i = (i); \
1031 __m128d __mask = (mask); \
1032 (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \
1033 (__v2di)__i, (__v2df)__mask, (s)); })
1034
1035#define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
1036 __m256d __a = (a); \
1037 double const *__m = (m); \
1038 __m256i __i = (i); \
1039 __m256d __mask = (mask); \
1040 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \
1041 (__v4di)__i, (__v4df)__mask, (s)); })
1042
1043#define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1044 __m128 __a = (a); \
1045 float const *__m = (m); \
1046 __m128i __i = (i); \
1047 __m128 __mask = (mask); \
1048 (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \
1049 (__v4si)__i, (__v4sf)__mask, (s)); })
1050
1051#define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1052 __m256 __a = (a); \
1053 float const *__m = (m); \
1054 __m256i __i = (i); \
1055 __m256 __mask = (mask); \
1056 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \
1057 (__v8si)__i, (__v8sf)__mask, (s)); })
1058
1059#define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1060 __m128 __a = (a); \
1061 float const *__m = (m); \
1062 __m128i __i = (i); \
1063 __m128 __mask = (mask); \
1064 (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \
1065 (__v2di)__i, (__v4sf)__mask, (s)); })
1066
1067#define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1068 __m128 __a = (a); \
1069 float const *__m = (m); \
1070 __m256i __i = (i); \
1071 __m128 __mask = (mask); \
1072 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \
1073 (__v4di)__i, (__v4sf)__mask, (s)); })
1074
1075#define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1076 __m128i __a = (a); \
1077 int const *__m = (m); \
1078 __m128i __i = (i); \
1079 __m128i __mask = (mask); \
1080 (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \
1081 (__v4si)__i, (__v4si)__mask, (s)); })
1082
1083#define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1084 __m256i __a = (a); \
1085 int const *__m = (m); \
1086 __m256i __i = (i); \
1087 __m256i __mask = (mask); \
1088 (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \
1089 (__v8si)__i, (__v8si)__mask, (s)); })
1090
1091#define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1092 __m128i __a = (a); \
1093 int const *__m = (m); \
1094 __m128i __i = (i); \
1095 __m128i __mask = (mask); \
1096 (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \
1097 (__v2di)__i, (__v4si)__mask, (s)); })
1098
1099#define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1100 __m128i __a = (a); \
1101 int const *__m = (m); \
1102 __m256i __i = (i); \
1103 __m128i __mask = (mask); \
1104 (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \
1105 (__v4di)__i, (__v4si)__mask, (s)); })
1106
1107#define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1108 __m128i __a = (a); \
Stephen Hines803c4682014-02-19 18:30:17 -08001109 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001110 __m128i __i = (i); \
1111 __m128i __mask = (mask); \
1112 (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \
1113 (__v4si)__i, (__v2di)__mask, (s)); })
1114
1115#define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1116 __m256i __a = (a); \
Stephen Hines803c4682014-02-19 18:30:17 -08001117 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001118 __m128i __i = (i); \
1119 __m256i __mask = (mask); \
1120 (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \
1121 (__v4si)__i, (__v4di)__mask, (s)); })
1122
1123#define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1124 __m128i __a = (a); \
Stephen Hines803c4682014-02-19 18:30:17 -08001125 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001126 __m128i __i = (i); \
1127 __m128i __mask = (mask); \
1128 (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \
1129 (__v2di)__i, (__v2di)__mask, (s)); })
1130
1131#define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1132 __m256i __a = (a); \
Stephen Hines803c4682014-02-19 18:30:17 -08001133 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001134 __m256i __i = (i); \
1135 __m256i __mask = (mask); \
1136 (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \
1137 (__v4di)__i, (__v4di)__mask, (s)); })
1138
1139#define _mm_i32gather_pd(m, i, s) __extension__ ({ \
1140 double const *__m = (m); \
1141 __m128i __i = (i); \
1142 (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \
1143 (const __v2df *)__m, (__v4si)__i, \
1144 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1145
