blob: 95744693ac7768eca0a849bc79f56a9fa973db8d [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); \
163 (__m256i)__builtin_ia32_pblendw256((__v16hi)__V1, (__v16hi)__V2, (M)); })
164
165static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700166_mm256_cmpeq_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800167{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700168 return (__m256i)((__v32qi)__a == (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800169}
170
171static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700172_mm256_cmpeq_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800173{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700174 return (__m256i)((__v16hi)__a == (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800175}
176
177static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700178_mm256_cmpeq_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800179{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700180 return (__m256i)((__v8si)__a == (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800181}
182
183static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700184_mm256_cmpeq_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800185{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700186 return (__m256i)(__a == __b);
Ying Wang60999142013-01-07 13:59:36 -0800187}
188
189static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700190_mm256_cmpgt_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800191{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700192 return (__m256i)((__v32qi)__a > (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800193}
194
195static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700196_mm256_cmpgt_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800197{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700198 return (__m256i)((__v16hi)__a > (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800199}
200
201static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700202_mm256_cmpgt_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800203{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700204 return (__m256i)((__v8si)__a > (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800205}
206
207static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700208_mm256_cmpgt_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800209{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700210 return (__m256i)(__a > __b);
Ying Wang60999142013-01-07 13:59:36 -0800211}
212
213static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700214_mm256_hadd_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800215{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700216 return (__m256i)__builtin_ia32_phaddw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800217}
218
219static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700220_mm256_hadd_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800221{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700222 return (__m256i)__builtin_ia32_phaddd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800223}
224
225static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700226_mm256_hadds_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800227{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700228 return (__m256i)__builtin_ia32_phaddsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800229}
230
231static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700232_mm256_hsub_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800233{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700234 return (__m256i)__builtin_ia32_phsubw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800235}
236
237static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700238_mm256_hsub_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800239{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700240 return (__m256i)__builtin_ia32_phsubd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800241}
242
243static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700244_mm256_hsubs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800245{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700246 return (__m256i)__builtin_ia32_phsubsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800247}
248
249static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700250_mm256_maddubs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800251{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700252 return (__m256i)__builtin_ia32_pmaddubsw256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800253}
254
255static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700256_mm256_madd_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800257{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700258 return (__m256i)__builtin_ia32_pmaddwd256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800259}
260
261static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700262_mm256_max_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800263{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700264 return (__m256i)__builtin_ia32_pmaxsb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800265}
266
267static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700268_mm256_max_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800269{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700270 return (__m256i)__builtin_ia32_pmaxsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800271}
272
273static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700274_mm256_max_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800275{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700276 return (__m256i)__builtin_ia32_pmaxsd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800277}
278
279static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700280_mm256_max_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800281{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700282 return (__m256i)__builtin_ia32_pmaxub256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800283}
284
