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#include <vppinfra/warnings.h>
54
55/* *INDENT-OFF* */
56WARN_OFF (stringop-overflow)
57/* *INDENT-ON* */
58
59static inline void
60clib_mov16 (u8 * dst, const u8 * src)
61{
62  __m128i xmm0;
63
64  xmm0 = _mm_loadu_si128 ((const __m128i *) src);
65  _mm_storeu_si128 ((__m128i *) dst, xmm0);
66}
67
68static inline void
69clib_mov32 (u8 * dst, const u8 * src)
70{
71  __m256i ymm0;
72
73  ymm0 = _mm256_loadu_si256 ((const __m256i *) src);
74  _mm256_storeu_si256 ((__m256i *) dst, ymm0);
75}
76
77static inline void
78clib_mov64 (u8 * dst, const u8 * src)
79{
80  __m512i zmm0;
81
82  zmm0 = _mm512_loadu_si512 ((const void *) src);
83  _mm512_storeu_si512 ((void *) dst, zmm0);
84}
85
86static inline void
87clib_mov128 (u8 * dst, const u8 * src)
88{
89  clib_mov64 (dst + 0 * 64, src + 0 * 64);
90  clib_mov64 (dst + 1 * 64, src + 1 * 64);
91}
92
93static inline void
94clib_mov256 (u8 * dst, const u8 * src)
95{
96  clib_mov128 (dst + 0 * 128, src + 0 * 128);
97  clib_mov128 (dst + 1 * 128, src + 1 * 128);
98}
99
100static inline void
101clib_mov128blocks (u8 * dst, const u8 * src, size_t n)
102{
103  __m512i zmm0, zmm1;
104
105  while (n >= 128)
106    {
107      zmm0 = _mm512_loadu_si512 ((const void *) (src + 0 * 64));
108      n -= 128;
109      zmm1 = _mm512_loadu_si512 ((const void *) (src + 1 * 64));
110      src = src + 128;
111      _mm512_storeu_si512 ((void *) (dst + 0 * 64), zmm0);
112      _mm512_storeu_si512 ((void *) (dst + 1 * 64), zmm1);
113      dst = dst + 128;
114    }
115}
116
117static inline void
118clib_mov512blocks (u8 * dst, const u8 * src, size_t n)
119{
120  __m512i zmm0, zmm1, zmm2, zmm3, zmm4, zmm5, zmm6, zmm7;
121
122  while (n >= 512)
123    {
124      zmm0 = _mm512_loadu_si512 ((const void *) (src + 0 * 64));
125      n -= 512;
126      zmm1 = _mm512_loadu_si512 ((const void *) (src + 1 * 64));
127      zmm2 = _mm512_loadu_si512 ((const void *) (src + 2 * 64));
128      zmm3 = _mm512_loadu_si512 ((const void *) (src + 3 * 64));
129      zmm4 = _mm512_loadu_si512 ((const void *) (src + 4 * 64));
130      zmm5 = _mm512_loadu_si512 ((const void *) (src + 5 * 64));
131      zmm6 = _mm512_loadu_si512 ((const void *) (src + 6 * 64));
132      zmm7 = _mm512_loadu_si512 ((const void *) (src + 7 * 64));
133      src = src + 512;
134      _mm512_storeu_si512 ((void *) (dst + 0 * 64), zmm0);
135      _mm512_storeu_si512 ((void *) (dst + 1 * 64), zmm1);
136      _mm512_storeu_si512 ((void *) (dst + 2 * 64), zmm2);
137      _mm512_storeu_si512 ((void *) (dst + 3 * 64), zmm3);
138      _mm512_storeu_si512 ((void *) (dst + 4 * 64), zmm4);
139      _mm512_storeu_si512 ((void *) (dst + 5 * 64), zmm5);
140      _mm512_storeu_si512 ((void *) (dst + 6 * 64), zmm6);
141      _mm512_storeu_si512 ((void *) (dst + 7 * 64), zmm7);
142      dst = dst + 512;
143    }
144}
145
146static inline void *
147clib_memcpy_fast (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	  *(u16 *) dstu = *(const u16 *) srcu;
