blob: 394fdfee96522ca4b4cd5cfe44f7a97f1b82f006 [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); \
Stephen Hinese65db132014-05-30 13:26:31 -0700163 (__m256d)__builtin_shufflevector((__v16hi)__V1, (__v16hi)__V2, \
164 (((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
545static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700546_mm256_slli_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800547{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700548 return (__m256i)__builtin_ia32_psllwi256((__v16hi)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800549}
550
551static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700552_mm256_sll_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800553{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700554 return (__m256i)__builtin_ia32_psllw256((__v16hi)__a, (__v8hi)__count);
Ying Wang60999142013-01-07 13:59:36 -0800555}
556
557static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700558_mm256_slli_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800559{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700560 return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800561}
562
563static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700564_mm256_sll_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800565{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700566 return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__count);
Ying Wang60999142013-01-07 13:59:36 -0800567}
568
569static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700570_mm256_slli_epi64(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800571{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700572 return __builtin_ia32_psllqi256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800573}
574
575static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700576_mm256_sll_epi64(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800577{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700578 return __builtin_ia32_psllq256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800579}
580
581static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700582_mm256_srai_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800583{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700584 return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800585}
586
587static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700588_mm256_sra_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800589{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700590 return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__count);
Ying Wang60999142013-01-07 13:59:36 -0800591}
592
593static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700594_mm256_srai_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800595{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700596 return (__m256i)__builtin_ia32_psradi256((__v8si)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800597}
598
599static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700600_mm256_sra_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800601{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700602 return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count);
Ying Wang60999142013-01-07 13:59:36 -0800603}
604
605#define _mm256_srli_si256(a, count) __extension__ ({ \
606 __m256i __a = (a); \
607 (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); })
608
609static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700610_mm256_srli_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800611{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700612 return (__m256i)__builtin_ia32_psrlwi256((__v16hi)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800613}
614
615static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700616_mm256_srl_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800617{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700618 return (__m256i)__builtin_ia32_psrlw256((__v16hi)__a, (__v8hi)__count);
Ying Wang60999142013-01-07 13:59:36 -0800619}
620
621static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700622_mm256_srli_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800623{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700624 return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800625}
626
627static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700628_mm256_srl_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800629{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700630 return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count);
Ying Wang60999142013-01-07 13:59:36 -0800631}
632
633static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700634_mm256_srli_epi64(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800635{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700636 return __builtin_ia32_psrlqi256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800637}
638
639static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700640_mm256_srl_epi64(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800641{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700642 return __builtin_ia32_psrlq256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800643}
644
645static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700646_mm256_sub_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800647{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700648 return (__m256i)((__v32qi)__a - (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800649}
650