285static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700286_mm256_max_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800287{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700288 return (__m256i)__builtin_ia32_pmaxuw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800289}
290
291static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700292_mm256_max_epu32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800293{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700294 return (__m256i)__builtin_ia32_pmaxud256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800295}
296
297static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700298_mm256_min_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800299{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700300 return (__m256i)__builtin_ia32_pminsb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800301}
302
303static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700304_mm256_min_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800305{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700306 return (__m256i)__builtin_ia32_pminsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800307}
308
309static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700310_mm256_min_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800311{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700312 return (__m256i)__builtin_ia32_pminsd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800313}
314
315static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700316_mm256_min_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800317{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700318 return (__m256i)__builtin_ia32_pminub256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800319}
320
321static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700322_mm256_min_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800323{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700324 return (__m256i)__builtin_ia32_pminuw256 ((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800325}
326
327static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700328_mm256_min_epu32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800329{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700330 return (__m256i)__builtin_ia32_pminud256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800331}
332
333static __inline__ int __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700334_mm256_movemask_epi8(__m256i __a)
Ying Wang60999142013-01-07 13:59:36 -0800335{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700336 return __builtin_ia32_pmovmskb256((__v32qi)__a);
Ying Wang60999142013-01-07 13:59:36 -0800337}
338
339static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
340_mm256_cvtepi8_epi16(__m128i __V)
341{
342 return (__m256i)__builtin_ia32_pmovsxbw256((__v16qi)__V);
343}
344
345static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
346_mm256_cvtepi8_epi32(__m128i __V)
347{
348 return (__m256i)__builtin_ia32_pmovsxbd256((__v16qi)__V);
349}
350
351static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
352_mm256_cvtepi8_epi64(__m128i __V)
353{
354 return (__m256i)__builtin_ia32_pmovsxbq256((__v16qi)__V);
355}
356
357static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
358_mm256_cvtepi16_epi32(__m128i __V)
359{
360 return (__m256i)__builtin_ia32_pmovsxwd256((__v8hi)__V);
361}
362
363static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
364_mm256_cvtepi16_epi64(__m128i __V)
365{
366 return (__m256i)__builtin_ia32_pmovsxwq256((__v8hi)__V);
367}
368
369static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
370_mm256_cvtepi32_epi64(__m128i __V)
371{
372 return (__m256i)__builtin_ia32_pmovsxdq256((__v4si)__V);
373}
374
375static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
376_mm256_cvtepu8_epi16(__m128i __V)
377{
378 return (__m256i)__builtin_ia32_pmovzxbw256((__v16qi)__V);
379}
380
381static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
382_mm256_cvtepu8_epi32(__m128i __V)
383{
384 return (__m256i)__builtin_ia32_pmovzxbd256((__v16qi)__V);
385}
386
387static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
388_mm256_cvtepu8_epi64(__m128i __V)
389{
390 return (__m256i)__builtin_ia32_pmovzxbq256((__v16qi)__V);
391}
392
393static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
394_mm256_cvtepu16_epi32(__m128i __V)
395{
396 return (__m256i)__builtin_ia32_pmovzxwd256((__v8hi)__V);
397}
398
399static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
400_mm256_cvtepu16_epi64(__m128i __V)
401{
402 return (__m256i)__builtin_ia32_pmovzxwq256((__v8hi)__V);
403}
404
405static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
406_mm256_cvtepu32_epi64(__m128i __V)
407{
408 return (__m256i)__builtin_ia32_pmovzxdq256((__v4si)__V);
409}
410
411static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700412_mm256_mul_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800413{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700414 return (__m256i)__builtin_ia32_pmuldq256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800415}
416
417static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700418_mm256_mulhrs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800419{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700420 return (__m256i)__builtin_ia32_pmulhrsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800421}
422
423static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700424_mm256_mulhi_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800425{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700426 return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800427}
428
429static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700430_mm256_mulhi_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800431{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700432 return (__m256i)__builtin_ia32_pmulhw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800433}
434
435static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700436_mm256_mullo_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800437{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700438 return (__m256i)((__v16hi)__a * (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800439}
