1 | /* |
2 | * Compute the CRC32 using a parallelized folding approach with the PCLMULQDQ |
3 | * instruction. |
4 | * |
5 | * A white paper describing this algorithm can be found at: |
6 | * http://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf |
7 | * |
8 | * Copyright (C) 2013 Intel Corporation. All rights reserved. |
9 | * Authors: |
10 | * Wajdi Feghali <wajdi.k.feghali@intel.com> |
11 | * Jim Guilford <james.guilford@intel.com> |
12 | * Vinodh Gopal <vinodh.gopal@intel.com> |
13 | * Erdinc Ozturk <erdinc.ozturk@intel.com> |
14 | * Jim Kukunas <james.t.kukunas@linux.intel.com> |
15 | * |
16 | * For conditions of distribution and use, see copyright notice in zlib.h |
17 | */ |
18 | |
19 | #ifdef X86_PCLMULQDQ_CRC |
20 | |
21 | #include "../../zbuild.h" |
22 | #include <inttypes.h> |
23 | #include <immintrin.h> |
24 | #include <wmmintrin.h> |
25 | |
26 | #include "crc_folding.h" |
27 | |
28 | ZLIB_INTERNAL void crc_fold_init(deflate_state *const s) { |
29 | /* CRC_SAVE */ |
30 | _mm_storeu_si128((__m128i *)s->crc0 + 0, _mm_cvtsi32_si128(0x9db42487)); |
31 | _mm_storeu_si128((__m128i *)s->crc0 + 1, _mm_setzero_si128()); |
32 | _mm_storeu_si128((__m128i *)s->crc0 + 2, _mm_setzero_si128()); |
33 | _mm_storeu_si128((__m128i *)s->crc0 + 3, _mm_setzero_si128()); |
34 | |
35 | s->strm->adler = 0; |
36 | } |
37 | |
38 | static void fold_1(__m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, __m128i *xmm_crc3) { |
39 | const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4, |
40 | 0x00000001, 0xc6e41596); |
41 | __m128i x_tmp3; |
42 | __m128 ps_crc0, ps_crc3, ps_res; |
43 | |
44 | x_tmp3 = *xmm_crc3; |
45 | |
46 | *xmm_crc3 = *xmm_crc0; |
47 | *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01); |
48 | *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10); |
49 | ps_crc0 = _mm_castsi128_ps(*xmm_crc0); |
50 | ps_crc3 = _mm_castsi128_ps(*xmm_crc3); |
51 | ps_res = _mm_xor_ps(ps_crc0, ps_crc3); |
52 | |
53 | *xmm_crc0 = *xmm_crc1; |
54 | *xmm_crc1 = *xmm_crc2; |
55 | *xmm_crc2 = x_tmp3; |
56 | *xmm_crc3 = _mm_castps_si128(ps_res); |
57 | } |
58 | |
59 | static void fold_2(__m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, __m128i *xmm_crc3) { |
60 | const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4, |
61 | 0x00000001, 0xc6e41596); |
62 | __m128i x_tmp3, x_tmp2; |
63 | __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3, ps_res31, ps_res20; |
64 | |
65 | x_tmp3 = *xmm_crc3; |
66 | x_tmp2 = *xmm_crc2; |
67 | |
68 | *xmm_crc3 = *xmm_crc1; |
69 | *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01); |
70 | *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10); |
71 | ps_crc3 = _mm_castsi128_ps(*xmm_crc3); |
72 | ps_crc1 = _mm_castsi128_ps(*xmm_crc1); |
73 | ps_res31 = _mm_xor_ps(ps_crc3, ps_crc1); |
74 | |
75 | *xmm_crc2 = *xmm_crc0; |
76 | *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01); |
77 | *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x10); |
78 | ps_crc0 = _mm_castsi128_ps(*xmm_crc0); |
