FD.io VPP  v20.01-48-g3e0dafb74
Vector Packet Processing
vector_avx2.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018 Cisco and/or its affiliates.
3  * Licensed under the Apache License, Version 2.0 (the "License");
4  * you may not use this file except in compliance with the License.
5  * You may obtain a copy of the License at:
6  *
7  * http://www.apache.org/licenses/LICENSE-2.0
8  *
9  * Unless required by applicable law or agreed to in writing, software
10  * distributed under the License is distributed on an "AS IS" BASIS,
11  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12  * See the License for the specific language governing permissions and
13  * limitations under the License.
14  */
15 
16 #ifndef included_vector_avx2_h
17 #define included_vector_avx2_h
18 
19 #include <vppinfra/clib.h>
20 #include <x86intrin.h>
21 
22 /* *INDENT-OFF* */
23 #define foreach_avx2_vec256i \
24  _(i,8,32,epi8) _(i,16,16,epi16) _(i,32,8,epi32) _(i,64,4,epi64)
25 #define foreach_avx2_vec256u \
26  _(u,8,32,epi8) _(u,16,16,epi16) _(u,32,8,epi32) _(u,64,4,epi64)
27 #define foreach_avx2_vec256f \
28  _(f,32,8,ps) _(f,64,4,pd)
29 
30 #define _mm256_set1_epi64 _mm256_set1_epi64x
31 
32 /* splat, load_unaligned, store_unaligned, is_all_zero, is_equal,
33  is_all_equal */
34 #define _(t, s, c, i) \
35 static_always_inline t##s##x##c \
36 t##s##x##c##_splat (t##s x) \
37 { return (t##s##x##c) _mm256_set1_##i (x); } \
38 \
39 static_always_inline t##s##x##c \
40 t##s##x##c##_load_unaligned (void *p) \
41 { return (t##s##x##c) _mm256_loadu_si256 (p); } \
42 \
43 static_always_inline void \
44 t##s##x##c##_store_unaligned (t##s##x##c v, void *p) \
45 { _mm256_storeu_si256 ((__m256i *) p, (__m256i) v); } \
46 \
47 static_always_inline int \
48 t##s##x##c##_is_all_zero (t##s##x##c x) \
49 { return _mm256_testz_si256 ((__m256i) x, (__m256i) x); } \
50 \
51 static_always_inline int \
52 t##s##x##c##_is_equal (t##s##x##c a, t##s##x##c b) \
53 { return t##s##x##c##_is_all_zero (a ^ b); } \
54 \
55 static_always_inline int \
56 t##s##x##c##_is_all_equal (t##s##x##c v, t##s x) \
57 { return t##s##x##c##_is_equal (v, t##s##x##c##_splat (x)); } \
58 \
59 static_always_inline t##s##x##c \
60 t##s##x##c##_interleave_lo (t##s##x##c a, t##s##x##c b) \
61 { return (t##s##x##c) _mm256_unpacklo_##i ((__m256i) a, (__m256i) b); } \
62 \
63 static_always_inline t##s##x##c \
64 t##s##x##c##_interleave_hi (t##s##x##c a, t##s##x##c b) \
65 { return (t##s##x##c) _mm256_unpackhi_##i ((__m256i) a, (__m256i) b); } \
66 
67 
69 #undef _
70 /* *INDENT-ON* */
71 
72 always_inline u32x8
73 u32x8_permute (u32x8 v, u32x8 idx)
74 {
75  return (u32x8) _mm256_permutevar8x32_epi32 ((__m256i) v, (__m256i) idx);
76 }
77 
78 /* _extract_lo, _extract_hi */
79 /* *INDENT-OFF* */
80 #define _(t1,t2) \
81 always_inline t1 \
82 t2##_extract_lo (t2 v) \
83 { return (t1) _mm256_extracti128_si256 ((__m256i) v, 0); } \
84 \
85 always_inline t1 \
86 t2##_extract_hi (t2 v) \
87 { return (t1) _mm256_extracti128_si256 ((__m256i) v, 1); } \
88 \
89 always_inline t2 \
90 t2##_insert_lo (t2 v1, t1 v2) \
91 { return (t2) _mm256_inserti128_si256 ((__m256i) v1, (__m128i) v2, 0); }\
92 \