440
441static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700442_mm256_mullo_epi32 (__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800443{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700444 return (__m256i)((__v8si)__a * (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800445}
446
447static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700448_mm256_mul_epu32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800449{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700450 return __builtin_ia32_pmuludq256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800451}
452
453static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700454_mm256_or_si256(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800455{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700456 return __a | __b;
Ying Wang60999142013-01-07 13:59:36 -0800457}
458
459static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700460_mm256_sad_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800461{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700462 return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800463}
464
465static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700466_mm256_shuffle_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800467{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700468 return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800469}
470
471#define _mm256_shuffle_epi32(a, imm) __extension__ ({ \
472 __m256i __a = (a); \
473 (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)_mm256_set1_epi32(0), \
474 (imm) & 0x3, ((imm) & 0xc) >> 2, \
475 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
476 4 + (((imm) & 0x03) >> 0), \
477 4 + (((imm) & 0x0c) >> 2), \
478 4 + (((imm) & 0x30) >> 4), \
479 4 + (((imm) & 0xc0) >> 6)); })
480
481#define _mm256_shufflehi_epi16(a, imm) __extension__ ({ \
482 __m256i __a = (a); \
483 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
484 0, 1, 2, 3, \
485 4 + (((imm) & 0x03) >> 0), \
486 4 + (((imm) & 0x0c) >> 2), \
487 4 + (((imm) & 0x30) >> 4), \
488 4 + (((imm) & 0xc0) >> 6), \
489 8, 9, 10, 11, \
490 12 + (((imm) & 0x03) >> 0), \
491 12 + (((imm) & 0x0c) >> 2), \
492 12 + (((imm) & 0x30) >> 4), \
493 12 + (((imm) & 0xc0) >> 6)); })
494
495#define _mm256_shufflelo_epi16(a, imm) __extension__ ({ \
496 __m256i __a = (a); \
497 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
498 (imm) & 0x3,((imm) & 0xc) >> 2, \
499 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
500 4, 5, 6, 7, \
501 8 + (((imm) & 0x03) >> 0), \
502 8 + (((imm) & 0x0c) >> 2), \
503 8 + (((imm) & 0x30) >> 4), \
504 8 + (((imm) & 0xc0) >> 6), \
505 12, 13, 14, 15); })
506
507static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700508_mm256_sign_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800509{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700510 return (__m256i)__builtin_ia32_psignb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800511}
512
513static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700514_mm256_sign_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800515{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700516 return (__m256i)__builtin_ia32_psignw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800517}
518
519static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700520_mm256_sign_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800521{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700522 return (__m256i)__builtin_ia32_psignd256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800523}
524
525#define _mm256_slli_si256(a, count) __extension__ ({ \
526 __m256i __a = (a); \
527 (__m256i)__builtin_ia32_pslldqi256(__a, (count)*8); })
528
529static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700530_mm256_slli_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800531{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700532 return (__m256i)__builtin_ia32_psllwi256((__v16hi)__a, __count);
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_sll_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800537{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700538 return (__m256i)__builtin_ia32_psllw256((__v16hi)__a, (__v8hi)__count);
Ying Wang60999142013-01-07 13:59:36 -0800539}
540
541static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700542_mm256_slli_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800543{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700544 return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800545}
546
547static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700548_mm256_sll_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800549{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700550 return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__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_slli_epi64(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800555{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700556 return __builtin_ia32_psllqi256(__a, __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_sll_epi64(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800561{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700562 return __builtin_ia32_psllq256(__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_srai_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800567{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700568 return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __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_sra_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800573{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700574 return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__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_srai_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800579{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700580 return (__m256i)__builtin_ia32_psradi256((__v8si)__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_sra_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800585{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700586 return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count);