79 | ps_crc2 = _mm_castsi128_ps(*xmm_crc2); |
80 | ps_res20 = _mm_xor_ps(ps_crc0, ps_crc2); |
81 | |
82 | *xmm_crc0 = x_tmp2; |
83 | *xmm_crc1 = x_tmp3; |
84 | *xmm_crc2 = _mm_castps_si128(ps_res20); |
85 | *xmm_crc3 = _mm_castps_si128(ps_res31); |
86 | } |
87 | |
88 | static void fold_3(__m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, __m128i *xmm_crc3) { |
89 | const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4, |
90 | 0x00000001, 0xc6e41596); |
91 | __m128i x_tmp3; |
92 | __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3, ps_res32, ps_res21, ps_res10; |
93 | |
94 | x_tmp3 = *xmm_crc3; |
95 | |
96 | *xmm_crc3 = *xmm_crc2; |
97 | *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x01); |
98 | *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x10); |
99 | ps_crc2 = _mm_castsi128_ps(*xmm_crc2); |
100 | ps_crc3 = _mm_castsi128_ps(*xmm_crc3); |
101 | ps_res32 = _mm_xor_ps(ps_crc2, ps_crc3); |
102 | |
103 | *xmm_crc2 = *xmm_crc1; |
104 | *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01); |
105 | *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x10); |
106 | ps_crc1 = _mm_castsi128_ps(*xmm_crc1); |
107 | ps_crc2 = _mm_castsi128_ps(*xmm_crc2); |
108 | ps_res21 = _mm_xor_ps(ps_crc1, ps_crc2); |
109 | |
110 | *xmm_crc1 = *xmm_crc0; |
111 | *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01); |
112 | *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x10); |
113 | ps_crc0 = _mm_castsi128_ps(*xmm_crc0); |
114 | ps_crc1 = _mm_castsi128_ps(*xmm_crc1); |
115 | ps_res10 = _mm_xor_ps(ps_crc0, ps_crc1); |
116 | |
117 | *xmm_crc0 = x_tmp3; |
118 | *xmm_crc1 = _mm_castps_si128(ps_res10); |
119 | *xmm_crc2 = _mm_castps_si128(ps_res21); |
120 | *xmm_crc3 = _mm_castps_si128(ps_res32); |
121 | } |
122 | |
123 | static void fold_4(__m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, __m128i *xmm_crc3) { |
124 | const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4, |
125 | 0x00000001, 0xc6e41596); |
126 | __m128i x_tmp0, x_tmp1, x_tmp2, x_tmp3; |
127 | __m128 ps_crc0, ps_crc1, ps_crc2, ps_crc3; |
128 | __m128 ps_t0, ps_t1, ps_t2, ps_t3; |
129 | __m128 ps_res0, ps_res1, ps_res2, ps_res3; |
130 | |
131 | x_tmp0 = *xmm_crc0; |
132 | x_tmp1 = *xmm_crc1; |
133 | x_tmp2 = *xmm_crc2; |
134 | x_tmp3 = *xmm_crc3; |
135 | |
136 | *xmm_crc0 = _mm_clmulepi64_si128(*xmm_crc0, xmm_fold4, 0x01); |
137 | x_tmp0 = _mm_clmulepi64_si128(x_tmp0, xmm_fold4, 0x10); |
138 | ps_crc0 = _mm_castsi128_ps(*xmm_crc0); |
139 | ps_t0 = _mm_castsi128_ps(x_tmp0); |
140 | ps_res0 = _mm_xor_ps(ps_crc0, ps_t0); |
141 | |
142 | *xmm_crc1 = _mm_clmulepi64_si128(*xmm_crc1, xmm_fold4, 0x01); |
143 | x_tmp1 = _mm_clmulepi64_si128(x_tmp1, xmm_fold4, 0x10); |
144 | ps_crc1 = _mm_castsi128_ps(*xmm_crc1); |
145 | ps_t1 = _mm_castsi128_ps(x_tmp1); |
146 | ps_res1 = _mm_xor_ps(ps_crc1, ps_t1); |
147 | |
148 | *xmm_crc2 = _mm_clmulepi64_si128(*xmm_crc2, xmm_fold4, 0x01); |
149 | x_tmp2 = _mm_clmulepi64_si128(x_tmp2, xmm_fold4, 0x10); |
150 | ps_crc2 = _mm_castsi128_ps(*xmm_crc2); |
