blob: cfa91410a338aa649f176f6506e67a32b4a3895f [file] [log] [blame]
Craig Topperdec792e2011-12-19 05:04:33 +00001/*===---- avx2intrin.h - AVX2 intrinsics -----------------------------------===
2 *
3 * Permission is hereby granted, free of charge, to any person obtaining a copy
4 * of this software and associated documentation files (the "Software"), to deal
5 * in the Software without restriction, including without limitation the rights
6 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7 * copies of the Software, and to permit persons to whom the Software is
8 * furnished to do so, subject to the following conditions:
9 *
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
12 *
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
19 * THE SOFTWARE.
20 *
21 *===-----------------------------------------------------------------------===
22 */
23
24#ifndef __IMMINTRIN_H
25#error "Never use <avx2intrin.h> directly; include <immintrin.h> instead."
26#endif
27
Richard Smith49e56442013-07-14 05:41:45 +000028#ifndef __AVX2INTRIN_H
29#define __AVX2INTRIN_H
30
Eric Christopher4d1851682015-06-17 07:09:20 +000031/* Define the default attributes for the functions in this file. */
Michael Kupersteine45af542015-06-30 13:36:19 +000032#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx2")))
Eric Christopher4d1851682015-06-17 07:09:20 +000033
Craig Topperdec792e2011-12-19 05:04:33 +000034/* SSE4 Multiple Packed Sums of Absolute Difference. */
35#define _mm256_mpsadbw_epu8(X, Y, M) __builtin_ia32_mpsadbw256((X), (Y), (M))
36
Michael Kupersteine45af542015-06-30 13:36:19 +000037static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000038_mm256_abs_epi8(__m256i __a)
Craig Topperdec792e2011-12-19 05:04:33 +000039{
David Blaikie3302f2b2013-01-16 23:08:36 +000040 return (__m256i)__builtin_ia32_pabsb256((__v32qi)__a);
Craig Topperdec792e2011-12-19 05:04:33 +000041}
42
Michael Kupersteine45af542015-06-30 13:36:19 +000043static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000044_mm256_abs_epi16(__m256i __a)
Craig Topperdec792e2011-12-19 05:04:33 +000045{
David Blaikie3302f2b2013-01-16 23:08:36 +000046 return (__m256i)__builtin_ia32_pabsw256((__v16hi)__a);
Craig Topperdec792e2011-12-19 05:04:33 +000047}
48
Michael Kupersteine45af542015-06-30 13:36:19 +000049static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000050_mm256_abs_epi32(__m256i __a)
Craig Topperdec792e2011-12-19 05:04:33 +000051{
David Blaikie3302f2b2013-01-16 23:08:36 +000052 return (__m256i)__builtin_ia32_pabsd256((__v8si)__a);
Craig Topperdec792e2011-12-19 05:04:33 +000053}
54
Michael Kupersteine45af542015-06-30 13:36:19 +000055static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000056_mm256_packs_epi16(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000057{
David Blaikie3302f2b2013-01-16 23:08:36 +000058 return (__m256i)__builtin_ia32_packsswb256((__v16hi)__a, (__v16hi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000059}
60
Michael Kupersteine45af542015-06-30 13:36:19 +000061static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000062_mm256_packs_epi32(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000063{
David Blaikie3302f2b2013-01-16 23:08:36 +000064 return (__m256i)__builtin_ia32_packssdw256((__v8si)__a, (__v8si)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000065}
66
Michael Kupersteine45af542015-06-30 13:36:19 +000067static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000068_mm256_packus_epi16(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000069{
David Blaikie3302f2b2013-01-16 23:08:36 +000070 return (__m256i)__builtin_ia32_packuswb256((__v16hi)__a, (__v16hi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000071}
72
Michael Kupersteine45af542015-06-30 13:36:19 +000073static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topperdec792e2011-12-19 05:04:33 +000074_mm256_packus_epi32(__m256i __V1, __m256i __V2)
75{
76 return (__m256i) __builtin_ia32_packusdw256((__v8si)__V1, (__v8si)__V2);
77}
78
Michael Kupersteine45af542015-06-30 13:36:19 +000079static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000080_mm256_add_epi8(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000081{
David Blaikie3302f2b2013-01-16 23:08:36 +000082 return (__m256i)((__v32qi)__a + (__v32qi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000083}
84
Michael Kupersteine45af542015-06-30 13:36:19 +000085static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000086_mm256_add_epi16(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000087{
David Blaikie3302f2b2013-01-16 23:08:36 +000088 return (__m256i)((__v16hi)__a + (__v16hi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000089}
90
Michael Kupersteine45af542015-06-30 13:36:19 +000091static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000092_mm256_add_epi32(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000093{
David Blaikie3302f2b2013-01-16 23:08:36 +000094 return (__m256i)((__v8si)__a + (__v8si)__b);
Craig Topperdec792e2011-12-19 05:04:33 +000095}
96
Michael Kupersteine45af542015-06-30 13:36:19 +000097static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +000098_mm256_add_epi64(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +000099{
David Blaikie3302f2b2013-01-16 23:08:36 +0000100 return __a + __b;
Craig Topperdec792e2011-12-19 05:04:33 +0000101}
102
Michael Kupersteine45af542015-06-30 13:36:19 +0000103static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000104_mm256_adds_epi8(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000105{
David Blaikie3302f2b2013-01-16 23:08:36 +0000106 return (__m256i)__builtin_ia32_paddsb256((__v32qi)__a, (__v32qi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000107}
108
Michael Kupersteine45af542015-06-30 13:36:19 +0000109static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000110_mm256_adds_epi16(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000111{
David Blaikie3302f2b2013-01-16 23:08:36 +0000112 return (__m256i)__builtin_ia32_paddsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000113}
114
Michael Kupersteine45af542015-06-30 13:36:19 +0000115static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000116_mm256_adds_epu8(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000117{
David Blaikie3302f2b2013-01-16 23:08:36 +0000118 return (__m256i)__builtin_ia32_paddusb256((__v32qi)__a, (__v32qi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000119}
120
Michael Kupersteine45af542015-06-30 13:36:19 +0000121static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000122_mm256_adds_epu16(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000123{
David Blaikie3302f2b2013-01-16 23:08:36 +0000124 return (__m256i)__builtin_ia32_paddusw256((__v16hi)__a, (__v16hi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000125}
126
127#define _mm256_alignr_epi8(a, b, n) __extension__ ({ \
128 __m256i __a = (a); \
129 __m256i __b = (b); \
130 (__m256i)__builtin_ia32_palignr256((__v32qi)__a, (__v32qi)__b, (n)); })
131
Michael Kupersteine45af542015-06-30 13:36:19 +0000132static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000133_mm256_and_si256(__m256i __a, __m256i __b)
Craig Toppera557e1c2011-12-19 09:03:48 +0000134{
David Blaikie3302f2b2013-01-16 23:08:36 +0000135 return __a & __b;
Craig Toppera557e1c2011-12-19 09:03:48 +0000136}
137
Michael Kupersteine45af542015-06-30 13:36:19 +0000138static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000139_mm256_andnot_si256(__m256i __a, __m256i __b)
Craig Toppera557e1c2011-12-19 09:03:48 +0000140{
David Blaikie3302f2b2013-01-16 23:08:36 +0000141 return ~__a & __b;
Craig Toppera557e1c2011-12-19 09:03:48 +0000142}
143
Michael Kupersteine45af542015-06-30 13:36:19 +0000144static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000145_mm256_avg_epu8(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000146{
David Blaikie3302f2b2013-01-16 23:08:36 +0000147 return (__m256i)__builtin_ia32_pavgb256((__v32qi)__a, (__v32qi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000148}
149
