xref: /openbmc/linux/arch/arm64/lib/xor-neon.c (revision a266ef69) 
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * arch/arm64/lib/xor-neon.c
4  *
5  * Authors: Jackie Liu <liuyun01@kylinos.cn>
6  * Copyright (C) 2018,Tianjin KYLIN Information Technology Co., Ltd.
7  */
8 
9 #include <linux/raid/xor.h>
10 #include <linux/module.h>
11 #include <asm/neon-intrinsics.h>
12 
13 void xor_arm64_neon_2(unsigned long bytes, unsigned long * __restrict p1,
14 	const unsigned long * __restrict p2)
15 {
16 	uint64_t *dp1 = (uint64_t *)p1;
17 	uint64_t *dp2 = (uint64_t *)p2;
18 
19 	register uint64x2_t v0, v1, v2, v3;
20 	long lines = bytes / (sizeof(uint64x2_t) * 4);
21 
22 	do {
23 		/* p1 ^= p2 */
24 		v0 = veorq_u64(vld1q_u64(dp1 +  0), vld1q_u64(dp2 +  0));
25 		v1 = veorq_u64(vld1q_u64(dp1 +  2), vld1q_u64(dp2 +  2));
26 		v2 = veorq_u64(vld1q_u64(dp1 +  4), vld1q_u64(dp2 +  4));
27 		v3 = veorq_u64(vld1q_u64(dp1 +  6), vld1q_u64(dp2 +  6));
28 
29 		/* store */
30 		vst1q_u64(dp1 +  0, v0);
31 		vst1q_u64(dp1 +  2, v1);
32 		vst1q_u64(dp1 +  4, v2);
33 		vst1q_u64(dp1 +  6, v3);
34 
35 		dp1 += 8;
36 		dp2 += 8;
37 	} while (--lines > 0);
38 }
39 
40 void xor_arm64_neon_3(unsigned long bytes, unsigned long * __restrict p1,
41 	const unsigned long * __restrict p2,
42 	const unsigned long * __restrict p3)
43 {
44 	uint64_t *dp1 = (uint64_t *)p1;
45 	uint64_t *dp2 = (uint64_t *)p2;
46 	uint64_t *dp3 = (uint64_t *)p3;
47 
48 	register uint64x2_t v0, v1, v2, v3;
49 	long lines = bytes / (sizeof(uint64x2_t) * 4);
50 
51 	do {
52 		/* p1 ^= p2 */
53 		v0 = veorq_u64(vld1q_u64(dp1 +  0), vld1q_u64(dp2 +  0));
54 		v1 = veorq_u64(vld1q_u64(dp1 +  2), vld1q_u64(dp2 +  2));
55 		v2 = veorq_u64(vld1q_u64(dp1 +  4), vld1q_u64(dp2 +  4));
56 		v3 = veorq_u64(vld1q_u64(dp1 +  6), vld1q_u64(dp2 +  6));
57 
58 		/* p1 ^= p3 */
59 		v0 = veorq_u64(v0, vld1q_u64(dp3 +  0));
60 		v1 = veorq_u64(v1, vld1q_u64(dp3 +  2));
61 		v2 = veorq_u64(v2, vld1q_u64(dp3 +  4));
62 		v3 = veorq_u64(v3, vld1q_u64(dp3 +  6));
63 
64 		/* store */
65 		vst1q_u64(dp1 +  0, v0);
66 		vst1q_u64(dp1 +  2, v1);
67 		vst1q_u64(dp1 +  4, v2);
68 		vst1q_u64(dp1 +  6, v3);
69 
70 		dp1 += 8;
71 		dp2 += 8;
72 		dp3 += 8;
73 	} while (--lines > 0);
74 }
75 
76 void xor_arm64_neon_4(unsigned long bytes, unsigned long * __restrict p1,
77 	const unsigned long * __restrict p2,
78 	const unsigned long * __restrict p3,
79 	const unsigned long * __restrict p4)
80 {
81 	uint64_t *dp1 = (uint64_t *)p1;
82 	uint64_t *dp2 = (uint64_t *)p2;
83 	uint64_t *dp3 = (uint64_t *)p3;
84 	uint64_t *dp4 = (uint64_t *)p4;
85 
86 	register uint64x2_t v0, v1, v2, v3;
87 	long lines = bytes / (sizeof(uint64x2_t) * 4);
88 
89 	do {
90 		/* p1 ^= p2 */
91 		v0 = veorq_u64(vld1q_u64(dp1 +  0), vld1q_u64(dp2 +  0));
92 		v1 = veorq_u64(vld1q_u64(dp1 +  2), vld1q_u64(dp2 +  2));
93 		v2 = veorq_u64(vld1q_u64(dp1 +  4), vld1q_u64(dp2 +  4));
94 		v3 = veorq_u64(vld1q_u64(dp1 +  6), vld1q_u64(dp2 +  6));
95 
96 		/* p1 ^= p3 */
97 		v0 = veorq_u64(v0, vld1q_u64(dp3 +  0));