151 | ps_t2 = _mm_castsi128_ps(x_tmp2); |
152 | ps_res2 = _mm_xor_ps(ps_crc2, ps_t2); |
153 | |
154 | *xmm_crc3 = _mm_clmulepi64_si128(*xmm_crc3, xmm_fold4, 0x01); |
155 | x_tmp3 = _mm_clmulepi64_si128(x_tmp3, xmm_fold4, 0x10); |
156 | ps_crc3 = _mm_castsi128_ps(*xmm_crc3); |
157 | ps_t3 = _mm_castsi128_ps(x_tmp3); |
158 | ps_res3 = _mm_xor_ps(ps_crc3, ps_t3); |
159 | |
160 | *xmm_crc0 = _mm_castps_si128(ps_res0); |
161 | *xmm_crc1 = _mm_castps_si128(ps_res1); |
162 | *xmm_crc2 = _mm_castps_si128(ps_res2); |
163 | *xmm_crc3 = _mm_castps_si128(ps_res3); |
164 | } |
165 | |
166 | static const unsigned ALIGNED_(32) pshufb_shf_table[60] = { |
167 | 0x84838281, 0x88878685, 0x8c8b8a89, 0x008f8e8d, /* shl 15 (16 - 1)/shr1 */ |
168 | 0x85848382, 0x89888786, 0x8d8c8b8a, 0x01008f8e, /* shl 14 (16 - 3)/shr2 */ |
169 | 0x86858483, 0x8a898887, 0x8e8d8c8b, 0x0201008f, /* shl 13 (16 - 4)/shr3 */ |
170 | 0x87868584, 0x8b8a8988, 0x8f8e8d8c, 0x03020100, /* shl 12 (16 - 4)/shr4 */ |
171 | 0x88878685, 0x8c8b8a89, 0x008f8e8d, 0x04030201, /* shl 11 (16 - 5)/shr5 */ |
172 | 0x89888786, 0x8d8c8b8a, 0x01008f8e, 0x05040302, /* shl 10 (16 - 6)/shr6 */ |
173 | 0x8a898887, 0x8e8d8c8b, 0x0201008f, 0x06050403, /* shl 9 (16 - 7)/shr7 */ |
174 | 0x8b8a8988, 0x8f8e8d8c, 0x03020100, 0x07060504, /* shl 8 (16 - 8)/shr8 */ |
175 | 0x8c8b8a89, 0x008f8e8d, 0x04030201, 0x08070605, /* shl 7 (16 - 9)/shr9 */ |
176 | 0x8d8c8b8a, 0x01008f8e, 0x05040302, 0x09080706, /* shl 6 (16 -10)/shr10*/ |
177 | 0x8e8d8c8b, 0x0201008f, 0x06050403, 0x0a090807, /* shl 5 (16 -11)/shr11*/ |
178 | 0x8f8e8d8c, 0x03020100, 0x07060504, 0x0b0a0908, /* shl 4 (16 -12)/shr12*/ |
179 | 0x008f8e8d, 0x04030201, 0x08070605, 0x0c0b0a09, /* shl 3 (16 -13)/shr13*/ |
180 | 0x01008f8e, 0x05040302, 0x09080706, 0x0d0c0b0a, /* shl 2 (16 -14)/shr14*/ |
181 | 0x0201008f, 0x06050403, 0x0a090807, 0x0e0d0c0b /* shl 1 (16 -15)/shr15*/ |
182 | }; |
183 | |
184 | static void partial_fold(const size_t len, __m128i *xmm_crc0, __m128i *xmm_crc1, __m128i *xmm_crc2, |
185 | __m128i *xmm_crc3, __m128i *xmm_crc_part) { |
186 | |
187 | const __m128i xmm_fold4 = _mm_set_epi32( 0x00000001, 0x54442bd4, |
188 | 0x00000001, 0xc6e41596); |
189 | const __m128i xmm_mask3 = _mm_set1_epi32(0x80808080); |
190 | |
191 | __m128i xmm_shl, xmm_shr, xmm_tmp1, xmm_tmp2, xmm_tmp3; |
192 | __m128i xmm_a0_0, xmm_a0_1; |
193 | __m128 ps_crc3, psa0_0, psa0_1, ps_res; |
194 | |
195 | xmm_shl = _mm_load_si128((__m128i *)pshufb_shf_table + (len - 1)); |
196 | xmm_shr = xmm_shl; |
197 | xmm_shr = _mm_xor_si128(xmm_shr, xmm_mask3); |
198 | |
199 | xmm_a0_0 = _mm_shuffle_epi8(*xmm_crc0, xmm_shl); |
200 | |
201 | *xmm_crc0 = _mm_shuffle_epi8(*xmm_crc0, xmm_shr); |
202 | xmm_tmp1 = _mm_shuffle_epi8(*xmm_crc1, xmm_shl); |
203 | *xmm_crc0 = _mm_or_si128(*xmm_crc0, xmm_tmp1); |
204 | |
205 | *xmm_crc1 = _mm_shuffle_epi8(*xmm_crc1, xmm_shr); |
206 | xmm_tmp2 = _mm_shuffle_epi8(*xmm_crc2, xmm_shl); |
207 | *xmm_crc1 = _mm_or_si128(*xmm_crc1, xmm_tmp2); |
208 | |