Michael Kupersteine45af542015-06-30 13:36:19 +0000150static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000151_mm256_avg_epu16(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000152{
David Blaikie3302f2b2013-01-16 23:08:36 +0000153 return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000154}
155
Michael Kupersteine45af542015-06-30 13:36:19 +0000156static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Toppera89747d2011-12-20 09:55:26 +0000157_mm256_blendv_epi8(__m256i __V1, __m256i __V2, __m256i __M)
158{
159 return (__m256i)__builtin_ia32_pblendvb256((__v32qi)__V1, (__v32qi)__V2,
160 (__v32qi)__M);
161}
162
163#define _mm256_blend_epi16(V1, V2, M) __extension__ ({ \
164 __m256i __V1 = (V1); \
165 __m256i __V2 = (V2); \
Ekaterina Romanovab929ad72015-04-10 02:39:45 +0000166 (__m256i)__builtin_shufflevector((__v16hi)__V1, (__v16hi)__V2, \
Filipe Cabecinhas5d289b42014-05-13 02:37:02 +0000167 (((M) & 0x01) ? 16 : 0), \
168 (((M) & 0x02) ? 17 : 1), \
169 (((M) & 0x04) ? 18 : 2), \
170 (((M) & 0x08) ? 19 : 3), \
171 (((M) & 0x10) ? 20 : 4), \
172 (((M) & 0x20) ? 21 : 5), \
173 (((M) & 0x40) ? 22 : 6), \
174 (((M) & 0x80) ? 23 : 7), \
175 (((M) & 0x01) ? 24 : 8), \
176 (((M) & 0x02) ? 25 : 9), \
177 (((M) & 0x04) ? 26 : 10), \
178 (((M) & 0x08) ? 27 : 11), \
179 (((M) & 0x10) ? 28 : 12), \
180 (((M) & 0x20) ? 29 : 13), \
181 (((M) & 0x40) ? 30 : 14), \
182 (((M) & 0x80) ? 31 : 15)); })
Craig Toppera89747d2011-12-20 09:55:26 +0000183
Michael Kupersteine45af542015-06-30 13:36:19 +0000184static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000185_mm256_cmpeq_epi8(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000186{
David Blaikie3302f2b2013-01-16 23:08:36 +0000187 return (__m256i)((__v32qi)__a == (__v32qi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000188}
189
Michael Kupersteine45af542015-06-30 13:36:19 +0000190static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000191_mm256_cmpeq_epi16(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000192{
David Blaikie3302f2b2013-01-16 23:08:36 +0000193 return (__m256i)((__v16hi)__a == (__v16hi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000194}
195
Michael Kupersteine45af542015-06-30 13:36:19 +0000196static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000197_mm256_cmpeq_epi32(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000198{
David Blaikie3302f2b2013-01-16 23:08:36 +0000199 return (__m256i)((__v8si)__a == (__v8si)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000200}
201
Michael Kupersteine45af542015-06-30 13:36:19 +0000202static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000203_mm256_cmpeq_epi64(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000204{
David Blaikie3302f2b2013-01-16 23:08:36 +0000205 return (__m256i)(__a == __b);
Craig Toppera89747d2011-12-20 09:55:26 +0000206}
207
Michael Kupersteine45af542015-06-30 13:36:19 +0000208static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000209_mm256_cmpgt_epi8(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000210{
David Blaikie3302f2b2013-01-16 23:08:36 +0000211 return (__m256i)((__v32qi)__a > (__v32qi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000212}
213
Michael Kupersteine45af542015-06-30 13:36:19 +0000214static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000215_mm256_cmpgt_epi16(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000216{
David Blaikie3302f2b2013-01-16 23:08:36 +0000217 return (__m256i)((__v16hi)__a > (__v16hi)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000218}
219
Michael Kupersteine45af542015-06-30 13:36:19 +0000220static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000221_mm256_cmpgt_epi32(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000222{
David Blaikie3302f2b2013-01-16 23:08:36 +0000223 return (__m256i)((__v8si)__a > (__v8si)__b);
Craig Toppera89747d2011-12-20 09:55:26 +0000224}
225
Michael Kupersteine45af542015-06-30 13:36:19 +0000226static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000227_mm256_cmpgt_epi64(__m256i __a, __m256i __b)
Craig Toppera89747d2011-12-20 09:55:26 +0000228{
David Blaikie3302f2b2013-01-16 23:08:36 +0000229 return (__m256i)(__a > __b);
Craig Toppera89747d2011-12-20 09:55:26 +0000230}
231
Michael Kupersteine45af542015-06-30 13:36:19 +0000232static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000233_mm256_hadd_epi16(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000234{
David Blaikie3302f2b2013-01-16 23:08:36 +0000235 return (__m256i)__builtin_ia32_phaddw256((__v16hi)__a, (__v16hi)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000236}
237
Michael Kupersteine45af542015-06-30 13:36:19 +0000238static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000239_mm256_hadd_epi32(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000240{
David Blaikie3302f2b2013-01-16 23:08:36 +0000241 return (__m256i)__builtin_ia32_phaddd256((__v8si)__a, (__v8si)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000242}
243
Michael Kupersteine45af542015-06-30 13:36:19 +0000244static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000245_mm256_hadds_epi16(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000246{
David Blaikie3302f2b2013-01-16 23:08:36 +0000247 return (__m256i)__builtin_ia32_phaddsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000248}
249
Michael Kupersteine45af542015-06-30 13:36:19 +0000250static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000251_mm256_hsub_epi16(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000252{
David Blaikie3302f2b2013-01-16 23:08:36 +0000253 return (__m256i)__builtin_ia32_phsubw256((__v16hi)__a, (__v16hi)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000254}
255
Michael Kupersteine45af542015-06-30 13:36:19 +0000256static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000257_mm256_hsub_epi32(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000258{
David Blaikie3302f2b2013-01-16 23:08:36 +0000259 return (__m256i)__builtin_ia32_phsubd256((__v8si)__a, (__v8si)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000260}
261
Michael Kupersteine45af542015-06-30 13:36:19 +0000262static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000263_mm256_hsubs_epi16(__m256i __a, __m256i __b)
Craig Topper3fe5ac42011-12-21 08:17:40 +0000264{
David Blaikie3302f2b2013-01-16 23:08:36 +0000265 return (__m256i)__builtin_ia32_phsubsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper3fe5ac42011-12-21 08:17:40 +0000266}
267
Michael Kupersteine45af542015-06-30 13:36:19 +0000268static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000269_mm256_maddubs_epi16(__m256i __a, __m256i __b)
Craig Toppera73baa82011-12-21 08:35:05 +0000270{
David Blaikie3302f2b2013-01-16 23:08:36 +0000271 return (__m256i)__builtin_ia32_pmaddubsw256((__v32qi)__a, (__v32qi)__b);
Craig Toppera73baa82011-12-21 08:35:05 +0000272}
273
Michael Kupersteine45af542015-06-30 13:36:19 +0000274static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000275_mm256_madd_epi16(__m256i __a, __m256i __b)
Craig Toppera73baa82011-12-21 08:35:05 +0000276{
David Blaikie3302f2b2013-01-16 23:08:36 +0000277 return (__m256i)__builtin_ia32_pmaddwd256((__v16hi)__a, (__v16hi)__b);
Craig Toppera73baa82011-12-21 08:35:05 +0000278}
279
Michael Kupersteine45af542015-06-30 13:36:19 +0000280static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000281_mm256_max_epi8(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000282{
David Blaikie3302f2b2013-01-16 23:08:36 +0000283 return (__m256i)__builtin_ia32_pmaxsb256((__v32qi)__a, (__v32qi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000284}
285
Michael Kupersteine45af542015-06-30 13:36:19 +0000286static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000287_mm256_max_epi16(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000288{
David Blaikie3302f2b2013-01-16 23:08:36 +0000289 return (__m256i)__builtin_ia32_pmaxsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000290}
291
Michael Kupersteine45af542015-06-30 13:36:19 +0000292static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000293_mm256_max_epi32(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000294{