Ying Wang60999142013-01-07 13:59:36 -0800587}
588
589#define _mm256_srli_si256(a, count) __extension__ ({ \
590 __m256i __a = (a); \
591 (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); })
592
593static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700594_mm256_srli_epi16(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800595{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700596 return (__m256i)__builtin_ia32_psrlwi256((__v16hi)__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_srl_epi16(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800601{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700602 return (__m256i)__builtin_ia32_psrlw256((__v16hi)__a, (__v8hi)__count);
Ying Wang60999142013-01-07 13:59:36 -0800603}
604
605static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700606_mm256_srli_epi32(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800607{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700608 return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800609}
610
611static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700612_mm256_srl_epi32(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800613{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700614 return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count);
Ying Wang60999142013-01-07 13:59:36 -0800615}
616
617static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700618_mm256_srli_epi64(__m256i __a, int __count)
Ying Wang60999142013-01-07 13:59:36 -0800619{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700620 return __builtin_ia32_psrlqi256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800621}
622
623static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700624_mm256_srl_epi64(__m256i __a, __m128i __count)
Ying Wang60999142013-01-07 13:59:36 -0800625{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700626 return __builtin_ia32_psrlq256(__a, __count);
Ying Wang60999142013-01-07 13:59:36 -0800627}
628
629static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700630_mm256_sub_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800631{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700632 return (__m256i)((__v32qi)__a - (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800633}
634
635static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700636_mm256_sub_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800637{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700638 return (__m256i)((__v16hi)__a - (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800639}
640
641static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700642_mm256_sub_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800643{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700644 return (__m256i)((__v8si)__a - (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800645}
646
647static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700648_mm256_sub_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800649{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700650 return __a - __b;
Ying Wang60999142013-01-07 13:59:36 -0800651}
652
653static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700654_mm256_subs_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800655{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700656 return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800657}
658
659static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700660_mm256_subs_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800661{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700662 return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800663}
664
665static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700666_mm256_subs_epu8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800667{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700668 return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800669}
670
671static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700672_mm256_subs_epu16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800673{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700674 return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b);
Ying Wang60999142013-01-07 13:59:36 -0800675}
676
677static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700678_mm256_unpackhi_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800679{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700680 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 -0800681}
682
683static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700684_mm256_unpackhi_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800685{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700686 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 -0800687}
688
689static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700690_mm256_unpackhi_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800691{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700692 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 -0800693}
694
695static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700696_mm256_unpackhi_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800697{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700698 return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3);
Ying Wang60999142013-01-07 13:59:36 -0800699}
700
701static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700702_mm256_unpacklo_epi8(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800703{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700704 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 -0800705}
706
707static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700708_mm256_unpacklo_epi16(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800709{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700710 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 -0800711}
712