209 | *xmm_crc2 = _mm_shuffle_epi8(*xmm_crc2, xmm_shr); |
210 | xmm_tmp3 = _mm_shuffle_epi8(*xmm_crc3, xmm_shl); |
211 | *xmm_crc2 = _mm_or_si128(*xmm_crc2, xmm_tmp3); |
212 | |
213 | *xmm_crc3 = _mm_shuffle_epi8(*xmm_crc3, xmm_shr); |
214 | *xmm_crc_part = _mm_shuffle_epi8(*xmm_crc_part, xmm_shl); |
215 | *xmm_crc3 = _mm_or_si128(*xmm_crc3, *xmm_crc_part); |
216 | |
217 | xmm_a0_1 = _mm_clmulepi64_si128(xmm_a0_0, xmm_fold4, 0x10); |
218 | xmm_a0_0 = _mm_clmulepi64_si128(xmm_a0_0, xmm_fold4, 0x01); |
219 | |
220 | ps_crc3 = _mm_castsi128_ps(*xmm_crc3); |
221 | psa0_0 = _mm_castsi128_ps(xmm_a0_0); |
222 | psa0_1 = _mm_castsi128_ps(xmm_a0_1); |
223 | |
224 | ps_res = _mm_xor_ps(ps_crc3, psa0_0); |
225 | ps_res = _mm_xor_ps(ps_res, psa0_1); |
226 | |
227 | *xmm_crc3 = _mm_castps_si128(ps_res); |
228 | } |
229 | |
230 | ZLIB_INTERNAL void crc_fold_copy(deflate_state *const s, unsigned char *dst, const unsigned char *src, long len) { |
231 | unsigned long algn_diff; |
232 | __m128i xmm_t0, xmm_t1, xmm_t2, xmm_t3; |
233 | char ALIGNED_(16) partial_buf[16] = { 0 }; |
234 | |
235 | /* CRC_LOAD */ |
236 | __m128i xmm_crc0 = _mm_loadu_si128((__m128i *)s->crc0 + 0); |
237 | __m128i xmm_crc1 = _mm_loadu_si128((__m128i *)s->crc0 + 1); |
238 | __m128i xmm_crc2 = _mm_loadu_si128((__m128i *)s->crc0 + 2); |
239 | __m128i xmm_crc3 = _mm_loadu_si128((__m128i *)s->crc0 + 3); |
240 | __m128i xmm_crc_part; |
241 | |
242 | if (len < 16) { |
243 | if (len == 0) |
244 | return; |
245 | |
246 | memcpy(partial_buf, src, len); |
247 | xmm_crc_part = _mm_loadu_si128((const __m128i *)partial_buf); |
248 | memcpy(dst, partial_buf, len); |
249 | goto partial; |
250 | } |
251 | |
252 | algn_diff = ((uintptr_t)0 - (uintptr_t)src) & 0xF; |
253 | if (algn_diff) { |
254 | xmm_crc_part = _mm_loadu_si128((__m128i *)src); |
255 | _mm_storeu_si128((__m128i *)dst, xmm_crc_part); |
256 | |
257 | dst += algn_diff; |
258 | src += algn_diff; |
259 | len -= algn_diff; |
260 | |
261 | partial_fold(algn_diff, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, &xmm_crc_part); |
262 | } else { |
263 | xmm_crc_part = _mm_setzero_si128(); |
264 | } |
265 | |
266 | while ((len -= 64) >= 0) { |
267 | /* CRC_LOAD */ |
268 | xmm_t0 = _mm_load_si128((__m128i *)src); |
269 | xmm_t1 = _mm_load_si128((__m128i *)src + 1); |
270 | xmm_t2 = _mm_load_si128((__m128i *)src + 2); |
271 | xmm_t3 = _mm_load_si128((__m128i *)src + 3); |
272 | |
273 | fold_4(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3); |
274 | |
275 | /* CRC_SAVE */ |
276 | _mm_storeu_si128((__m128i *)dst, xmm_t0); |
277 | _mm_storeu_si128((__m128i *)dst + 1, xmm_t1); |
278 | _mm_storeu_si128((__m128i *)dst + 2, xmm_t2); |
279 | _mm_storeu_si128((__m128i *)dst + 3, xmm_t3); |
280 | |
281 | xmm_crc0 = _mm_xor_si128(xmm_crc0, xmm_t0); |
282 | xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_t1); |
283 | xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t2); |
284 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t3); |
285 | |
286 | src += 64; |