651static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700652_mm256_sub_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800653{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700654 return (__m256i)((__v16hi)__a - (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800655}
656
657static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700658_mm256_sub_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800659{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700660 return (__m256i)((__v8si)__a - (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800661}
662
663static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700664_mm256_sub_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800665{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700666 return __a - __b;
Ying Wang60999142013-01-07 13:59:36 -0800667}
668
669static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700670_mm256_subs_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800671{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700672 return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800673}
674
675static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700676_mm256_subs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800677{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700678 return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800679}
680
681static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700682_mm256_subs_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800683{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700684 return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800685}
686
687static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700688_mm256_subs_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800689{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700690 return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800691}
692
693static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700694_mm256_unpackhi_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800695{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700696 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 -0800697}
698
699static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700700_mm256_unpackhi_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800701{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700702 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 -0800703}
704
705static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700706_mm256_unpackhi_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800707{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700708 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 -0800709}
710
711static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700712_mm256_unpackhi_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800713{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700714 return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3);
Ying Wang60999142013-01-07 13:59:36 -0800715}
716
717static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700718_mm256_unpacklo_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800719{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700720 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 -0800721}
722
723static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700724_mm256_unpacklo_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800725{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700726 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 -0800727}
728
729static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700730_mm256_unpacklo_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800731{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700732 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 -0800733}
734
735static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700736_mm256_unpacklo_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800737{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700738 return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2);
Ying Wang60999142013-01-07 13:59:36 -0800739}
740
741static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700742_mm256_xor_si256(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800743{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700744 return __a ^ __b;
Ying Wang60999142013-01-07 13:59:36 -0800745}
746
747static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
748_mm256_stream_load_si256(__m256i *__V)
749{
750 return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V);
751}
752
753static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
754_mm_broadcastss_ps(__m128 __X)
755{
756 return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X);
757}
758
759static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
760_mm256_broadcastss_ps(__m128 __X)
761{
762 return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X);
763}
764
765static __inline__ __m256d __attribute__((__always_inline__, __nodebug__))
766_mm256_broadcastsd_pd(__m128d __X)
767{
768 return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X);
769}
770
771static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hines803c4682014-02-19 18:30:17 -0800772_mm256_broadcastsi128_si256(__m128i __X)
Ying Wang60999142013-01-07 13:59:36 -0800773{
Stephen Hines803c4682014-02-19 18:30:17 -0800774 return (__m256i)__builtin_ia32_vbroadcastsi256(__X);
Ying Wang60999142013-01-07 13:59:36 -0800775}
776
777#define _mm_blend_epi32(V1, V2, M) __extension__ ({ \
778 __m128i __V1 = (V1); \
779 __m128i __V2 = (V2); \