David Blaikie3302f2b2013-01-16 23:08:36 +0000295 return (__m256i)__builtin_ia32_pmaxsd256((__v8si)__a, (__v8si)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000296}
297
Michael Kupersteine45af542015-06-30 13:36:19 +0000298static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000299_mm256_max_epu8(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000300{
David Blaikie3302f2b2013-01-16 23:08:36 +0000301 return (__m256i)__builtin_ia32_pmaxub256((__v32qi)__a, (__v32qi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000302}
303
Michael Kupersteine45af542015-06-30 13:36:19 +0000304static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000305_mm256_max_epu16(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000306{
David Blaikie3302f2b2013-01-16 23:08:36 +0000307 return (__m256i)__builtin_ia32_pmaxuw256((__v16hi)__a, (__v16hi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000308}
309
Michael Kupersteine45af542015-06-30 13:36:19 +0000310static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000311_mm256_max_epu32(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000312{
David Blaikie3302f2b2013-01-16 23:08:36 +0000313 return (__m256i)__builtin_ia32_pmaxud256((__v8si)__a, (__v8si)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000314}
315
Michael Kupersteine45af542015-06-30 13:36:19 +0000316static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000317_mm256_min_epi8(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000318{
David Blaikie3302f2b2013-01-16 23:08:36 +0000319 return (__m256i)__builtin_ia32_pminsb256((__v32qi)__a, (__v32qi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000320}
321
Michael Kupersteine45af542015-06-30 13:36:19 +0000322static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000323_mm256_min_epi16(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000324{
David Blaikie3302f2b2013-01-16 23:08:36 +0000325 return (__m256i)__builtin_ia32_pminsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000326}
327
Michael Kupersteine45af542015-06-30 13:36:19 +0000328static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000329_mm256_min_epi32(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000330{
David Blaikie3302f2b2013-01-16 23:08:36 +0000331 return (__m256i)__builtin_ia32_pminsd256((__v8si)__a, (__v8si)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000332}
333
Michael Kupersteine45af542015-06-30 13:36:19 +0000334static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000335_mm256_min_epu8(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000336{
David Blaikie3302f2b2013-01-16 23:08:36 +0000337 return (__m256i)__builtin_ia32_pminub256((__v32qi)__a, (__v32qi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000338}
339
Michael Kupersteine45af542015-06-30 13:36:19 +0000340static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000341_mm256_min_epu16(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000342{
David Blaikie3302f2b2013-01-16 23:08:36 +0000343 return (__m256i)__builtin_ia32_pminuw256 ((__v16hi)__a, (__v16hi)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000344}
345
Michael Kupersteine45af542015-06-30 13:36:19 +0000346static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000347_mm256_min_epu32(__m256i __a, __m256i __b)
Craig Topper1f2460a2011-12-22 09:18:58 +0000348{
David Blaikie3302f2b2013-01-16 23:08:36 +0000349 return (__m256i)__builtin_ia32_pminud256((__v8si)__a, (__v8si)__b);
Craig Topper1f2460a2011-12-22 09:18:58 +0000350}
351
Michael Kupersteine45af542015-06-30 13:36:19 +0000352static __inline__ int __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000353_mm256_movemask_epi8(__m256i __a)
Craig Topper1f2460a2011-12-22 09:18:58 +0000354{
David Blaikie3302f2b2013-01-16 23:08:36 +0000355 return __builtin_ia32_pmovmskb256((__v32qi)__a);
Craig Topper1f2460a2011-12-22 09:18:58 +0000356}
357
Michael Kupersteine45af542015-06-30 13:36:19 +0000358static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000359_mm256_cvtepi8_epi16(__m128i __V)
360{
361 return (__m256i)__builtin_ia32_pmovsxbw256((__v16qi)__V);
362}
363
Michael Kupersteine45af542015-06-30 13:36:19 +0000364static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000365_mm256_cvtepi8_epi32(__m128i __V)
366{
367 return (__m256i)__builtin_ia32_pmovsxbd256((__v16qi)__V);
368}
369
Michael Kupersteine45af542015-06-30 13:36:19 +0000370static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000371_mm256_cvtepi8_epi64(__m128i __V)
372{
373 return (__m256i)__builtin_ia32_pmovsxbq256((__v16qi)__V);
374}
375
Michael Kupersteine45af542015-06-30 13:36:19 +0000376static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000377_mm256_cvtepi16_epi32(__m128i __V)
378{
379 return (__m256i)__builtin_ia32_pmovsxwd256((__v8hi)__V);
380}
381
Michael Kupersteine45af542015-06-30 13:36:19 +0000382static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000383_mm256_cvtepi16_epi64(__m128i __V)
384{
385 return (__m256i)__builtin_ia32_pmovsxwq256((__v8hi)__V);
386}
387
Michael Kupersteine45af542015-06-30 13:36:19 +0000388static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000389_mm256_cvtepi32_epi64(__m128i __V)
390{
391 return (__m256i)__builtin_ia32_pmovsxdq256((__v4si)__V);
392}
393
Michael Kupersteine45af542015-06-30 13:36:19 +0000394static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000395_mm256_cvtepu8_epi16(__m128i __V)
396{
397 return (__m256i)__builtin_ia32_pmovzxbw256((__v16qi)__V);
398}
399
Michael Kupersteine45af542015-06-30 13:36:19 +0000400static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000401_mm256_cvtepu8_epi32(__m128i __V)
402{
403 return (__m256i)__builtin_ia32_pmovzxbd256((__v16qi)__V);
404}
405
Michael Kupersteine45af542015-06-30 13:36:19 +0000406static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000407_mm256_cvtepu8_epi64(__m128i __V)
408{
409 return (__m256i)__builtin_ia32_pmovzxbq256((__v16qi)__V);
410}
411
Michael Kupersteine45af542015-06-30 13:36:19 +0000412static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000413_mm256_cvtepu16_epi32(__m128i __V)
414{
415 return (__m256i)__builtin_ia32_pmovzxwd256((__v8hi)__V);
416}
417
Michael Kupersteine45af542015-06-30 13:36:19 +0000418static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000419_mm256_cvtepu16_epi64(__m128i __V)
420{
421 return (__m256i)__builtin_ia32_pmovzxwq256((__v8hi)__V);
422}
423
Michael Kupersteine45af542015-06-30 13:36:19 +0000424static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper1f2460a2011-12-22 09:18:58 +0000425_mm256_cvtepu32_epi64(__m128i __V)
426{
427 return (__m256i)__builtin_ia32_pmovzxdq256((__v4si)__V);
428}
429
Michael Kupersteine45af542015-06-30 13:36:19 +0000430static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000431_mm256_mul_epi32(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000432{
David Blaikie3302f2b2013-01-16 23:08:36 +0000433 return (__m256i)__builtin_ia32_pmuldq256((__v8si)__a, (__v8si)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000434}
435
Michael Kupersteine45af542015-06-30 13:36:19 +0000436static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000437_mm256_mulhrs_epi16(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000438{
David Blaikie3302f2b2013-01-16 23:08:36 +0000439 return (__m256i)__builtin_ia32_pmulhrsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000440}
441
Michael Kupersteine45af542015-06-30 13:36:19 +0000442static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000443_mm256_mulhi_epu16(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000444{
David Blaikie3302f2b2013-01-16 23:08:36 +0000445 return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000446}
447
Michael Kupersteine45af542015-06-30 13:36:19 +0000448static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000449_mm256_mulhi_epi16(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000450{
David Blaikie3302f2b2013-01-16 23:08:36 +0000451 return (__m256i)__builtin_ia32_pmulhw256((__v16hi)__a, (__v16hi)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000452}