93 always_inline t2 \
94 t2##_insert_hi (t2 v1, t1 v2) \
95 { return (t2) _mm256_inserti128_si256 ((__m256i) v1, (__m128i) v2, 1); }\
96 
97 _(u8x16, u8x32)
98 _(u16x8, u16x16)
99 _(u32x4, u32x8)
100 _(u64x2, u64x4)
101 #undef _
102 /* *INDENT-ON* */
103 
104 
105 
106 
108 u8x32_msb_mask (u8x32 v)
109 {
110  return _mm256_movemask_epi8 ((__m256i) v);
111 }
112 
113 /* _extend_to_ */
114 /* *INDENT-OFF* */
115 #define _(f,t,i) \
116 static_always_inline t \
117 f##_extend_to_##t (f x) \
118 { return (t) _mm256_cvt##i ((__m128i) x); }
119 
120 _(u16x8, u32x8, epu16_epi32)
121 _(u16x8, u64x4, epu16_epi64)
122 _(u32x4, u64x4, epu32_epi64)
123 _(u8x16, u16x16, epu8_epi64)
124 _(u8x16, u32x8, epu8_epi32)
125 _(u8x16, u64x4, epu8_epi64)
126 _(i16x8, i32x8, epi16_epi32)
127 _(i16x8, i64x4, epi16_epi64)
128 _(i32x4, i64x4, epi32_epi64)
129 _(i8x16, i16x16, epi8_epi64)
130 _(i8x16, i32x8, epi8_epi32)
131 _(i8x16, i64x4, epi8_epi64)
132 #undef _
133 /* *INDENT-ON* */
134 
137 {
138  u8x32 swap = {
139  3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12,
140  3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12
141  };
142  return (u32x8) _mm256_shuffle_epi8 ((__m256i) v, (__m256i) swap);
143 }
144 
147 {
148  u8x32 swap = {
149  1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
150  1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14
151  };
152  return (u16x16) _mm256_shuffle_epi8 ((__m256i) v, (__m256i) swap);
153 }
154 
156 u32x8_hadd (u32x8 v1, u32x8 v2)
157 {
158  return (u32x8) _mm256_hadd_epi32 ((__m256i) v1, (__m256i) v2);
159 }
160 
163 {
164  const u16x16 masks[17] = {
165  {0},
166  {-1},
167  {-1, -1},
168  {-1, -1, -1},
169  {-1, -1, -1, -1},
170  {-1, -1, -1, -1, -1},
171  {-1, -1, -1, -1, -1, -1},
172  {-1, -1, -1, -1, -1, -1, -1},
173  {-1, -1, -1, -1, -1, -1, -1, -1},
174  {-1, -1, -1, -1, -1, -1, -1, -1, -1},
175  {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1},
176  {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1},
177  {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1},
178  {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1},
179  {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1},
180  {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1},
181  {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1},
182  };
183 
184  ASSERT (n_last < 17);
185 
186  return v & masks[16 - n_last];
187 }
188 
191 {
192  return (f32x8) _mm256_cvtepi32_ps ((__m256i) v);
193 }
194 
197 {
198  return (u32x8) _mm256_cvttps_epi32 ((__m256) v);
199 }
200 
201 #define u32x8_blend(a,b,m) \
202  (u32x8) _mm256_blend_epi32 ((__m256i) a, (__m256i) b, m)
203 
204 #define u16x16_blend(v1, v2, mask) \
205  (u16x16) _mm256_blend_epi16 ((__m256i) (v1), (__m256i) (v2), mask)
206 
208 u64x4_gather (void *p0, void *p1, void *p2, void *p3)
209 {
210  u64x4 r = {
211  *(u64 *) p0, *(u64 *) p1, *(u64 *) p2, *(u64 *) p3
212  };
213  return r;
214 }
215 
217 u32x8_gather (void *p0, void *p1, void *p2, void *p3, void *p4, void *p5,
218  void *p6, void *p7)
219 {
220  u32x8 r = {
221  *(u32 *) p0, *(u32 *) p1, *(u32 *) p2, *(u32 *) p3,
222  *(u32 *) p4, *(u32 *) p5, *(u32 *) p6, *(u32 *) p7,
223  };
