FD.io VPP  v20.01-48-g3e0dafb74
Vector Packet Processing
memcpy_avx512.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_avx512_h
49 #define included_clib_memcpy_avx512_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  __m512i zmm0;
76 
77  zmm0 = _mm512_loadu_si512 ((const void *) src);
78  _mm512_storeu_si512 ((void *) dst, zmm0);
79 }
80 
81 static inline void
82 clib_mov128 (u8 * dst, const u8 * src)
83 {
84  clib_mov64 (dst + 0 * 64, src + 0 * 64);
85  clib_mov64 (dst + 1 * 64, src + 1 * 64);
86 }
87 
88 static inline void
89 clib_mov256 (u8 * dst, const u8 * src)
90 {
91  clib_mov128 (dst + 0 * 128, src + 0 * 128);
92  clib_mov128 (dst + 1 * 128, src + 1 * 128);
93 }
94 
95 static inline void
96 clib_mov128blocks (u8 * dst, const u8 * src, size_t n)
97 {
98  __m512i zmm0, zmm1;
99 
100  while (n >= 128)
101  {
102  zmm0 = _mm512_loadu_si512 ((const void *) (src + 0 * 64));
103  n -= 128;
104  zmm1 = _mm512_loadu_si512 ((const void *) (src + 1 * 64));
105  src = src + 128;
106  _mm512_storeu_si512 ((void *) (dst + 0 * 64), zmm0);
107  _mm512_storeu_si512 ((void *) (dst + 1 * 64), zmm1);
108  dst = dst + 128;
109  }
110 }
111 
112 static inline void
113 clib_mov512blocks (u8 * dst, const u8 * src, size_t n)
114 {
115  __m512i zmm0, zmm1, zmm2, zmm3, zmm4, zmm5, zmm6, zmm7;
116 
117  while (n >= 512)
118  {
119  zmm0 = _mm512_loadu_si512 ((const void *) (src + 0 * 64));
120  n -= 512;
121  zmm1 = _mm512_loadu_si512 ((const void *) (src + 1 * 64));
122  zmm2 = _mm512_loadu_si512 ((const void *) (src + 2 * 64));
123  zmm3 = _mm512_loadu_si512 ((const void *) (src + 3 * 64));
124  zmm4 = _mm512_loadu_si512 ((const void *) (src + 4 * 64));
125  zmm5 = _mm512_loadu_si512 ((const void *) (src + 5 * 64));
126  zmm6 = _mm512_loadu_si512 ((const void *) (src + 6 * 64));
127  zmm7 = _mm512_loadu_si512 ((const void *) (src + 7 * 64));
128  src = src + 512;
129  _mm512_storeu_si512 ((void *) (dst + 0 * 64), zmm0);
130  _mm512_storeu_si512 ((void *) (dst + 1 * 64), zmm1);
131  _mm512_storeu_si512 ((void *) (dst + 2 * 64), zmm2);
132  _mm512_storeu_si512 ((void *) (dst + 3 * 64), zmm3);
133  _mm512_storeu_si512 ((void *) (dst + 4 * 64), zmm4);
134  _mm512_storeu_si512 ((void *) (dst + 5 * 64), zmm5);
135  _mm512_storeu_si512 ((void *) (dst + 6 * 64), zmm6);
136  _mm512_storeu_si512 ((void *) (dst + 7 * 64), zmm7);
137  dst = dst + 512;
138  }
139 }
140 
141 static inline void *
142 clib_memcpy_fast (void *dst, const void *src, size_t n)
143 {
144  uword dstu = (uword) dst;
145  uword srcu = (uword) src;
146  void *ret = dst;
147  size_t dstofss;
148  size_t bits;
149 
150  /**
151  * Copy less than 16 bytes
152  */
153  if (n < 16)
154  {
155  if (n & 0x01)
156  {
157  *(u8 *) dstu = *(const u8 *) srcu;
158  srcu = (uword) ((const u8 *) srcu + 1);
159  dstu = (uword) ((u8 *) dstu + 1);
160  }
161  if (n & 0x02)
162  {
163  *(u16 *) dstu = *(const u16 *) srcu;
164  srcu = (uword) ((const u16 *) srcu + 1);
165  dstu = (uword) ((u16 *) dstu + 1);
166  }
167  if (n & 0x04)
168  {
169  *(u32 *) dstu = *(const u32 *) srcu;
170  srcu = (uword) ((const u32 *) srcu + 1);
171  dstu = (uword) ((u32 *) dstu + 1);
172  }
173  if (n & 0x08)
174  *(u64 *) dstu = *(const u64 *) srcu;
175  return ret;
176  }
177 
178  /**