1146#define _mm256_i32gather_pd(m, i, s) __extension__ ({ \
1147 double const *__m = (m); \
1148 __m128i __i = (i); \
1149 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \
1150 (const __v4df *)__m, (__v4si)__i, \
1151 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1152
1153#define _mm_i64gather_pd(m, i, s) __extension__ ({ \
1154 double const *__m = (m); \
1155 __m128i __i = (i); \
1156 (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \
1157 (const __v2df *)__m, (__v2di)__i, \
1158 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1159
1160#define _mm256_i64gather_pd(m, i, s) __extension__ ({ \
1161 double const *__m = (m); \
1162 __m256i __i = (i); \
1163 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \
1164 (const __v4df *)__m, (__v4di)__i, \
1165 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1166
1167#define _mm_i32gather_ps(m, i, s) __extension__ ({ \
1168 float const *__m = (m); \
1169 __m128i __i = (i); \
1170 (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \
1171 (const __v4sf *)__m, (__v4si)__i, \
1172 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1173
1174#define _mm256_i32gather_ps(m, i, s) __extension__ ({ \
1175 float const *__m = (m); \
1176 __m256i __i = (i); \
1177 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \
1178 (const __v8sf *)__m, (__v8si)__i, \
1179 (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); })
1180
1181#define _mm_i64gather_ps(m, i, s) __extension__ ({ \
1182 float const *__m = (m); \
1183 __m128i __i = (i); \
1184 (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \
1185 (const __v4sf *)__m, (__v2di)__i, \
1186 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1187
1188#define _mm256_i64gather_ps(m, i, s) __extension__ ({ \
1189 float const *__m = (m); \
1190 __m256i __i = (i); \
1191 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \
1192 (const __v4sf *)__m, (__v4di)__i, \
1193 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1194
1195#define _mm_i32gather_epi32(m, i, s) __extension__ ({ \
1196 int const *__m = (m); \
1197 __m128i __i = (i); \
1198 (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \
1199 (const __v4si *)__m, (__v4si)__i, \
1200 (__v4si)_mm_set1_epi32(-1), (s)); })
1201
1202#define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \
1203 int const *__m = (m); \
1204 __m256i __i = (i); \
1205 (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \
1206 (const __v8si *)__m, (__v8si)__i, \
1207 (__v8si)_mm256_set1_epi32(-1), (s)); })
1208
1209#define _mm_i64gather_epi32(m, i, s) __extension__ ({ \
1210 int const *__m = (m); \
1211 __m128i __i = (i); \
1212 (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \
1213 (const __v4si *)__m, (__v2di)__i, \
1214 (__v4si)_mm_set1_epi32(-1), (s)); })
1215
1216#define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \
1217 int const *__m = (m); \
1218 __m256i __i = (i); \
1219 (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \
1220 (const __v4si *)__m, (__v4di)__i, \
1221 (__v4si)_mm_set1_epi32(-1), (s)); })
1222
1223#define _mm_i32gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines803c4682014-02-19 18:30:17 -08001224 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001225 __m128i __i = (i); \
1226 (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \
1227 (const __v2di *)__m, (__v4si)__i, \
1228 (__v2di)_mm_set1_epi64x(-1), (s)); })
1229
1230#define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines803c4682014-02-19 18:30:17 -08001231 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001232 __m128i __i = (i); \
1233 (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \
1234 (const __v4di *)__m, (__v4si)__i, \
1235 (__v4di)_mm256_set1_epi64x(-1), (s)); })
1236
1237#define _mm_i64gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines803c4682014-02-19 18:30:17 -08001238 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001239 __m128i __i = (i); \
1240 (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \
1241 (const __v2di *)__m, (__v2di)__i, \
1242 (__v2di)_mm_set1_epi64x(-1), (s)); })
1243
1244#define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines803c4682014-02-19 18:30:17 -08001245 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001246 __m256i __i = (i); \
1247 (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \
1248 (const __v4di *)__m, (__v4di)__i, \
1249 (__v4di)_mm256_set1_epi64x(-1), (s)); })
Stephen Hines996e4dc2013-08-13 01:04:14 -07001250
1251#endif /* __AVX2INTRIN_H */