453
Michael Kupersteine45af542015-06-30 13:36:19 +0000454static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000455_mm256_mullo_epi16(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000456{
David Blaikie3302f2b2013-01-16 23:08:36 +0000457 return (__m256i)((__v16hi)__a * (__v16hi)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000458}
459
Michael Kupersteine45af542015-06-30 13:36:19 +0000460static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000461_mm256_mullo_epi32 (__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000462{
David Blaikie3302f2b2013-01-16 23:08:36 +0000463 return (__m256i)((__v8si)__a * (__v8si)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000464}
465
Michael Kupersteine45af542015-06-30 13:36:19 +0000466static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000467_mm256_mul_epu32(__m256i __a, __m256i __b)
Craig Topper235a3652011-12-23 08:31:16 +0000468{
David Blaikie3302f2b2013-01-16 23:08:36 +0000469 return __builtin_ia32_pmuludq256((__v8si)__a, (__v8si)__b);
Craig Topper235a3652011-12-23 08:31:16 +0000470}
471
Michael Kupersteine45af542015-06-30 13:36:19 +0000472static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000473_mm256_or_si256(__m256i __a, __m256i __b)
Craig Toppera557e1c2011-12-19 09:03:48 +0000474{
David Blaikie3302f2b2013-01-16 23:08:36 +0000475 return __a | __b;
Craig Toppera557e1c2011-12-19 09:03:48 +0000476}
477
Michael Kupersteine45af542015-06-30 13:36:19 +0000478static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000479_mm256_sad_epu8(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000480{
David Blaikie3302f2b2013-01-16 23:08:36 +0000481 return __builtin_ia32_psadbw256((__v32qi)__a, (__v32qi)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000482}
483
Michael Kupersteine45af542015-06-30 13:36:19 +0000484static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000485_mm256_shuffle_epi8(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000486{
David Blaikie3302f2b2013-01-16 23:08:36 +0000487 return (__m256i)__builtin_ia32_pshufb256((__v32qi)__a, (__v32qi)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000488}
489
490#define _mm256_shuffle_epi32(a, imm) __extension__ ({ \
491 __m256i __a = (a); \
492 (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)_mm256_set1_epi32(0), \
493 (imm) & 0x3, ((imm) & 0xc) >> 2, \
494 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
495 4 + (((imm) & 0x03) >> 0), \
496 4 + (((imm) & 0x0c) >> 2), \
497 4 + (((imm) & 0x30) >> 4), \
498 4 + (((imm) & 0xc0) >> 6)); })
499
500#define _mm256_shufflehi_epi16(a, imm) __extension__ ({ \
501 __m256i __a = (a); \
502 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
503 0, 1, 2, 3, \
504 4 + (((imm) & 0x03) >> 0), \
505 4 + (((imm) & 0x0c) >> 2), \
506 4 + (((imm) & 0x30) >> 4), \
507 4 + (((imm) & 0xc0) >> 6), \
508 8, 9, 10, 11, \
509 12 + (((imm) & 0x03) >> 0), \
510 12 + (((imm) & 0x0c) >> 2), \
511 12 + (((imm) & 0x30) >> 4), \
512 12 + (((imm) & 0xc0) >> 6)); })
513
514#define _mm256_shufflelo_epi16(a, imm) __extension__ ({ \
515 __m256i __a = (a); \
516 (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)_mm256_set1_epi16(0), \
517 (imm) & 0x3,((imm) & 0xc) >> 2, \
518 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
519 4, 5, 6, 7, \
520 8 + (((imm) & 0x03) >> 0), \
521 8 + (((imm) & 0x0c) >> 2), \
522 8 + (((imm) & 0x30) >> 4), \
523 8 + (((imm) & 0xc0) >> 6), \
524 12, 13, 14, 15); })
525
Michael Kupersteine45af542015-06-30 13:36:19 +0000526static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000527_mm256_sign_epi8(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000528{
David Blaikie3302f2b2013-01-16 23:08:36 +0000529 return (__m256i)__builtin_ia32_psignb256((__v32qi)__a, (__v32qi)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000530}
531
Michael Kupersteine45af542015-06-30 13:36:19 +0000532static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000533_mm256_sign_epi16(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000534{
David Blaikie3302f2b2013-01-16 23:08:36 +0000535 return (__m256i)__builtin_ia32_psignw256((__v16hi)__a, (__v16hi)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000536}
537
Michael Kupersteine45af542015-06-30 13:36:19 +0000538static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000539_mm256_sign_epi32(__m256i __a, __m256i __b)
Craig Topperf4bb9522011-12-24 03:28:57 +0000540{
David Blaikie3302f2b2013-01-16 23:08:36 +0000541 return (__m256i)__builtin_ia32_psignd256((__v8si)__a, (__v8si)__b);
Craig Topperf4bb9522011-12-24 03:28:57 +0000542}
543
544#define _mm256_slli_si256(a, count) __extension__ ({ \
545 __m256i __a = (a); \
546 (__m256i)__builtin_ia32_pslldqi256(__a, (count)*8); })
547
Michael Kuperstein61681832015-05-19 13:05:46 +0000548#define _mm256_bslli_epi128(a, count) _mm256_slli_si256((a), (count))
549
Michael Kupersteine45af542015-06-30 13:36:19 +0000550static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000551_mm256_slli_epi16(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000552{
David Blaikie3302f2b2013-01-16 23:08:36 +0000553 return (__m256i)__builtin_ia32_psllwi256((__v16hi)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000554}
555
Michael Kupersteine45af542015-06-30 13:36:19 +0000556static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000557_mm256_sll_epi16(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000558{
David Blaikie3302f2b2013-01-16 23:08:36 +0000559 return (__m256i)__builtin_ia32_psllw256((__v16hi)__a, (__v8hi)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000560}
561
Michael Kupersteine45af542015-06-30 13:36:19 +0000562static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000563_mm256_slli_epi32(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000564{
David Blaikie3302f2b2013-01-16 23:08:36 +0000565 return (__m256i)__builtin_ia32_pslldi256((__v8si)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000566}
567
Michael Kupersteine45af542015-06-30 13:36:19 +0000568static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000569_mm256_sll_epi32(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000570{
David Blaikie3302f2b2013-01-16 23:08:36 +0000571 return (__m256i)__builtin_ia32_pslld256((__v8si)__a, (__v4si)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000572}
573
Michael Kupersteine45af542015-06-30 13:36:19 +0000574static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000575_mm256_slli_epi64(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000576{
David Blaikie3302f2b2013-01-16 23:08:36 +0000577 return __builtin_ia32_psllqi256(__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000578}
579
Michael Kupersteine45af542015-06-30 13:36:19 +0000580static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000581_mm256_sll_epi64(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000582{
David Blaikie3302f2b2013-01-16 23:08:36 +0000583 return __builtin_ia32_psllq256(__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000584}
585
Michael Kupersteine45af542015-06-30 13:36:19 +0000586static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000587_mm256_srai_epi16(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000588{
David Blaikie3302f2b2013-01-16 23:08:36 +0000589 return (__m256i)__builtin_ia32_psrawi256((__v16hi)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000590}
591
Michael Kupersteine45af542015-06-30 13:36:19 +0000592static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000593_mm256_sra_epi16(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000594{
David Blaikie3302f2b2013-01-16 23:08:36 +0000595 return (__m256i)__builtin_ia32_psraw256((__v16hi)__a, (__v8hi)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000596}
597
Michael Kupersteine45af542015-06-30 13:36:19 +0000598static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000599_mm256_srai_epi32(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000600{
David Blaikie3302f2b2013-01-16 23:08:36 +0000601 return (__m256i)__builtin_ia32_psradi256((__v8si)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000602}