287 | dst += 64; |
288 | } |
289 | |
290 | /* |
291 | * len = num bytes left - 64 |
292 | */ |
293 | if (len + 16 >= 0) { |
294 | len += 16; |
295 | |
296 | xmm_t0 = _mm_load_si128((__m128i *)src); |
297 | xmm_t1 = _mm_load_si128((__m128i *)src + 1); |
298 | xmm_t2 = _mm_load_si128((__m128i *)src + 2); |
299 | |
300 | fold_3(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3); |
301 | |
302 | _mm_storeu_si128((__m128i *)dst, xmm_t0); |
303 | _mm_storeu_si128((__m128i *)dst + 1, xmm_t1); |
304 | _mm_storeu_si128((__m128i *)dst + 2, xmm_t2); |
305 | |
306 | xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_t0); |
307 | xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t1); |
308 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t2); |
309 | |
310 | if (len == 0) |
311 | goto done; |
312 | |
313 | dst += 48; |
314 | memcpy(&xmm_crc_part, (__m128i *)src + 3, len); |
315 | } else if (len + 32 >= 0) { |
316 | len += 32; |
317 | |
318 | xmm_t0 = _mm_load_si128((__m128i *)src); |
319 | xmm_t1 = _mm_load_si128((__m128i *)src + 1); |
320 | |
321 | fold_2(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3); |
322 | |
323 | _mm_storeu_si128((__m128i *)dst, xmm_t0); |
324 | _mm_storeu_si128((__m128i *)dst + 1, xmm_t1); |
325 | |
326 | xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_t0); |
327 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t1); |
328 | |
329 | if (len == 0) |
330 | goto done; |
331 | |
332 | dst += 32; |
333 | memcpy(&xmm_crc_part, (__m128i *)src + 2, len); |
334 | } else if (len + 48 >= 0) { |
335 | len += 48; |
336 | |
337 | xmm_t0 = _mm_load_si128((__m128i *)src); |
338 | |
339 | fold_1(&xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3); |
340 | |
341 | _mm_storeu_si128((__m128i *)dst, xmm_t0); |
342 | |
343 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_t0); |
344 | |
345 | if (len == 0) |
346 | goto done; |
347 | |
348 | dst += 16; |
349 | memcpy(&xmm_crc_part, (__m128i *)src + 1, len); |
350 | } else { |
351 | len += 64; |
352 | if (len == 0) |
353 | goto done; |
354 | memcpy(&xmm_crc_part, src, len); |
355 | } |
356 | |
357 | _mm_storeu_si128((__m128i *)partial_buf, xmm_crc_part); |
358 | memcpy(dst, partial_buf, len); |
359 | |
360 | partial: |
361 | partial_fold(len, &xmm_crc0, &xmm_crc1, &xmm_crc2, &xmm_crc3, &xmm_crc_part); |
362 | done: |
363 | /* CRC_SAVE */ |
364 | _mm_storeu_si128((__m128i *)s->crc0 + 0, xmm_crc0); |
365 | _mm_storeu_si128((__m128i *)s->crc0 + 1, xmm_crc1); |
366 | _mm_storeu_si128((__m128i *)s->crc0 + 2, xmm_crc2); |
367 | _mm_storeu_si128((__m128i *)s->crc0 + 3, xmm_crc3); |
368 | _mm_storeu_si128((__m128i *)s->crc0 + 4, xmm_crc_part); |
369 | } |
370 | |
371 | static const unsigned ALIGNED_(16) crc_k[] = { |
372 | 0xccaa009e, 0x00000000, /* rk1 */ |
373 | 0x751997d0, 0x00000001, /* rk2 */ |
374 | 0xccaa009e, 0x00000000, /* rk5 */ |
375 | 0x63cd6124, 0x00000001, /* rk6 */ |
376 | 0xf7011640, 0x00000001, /* rk7 */ |
377 | 0xdb710640, 0x00000001 /* rk8 */ |
378 | }; |
379 | |
