FD.io VPP  v17.01.1-3-gc6833f8
Vector Packet Processing
memcpy_avx.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016 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  * BSD LICENSE
17  *
18  * Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
19  * All rights reserved.
20  *
21  * Redistribution and use in source and binary forms, with or without
22  * modification, are permitted provided that the following conditions
23  * are met:
24  *
25  * * Redistributions of source code must retain the above copyright
26  * notice, this list of conditions and the following disclaimer.
27  * * Redistributions in binary form must reproduce the above copyright
28  * notice, this list of conditions and the following disclaimer in
29  * the documentation and/or other materials provided with the
30  * distribution.
31  * * Neither the name of Intel Corporation nor the names of its
32  * contributors may be used to endorse or promote products derived
33  * from this software without specific prior written permission.
34  *
35  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
36  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
37  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
38  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
39  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
40  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
41  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
42  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
43  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
44  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
45  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
46  */
47 
48 #ifndef included_clib_memcpy_avx_h
49 #define included_clib_memcpy_avx_h
50 
51 #include <stdint.h>
52 #include <x86intrin.h>
53 
54 typedef u8 u8x16u __attribute__ ((vector_size (16), aligned (1)));
55 typedef u8 u8x32u __attribute__ ((vector_size (32), aligned (1)));
56 
57 static inline void
58 clib_mov16 (u8 * dst, const u8 * src)
59 {
60  *(u8x16u *) dst = *(u8x16u *) src;
61 }
62 
63 static inline void
64 clib_mov32 (u8 * dst, const u8 * src)
65 {
66  *(u8x32u *) dst = *(u8x32u *) src;
67 }
68 
69 static inline void
70 clib_mov64 (u8 * dst, const u8 * src)
71 {
72  clib_mov32 ((u8 *) dst + 0 * 32, (const u8 *) src + 0 * 32);
73  clib_mov32 ((u8 *) dst + 1 * 32, (const u8 *) src + 1 * 32);
74 }
75 
76 static inline void
77 clib_mov128 (u8 * dst, const u8 * src)
78 {
79  clib_mov64 ((u8 *) dst + 0 * 64, (const u8 *) src + 0 * 64);
80  clib_mov64 ((u8 *) dst + 1 * 64, (const u8 *) src + 1 * 64);
81 }
82 
83 static inline void
84 clib_mov256 (u8 * dst, const u8 * src)
85 {
86  clib_mov128 ((u8 *) dst + 0 * 128, (const u8 *) src + 0 * 128);
87  clib_mov128 ((u8 *) dst + 1 * 128, (const u8 *) src + 1 * 128);
88 }
89 
90 static inline void
91 clib_mov64blocks (u8 * dst, const u8 * src, size_t n)
92 {
93  __m256i ymm0, ymm1;
94 
95  while (n >= 64)
96  {
97  ymm0 =
98  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 0 * 32));
99  n -= 64;
100  ymm1 =
101  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 1 * 32));
102  src = (const u8 *) src + 64;
103  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 0 * 32), ymm0);
104  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 1 * 32), ymm1);
105  dst = (u8 *) dst + 64;
106  }
107 }
108 
109 static inline void
110 clib_mov256blocks (u8 * dst, const u8 * src, size_t n)
111 {
112  __m256i ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7;
113 
114  while (n >= 256)
115  {
116  ymm0 =
117  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 0 * 32));
118  n -= 256;
119  ymm1 =
120  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 1 * 32));
121  ymm2 =
122  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 2 * 32));
123  ymm3 =
124  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 3 * 32));
125  ymm4 =
126  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 4 * 32));
127  ymm5 =
128  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 5 * 32));
129  ymm6 =
130  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 6 * 32));
131  ymm7 =
132  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 7 * 32));
133  src = (const u8 *) src + 256;
134  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 0 * 32), ymm0);
135  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 1 * 32), ymm1);
136  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 2 * 32), ymm2);
137  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 3 * 32), ymm3);