603
Michael Kupersteine45af542015-06-30 13:36:19 +0000604static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000605_mm256_sra_epi32(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000606{
David Blaikie3302f2b2013-01-16 23:08:36 +0000607 return (__m256i)__builtin_ia32_psrad256((__v8si)__a, (__v4si)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000608}
609
610#define _mm256_srli_si256(a, count) __extension__ ({ \
611 __m256i __a = (a); \
612 (__m256i)__builtin_ia32_psrldqi256(__a, (count)*8); })
613
Michael Kuperstein61681832015-05-19 13:05:46 +0000614#define _mm256_bsrli_epi128(a, count) _mm256_srli_si256((a), (count))
615
Michael Kupersteine45af542015-06-30 13:36:19 +0000616static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000617_mm256_srli_epi16(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000618{
David Blaikie3302f2b2013-01-16 23:08:36 +0000619 return (__m256i)__builtin_ia32_psrlwi256((__v16hi)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000620}
621
Michael Kupersteine45af542015-06-30 13:36:19 +0000622static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000623_mm256_srl_epi16(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000624{
David Blaikie3302f2b2013-01-16 23:08:36 +0000625 return (__m256i)__builtin_ia32_psrlw256((__v16hi)__a, (__v8hi)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000626}
627
Michael Kupersteine45af542015-06-30 13:36:19 +0000628static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000629_mm256_srli_epi32(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000630{
David Blaikie3302f2b2013-01-16 23:08:36 +0000631 return (__m256i)__builtin_ia32_psrldi256((__v8si)__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000632}
633
Michael Kupersteine45af542015-06-30 13:36:19 +0000634static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000635_mm256_srl_epi32(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000636{
David Blaikie3302f2b2013-01-16 23:08:36 +0000637 return (__m256i)__builtin_ia32_psrld256((__v8si)__a, (__v4si)__count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000638}
639
Michael Kupersteine45af542015-06-30 13:36:19 +0000640static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000641_mm256_srli_epi64(__m256i __a, int __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000642{
David Blaikie3302f2b2013-01-16 23:08:36 +0000643 return __builtin_ia32_psrlqi256(__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000644}
645
Michael Kupersteine45af542015-06-30 13:36:19 +0000646static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000647_mm256_srl_epi64(__m256i __a, __m128i __count)
Craig Topperf4bb9522011-12-24 03:28:57 +0000648{
David Blaikie3302f2b2013-01-16 23:08:36 +0000649 return __builtin_ia32_psrlq256(__a, __count);
Craig Topperf4bb9522011-12-24 03:28:57 +0000650}
651
Michael Kupersteine45af542015-06-30 13:36:19 +0000652static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000653_mm256_sub_epi8(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +0000654{
David Blaikie3302f2b2013-01-16 23:08:36 +0000655 return (__m256i)((__v32qi)__a - (__v32qi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +0000656}
657
Michael Kupersteine45af542015-06-30 13:36:19 +0000658static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000659_mm256_sub_epi16(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +0000660{
David Blaikie3302f2b2013-01-16 23:08:36 +0000661 return (__m256i)((__v16hi)__a - (__v16hi)__b);
Craig Topperdec792e2011-12-19 05:04:33 +0000662}
663
Michael Kupersteine45af542015-06-30 13:36:19 +0000664static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000665_mm256_sub_epi32(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +0000666{
David Blaikie3302f2b2013-01-16 23:08:36 +0000667 return (__m256i)((__v8si)__a - (__v8si)__b);
Craig Topperdec792e2011-12-19 05:04:33 +0000668}
669
Michael Kupersteine45af542015-06-30 13:36:19 +0000670static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000671_mm256_sub_epi64(__m256i __a, __m256i __b)
Craig Topperdec792e2011-12-19 05:04:33 +0000672{
David Blaikie3302f2b2013-01-16 23:08:36 +0000673 return __a - __b;
Craig Topperdec792e2011-12-19 05:04:33 +0000674}
Craig Topper94aba2c2011-12-19 07:03:25 +0000675
Michael Kupersteine45af542015-06-30 13:36:19 +0000676static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000677_mm256_subs_epi8(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000678{
David Blaikie3302f2b2013-01-16 23:08:36 +0000679 return (__m256i)__builtin_ia32_psubsb256((__v32qi)__a, (__v32qi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000680}
681
Michael Kupersteine45af542015-06-30 13:36:19 +0000682static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000683_mm256_subs_epi16(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000684{
David Blaikie3302f2b2013-01-16 23:08:36 +0000685 return (__m256i)__builtin_ia32_psubsw256((__v16hi)__a, (__v16hi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000686}
687
Michael Kupersteine45af542015-06-30 13:36:19 +0000688static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000689_mm256_subs_epu8(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000690{
David Blaikie3302f2b2013-01-16 23:08:36 +0000691 return (__m256i)__builtin_ia32_psubusb256((__v32qi)__a, (__v32qi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000692}
693
Michael Kupersteine45af542015-06-30 13:36:19 +0000694static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000695_mm256_subs_epu16(__m256i __a, __m256i __b)
Craig Topper94aba2c2011-12-19 07:03:25 +0000696{
David Blaikie3302f2b2013-01-16 23:08:36 +0000697 return (__m256i)__builtin_ia32_psubusw256((__v16hi)__a, (__v16hi)__b);
Craig Topper94aba2c2011-12-19 07:03:25 +0000698}
699
Michael Kupersteine45af542015-06-30 13:36:19 +0000700static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000701_mm256_unpackhi_epi8(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000702{
David Blaikie3302f2b2013-01-16 23:08:36 +0000703 return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 8, 32+8, 9, 32+9, 10, 32+10, 11, 32+11, 12, 32+12, 13, 32+13, 14, 32+14, 15, 32+15, 24, 32+24, 25, 32+25, 26, 32+26, 27, 32+27, 28, 32+28, 29, 32+29, 30, 32+30, 31, 32+31);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000704}
705
Michael Kupersteine45af542015-06-30 13:36:19 +0000706static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000707_mm256_unpackhi_epi16(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000708{
David Blaikie3302f2b2013-01-16 23:08:36 +0000709 return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000710}
711
Michael Kupersteine45af542015-06-30 13:36:19 +0000712static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000713_mm256_unpackhi_epi32(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000714{
David Blaikie3302f2b2013-01-16 23:08:36 +0000715 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 2, 8+2, 3, 8+3, 6, 8+6, 7, 8+7);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000716}
717
Michael Kupersteine45af542015-06-30 13:36:19 +0000718static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000719_mm256_unpackhi_epi64(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000720{
David Blaikie3302f2b2013-01-16 23:08:36 +0000721 return (__m256i)__builtin_shufflevector(__a, __b, 1, 4+1, 3, 4+3);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000722}
723
Michael Kupersteine45af542015-06-30 13:36:19 +0000724static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000725_mm256_unpacklo_epi8(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000726{
David Blaikie3302f2b2013-01-16 23:08:36 +0000727 return (__m256i)__builtin_shufflevector((__v32qi)__a, (__v32qi)__b, 0, 32+0, 1, 32+1, 2, 32+2, 3, 32+3, 4, 32+4, 5, 32+5, 6, 32+6, 7, 32+7, 16, 32+16, 17, 32+17, 18, 32+18, 19, 32+19, 20, 32+20, 21, 32+21, 22, 32+22, 23, 32+23);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000728}
729
Michael Kupersteine45af542015-06-30 13:36:19 +0000730static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000731_mm256_unpacklo_epi16(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000732{