380 | static const unsigned ALIGNED_(16) crc_mask[4] = { |
381 | 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000 |
382 | }; |
383 | |
384 | static const unsigned ALIGNED_(16) crc_mask2[4] = { |
385 | 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF |
386 | }; |
387 | |
388 | uint32_t ZLIB_INTERNAL crc_fold_512to32(deflate_state *const s) { |
389 | const __m128i xmm_mask = _mm_load_si128((__m128i *)crc_mask); |
390 | const __m128i xmm_mask2 = _mm_load_si128((__m128i *)crc_mask2); |
391 | |
392 | uint32_t crc; |
393 | __m128i x_tmp0, x_tmp1, x_tmp2, crc_fold; |
394 | |
395 | /* CRC_LOAD */ |
396 | __m128i xmm_crc0 = _mm_loadu_si128((__m128i *)s->crc0 + 0); |
397 | __m128i xmm_crc1 = _mm_loadu_si128((__m128i *)s->crc0 + 1); |
398 | __m128i xmm_crc2 = _mm_loadu_si128((__m128i *)s->crc0 + 2); |
399 | __m128i xmm_crc3 = _mm_loadu_si128((__m128i *)s->crc0 + 3); |
400 | |
401 | /* |
402 | * k1 |
403 | */ |
404 | crc_fold = _mm_load_si128((__m128i *)crc_k); |
405 | |
406 | x_tmp0 = _mm_clmulepi64_si128(xmm_crc0, crc_fold, 0x10); |
407 | xmm_crc0 = _mm_clmulepi64_si128(xmm_crc0, crc_fold, 0x01); |
408 | xmm_crc1 = _mm_xor_si128(xmm_crc1, x_tmp0); |
409 | xmm_crc1 = _mm_xor_si128(xmm_crc1, xmm_crc0); |
410 | |
411 | x_tmp1 = _mm_clmulepi64_si128(xmm_crc1, crc_fold, 0x10); |
412 | xmm_crc1 = _mm_clmulepi64_si128(xmm_crc1, crc_fold, 0x01); |
413 | xmm_crc2 = _mm_xor_si128(xmm_crc2, x_tmp1); |
414 | xmm_crc2 = _mm_xor_si128(xmm_crc2, xmm_crc1); |
415 | |
416 | x_tmp2 = _mm_clmulepi64_si128(xmm_crc2, crc_fold, 0x10); |
417 | xmm_crc2 = _mm_clmulepi64_si128(xmm_crc2, crc_fold, 0x01); |
418 | xmm_crc3 = _mm_xor_si128(xmm_crc3, x_tmp2); |
419 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2); |
420 | |
421 | /* |
422 | * k5 |
423 | */ |
424 | crc_fold = _mm_load_si128((__m128i *)crc_k + 1); |
425 | |
426 | xmm_crc0 = xmm_crc3; |
427 | xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0); |
428 | xmm_crc0 = _mm_srli_si128(xmm_crc0, 8); |
429 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc0); |
430 | |
431 | xmm_crc0 = xmm_crc3; |
432 | xmm_crc3 = _mm_slli_si128(xmm_crc3, 4); |
433 | xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0x10); |
434 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc0); |
435 | xmm_crc3 = _mm_and_si128(xmm_crc3, xmm_mask2); |
436 | |
437 | /* |
438 | * k7 |
439 | */ |
440 | xmm_crc1 = xmm_crc3; |
441 | xmm_crc2 = xmm_crc3; |
442 | crc_fold = _mm_load_si128((__m128i *)crc_k + 2); |
443 | |
444 | xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0); |
445 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2); |
446 | xmm_crc3 = _mm_and_si128(xmm_crc3, xmm_mask); |
447 | |
448 | xmm_crc2 = xmm_crc3; |
449 | xmm_crc3 = _mm_clmulepi64_si128(xmm_crc3, crc_fold, 0x10); |
450 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc2); |
451 | xmm_crc3 = _mm_xor_si128(xmm_crc3, xmm_crc1); |
452 | |
453 | crc = _mm_extract_epi32(xmm_crc3, 2); |
454 | return ~crc; |
455 | } |
456 | |
457 | #endif |
458 | |
459 | |