FD.io VPP  v17.07.01-10-g3be13f0
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 static inline void
55 clib_mov16 (u8 * dst, const u8 * src)
56 {
57  __m128i xmm0;
58 
59  xmm0 = _mm_loadu_si128 ((const __m128i *) src);
60  _mm_storeu_si128 ((__m128i *) dst, xmm0);
61 }
62 
63 static inline void
64 clib_mov32 (u8 * dst, const u8 * src)
65 {
66  __m256i ymm0;
67 
68  ymm0 = _mm256_loadu_si256 ((const __m256i *) src);
69  _mm256_storeu_si256 ((__m256i *) dst, ymm0);
70 }
71 
72 static inline void
73 clib_mov64 (u8 * dst, const u8 * src)
74 {
75  clib_mov32 ((u8 *) dst + 0 * 32, (const u8 *) src + 0 * 32);
76  clib_mov32 ((u8 *) dst + 1 * 32, (const u8 *) src + 1 * 32);
77 }
78 
79 static inline void
80 clib_mov128 (u8 * dst, const u8 * src)
81 {
82  clib_mov64 ((u8 *) dst + 0 * 64, (const u8 *) src + 0 * 64);
83  clib_mov64 ((u8 *) dst + 1 * 64, (const u8 *) src + 1 * 64);
84 }
85 
86 static inline void
87 clib_mov256 (u8 * dst, const u8 * src)
88 {
89  clib_mov128 ((u8 *) dst + 0 * 128, (const u8 *) src + 0 * 128);
90  clib_mov128 ((u8 *) dst + 1 * 128, (const u8 *) src + 1 * 128);
91 }
92 
93 static inline void
94 clib_mov64blocks (u8 * dst, const u8 * src, size_t n)
95 {
96  __m256i ymm0, ymm1;
97 
98  while (n >= 64)
99  {
100  ymm0 =
101  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 0 * 32));
102  n -= 64;
103  ymm1 =
104  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 1 * 32));
105  src = (const u8 *) src + 64;
106  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 0 * 32), ymm0);
107  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 1 * 32), ymm1);
108  dst = (u8 *) dst + 64;
109  }
110 }
111 
112 static inline void
113 clib_mov256blocks (u8 * dst, const u8 * src, size_t n)
114 {
115  __m256i ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7;
116 
117  while (n >= 256)
118  {
119  ymm0 =
120  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 0 * 32));
121  n -= 256;
122  ymm1 =
123  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 1 * 32));
124  ymm2 =
125  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 2 * 32));
126  ymm3 =
127  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 3 * 32));
128  ymm4 =
129  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 4 * 32));
130  ymm5 =
131  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 5 * 32));
132  ymm6 =
133  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 6 * 32));
134  ymm7 =
135  _mm256_loadu_si256 ((const __m256i *) ((const u8 *) src + 7 * 32));
136  src = (const u8 *) src + 256;
137  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 0 * 32), ymm0);
138  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 1 * 32), ymm1);
139  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 2 * 32), ymm2);
140  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 3 * 32), ymm3);
141  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 4 * 32), ymm4);
142  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 5 * 32), ymm5);
143  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 6 * 32), ymm6);
144  _mm256_storeu_si256 ((__m256i *) ((u8 *) dst + 7 * 32), ymm7);
145  dst = (u8 *) dst + 256;
146  }
147 }
148 
149 static inline void *
150 clib_memcpy (void *dst, const void *src, size_t n)
151 {
152  uword dstu = (uword) dst;
153  uword srcu = (uword) src;
154  void *ret = dst;
155  size_t dstofss;
156  size_t bits;
157 
158  /**
159  * Copy less than 16 bytes
160  */
161  if (n < 16)
162  {
163  if (n & 0x01)
164  {
165  *(u8 *) dstu = *(const u8 *) srcu;
166  srcu = (uword) ((const u8 *) srcu + 1);