Stephen Hinese65db132014-05-30 13:26:31 -0700780 (__m128i)__builtin_shufflevector((__v4si)__V1, (__v4si)__V2, \
781 (((M) & 0x01) ? 4 : 0), \
782 (((M) & 0x02) ? 5 : 1), \
783 (((M) & 0x04) ? 6 : 2), \
784 (((M) & 0x08) ? 7 : 3)); })
Ying Wang60999142013-01-07 13:59:36 -0800785
786#define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \
787 __m256i __V1 = (V1); \
788 __m256i __V2 = (V2); \
Stephen Hinese65db132014-05-30 13:26:31 -0700789 (__m256i)__builtin_shufflevector((__v8si)__V1, (__v8si)__V2, \
790 (((M) & 0x01) ? 8 : 0), \
791 (((M) & 0x02) ? 9 : 1), \
792 (((M) & 0x04) ? 10 : 2), \
793 (((M) & 0x08) ? 11 : 3), \
794 (((M) & 0x10) ? 12 : 4), \
795 (((M) & 0x20) ? 13 : 5), \
796 (((M) & 0x40) ? 14 : 6), \
797 (((M) & 0x80) ? 15 : 7)); })
Ying Wang60999142013-01-07 13:59:36 -0800798
799static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
800_mm256_broadcastb_epi8(__m128i __X)
801{
802 return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X);
803}
804
805static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
806_mm256_broadcastw_epi16(__m128i __X)
807{
808 return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X);
809}
810
811static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
812_mm256_broadcastd_epi32(__m128i __X)
813{
814 return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X);
815}
816
817static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
818_mm256_broadcastq_epi64(__m128i __X)
819{
820 return (__m256i)__builtin_ia32_pbroadcastq256(__X);
821}
822
823static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
824_mm_broadcastb_epi8(__m128i __X)
825{
826 return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X);
827}
828
829static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
830_mm_broadcastw_epi16(__m128i __X)
831{
832 return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X);
833}
834
835
836static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
837_mm_broadcastd_epi32(__m128i __X)
838{
839 return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X);
840}
841
842static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
843_mm_broadcastq_epi64(__m128i __X)
844{
845 return (__m128i)__builtin_ia32_pbroadcastq128(__X);
846}
847
848static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700849_mm256_permutevar8x32_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800850{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700851 return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800852}
853
854#define _mm256_permute4x64_pd(V, M) __extension__ ({ \
855 __m256d __V = (V); \
856 (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \
857 (M) & 0x3, ((M) & 0xc) >> 2, \
858 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
859
860static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700861_mm256_permutevar8x32_ps(__m256 __a, __m256 __b)
Ying Wang60999142013-01-07 13:59:36 -0800862{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700863 return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b);
Ying Wang60999142013-01-07 13:59:36 -0800864}
865
866#define _mm256_permute4x64_epi64(V, M) __extension__ ({ \
867 __m256i __V = (V); \
868 (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \
869 (M) & 0x3, ((M) & 0xc) >> 2, \
870 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
871
872#define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \
873 __m256i __V1 = (V1); \
874 __m256i __V2 = (V2); \
875 (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); })
876
877#define _mm256_extracti128_si256(A, O) __extension__ ({ \
878 __m256i __A = (A); \
879 (__m128i)__builtin_ia32_extract128i256(__A, (O)); })
880
881#define _mm256_inserti128_si256(V1, V2, O) __extension__ ({ \
882 __m256i __V1 = (V1); \
883 __m128i __V2 = (V2); \
884 (__m256i)__builtin_ia32_insert128i256(__V1, __V2, (O)); })
885
886static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
887_mm256_maskload_epi32(int const *__X, __m256i __M)
888{
889 return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M);
890}
891
892static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
893_mm256_maskload_epi64(long long const *__X, __m256i __M)
894{
895 return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M);
896}
897
898static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
899_mm_maskload_epi32(int const *__X, __m128i __M)
900{
901 return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M);
902}
903
904static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
905_mm_maskload_epi64(long long const *__X, __m128i __M)
906{
907 return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M);
908}
909
910static __inline__ void __attribute__((__always_inline__, __nodebug__))
911_mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y)
912{
913 __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y);
914}
915
916static __inline__ void __attribute__((__always_inline__, __nodebug__))
917_mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y)
918{
919 __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y);
920}
921
922static __inline__ void __attribute__((__always_inline__, __nodebug__))
923_mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y)
924{
925 __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y);
926}
927
928static __inline__ void __attribute__((__always_inline__, __nodebug__))
929_mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y)
930{
931 __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y);
932}
933
934static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
935_mm256_sllv_epi32(__m256i __X, __m256i __Y)
936{
937 return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y);
938}