138  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 4 * 32), ymm4);
139  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 5 * 32), ymm5);
140  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 6 * 32), ymm6);
141  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 7 * 32), ymm7);
142  dst = (u8 *) dst + 256;
143  }
144 }
145 
146 static inline void *
147 clib_memcpy (void *dst, const void *src, size_t n)
148 {
149  uword dstu = (uword) dst;
150  uword srcu = (uword) src;
151  void *ret = dst;
152  size_t dstofss;
153  size_t bits;
154 
155  /**
156  * Copy less than 16 bytes
157  */
158  if (n < 16)
159  {
160  if (n & 0x01)
161  {
162  *(u8 *) dstu = *(const u8 *) srcu;
163  srcu = (uword) ((const u8 *) srcu + 1);
164  dstu = (uword) ((u8 *) dstu + 1);
165  }
166  if (n & 0x02)
167  {
168  *(uint16_t *) dstu = *(const uint16_t *) srcu;
169  srcu = (uword) ((const uint16_t *) srcu + 1);
170  dstu = (uword) ((uint16_t *) dstu + 1);
171  }
172  if (n & 0x04)
173  {
174  *(uint32_t *) dstu = *(const uint32_t *) srcu;
175  srcu = (uword) ((const uint32_t *) srcu + 1);
176  dstu = (uword) ((uint32_t *) dstu + 1);
177  }
178  if (n & 0x08)
179  {
180  *(uint64_t *) dstu = *(const uint64_t *) srcu;
181  }
182  return ret;
183  }
184 
185  /**
186  * Fast way when copy size doesn't exceed 512 bytes
187  */
188  if (n <= 32)
189  {
190  clib_mov16 ((u8 *) dst, (const u8 *) src);
191  clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
192  return ret;
193  }
194  if (n <= 64)
195  {
196  clib_mov32 ((u8 *) dst, (const u8 *) src);
197  clib_mov32 ((u8 *) dst - 32 + n, (const u8 *) src - 32 + n);
198  return ret;
199  }
200  if (n <= 512)
201  {
202  if (n >= 256)
203  {
204  n -= 256;
205  clib_mov256 ((u8 *) dst, (const u8 *) src);
206  src = (const u8 *) src + 256;
207  dst = (u8 *) dst + 256;
208  }
209  if (n >= 128)
210  {
211  n -= 128;
212  clib_mov128 ((u8 *) dst, (const u8 *) src);
213  src = (const u8 *) src + 128;
214  dst = (u8 *) dst + 128;
215  }
216  if (n >= 64)
217  {
218  n -= 64;
219  clib_mov64 ((u8 *) dst, (const u8 *) src);
220  src = (const u8 *) src + 64;
221  dst = (u8 *) dst + 64;
222  }
223  COPY_BLOCK_64_BACK31:
224  if (n > 32)
225  {
226  clib_mov32 ((u8 *) dst, (const u8 *) src);
227  clib_mov32 ((u8 *) dst - 32 + n, (const u8 *) src - 32 + n);
228  return ret;
229  }
230  if (n > 0)
231  {
232  clib_mov32 ((u8 *) dst - 32 + n, (const u8 *) src - 32 + n);
233  }
234  return ret;
235  }
236 
237  /**
238  * Make store aligned when copy size exceeds 512 bytes
239  */
240  dstofss = (uword) dst & 0x1F;
241  if (dstofss > 0)
242  {
243  dstofss = 32 - dstofss;
244  n -= dstofss;
245  clib_mov32 ((u8 *) dst, (const u8 *) src);
246  src = (const u8 *) src + dstofss;
247  dst = (u8 *) dst + dstofss;
248  }
249 
250  /**
251  * Copy 256-byte blocks.
252  * Use copy block function for better instruction order control,
253  * which is important when load is unaligned.
254  */
255  clib_mov256blocks ((u8 *) dst, (const u8 *) src, n);
256  bits = n;
257  n = n & 255;
258  bits -= n;
259  src = (const u8 *) src + bits;
260  dst = (u8 *) dst + bits;
261 
262  /**
263  * Copy 64-byte blocks.
264  * Use copy block function for better instruction order control,
265  * which is important when load is unaligned.
266  */
267  if (n >= 64)
268  {
269  clib_mov64blocks ((u8 *) dst, (const u8 *) src, n);
270  bits = n;
271  n = n & 63;
272  bits -= n;
273  src = (const u8 *) src + bits;
274  dst = (u8 *) dst + bits;
275  }
276 
277  /**
278  * Copy whatever left
279  */
280  goto COPY_BLOCK_64_BACK31;
281 }
282 
283 
284 #endif /* included_clib_mamcpy_avx_h */
285 
286 
287 /*
288  * fd.io coding-style-patch-verification: ON
289  *
290  * Local Variables:
291  * eval: (c-set-style "gnu")
292  * End:
293  */
static void clib_mov64(u8 *dst, const u8 *src)
Definition: memcpy_avx.h:70
u8 u8x32u
Definition: memcpy_avx.h:55
static void * clib_memcpy(void *dst, const void *src, size_t n)
Definition: memcpy_avx.h:147
u8 u8x16u
Definition: memcpy_avx.h:54
u64 uword
Definition: types.h:112
unsigned char u8
Definition: types.h:56
static void clib_mov64blocks(u8 *dst, const u8 *src, size_t n)
Definition: memcpy_avx.h:91
static void clib_mov16(u8 *dst, const u8 *src)
Definition: memcpy_avx.h:58
static void clib_mov256blocks(u8 *dst, const u8 *src, size_t n)
Definition: memcpy_avx.h:110
static void clib_mov32(u8 *dst, const u8 *src)
Definition: memcpy_avx.h:64
static void clib_mov128(u8 *dst, const u8 *src)
Definition: memcpy_avx.h:77
static void clib_mov256(u8 *dst, const u8 *src)
Definition: memcpy_avx.h:84