713static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700714_mm256_unpacklo_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800715{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700716 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 -0800717}
718
719static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700720_mm256_unpacklo_epi64(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800721{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700722 return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2);
Ying Wang60999142013-01-07 13:59:36 -0800723}
724
725static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700726_mm256_xor_si256(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800727{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700728 return __a ^ __b;
Ying Wang60999142013-01-07 13:59:36 -0800729}
730
731static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
732_mm256_stream_load_si256(__m256i *__V)
733{
734 return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V);
735}
736
737static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
738_mm_broadcastss_ps(__m128 __X)
739{
740 return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X);
741}
742
743static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
744_mm256_broadcastss_ps(__m128 __X)
745{
746 return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X);
747}
748
749static __inline__ __m256d __attribute__((__always_inline__, __nodebug__))
750_mm256_broadcastsd_pd(__m128d __X)
751{
752 return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X);
753}
754
755static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hines73a76582014-02-13 02:55:05 -0800756_mm256_broadcastsi128_si256(__m128i __X)
Ying Wang60999142013-01-07 13:59:36 -0800757{
Stephen Hines73a76582014-02-13 02:55:05 -0800758 return (__m256i)__builtin_ia32_vbroadcastsi256(__X);
Ying Wang60999142013-01-07 13:59:36 -0800759}
760
761#define _mm_blend_epi32(V1, V2, M) __extension__ ({ \
762 __m128i __V1 = (V1); \
763 __m128i __V2 = (V2); \
764 (__m128i)__builtin_ia32_pblendd128((__v4si)__V1, (__v4si)__V2, (M)); })
765
766#define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \
767 __m256i __V1 = (V1); \
768 __m256i __V2 = (V2); \
769 (__m256i)__builtin_ia32_pblendd256((__v8si)__V1, (__v8si)__V2, (M)); })
770
771static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
772_mm256_broadcastb_epi8(__m128i __X)
773{
774 return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X);
775}
776
777static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
778_mm256_broadcastw_epi16(__m128i __X)
779{
780 return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X);
781}
782
783static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
784_mm256_broadcastd_epi32(__m128i __X)
785{
786 return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X);
787}
788
789static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
790_mm256_broadcastq_epi64(__m128i __X)
791{
792 return (__m256i)__builtin_ia32_pbroadcastq256(__X);
793}
794
795static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
796_mm_broadcastb_epi8(__m128i __X)
797{
798 return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X);
799}
800
801static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
802_mm_broadcastw_epi16(__m128i __X)
803{
804 return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X);
805}
806
807
808static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
809_mm_broadcastd_epi32(__m128i __X)
810{
811 return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X);
812}
813
814static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
815_mm_broadcastq_epi64(__m128i __X)
816{
817 return (__m128i)__builtin_ia32_pbroadcastq128(__X);
818}
819
820static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700821_mm256_permutevar8x32_epi32(__m256i __a, __m256i __b)
Ying Wang60999142013-01-07 13:59:36 -0800822{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700823 return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b);
Ying Wang60999142013-01-07 13:59:36 -0800824}
825
826#define _mm256_permute4x64_pd(V, M) __extension__ ({ \
827 __m256d __V = (V); \
828 (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \
829 (M) & 0x3, ((M) & 0xc) >> 2, \
830 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
831
832static __inline__ __m256 __attribute__((__always_inline__, __nodebug__))
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700833_mm256_permutevar8x32_ps(__m256 __a, __m256 __b)
Ying Wang60999142013-01-07 13:59:36 -0800834{
Stephen Hinesc6ee7df2013-04-02 18:41:57 -0700835 return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b);
Ying Wang60999142013-01-07 13:59:36 -0800836}
837
838#define _mm256_permute4x64_epi64(V, M) __extension__ ({ \
839 __m256i __V = (V); \
840 (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \
841 (M) & 0x3, ((M) & 0xc) >> 2, \
842 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
843
844#define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \
845 __m256i __V1 = (V1); \
846 __m256i __V2 = (V2); \
847 (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); })
848
849#define _mm256_extracti128_si256(A, O) __extension__ ({ \
850 __m256i __A = (A); \
851 (__m128i)__builtin_ia32_extract128i256(__A, (O)); })
852
853#define _mm256_inserti128_si256(V1, V2, O) __extension__ ({ \
854 __m256i __V1 = (V1); \
855 __m128i __V2 = (V2); \
856 (__m256i)__builtin_ia32_insert128i256(__V1, __V2, (O)); })
857
858static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
859_mm256_maskload_epi32(int const *__X, __m256i __M)
860{
861 return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M);
862}
863
864static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
865_mm256_maskload_epi64(long long const *__X, __m256i __M)
866{
867 return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M);
868}
869
870static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
871_mm_maskload_epi32(int const *__X, __m128i __M)
872{
873 return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M);
874}
875
876static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
877_mm_maskload_epi64(long long const *__X, __m128i __M)
878{