David Blaikie3302f2b2013-01-16 23:08:36 +0000733 return (__m256i)__builtin_shufflevector((__v16hi)__a, (__v16hi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000734}
735
Michael Kupersteine45af542015-06-30 13:36:19 +0000736static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000737_mm256_unpacklo_epi32(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000738{
David Blaikie3302f2b2013-01-16 23:08:36 +0000739 return (__m256i)__builtin_shufflevector((__v8si)__a, (__v8si)__b, 0, 8+0, 1, 8+1, 4, 8+4, 5, 8+5);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000740}
741
Michael Kupersteine45af542015-06-30 13:36:19 +0000742static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000743_mm256_unpacklo_epi64(__m256i __a, __m256i __b)
Craig Toppera6fdbd12011-12-24 03:58:43 +0000744{
David Blaikie3302f2b2013-01-16 23:08:36 +0000745 return (__m256i)__builtin_shufflevector(__a, __b, 0, 4+0, 2, 4+2);
Craig Toppera6fdbd12011-12-24 03:58:43 +0000746}
747
Michael Kupersteine45af542015-06-30 13:36:19 +0000748static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000749_mm256_xor_si256(__m256i __a, __m256i __b)
Craig Toppera557e1c2011-12-19 09:03:48 +0000750{
David Blaikie3302f2b2013-01-16 23:08:36 +0000751 return __a ^ __b;
Craig Toppera557e1c2011-12-19 09:03:48 +0000752}
Craig Topper94798952011-12-24 05:19:29 +0000753
Michael Kupersteine45af542015-06-30 13:36:19 +0000754static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper94798952011-12-24 05:19:29 +0000755_mm256_stream_load_si256(__m256i *__V)
756{
757 return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V);
758}
759
Michael Kupersteine45af542015-06-30 13:36:19 +0000760static __inline__ __m128 __DEFAULT_FN_ATTRS
Craig Topper94798952011-12-24 05:19:29 +0000761_mm_broadcastss_ps(__m128 __X)
762{
763 return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X);
764}
765
Michael Kupersteine45af542015-06-30 13:36:19 +0000766static __inline__ __m128d __DEFAULT_FN_ATTRS
Michael Kuperstein877f3cb2015-05-19 14:49:14 +0000767_mm_broadcastsd_pd(__m128d __a)
768{
769 return __builtin_shufflevector(__a, __a, 0, 0);
770}
771
Michael Kupersteine45af542015-06-30 13:36:19 +0000772static __inline__ __m256 __DEFAULT_FN_ATTRS
Craig Topper94798952011-12-24 05:19:29 +0000773_mm256_broadcastss_ps(__m128 __X)
774{
775 return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X);
776}
777
Michael Kupersteine45af542015-06-30 13:36:19 +0000778static __inline__ __m256d __DEFAULT_FN_ATTRS
Craig Topper94798952011-12-24 05:19:29 +0000779_mm256_broadcastsd_pd(__m128d __X)
780{
781 return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X);
782}
783
Michael Kupersteine45af542015-06-30 13:36:19 +0000784static __inline__ __m256i __DEFAULT_FN_ATTRS
Juergen Ributzka2c2dbf42013-08-17 16:40:09 +0000785_mm256_broadcastsi128_si256(__m128i __X)
Craig Topper94798952011-12-24 05:19:29 +0000786{
Juergen Ributzka9baa03f2015-03-03 17:22:53 +0000787 return (__m256i)__builtin_shufflevector(__X, __X, 0, 1, 0, 1);
Craig Topper94798952011-12-24 05:19:29 +0000788}
789
790#define _mm_blend_epi32(V1, V2, M) __extension__ ({ \
791 __m128i __V1 = (V1); \
792 __m128i __V2 = (V2); \
Filipe Cabecinhas5d289b42014-05-13 02:37:02 +0000793 (__m128i)__builtin_shufflevector((__v4si)__V1, (__v4si)__V2, \
794 (((M) & 0x01) ? 4 : 0), \
795 (((M) & 0x02) ? 5 : 1), \
796 (((M) & 0x04) ? 6 : 2), \
797 (((M) & 0x08) ? 7 : 3)); })
Craig Topper94798952011-12-24 05:19:29 +0000798
799#define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \
800 __m256i __V1 = (V1); \
801 __m256i __V2 = (V2); \
Filipe Cabecinhas5d289b42014-05-13 02:37:02 +0000802 (__m256i)__builtin_shufflevector((__v8si)__V1, (__v8si)__V2, \
803 (((M) & 0x01) ? 8 : 0), \
804 (((M) & 0x02) ? 9 : 1), \
805 (((M) & 0x04) ? 10 : 2), \
806 (((M) & 0x08) ? 11 : 3), \
807 (((M) & 0x10) ? 12 : 4), \
808 (((M) & 0x20) ? 13 : 5), \
809 (((M) & 0x40) ? 14 : 6), \
810 (((M) & 0x80) ? 15 : 7)); })
Craig Topper9f009482011-12-24 07:55:14 +0000811
Michael Kupersteine45af542015-06-30 13:36:19 +0000812static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper9f009482011-12-24 07:55:14 +0000813_mm256_broadcastb_epi8(__m128i __X)
814{
815 return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X);
816}
817
Michael Kupersteine45af542015-06-30 13:36:19 +0000818static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper9f009482011-12-24 07:55:14 +0000819_mm256_broadcastw_epi16(__m128i __X)
820{
821 return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X);
822}
823
Michael Kupersteine45af542015-06-30 13:36:19 +0000824static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper9f009482011-12-24 07:55:14 +0000825_mm256_broadcastd_epi32(__m128i __X)
826{
827 return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X);
828}
829
Michael Kupersteine45af542015-06-30 13:36:19 +0000830static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper9f009482011-12-24 07:55:14 +0000831_mm256_broadcastq_epi64(__m128i __X)
832{
833 return (__m256i)__builtin_ia32_pbroadcastq256(__X);
834}
835
Michael Kupersteine45af542015-06-30 13:36:19 +0000836static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper9f009482011-12-24 07:55:14 +0000837_mm_broadcastb_epi8(__m128i __X)
838{
839 return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X);
840}
841
Michael Kupersteine45af542015-06-30 13:36:19 +0000842static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper9f009482011-12-24 07:55:14 +0000843_mm_broadcastw_epi16(__m128i __X)
844{
845 return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X);
846}
847
848
Michael Kupersteine45af542015-06-30 13:36:19 +0000849static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper9f009482011-12-24 07:55:14 +0000850_mm_broadcastd_epi32(__m128i __X)
851{
852 return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X);
853}
854
Michael Kupersteine45af542015-06-30 13:36:19 +0000855static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper9f009482011-12-24 07:55:14 +0000856_mm_broadcastq_epi64(__m128i __X)
857{
858 return (__m128i)__builtin_ia32_pbroadcastq128(__X);
859}
860
Michael Kupersteine45af542015-06-30 13:36:19 +0000861static __inline__ __m256i __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000862_mm256_permutevar8x32_epi32(__m256i __a, __m256i __b)
Craig Topper9f009482011-12-24 07:55:14 +0000863{
David Blaikie3302f2b2013-01-16 23:08:36 +0000864 return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b);
Craig Topper9f009482011-12-24 07:55:14 +0000865}
866
867#define _mm256_permute4x64_pd(V, M) __extension__ ({ \
868 __m256d __V = (V); \
Craig Topper8e578552012-04-15 22:18:10 +0000869 (__m256d)__builtin_shufflevector((__v4df)__V, (__v4df) _mm256_setzero_pd(), \
870 (M) & 0x3, ((M) & 0xc) >> 2, \
871 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
Craig Topper9f009482011-12-24 07:55:14 +0000872
Michael Kupersteine45af542015-06-30 13:36:19 +0000873static __inline__ __m256 __DEFAULT_FN_ATTRS
David Blaikie3302f2b2013-01-16 23:08:36 +0000874_mm256_permutevar8x32_ps(__m256 __a, __m256 __b)
Craig Topper9f009482011-12-24 07:55:14 +0000875{
David Blaikie3302f2b2013-01-16 23:08:36 +0000876 return (__m256)__builtin_ia32_permvarsf256((__v8sf)__a, (__v8sf)__b);
Craig Topper9f009482011-12-24 07:55:14 +0000877}
878
879#define _mm256_permute4x64_epi64(V, M) __extension__ ({ \
880 __m256i __V = (V); \
Craig Topper8e578552012-04-15 22:18:10 +0000881 (__m256i)__builtin_shufflevector((__v4di)__V, (__v4di) _mm256_setzero_si256(), \
882 (M) & 0x3, ((M) & 0xc) >> 2, \
883 ((M) & 0x30) >> 4, ((M) & 0xc0) >> 6); })
Craig Topper9f009482011-12-24 07:55:14 +0000884
885#define _mm256_permute2x128_si256(V1, V2, M) __extension__ ({ \
886 __m256i __V1 = (V1); \
887 __m256i __V2 = (V2); \
Craig Topper26e74e52012-04-17 05:16:56 +0000888 (__m256i)__builtin_ia32_permti256(__V1, __V2, (M)); })
Craig Topper175543a2011-12-24 17:20:15 +0000889
Sanjay Patel0a6da5d2015-03-12 21:54:24 +0000890#define _mm256_extracti128_si256(V, M) __extension__ ({ \