939
940static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
941_mm_sllv_epi32(__m128i __X, __m128i __Y)
942{
943 return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y);
944}
945
946static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
947_mm256_sllv_epi64(__m256i __X, __m256i __Y)
948{
949 return (__m256i)__builtin_ia32_psllv4di(__X, __Y);
950}
951
952static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
953_mm_sllv_epi64(__m128i __X, __m128i __Y)
954{
955 return (__m128i)__builtin_ia32_psllv2di(__X, __Y);
956}
957
958static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
959_mm256_srav_epi32(__m256i __X, __m256i __Y)
960{
961 return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y);
962}
963
964static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
965_mm_srav_epi32(__m128i __X, __m128i __Y)
966{
967 return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y);
968}
969
970static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
971_mm256_srlv_epi32(__m256i __X, __m256i __Y)
972{
973 return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y);
974}
975
976static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
977_mm_srlv_epi32(__m128i __X, __m128i __Y)
978{
979 return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y);
980}
981
982static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
983_mm256_srlv_epi64(__m256i __X, __m256i __Y)
984{
985 return (__m256i)__builtin_ia32_psrlv4di(__X, __Y);
986}
987
988static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
989_mm_srlv_epi64(__m128i __X, __m128i __Y)
990{
991 return (__m128i)__builtin_ia32_psrlv2di(__X, __Y);
992}
993
994#define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
995 __m128d __a = (a); \
996 double const *__m = (m); \
997 __m128i __i = (i); \
998 __m128d __mask = (mask); \
999 (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \
1000 (__v4si)__i, (__v2df)__mask, (s)); })
1001
1002#define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
1003 __m256d __a = (a); \
1004 double const *__m = (m); \
1005 __m128i __i = (i); \
1006 __m256d __mask = (mask); \
1007 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \
1008 (__v4si)__i, (__v4df)__mask, (s)); })
1009
1010#define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
1011 __m128d __a = (a); \
1012 double const *__m = (m); \
1013 __m128i __i = (i); \
1014 __m128d __mask = (mask); \
1015 (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \
1016 (__v2di)__i, (__v2df)__mask, (s)); })
1017
1018#define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
1019 __m256d __a = (a); \
1020 double const *__m = (m); \
1021 __m256i __i = (i); \
1022 __m256d __mask = (mask); \
1023 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \
1024 (__v4di)__i, (__v4df)__mask, (s)); })
1025
1026#define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1027 __m128 __a = (a); \
1028 float const *__m = (m); \
1029 __m128i __i = (i); \
1030 __m128 __mask = (mask); \
1031 (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \
1032 (__v4si)__i, (__v4sf)__mask, (s)); })
1033
1034#define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1035 __m256 __a = (a); \
1036 float const *__m = (m); \
1037 __m256i __i = (i); \
1038 __m256 __mask = (mask); \
1039 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \
1040 (__v8si)__i, (__v8sf)__mask, (s)); })
1041
1042#define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1043 __m128 __a = (a); \
1044 float const *__m = (m); \
1045 __m128i __i = (i); \
1046 __m128 __mask = (mask); \
1047 (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \
1048 (__v2di)__i, (__v4sf)__mask, (s)); })
1049
1050#define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1051 __m128 __a = (a); \
1052 float const *__m = (m); \
1053 __m256i __i = (i); \
1054 __m128 __mask = (mask); \
1055 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \
1056 (__v4di)__i, (__v4sf)__mask, (s)); })
1057
1058#define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1059 __m128i __a = (a); \
1060 int const *__m = (m); \
1061 __m128i __i = (i); \
1062 __m128i __mask = (mask); \
1063 (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \
1064 (__v4si)__i, (__v4si)__mask, (s)); })
1065
1066#define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1067 __m256i __a = (a); \
1068 int const *__m = (m); \
1069 __m256i __i = (i); \
1070 __m256i __mask = (mask); \
1071 (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \
1072 (__v8si)__i, (__v8si)__mask, (s)); })
1073
1074#define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1075 __m128i __a = (a); \
1076 int const *__m = (m); \
1077 __m128i __i = (i); \
1078 __m128i __mask = (mask); \
1079 (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \
1080 (__v2di)__i, (__v4si)__mask, (s)); })
1081
1082#define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1083 __m128i __a = (a); \
1084 int const *__m = (m); \
1085 __m256i __i = (i); \
1086 __m128i __mask = (mask); \
1087 (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \
1088 (__v4di)__i, (__v4si)__mask, (s)); })
1089
1090#define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1091 __m128i __a = (a); \
Stephen Hines803c4682014-02-19 18:30:17 -08001092 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001093 __m128i __i = (i); \
1094 __m128i __mask = (mask); \
1095 (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \
1096 (__v4si)__i, (__v2di)__mask, (s)); })
1097
1098#define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1099 __m256i __a = (a); \
Stephen Hines803c4682014-02-19 18:30:17 -08001100 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001101 __m128i __i = (i); \