179  * Fast way when copy size doesn't exceed 512 bytes
180  */
181  if (n <= 32)
182  {
183  clib_mov16 ((u8 *) dst, (const u8 *) src);
184  clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
185  return ret;
186  }
187  if (n <= 64)
188  {
189  clib_mov32 ((u8 *) dst, (const u8 *) src);
190  clib_mov32 ((u8 *) dst - 32 + n, (const u8 *) src - 32 + n);
191  return ret;
192  }
193  if (n <= 512)
194  {
195  if (n >= 256)
196  {
197  n -= 256;
198  clib_mov256 ((u8 *) dst, (const u8 *) src);
199  src = (const u8 *) src + 256;
200  dst = (u8 *) dst + 256;
201  }
202  if (n >= 128)
203  {
204  n -= 128;
205  clib_mov128 ((u8 *) dst, (const u8 *) src);
206  src = (const u8 *) src + 128;
207  dst = (u8 *) dst + 128;
208  }
209  COPY_BLOCK_128_BACK63:
210  if (n > 64)
211  {
212  clib_mov64 ((u8 *) dst, (const u8 *) src);
213  clib_mov64 ((u8 *) dst - 64 + n, (const u8 *) src - 64 + n);
214  return ret;
215  }
216  if (n > 0)
217  clib_mov64 ((u8 *) dst - 64 + n, (const u8 *) src - 64 + n);
218  return ret;
219  }
220 
221  /**
222  * Make store aligned when copy size exceeds 512 bytes
223  */
224  dstofss = (uword) dst & 0x3F;
225  if (dstofss > 0)
226  {
227  dstofss = 64 - dstofss;
228  n -= dstofss;
229  clib_mov64 ((u8 *) dst, (const u8 *) src);
230  src = (const u8 *) src + dstofss;
231  dst = (u8 *) dst + dstofss;
232  }
233 
234  /**
235  * Copy 512-byte blocks.
236  * Use copy block function for better instruction order control,
237  * which is important when load is unaligned.
238  */
239  clib_mov512blocks ((u8 *) dst, (const u8 *) src, n);
240  bits = n;
241  n = n & 511;
242  bits -= n;
243  src = (const u8 *) src + bits;
244  dst = (u8 *) dst + bits;
245 
246  /**
247  * Copy 128-byte blocks.
248  * Use copy block function for better instruction order control,
249  * which is important when load is unaligned.
250  */
251  if (n >= 128)
252  {
253  clib_mov128blocks ((u8 *) dst, (const u8 *) src, n);
254  bits = n;
255  n = n & 127;
256  bits -= n;
257  src = (const u8 *) src + bits;
258  dst = (u8 *) dst + bits;
259  }
260 
261  /**
262  * Copy whatever left
263  */
264  goto COPY_BLOCK_128_BACK63;
265 }
266 
267 
268 #endif /* included_clib_memcpy_avx512_h */
269 
270 
271 /*
272  * fd.io coding-style-patch-verification: ON
273  *
274  * Local Variables:
275  * eval: (c-set-style "gnu")
276  * End:
277  */
unsigned long u64
Definition: types.h:89
static void clib_mov16(u8 *dst, const u8 *src)
Definition: memcpy_avx512.h:55
vl_api_address_t src
Definition: gre.api:60
static void clib_mov64(u8 *dst, const u8 *src)
Definition: memcpy_avx512.h:73
unsigned char u8
Definition: types.h:56
static void clib_mov32(u8 *dst, const u8 *src)
Definition: memcpy_avx512.h:64
unsigned int u32
Definition: types.h:88
static void clib_mov256(u8 *dst, const u8 *src)
Definition: memcpy_avx512.h:89
unsigned short u16
Definition: types.h:57
static void clib_mov128(u8 *dst, const u8 *src)
Definition: memcpy_avx512.h:82
vl_api_address_t dst
Definition: gre.api:61
static void clib_mov512blocks(u8 *dst, const u8 *src, size_t n)
static void clib_mov128blocks(u8 *dst, const u8 *src, size_t n)
Definition: memcpy_avx512.h:96
static void * clib_memcpy_fast(void *dst, const void *src, size_t n)
u64 uword
Definition: types.h:112