891 (__m128i)__builtin_shufflevector( \
892 (__v4di)(V), \
893 (__v4di)(_mm256_setzero_si256()), \
894 (((M) & 1) ? 2 : 0), \
895 (((M) & 1) ? 3 : 1) );})
Craig Topper175543a2011-12-24 17:20:15 +0000896
Sanjay Patel0a6da5d2015-03-12 21:54:24 +0000897#define _mm256_inserti128_si256(V1, V2, M) __extension__ ({ \
898 (__m256i)__builtin_shufflevector( \
899 (__v4di)(V1), \
900 (__v4di)_mm256_castsi128_si256((__m128i)(V2)), \
901 (((M) & 1) ? 0 : 4), \
902 (((M) & 1) ? 1 : 5), \
903 (((M) & 1) ? 4 : 2), \
904 (((M) & 1) ? 5 : 3) );})
Craig Topper175543a2011-12-24 17:20:15 +0000905
Michael Kupersteine45af542015-06-30 13:36:19 +0000906static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000907_mm256_maskload_epi32(int const *__X, __m256i __M)
908{
909 return (__m256i)__builtin_ia32_maskloadd256((const __v8si *)__X, (__v8si)__M);
910}
911
Michael Kupersteine45af542015-06-30 13:36:19 +0000912static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000913_mm256_maskload_epi64(long long const *__X, __m256i __M)
914{
915 return (__m256i)__builtin_ia32_maskloadq256((const __v4di *)__X, __M);
916}
917
Michael Kupersteine45af542015-06-30 13:36:19 +0000918static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000919_mm_maskload_epi32(int const *__X, __m128i __M)
920{
921 return (__m128i)__builtin_ia32_maskloadd((const __v4si *)__X, (__v4si)__M);
922}
923
Michael Kupersteine45af542015-06-30 13:36:19 +0000924static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000925_mm_maskload_epi64(long long const *__X, __m128i __M)
926{
927 return (__m128i)__builtin_ia32_maskloadq((const __v2di *)__X, (__v2di)__M);
928}
929
Michael Kupersteine45af542015-06-30 13:36:19 +0000930static __inline__ void __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000931_mm256_maskstore_epi32(int *__X, __m256i __M, __m256i __Y)
932{
933 __builtin_ia32_maskstored256((__v8si *)__X, (__v8si)__M, (__v8si)__Y);
934}
935
Michael Kupersteine45af542015-06-30 13:36:19 +0000936static __inline__ void __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000937_mm256_maskstore_epi64(long long *__X, __m256i __M, __m256i __Y)
938{
939 __builtin_ia32_maskstoreq256((__v4di *)__X, __M, __Y);
940}
941
Michael Kupersteine45af542015-06-30 13:36:19 +0000942static __inline__ void __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000943_mm_maskstore_epi32(int *__X, __m128i __M, __m128i __Y)
944{
945 __builtin_ia32_maskstored((__v4si *)__X, (__v4si)__M, (__v4si)__Y);
946}
947
Michael Kupersteine45af542015-06-30 13:36:19 +0000948static __inline__ void __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000949_mm_maskstore_epi64(long long *__X, __m128i __M, __m128i __Y)
950{
951 __builtin_ia32_maskstoreq(( __v2di *)__X, __M, __Y);
952}
953
Michael Kupersteine45af542015-06-30 13:36:19 +0000954static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000955_mm256_sllv_epi32(__m256i __X, __m256i __Y)
956{
957 return (__m256i)__builtin_ia32_psllv8si((__v8si)__X, (__v8si)__Y);
958}
959
Michael Kupersteine45af542015-06-30 13:36:19 +0000960static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000961_mm_sllv_epi32(__m128i __X, __m128i __Y)
962{
963 return (__m128i)__builtin_ia32_psllv4si((__v4si)__X, (__v4si)__Y);
964}
965
Michael Kupersteine45af542015-06-30 13:36:19 +0000966static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000967_mm256_sllv_epi64(__m256i __X, __m256i __Y)
968{
969 return (__m256i)__builtin_ia32_psllv4di(__X, __Y);
970}
971
Michael Kupersteine45af542015-06-30 13:36:19 +0000972static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000973_mm_sllv_epi64(__m128i __X, __m128i __Y)
974{
975 return (__m128i)__builtin_ia32_psllv2di(__X, __Y);
976}
977
Michael Kupersteine45af542015-06-30 13:36:19 +0000978static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000979_mm256_srav_epi32(__m256i __X, __m256i __Y)
980{
981 return (__m256i)__builtin_ia32_psrav8si((__v8si)__X, (__v8si)__Y);
982}
983
Michael Kupersteine45af542015-06-30 13:36:19 +0000984static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000985_mm_srav_epi32(__m128i __X, __m128i __Y)
986{
987 return (__m128i)__builtin_ia32_psrav4si((__v4si)__X, (__v4si)__Y);
988}
989
Michael Kupersteine45af542015-06-30 13:36:19 +0000990static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000991_mm256_srlv_epi32(__m256i __X, __m256i __Y)
992{
993 return (__m256i)__builtin_ia32_psrlv8si((__v8si)__X, (__v8si)__Y);
994}
995
Michael Kupersteine45af542015-06-30 13:36:19 +0000996static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +0000997_mm_srlv_epi32(__m128i __X, __m128i __Y)
998{
999 return (__m128i)__builtin_ia32_psrlv4si((__v4si)__X, (__v4si)__Y);
1000}
1001
Michael Kupersteine45af542015-06-30 13:36:19 +00001002static __inline__ __m256i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +00001003_mm256_srlv_epi64(__m256i __X, __m256i __Y)
1004{
1005 return (__m256i)__builtin_ia32_psrlv4di(__X, __Y);
1006}
1007
Michael Kupersteine45af542015-06-30 13:36:19 +00001008static __inline__ __m128i __DEFAULT_FN_ATTRS
Craig Topper175543a2011-12-24 17:20:15 +00001009_mm_srlv_epi64(__m128i __X, __m128i __Y)
1010{
1011 return (__m128i)__builtin_ia32_psrlv2di(__X, __Y);
1012}
Manman Renadd5e9e2012-06-26 19:55:09 +00001013
1014#define _mm_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
1015 __m128d __a = (a); \
1016 double const *__m = (m); \
1017 __m128i __i = (i); \
1018 __m128d __mask = (mask); \
1019 (__m128d)__builtin_ia32_gatherd_pd((__v2df)__a, (const __v2df *)__m, \
1020 (__v4si)__i, (__v2df)__mask, (s)); })
1021
1022#define _mm256_mask_i32gather_pd(a, m, i, mask, s) __extension__ ({ \
1023 __m256d __a = (a); \
1024 double const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001025 __m128i __i = (i); \
Manman Renadd5e9e2012-06-26 19:55:09 +00001026 __m256d __mask = (mask); \
1027 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)__a, (const __v4df *)__m, \
Manman Ren86c32502012-06-29 00:54:35 +00001028 (__v4si)__i, (__v4df)__mask, (s)); })
Manman Renadd5e9e2012-06-26 19:55:09 +00001029
1030#define _mm_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
1031 __m128d __a = (a); \
1032 double const *__m = (m); \
1033 __m128i __i = (i); \
1034 __m128d __mask = (mask); \
1035 (__m128d)__builtin_ia32_gatherq_pd((__v2df)__a, (const __v2df *)__m, \
1036 (__v2di)__i, (__v2df)__mask, (s)); })
1037
1038#define _mm256_mask_i64gather_pd(a, m, i, mask, s) __extension__ ({ \
1039 __m256d __a = (a); \
1040 double const *__m = (m); \
1041 __m256i __i = (i); \
1042 __m256d __mask = (mask); \
1043 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)__a, (const __v4df *)__m, \
1044 (__v4di)__i, (__v4df)__mask, (s)); })
1045
1046#define _mm_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1047 __m128 __a = (a); \
1048 float const *__m = (m); \
1049 __m128i __i = (i); \
1050 __m128 __mask = (mask); \
1051 (__m128)__builtin_ia32_gatherd_ps((__v4sf)__a, (const __v4sf *)__m, \
1052 (__v4si)__i, (__v4sf)__mask, (s)); })
1053
1054#define _mm256_mask_i32gather_ps(a, m, i, mask, s) __extension__ ({ \
1055 __m256 __a = (a); \
1056 float const *__m = (m); \
1057 __m256i __i = (i); \
1058 __m256 __mask = (mask); \
1059 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)__a, (const __v8sf *)__m, \
1060 (__v8si)__i, (__v8sf)__mask, (s)); })
1061
1062#define _mm_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
1063 __m128 __a = (a); \
1064 float const *__m = (m); \
1065 __m128i __i = (i); \
1066 __m128 __mask = (mask); \
1067 (__m128)__builtin_ia32_gatherq_ps((__v4sf)__a, (const __v4sf *)__m, \
1068 (__v2di)__i, (__v4sf)__mask, (s)); })
1069
1070#define _mm256_mask_i64gather_ps(a, m, i, mask, s) __extension__ ({ \
Manman Ren86c32502012-06-29 00:54:35 +00001071 __m128 __a = (a); \
Manman Renadd5e9e2012-06-26 19:55:09 +00001072 float const *__m = (m); \
1073 __m256i __i = (i); \
Manman Ren86c32502012-06-29 00:54:35 +00001074 __m128 __mask = (mask); \
1075 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)__a, (const __v4sf *)__m, \
1076 (__v4di)__i, (__v4sf)__mask, (s)); })
1077