167  dstu = (uword) ((u8 *) dstu + 1);
168  }
169  if (n & 0x02)
170  {
171  *(uint16_t *) dstu = *(const uint16_t *) srcu;
172  srcu = (uword) ((const uint16_t *) srcu + 1);
173  dstu = (uword) ((uint16_t *) dstu + 1);
174  }
175  if (n & 0x04)
176  {
177  *(uint32_t *) dstu = *(const uint32_t *) srcu;
178  srcu = (uword) ((const uint32_t *) srcu + 1);
179  dstu = (uword) ((uint32_t *) dstu + 1);
180  }
181  if (n & 0x08)
182  {
183  *(uint64_t *) dstu = *(const uint64_t *) srcu;
184  }
185  return ret;
186  }
187 
188  /**
189  * Fast way when copy size doesn't exceed 512 bytes
190  */
191  if (n <= 32)
192  {
193  clib_mov16 ((u8 *) dst, (const u8 *) src);
194  clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
195  return ret;
196  }
197  if (n <= 64)
198  {
199  clib_mov32 ((u8 *) dst, (const u8 *) src);
200  clib_mov32 ((u8 *) dst - 32 + n, (const u8 *) src - 32 + n);
201  return ret;
202  }
203  if (n <= 512)
204  {
205  if (n >= 256)
206  {
207  n -= 256;
208  clib_mov256 ((u8 *) dst, (const u8 *) src);
209  src = (const u8 *) src + 256;
210  dst = (u8 *) dst + 256;
211  }
212  if (n >= 128)
213  {
214  n -= 128;
215  clib_mov128 ((u8 *) dst, (const u8 *) src);
216  src = (const u8 *) src + 128;
217  dst = (u8 *) dst + 128;
218  }
219  if (n >= 64)
220  {
221  n -= 64;
222  clib_mov64 ((u8 *) dst, (const u8 *) src);
223  src = (const u8 *) src + 64;
224  dst = (u8 *) dst + 64;
225  }
226  COPY_BLOCK_64_BACK31:
227  if (n > 32)
228  {
229  clib_mov32 ((u8 *) dst, (const u8 *) src);
230  clib_mov32 ((u8 *) dst - 32 + n, (const u8 *) src - 32 + n);
231  return ret;
232  }
233  if (n > 0)
234  {
235  clib_mov32 ((u8 *) dst - 32 + n, (const u8 *) src - 32 + n);
236  }
237  return ret;
238  }
239 
240  /**
241  * Make store aligned when copy size exceeds 512 bytes
242  */
243  dstofss = (uword) dst & 0x1F;
244  if (dstofss > 0)
245  {
246  dstofss = 32 - dstofss;
247  n -= dstofss;
248  clib_mov32 ((u8 *) dst, (const u8 *) src);
249  src = (const u8 *) src + dstofss;
250  dst = (u8 *) dst + dstofss;
251  }
252 
253  /**
254  * Copy 256-byte blocks.
255  * Use copy block function for better instruction order control,
256  * which is important when load is unaligned.
257  */
258  clib_mov256blocks ((u8 *) dst, (const u8 *) src, n);
259  bits = n;
260  n = n & 255;
261  bits -= n;
262  src = (const u8 *) src + bits;
263  dst = (u8 *) dst + bits;
264 
265  /**
266  * Copy 64-byte blocks.
267  * Use copy block function for better instruction order control,
268  * which is important when load is unaligned.
269  */
270  if (n >= 64)
271  {
272  clib_mov64blocks ((u8 *) dst, (const u8 *) src, n);
273  bits = n;
274  n = n & 63;
275  bits -= n;
276  src = (const u8 *) src + bits;
277  dst = (u8 *) dst + bits;
278  }
279 
280  /**
281  * Copy whatever left
282  */
283  goto COPY_BLOCK_64_BACK31;
284 }
285 
286 
287 #endif /* included_clib_mamcpy_avx_h */
288 
289 
290 /*
291  * fd.io coding-style-patch-verification: ON
292  *
293  * Local Variables:
294  * eval: (c-set-style "gnu")
295  * End:
296  */
static void clib_mov64(u8 *dst, const u8 *src)
Definition: memcpy_avx.h:73
static void * clib_memcpy(void *dst, const void *src, size_t n)
Definition: memcpy_avx.h:150
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:94
static void clib_mov16(u8 *dst, const u8 *src)
Definition: memcpy_avx.h:55
static void clib_mov256blocks(u8 *dst, const u8 *src, size_t n)
Definition: memcpy_avx.h:113
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:80
static void clib_mov256(u8 *dst, const u8 *src)
Definition: memcpy_avx.h:87