879 return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M);
880}
881
882static __inline__ void __attribute__((__always_inline__, __nodebug__))
883_mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y)
884{
885 __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y);
886}
887
888static __inline__ void __attribute__((__always_inline__, __nodebug__))
889_mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y)
890{
891 __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y);
892}
893
894static __inline__ void __attribute__((__always_inline__, __nodebug__))
895_mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y)
896{
897 __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y);
898}
899
900static __inline__ void __attribute__((__always_inline__, __nodebug__))
901_mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y)
902{
903 __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y);
904}
905
906static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
907_mm256_sllv_epi32(__m256i __X, __m256i __Y)
908{
909 return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y);
910}
911
912static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
913_mm_sllv_epi32(__m128i __X, __m128i __Y)
914{
915 return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y);
916}
917
918static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
919_mm256_sllv_epi64(__m256i __X, __m256i __Y)
920{
921 return (__m256i)__builtin_ia32_psllv4di(__X, __Y);
922}
923
924static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
925_mm_sllv_epi64(__m128i __X, __m128i __Y)
926{
927 return (__m128i)__builtin_ia32_psllv2di(__X, __Y);
928}
929
930static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
931_mm256_srav_epi32(__m256i __X, __m256i __Y)
932{
933 return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y);
934}
935
936static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
937_mm_srav_epi32(__m128i __X, __m128i __Y)
938{
939 return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y);
940}
941
942static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
943_mm256_srlv_epi32(__m256i __X, __m256i __Y)
944{
945 return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y);
946}
947
948static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
949_mm_srlv_epi32(__m128i __X, __m128i __Y)
950{
951 return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y);
952}
953
954static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
955_mm256_srlv_epi64(__m256i __X, __m256i __Y)
956{
957 return (__m256i)__builtin_ia32_psrlv4di(__X, __Y);
958}
959
960static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
961_mm_srlv_epi64(__m128i __X, __m128i __Y)
962{
963 return (__m128i)__builtin_ia32_psrlv2di(__X, __Y);
964}
965
966#define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
967 __m128d __a = (a); \
968 double const *__m = (m); \
969 __m128i __i = (i); \
970 __m128d __mask = (mask); \
971 (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \
972 (__v4si)__i, (__v2df)__mask, (s)); })
973
974#define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
975 __m256d __a = (a); \
976 double const *__m = (m); \
977 __m128i __i = (i); \
978 __m256d __mask = (mask); \
979 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \
980 (__v4si)__i, (__v4df)__mask, (s)); })
981
982#define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
983 __m128d __a = (a); \
984 double const *__m = (m); \
985 __m128i __i = (i); \
986 __m128d __mask = (mask); \
987 (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \
988 (__v2di)__i, (__v2df)__mask, (s)); })
989
990#define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
991 __m256d __a = (a); \
992 double const *__m = (m); \
993 __m256i __i = (i); \
994 __m256d __mask = (mask); \
995 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \
996 (__v4di)__i, (__v4df)__mask, (s)); })
997
998#define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
999 __m128 __a = (a); \
1000 float const *__m = (m); \
1001 __m128i __i = (i); \
1002 __m128 __mask = (mask); \
1003 (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \
1004 (__v4si)__i, (__v4sf)__mask, (s)); })
1005
1006#define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1007 __m256 __a = (a); \
1008 float const *__m = (m); \
1009 __m256i __i = (i); \
1010 __m256 __mask = (mask); \
1011 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \
1012 (__v8si)__i, (__v8sf)__mask, (s)); })
1013
1014#define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1015 __m128 __a = (a); \
1016 float const *__m = (m); \
1017 __m128i __i = (i); \
1018 __m128 __mask = (mask); \
1019 (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \
1020 (__v2di)__i, (__v4sf)__mask, (s)); })
1021
1022#define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1023 __m128 __a = (a); \
1024 float const *__m = (m); \
1025 __m256i __i = (i); \
1026 __m128 __mask = (mask); \
1027 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \
1028 (__v4di)__i, (__v4sf)__mask, (s)); })
1029
1030#define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1031 __m128i __a = (a); \
1032 int const *__m = (m); \
1033 __m128i __i = (i); \
1034 __m128i __mask = (mask); \
1035 (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \
1036 (__v4si)__i, (__v4si)__mask, (s)); })
1037
1038#define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1039 __m256i __a = (a); \
1040 int const *__m = (m); \
1041 __m256i __i = (i); \
1042 __m256i __mask = (mask); \
1043 (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \
1044 (__v8si)__i, (__v8si)__mask, (s)); })
1045
1046#define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1047 __m128i __a = (a); \
1048 int const *__m = (m); \
1049 __m128i __i = (i); \
1050 __m128i __mask = (mask); \
1051 (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \
1052 (__v2di)__i, (__v4si)__mask, (s)); })
1053
1054#define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1055 __m128i __a = (a); \
1056 int const *__m = (m); \
1057 __m256i __i = (i); \
1058 __m128i __mask = (mask); \
1059 (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \
1060 (__v4di)__i, (__v4si)__mask, (s)); })
1061