1078#define _mm_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1079 __m128i __a = (a); \
1080 int const *__m = (m); \
1081 __m128i __i = (i); \
1082 __m128i __mask = (mask); \
1083 (__m128i)__builtin_ia32_gatherd_d((__v4si)__a, (const __v4si *)__m, \
1084 (__v4si)__i, (__v4si)__mask, (s)); })
1085
1086#define _mm256_mask_i32gather_epi32(a, m, i, mask, s) __extension__ ({ \
1087 __m256i __a = (a); \
1088 int const *__m = (m); \
1089 __m256i __i = (i); \
1090 __m256i __mask = (mask); \
1091 (__m256i)__builtin_ia32_gatherd_d256((__v8si)__a, (const __v8si *)__m, \
1092 (__v8si)__i, (__v8si)__mask, (s)); })
1093
1094#define _mm_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1095 __m128i __a = (a); \
1096 int const *__m = (m); \
1097 __m128i __i = (i); \
1098 __m128i __mask = (mask); \
1099 (__m128i)__builtin_ia32_gatherq_d((__v4si)__a, (const __v4si *)__m, \
1100 (__v2di)__i, (__v4si)__mask, (s)); })
1101
1102#define _mm256_mask_i64gather_epi32(a, m, i, mask, s) __extension__ ({ \
1103 __m128i __a = (a); \
1104 int const *__m = (m); \
1105 __m256i __i = (i); \
1106 __m128i __mask = (mask); \
1107 (__m128i)__builtin_ia32_gatherq_d256((__v4si)__a, (const __v4si *)__m, \
1108 (__v4di)__i, (__v4si)__mask, (s)); })
1109
1110#define _mm_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1111 __m128i __a = (a); \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001112 long long const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001113 __m128i __i = (i); \
1114 __m128i __mask = (mask); \
1115 (__m128i)__builtin_ia32_gatherd_q((__v2di)__a, (const __v2di *)__m, \
1116 (__v4si)__i, (__v2di)__mask, (s)); })
1117
1118#define _mm256_mask_i32gather_epi64(a, m, i, mask, s) __extension__ ({ \
1119 __m256i __a = (a); \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001120 long long const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001121 __m128i __i = (i); \
1122 __m256i __mask = (mask); \
1123 (__m256i)__builtin_ia32_gatherd_q256((__v4di)__a, (const __v4di *)__m, \
1124 (__v4si)__i, (__v4di)__mask, (s)); })
1125
1126#define _mm_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1127 __m128i __a = (a); \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001128 long long const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001129 __m128i __i = (i); \
1130 __m128i __mask = (mask); \
1131 (__m128i)__builtin_ia32_gatherq_q((__v2di)__a, (const __v2di *)__m, \
1132 (__v2di)__i, (__v2di)__mask, (s)); })
1133
1134#define _mm256_mask_i64gather_epi64(a, m, i, mask, s) __extension__ ({ \
1135 __m256i __a = (a); \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001136 long long const *__m = (m); \
Manman Ren86c32502012-06-29 00:54:35 +00001137 __m256i __i = (i); \
1138 __m256i __mask = (mask); \
1139 (__m256i)__builtin_ia32_gatherq_q256((__v4di)__a, (const __v4di *)__m, \
1140 (__v4di)__i, (__v4di)__mask, (s)); })
Manman Renf865ba02012-06-29 05:19:13 +00001141
1142#define _mm_i32gather_pd(m, i, s) __extension__ ({ \
1143 double const *__m = (m); \
1144 __m128i __i = (i); \
1145 (__m128d)__builtin_ia32_gatherd_pd((__v2df)_mm_setzero_pd(), \
1146 (const __v2df *)__m, (__v4si)__i, \
1147 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1148
1149#define _mm256_i32gather_pd(m, i, s) __extension__ ({ \
1150 double const *__m = (m); \
1151 __m128i __i = (i); \
1152 (__m256d)__builtin_ia32_gatherd_pd256((__v4df)_mm256_setzero_pd(), \
1153 (const __v4df *)__m, (__v4si)__i, \
1154 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1155
1156#define _mm_i64gather_pd(m, i, s) __extension__ ({ \
1157 double const *__m = (m); \
1158 __m128i __i = (i); \
1159 (__m128d)__builtin_ia32_gatherq_pd((__v2df)_mm_setzero_pd(), \
1160 (const __v2df *)__m, (__v2di)__i, \
1161 (__v2df)_mm_set1_pd((double)(long long int)-1), (s)); })
1162
1163#define _mm256_i64gather_pd(m, i, s) __extension__ ({ \
1164 double const *__m = (m); \
1165 __m256i __i = (i); \
1166 (__m256d)__builtin_ia32_gatherq_pd256((__v4df)_mm256_setzero_pd(), \
1167 (const __v4df *)__m, (__v4di)__i, \
1168 (__v4df)_mm256_set1_pd((double)(long long int)-1), (s)); })
1169
1170#define _mm_i32gather_ps(m, i, s) __extension__ ({ \
1171 float const *__m = (m); \
1172 __m128i __i = (i); \
1173 (__m128)__builtin_ia32_gatherd_ps((__v4sf)_mm_setzero_ps(), \
1174 (const __v4sf *)__m, (__v4si)__i, \
1175 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1176
1177#define _mm256_i32gather_ps(m, i, s) __extension__ ({ \
1178 float const *__m = (m); \
1179 __m256i __i = (i); \
1180 (__m256)__builtin_ia32_gatherd_ps256((__v8sf)_mm256_setzero_ps(), \
1181 (const __v8sf *)__m, (__v8si)__i, \
1182 (__v8sf)_mm256_set1_ps((float)(int)-1), (s)); })
1183
1184#define _mm_i64gather_ps(m, i, s) __extension__ ({ \
1185 float const *__m = (m); \
1186 __m128i __i = (i); \
1187 (__m128)__builtin_ia32_gatherq_ps((__v4sf)_mm_setzero_ps(), \
1188 (const __v4sf *)__m, (__v2di)__i, \
1189 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1190
1191#define _mm256_i64gather_ps(m, i, s) __extension__ ({ \
1192 float const *__m = (m); \
1193 __m256i __i = (i); \
1194 (__m128)__builtin_ia32_gatherq_ps256((__v4sf)_mm_setzero_ps(), \
1195 (const __v4sf *)__m, (__v4di)__i, \
1196 (__v4sf)_mm_set1_ps((float)(int)-1), (s)); })
1197
1198#define _mm_i32gather_epi32(m, i, s) __extension__ ({ \
1199 int const *__m = (m); \
1200 __m128i __i = (i); \
1201 (__m128i)__builtin_ia32_gatherd_d((__v4si)_mm_setzero_si128(), \
1202 (const __v4si *)__m, (__v4si)__i, \
1203 (__v4si)_mm_set1_epi32(-1), (s)); })
1204
1205#define _mm256_i32gather_epi32(m, i, s) __extension__ ({ \
1206 int const *__m = (m); \
1207 __m256i __i = (i); \
1208 (__m256i)__builtin_ia32_gatherd_d256((__v8si)_mm256_setzero_si256(), \
1209 (const __v8si *)__m, (__v8si)__i, \
1210 (__v8si)_mm256_set1_epi32(-1), (s)); })
1211
1212#define _mm_i64gather_epi32(m, i, s) __extension__ ({ \
1213 int const *__m = (m); \
1214 __m128i __i = (i); \
1215 (__m128i)__builtin_ia32_gatherq_d((__v4si)_mm_setzero_si128(), \
1216 (const __v4si *)__m, (__v2di)__i, \
1217 (__v4si)_mm_set1_epi32(-1), (s)); })
1218
1219#define _mm256_i64gather_epi32(m, i, s) __extension__ ({ \
1220 int const *__m = (m); \
1221 __m256i __i = (i); \
1222 (__m128i)__builtin_ia32_gatherq_d256((__v4si)_mm_setzero_si128(), \
1223 (const __v4si *)__m, (__v4di)__i, \
1224 (__v4si)_mm_set1_epi32(-1), (s)); })
1225
1226#define _mm_i32gather_epi64(m, i, s) __extension__ ({ \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001227 long long const *__m = (m); \
Manman Renf865ba02012-06-29 05:19:13 +00001228 __m128i __i = (i); \
1229 (__m128i)__builtin_ia32_gatherd_q((__v2di)_mm_setzero_si128(), \
1230 (const __v2di *)__m, (__v4si)__i, \
1231 (__v2di)_mm_set1_epi64x(-1), (s)); })
1232
1233#define _mm256_i32gather_epi64(m, i, s) __extension__ ({ \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001234 long long const *__m = (m); \
Manman Renf865ba02012-06-29 05:19:13 +00001235 __m128i __i = (i); \
1236 (__m256i)__builtin_ia32_gatherd_q256((__v4di)_mm256_setzero_si256(), \
1237 (const __v4di *)__m, (__v4si)__i, \
1238 (__v4di)_mm256_set1_epi64x(-1), (s)); })
1239
1240#define _mm_i64gather_epi64(m, i, s) __extension__ ({ \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001241 long long const *__m = (m); \
Manman Renf865ba02012-06-29 05:19:13 +00001242 __m128i __i = (i); \
1243 (__m128i)__builtin_ia32_gatherq_q((__v2di)_mm_setzero_si128(), \
1244 (const __v2di *)__m, (__v2di)__i, \
1245 (__v2di)_mm_set1_epi64x(-1), (s)); })
1246
1247#define _mm256_i64gather_epi64(m, i, s) __extension__ ({ \
Eli Friedman3cd55f42013-09-23 23:52:04 +00001248 long long const *__m = (m); \
Manman Renf865ba02012-06-29 05:19:13 +00001249 __m256i __i = (i); \
1250 (__m256i)__builtin_ia32_gatherq_q256((__v4di)_mm256_setzero_si256(), \
1251 (const __v4di *)__m, (__v4di)__i, \
1252 (__v4di)_mm256_set1_epi64x(-1), (s)); })
Richard Smith49e56442013-07-14 05:41:45 +00001253
Michael Kupersteine45af542015-06-30 13:36:19 +00001254#undef __DEFAULT_FN_ATTRS
Eric Christopher4d1851682015-06-17 07:09:20 +00001255
Richard Smith49e56442013-07-14 05:41:45 +00001256#endif /* __AVX2INTRIN_H */