169	  srcu = (uword) ((const u16 *) srcu + 1);
170	  dstu = (uword) ((u16 *) dstu + 1);
171	}
172      if (n & 0x04)
173	{
174	  *(u32 *) dstu = *(const u32 *) srcu;
175	  srcu = (uword) ((const u32 *) srcu + 1);
176	  dstu = (uword) ((u32 *) dstu + 1);
177	}
178      if (n & 0x08)
179	*(u64 *) dstu = *(const u64 *) srcu;
180      return ret;
181    }
182
183	/**
184         * Fast way when copy size doesn't exceed 512 bytes
185         */
186  if (n <= 32)
187    {
188      clib_mov16 ((u8 *) dst, (const u8 *) src);
189      clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
190      return ret;
191    }
192  if (n <= 64)
193    {
194      clib_mov32 ((u8 *) dst, (const u8 *) src);
195      clib_mov32 ((u8 *) dst - 32 + n, (const u8 *) src - 32 + n);
196      return ret;
197    }
198  if (n <= 512)
199    {
200      if (n >= 256)
201	{
202	  n -= 256;
203	  clib_mov256 ((u8 *) dst, (const u8 *) src);
204	  src = (const u8 *) src + 256;
205	  dst = (u8 *) dst + 256;
206	}
207      if (n >= 128)
208	{
209	  n -= 128;
210	  clib_mov128 ((u8 *) dst, (const u8 *) src);
211	  src = (const u8 *) src + 128;
212	  dst = (u8 *) dst + 128;
213	}
214    COPY_BLOCK_128_BACK63:
215      if (n > 64)
216	{
217	  clib_mov64 ((u8 *) dst, (const u8 *) src);
218	  clib_mov64 ((u8 *) dst - 64 + n, (const u8 *) src - 64 + n);
219	  return ret;
220	}
221      if (n > 0)
222	clib_mov64 ((u8 *) dst - 64 + n, (const u8 *) src - 64 + n);
223      return ret;
224    }
225
226	/**
227         * Make store aligned when copy size exceeds 512 bytes
228         */
229  dstofss = (uword) dst & 0x3F;
230  if (dstofss > 0)
231    {
232      dstofss = 64 - dstofss;
233      n -= dstofss;
234      clib_mov64 ((u8 *) dst, (const u8 *) src);
235      src = (const u8 *) src + dstofss;
236      dst = (u8 *) dst + dstofss;
237    }
238
239	/**
240         * Copy 512-byte blocks.
241         * Use copy block function for better instruction order control,
242         * which is important when load is unaligned.
243         */
244  clib_mov512blocks ((u8 *) dst, (const u8 *) src, n);
245  bits = n;
246  n = n & 511;
247  bits -= n;
248  src = (const u8 *) src + bits;
249  dst = (u8 *) dst + bits;
250
251	/**
252         * Copy 128-byte blocks.
253         * Use copy block function for better instruction order control,
254         * which is important when load is unaligned.
255         */
256  if (n >= 128)
257    {
258      clib_mov128blocks ((u8 *) dst, (const u8 *) src, n);
259      bits = n;
260      n = n & 127;
261      bits -= n;
262      src = (const u8 *) src + bits;
263      dst = (u8 *) dst + bits;
264    }
265
266	/**
267         * Copy whatever left
268         */
269  goto COPY_BLOCK_128_BACK63;
270}
271
272/* *INDENT-OFF* */
273WARN_ON (stringop-overflow)
274/* *INDENT-ON* */
275
276#endif /* included_clib_memcpy_avx512_h */
277
278
279/*
280 * fd.io coding-style-patch-verification: ON
281 *
282 * Local Variables:
283 * eval: (c-set-style "gnu")
284 * End:
285 */
286