1062#define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1063 __m128i __a = (a); \
Stephen Hines73a76582014-02-13 02:55:05 -08001064 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001065 __m128i __i = (i); \
1066 __m128i __mask = (mask); \
1067 (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \
1068 (__v4si)__i, (__v2di)__mask, (s)); })
1069
1070#define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1071 __m256i __a = (a); \
Stephen Hines73a76582014-02-13 02:55:05 -08001072 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001073 __m128i __i = (i); \
1074 __m256i __mask = (mask); \
1075 (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \
1076 (__v4si)__i, (__v4di)__mask, (s)); })
1077
1078#define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1079 __m128i __a = (a); \
Stephen Hines73a76582014-02-13 02:55:05 -08001080 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001081 __m128i __i = (i); \
1082 __m128i __mask = (mask); \
1083 (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \
1084 (__v2di)__i, (__v2di)__mask, (s)); })
1085
1086#define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1087 __m256i __a = (a); \
Stephen Hines73a76582014-02-13 02:55:05 -08001088 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001089 __m256i __i = (i); \
1090 __m256i __mask = (mask); \
1091 (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \
1092 (__v4di)__i, (__v4di)__mask, (s)); })
1093
1094#define _mm_i32gather_pd(m, i, s) __extension__ ({ \
1095 double const *__m = (m); \
1096 __m128i __i = (i); \
1097 (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \
1098 (const __v2df *)__m, (__v4si)__i, \
1099 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1100
1101#define _mm256_i32gather_pd(m, i, s) __extension__ ({ \
1102 double const *__m = (m); \
1103 __m128i __i = (i); \
1104 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \
1105 (const __v4df *)__m, (__v4si)__i, \
1106 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1107
1108#define _mm_i64gather_pd(m, i, s) __extension__ ({ \
1109 double const *__m = (m); \
1110 __m128i __i = (i); \
1111 (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \
1112 (const __v2df *)__m, (__v2di)__i, \
1113 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1114
1115#define _mm256_i64gather_pd(m, i, s) __extension__ ({ \
1116 double const *__m = (m); \
1117 __m256i __i = (i); \
1118 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \
1119 (const __v4df *)__m, (__v4di)__i, \
1120 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1121
1122#define _mm_i32gather_ps(m, i, s) __extension__ ({ \
1123 float const *__m = (m); \
1124 __m128i __i = (i); \
1125 (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \
1126 (const __v4sf *)__m, (__v4si)__i, \
1127 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1128
1129#define _mm256_i32gather_ps(m, i, s) __extension__ ({ \
1130 float const *__m = (m); \
1131 __m256i __i = (i); \
1132 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \
1133 (const __v8sf *)__m, (__v8si)__i, \
1134 (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); })
1135
1136#define _mm_i64gather_ps(m, i, s) __extension__ ({ \
1137 float const *__m = (m); \
1138 __m128i __i = (i); \
1139 (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \
1140 (const __v4sf *)__m, (__v2di)__i, \
1141 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1142
1143#define _mm256_i64gather_ps(m, i, s) __extension__ ({ \
1144 float const *__m = (m); \
1145 __m256i __i = (i); \
1146 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \
1147 (const __v4sf *)__m, (__v4di)__i, \
1148 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1149
1150#define _mm_i32gather_epi32(m, i, s) __extension__ ({ \
1151 int const *__m = (m); \
1152 __m128i __i = (i); \
1153 (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \
1154 (const __v4si *)__m, (__v4si)__i, \
1155 (__v4si)_mm_set1_epi32(-1), (s)); })
1156
1157#define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \
1158 int const *__m = (m); \
1159 __m256i __i = (i); \
1160 (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \
1161 (const __v8si *)__m, (__v8si)__i, \
1162 (__v8si)_mm256_set1_epi32(-1), (s)); })
1163
1164#define _mm_i64gather_epi32(m, i, s) __extension__ ({ \
1165 int const *__m = (m); \
1166 __m128i __i = (i); \
1167 (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \
1168 (const __v4si *)__m, (__v2di)__i, \
1169 (__v4si)_mm_set1_epi32(-1), (s)); })
1170
1171#define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \
1172 int const *__m = (m); \
1173 __m256i __i = (i); \
1174 (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \
1175 (const __v4si *)__m, (__v4di)__i, \
1176 (__v4si)_mm_set1_epi32(-1), (s)); })
1177
1178#define _mm_i32gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines73a76582014-02-13 02:55:05 -08001179 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001180 __m128i __i = (i); \
1181 (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \
1182 (const __v2di *)__m, (__v4si)__i, \
1183 (__v2di)_mm_set1_epi64x(-1), (s)); })
1184
1185#define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines73a76582014-02-13 02:55:05 -08001186 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001187 __m128i __i = (i); \
1188 (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \
1189 (const __v4di *)__m, (__v4si)__i, \
1190 (__v4di)_mm256_set1_epi64x(-1), (s)); })
1191
1192#define _mm_i64gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines73a76582014-02-13 02:55:05 -08001193 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001194 __m128i __i = (i); \
1195 (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \
1196 (const __v2di *)__m, (__v2di)__i, \
1197 (__v2di)_mm_set1_epi64x(-1), (s)); })
1198
1199#define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \
Stephen Hines73a76582014-02-13 02:55:05 -08001200 long long const *__m = (m); \
Ying Wang60999142013-01-07 13:59:36 -08001201 __m256i __i = (i); \
1202 (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \
1203 (const __v4di *)__m, (__v4di)__i, \
1204 (__v4di)_mm256_set1_epi64x(-1), (s)); })
Stephen Hines996e4dc2013-08-13 01:04:14 -07001205
1206#endif /* __AVX2INTRIN_H */