1102 __m256i __mask = (mask); \
1103 (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \
1104 (__v4si)__i, (__v4di)__mask, (s)); })
1105
1106#define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1107 __m128i __a = (a); \
Stephen Hines803c4682014-02-19 18:30:17 -08001108 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001109 __m128i __i = (i); \
1110 __m128i __mask = (mask); \
1111 (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \
1112 (__v2di)__i, (__v2di)__mask, (s)); })
1113
1114#define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1115 __m256i __a = (a); \
Stephen Hines803c4682014-02-19 18:30:17 -08001116 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001117 __m256i __i = (i); \
1118 __m256i __mask = (mask); \
1119 (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \
1120 (__v4di)__i, (__v4di)__mask, (s)); })
1121
1122#define _mm_i32gather_pd(m, i, s) __extension__ ({ \
1123 double const *__m = (m); \
1124 __m128i __i = (i); \
1125 (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \
1126 (const __v2df *)__m, (__v4si)__i, \
1127 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1128
1129#define _mm256_i32gather_pd(m, i, s) __extension__ ({ \
1130 double const *__m = (m); \
1131 __m128i __i = (i); \
1132 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \
1133 (const __v4df *)__m, (__v4si)__i, \
1134 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1135
1136#define _mm_i64gather_pd(m, i, s) __extension__ ({ \
1137 double const *__m = (m); \
1138 __m128i __i = (i); \
1139 (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \
1140 (const __v2df *)__m, (__v2di)__i, \
1141 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1142
1143#define _mm256_i64gather_pd(m, i, s) __extension__ ({ \
1144 double const *__m = (m); \
1145 __m256i __i = (i); \
1146 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \
1147 (const __v4df *)__m, (__v4di)__i, \
1148 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1149
1150#define _mm_i32gather_ps(m, i, s) __extension__ ({ \
1151 float const *__m = (m); \
1152 __m128i __i = (i); \
1153 (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \
1154 (const __v4sf *)__m, (__v4si)__i, \
1155 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1156
1157#define _mm256_i32gather_ps(m, i, s) __extension__ ({ \
1158 float const *__m = (m); \
1159 __m256i __i = (i); \
1160 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \
1161 (const __v8sf *)__m, (__v8si)__i, \
1162 (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); })
1163
1164#define _mm_i64gather_ps(m, i, s) __extension__ ({ \
1165 float const *__m = (m); \
1166 __m128i __i = (i); \
1167 (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \
1168 (const __v4sf *)__m, (__v2di)__i, \
1169 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1170
1171#define _mm256_i64gather_ps(m, i, s) __extension__ ({ \
1172 float const *__m = (m); \
1173 __m256i __i = (i); \
1174 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \
1175 (const __v4sf *)__m, (__v4di)__i, \
1176 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1177
1178#define _mm_i32gather_epi32(m, i, s) __extension__ ({ \
1179 int const *__m = (m); \
1180 __m128i __i = (i); \
1181 (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \
1182 (const __v4si *)__m, (__v4si)__i, \
1183 (__v4si)_mm_set1_epi32(-1), (s)); })
1184
1185#define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \
1186 int const *__m = (m); \
1187 __m256i __i = (i); \
1188 (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \
1189 (const __v8si *)__m, (__v8si)__i, \
1190 (__v8si)_mm256_set1_epi32(-1), (s)); })
1191
1192#define _mm_i64gather_epi32(m, i, s) __extension__ ({ \
1193 int const *__m = (m); \
1194 __m128i __i = (i); \
1195 (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \
1196 (const __v4si *)__m, (__v2di)__i, \
1197 (__v4si)_mm_set1_epi32(-1), (s)); })
1198
1199#define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \
1200 int const *__m = (m); \
1201 __m256i __i = (i); \
1202 (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \
1203 (const __v4si *)__m, (__v4di)__i, \
1204 (__v4si)_mm_set1_epi32(-1), (s)); })
1205
1206#define _mm_i32gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines803c4682014-02-19 18:30:17 -08001207 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001208 __m128i __i = (i); \
1209 (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \
1210 (const __v2di *)__m, (__v4si)__i, \
1211 (__v2di)_mm_set1_epi64x(-1), (s)); })
1212
1213#define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines803c4682014-02-19 18:30:17 -08001214 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001215 __m128i __i = (i); \
1216 (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \
1217 (const __v4di *)__m, (__v4si)__i, \
1218 (__v4di)_mm256_set1_epi64x(-1), (s)); })
1219
1220#define _mm_i64gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines803c4682014-02-19 18:30:17 -08001221 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001222 __m128i __i = (i); \
1223 (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \
1224 (const __v2di *)__m, (__v2di)__i, \
1225 (__v2di)_mm_set1_epi64x(-1), (s)); })
1226
1227#define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines803c4682014-02-19 18:30:17 -08001228 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001229 __m256i __i = (i); \
1230 (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \
1231 (const __v4di *)__m, (__v4di)__i, \
1232 (__v4di)_mm256_set1_epi64x(-1), (s)); })
Stephen Hines996e4dc2013-08-13 01:04:14 -07001233
1234#endif /* __AVX2INTRIN_H */