224  return r;
225 }
226 
227 
229 u64x4_scatter (u64x4 r, void *p0, void *p1, void *p2, void *p3)
230 {
231  *(u64 *) p0 = r[0];
232  *(u64 *) p1 = r[1];
233  *(u64 *) p2 = r[2];
234  *(u64 *) p3 = r[3];
235 }
236 
238 u32x8_scatter (u32x8 r, void *p0, void *p1, void *p2, void *p3, void *p4,
239  void *p5, void *p6, void *p7)
240 {
241  *(u32 *) p0 = r[0];
242  *(u32 *) p1 = r[1];
243  *(u32 *) p2 = r[2];
244  *(u32 *) p3 = r[3];
245  *(u32 *) p4 = r[4];
246  *(u32 *) p5 = r[5];
247  *(u32 *) p6 = r[6];
248  *(u32 *) p7 = r[7];
249 }
250 
252 u64x4_scatter_one (u64x4 r, int index, void *p)
253 {
254  *(u64 *) p = r[index];
255 }
256 
258 u32x8_scatter_one (u32x8 r, int index, void *p)
259 {
260  *(u32 *) p = r[index];
261 }
262 
264 u8x32_is_greater (u8x32 v1, u8x32 v2)
265 {
266  return (u8x32) _mm256_cmpgt_epi8 ((__m256i) v1, (__m256i) v2);
267 }
268 
270 u8x32_blend (u8x32 v1, u8x32 v2, u8x32 mask)
271 {
272  return (u8x32) _mm256_blendv_epi8 ((__m256i) v1, (__m256i) v2,
273  (__m256i) mask);
274 }
275 
276 #define u32x8_permute_lanes(a, b, m) \
277  (u32x8) _mm256_permute2x128_si256 ((__m256i) a, (__m256i) b, m)
278 #define u64x4_permute_lanes(a, b, m) \
279  (u64x4) _mm256_permute2x128_si256 ((__m256i) a, (__m256i) b, m)
280 
282 u32x8_min (u32x8 a, u32x8 b)
283 {
284  return (u32x8) _mm256_min_epu32 ((__m256i) a, (__m256i) b);
285 }
286 
289 {
290  return u32x4_min_scalar (u32x4_min (u32x8_extract_lo (v),
291  u32x8_extract_hi (v)));
292 }
293 
295 u32x8_transpose (u32x8 a[8])
296 {
297  u64x4 r[8], x, y;
298 
299  r[0] = (u64x4) u32x8_interleave_lo (a[0], a[1]);
300  r[1] = (u64x4) u32x8_interleave_hi (a[0], a[1]);
301  r[2] = (u64x4) u32x8_interleave_lo (a[2], a[3]);
302  r[3] = (u64x4) u32x8_interleave_hi (a[2], a[3]);
303  r[4] = (u64x4) u32x8_interleave_lo (a[4], a[5]);
304  r[5] = (u64x4) u32x8_interleave_hi (a[4], a[5]);
305  r[6] = (u64x4) u32x8_interleave_lo (a[6], a[7]);
306  r[7] = (u64x4) u32x8_interleave_hi (a[6], a[7]);
307 
308  x = u64x4_interleave_lo (r[0], r[2]);
309  y = u64x4_interleave_lo (r[4], r[6]);
310  a[0] = u32x8_permute_lanes (x, y, 0x20);
311  a[4] = u32x8_permute_lanes (x, y, 0x31);
312 
313  x = u64x4_interleave_hi (r[0], r[2]);
314  y = u64x4_interleave_hi (r[4], r[6]);
315  a[1] = u32x8_permute_lanes (x, y, 0x20);
316  a[5] = u32x8_permute_lanes (x, y, 0x31);
317 
318  x = u64x4_interleave_lo (r[1], r[3]);
319  y = u64x4_interleave_lo (r[5], r[7]);
320  a[2] = u32x8_permute_lanes (x, y, 0x20);
321  a[6] = u32x8_permute_lanes (x, y, 0x31);
322 
323  x = u64x4_interleave_hi (r[1], r[3]);
324  y = u64x4_interleave_hi (r[5], r[7]);
325  a[3] = u32x8_permute_lanes (x, y, 0x20);
326  a[7] = u32x8_permute_lanes (x, y, 0x31);
327 }
328 
331 {
332  u64x4 r[4];
333 
334  r[0] = u64x4_interleave_lo (a[0], a[1]);
335  r[1] = u64x4_interleave_hi (a[0], a[1]);
336  r[2] = u64x4_interleave_lo (a[2], a[3]);
337  r[3] = u64x4_interleave_hi (a[2], a[3]);
338 
339  a[0] = u64x4_permute_lanes (r[0], r[2], 0x20);
340  a[1] = u64x4_permute_lanes (r[1], r[3], 0x20);
341  a[2] = u64x4_permute_lanes (r[0], r[2], 0x31);
342  a[3] = u64x4_permute_lanes (r[1], r[3], 0x31);
343 }
344 
345 #endif /* included_vector_avx2_h */
346 
347 /*
348  * fd.io coding-style-patch-verification: ON
349  *
350  * Local Variables:
351  * eval: (c-set-style "gnu")
352  * End:
353  */
static_always_inline u32 u32x4_min_scalar(u32x4 v)
Definition: vector_sse42.h:579
a
Definition: bitmap.h:538
unsigned long u64
Definition: types.h:89
u16x16 u64x4 static_always_inline u32 u8x32_msb_mask(u8x32 v)
Definition: vector_avx2.h:108
foreach_avx2_vec256i static foreach_avx2_vec256u u32x8 u32x8_permute(u32x8 v, u32x8 idx)
Definition: vector_avx2.h:73
static_always_inline u32 u32x8_min_scalar(u32x8 v)
Definition: vector_avx2.h:288
static_always_inline u16x16 u16x16_mask_last(u16x16 v, u8 n_last)
Definition: vector_avx2.h:162
unsigned char u8
Definition: types.h:56
static_always_inline void u64x4_scatter(u64x4 r, void *p0, void *p1, void *p2, void *p3)
Definition: vector_avx2.h:229
epu16_epi64 epu8_epi64 epu8_epi64 epi16_epi64 i16x16
Definition: vector_avx2.h:129
#define static_always_inline
Definition: clib.h:99
static_always_inline void u64x4_scatter_one(u64x4 r, int index, void *p)
Definition: vector_avx2.h:252
unsigned int u32
Definition: types.h:88
#define foreach_avx2_vec256i
Definition: vector_avx2.h:23
i32x4
epu8_epi32 epu16_epi32 u64x2
Definition: vector_sse42.h:674
epu16_epi64 epu8_epi64 epu8_epi64 epi16_epi64 epi8_epi64 epi8_epi64 static_always_inline u32x8 u32x8_byte_swap(u32x8 v)
Definition: vector_avx2.h:136
static_always_inline void u32x8_scatter(u32x8 r, void *p0, void *p1, void *p2, void *p3, void *p4, void *p5, void *p6, void *p7)
Definition: vector_avx2.h:238
static_always_inline u64x4 u64x4_gather(void *p0, void *p1, void *p2, void *p3)
Definition: vector_avx2.h:208
#define always_inline
Definition: ipsec.h:28
static_always_inline void u32x8_transpose(u32x8 a[8])
Definition: vector_avx2.h:295
static_always_inline u16x16 u16x16_byte_swap(u16x16 v)
Definition: vector_avx2.h:146
static_always_inline void u32x8_scatter_one(u32x8 r, int index, void *p)
Definition: vector_avx2.h:258
#define ASSERT(truth)
static_always_inline u32x8 u32x8_hadd(u32x8 v1, u32x8 v2)
Definition: vector_avx2.h:156
vmrglw i16x8
static_always_inline u32x8 u32x8_min(u32x8 a, u32x8 b)
Definition: vector_avx2.h:282
static_always_inline u8x32 u8x32_is_greater(u8x32 v1, u8x32 v2)
Definition: vector_avx2.h:264
#define u64x4_permute_lanes(a, b, m)
Definition: vector_avx2.h:278
#define foreach_avx2_vec256u
Definition: vector_avx2.h:25
static_always_inline u32x8 u32x8_from_f32x8(f32x8 v)
Definition: vector_avx2.h:196
epu16_epi64 epu8_epi64 epu8_epi64 i64x4
Definition: vector_avx2.h:127
u64x4
Definition: vector_avx2.h:121
#define u32x8_permute_lanes(a, b, m)
Definition: vector_avx2.h:276
unsigned long long u32x4
Definition: ixge.c:28
static_always_inline void u64x4_transpose(u64x4 a[8])
Definition: vector_avx2.h:330
epu16_epi64 u16x16
Definition: vector_avx2.h:123
static_always_inline f32x8 f32x8_from_u32x8(u32x8 v)
Definition: vector_avx2.h:190
static_always_inline u8x32 u8x32_blend(u8x32 v1, u8x32 v2, u8x32 mask)
Definition: vector_avx2.h:270
static_always_inline u32x8 u32x8_gather(void *p0, void *p1, void *p2, void *p3, void *p4, void *p5, void *p6, void *p7)
Definition: vector_avx2.h:217