98 		v1 = veorq_u64(v1, vld1q_u64(dp3 +  2));
99 		v2 = veorq_u64(v2, vld1q_u64(dp3 +  4));
100 		v3 = veorq_u64(v3, vld1q_u64(dp3 +  6));
101 
102 		/* p1 ^= p4 */
103 		v0 = veorq_u64(v0, vld1q_u64(dp4 +  0));
104 		v1 = veorq_u64(v1, vld1q_u64(dp4 +  2));
105 		v2 = veorq_u64(v2, vld1q_u64(dp4 +  4));
106 		v3 = veorq_u64(v3, vld1q_u64(dp4 +  6));
107 
108 		/* store */
109 		vst1q_u64(dp1 +  0, v0);
110 		vst1q_u64(dp1 +  2, v1);
111 		vst1q_u64(dp1 +  4, v2);
112 		vst1q_u64(dp1 +  6, v3);
113 
114 		dp1 += 8;
115 		dp2 += 8;
116 		dp3 += 8;
117 		dp4 += 8;
118 	} while (--lines > 0);
119 }
120 
121 void xor_arm64_neon_5(unsigned long bytes, unsigned long * __restrict p1,
122 	const unsigned long * __restrict p2,
123 	const unsigned long * __restrict p3,
124 	const unsigned long * __restrict p4,
125 	const unsigned long * __restrict p5)
126 {
127 	uint64_t *dp1 = (uint64_t *)p1;
128 	uint64_t *dp2 = (uint64_t *)p2;
129 	uint64_t *dp3 = (uint64_t *)p3;
130 	uint64_t *dp4 = (uint64_t *)p4;
131 	uint64_t *dp5 = (uint64_t *)p5;
132 
133 	register uint64x2_t v0, v1, v2, v3;
134 	long lines = bytes / (sizeof(uint64x2_t) * 4);
135 
136 	do {
137 		/* p1 ^= p2 */
138 		v0 = veorq_u64(vld1q_u64(dp1 +  0), vld1q_u64(dp2 +  0));
139 		v1 = veorq_u64(vld1q_u64(dp1 +  2), vld1q_u64(dp2 +  2));
140 		v2 = veorq_u64(vld1q_u64(dp1 +  4), vld1q_u64(dp2 +  4));
141 		v3 = veorq_u64(vld1q_u64(dp1 +  6), vld1q_u64(dp2 +  6));
142 
143 		/* p1 ^= p3 */
144 		v0 = veorq_u64(v0, vld1q_u64(dp3 +  0));
145 		v1 = veorq_u64(v1, vld1q_u64(dp3 +  2));
146 		v2 = veorq_u64(v2, vld1q_u64(dp3 +  4));
147 		v3 = veorq_u64(v3, vld1q_u64(dp3 +  6));
148 
149 		/* p1 ^= p4 */
150 		v0 = veorq_u64(v0, vld1q_u64(dp4 +  0));
151 		v1 = veorq_u64(v1, vld1q_u64(dp4 +  2));
152 		v2 = veorq_u64(v2, vld1q_u64(dp4 +  4));
153 		v3 = veorq_u64(v3, vld1q_u64(dp4 +  6));
154 
155 		/* p1 ^= p5 */
156 		v0 = veorq_u64(v0, vld1q_u64(dp5 +  0));
157 		v1 = veorq_u64(v1, vld1q_u64(dp5 +  2));
158 		v2 = veorq_u64(v2, vld1q_u64(dp5 +  4));
159 		v3 = veorq_u64(v3, vld1q_u64(dp5 +  6));
160 
161 		/* store */
162 		vst1q_u64(dp1 +  0, v0);
163 		vst1q_u64(dp1 +  2, v1);
164 		vst1q_u64(dp1 +  4, v2);
165 		vst1q_u64(dp1 +  6, v3);
166 
167 		dp1 += 8;
168 		dp2 += 8;
169 		dp3 += 8;
170 		dp4 += 8;
171 		dp5 += 8;
172 	} while (--lines > 0);
173 }
174 
175 struct xor_block_template xor_block_inner_neon __ro_after_init = {
176 	.name	= "__inner_neon__",
177 	.do_2	= xor_arm64_neon_2,
178 	.do_3	= xor_arm64_neon_3,
179 	.do_4	= xor_arm64_neon_4,
180 	.do_5	= xor_arm64_neon_5,
181 };
182 EXPORT_SYMBOL(xor_block_inner_neon);
183 
184 static inline uint64x2_t eor3(uint64x2_t p, uint64x2_t q, uint64x2_t r)
185 {
186 	uint64x2_t res;
187 
188 	asm(ARM64_ASM_PREAMBLE ".arch_extension sha3\n"
189 	    "eor3 %0.16b, %1.16b, %2.16b, %3.16b"
190 	    : "=w"(res) : "w"(p), "w"(q), "w"(r));
191 	return res;
192 }
193 
194 static void xor_arm64_eor3_3(unsigned long bytes,
195 	unsigned long * __restrict p1,
196 	const unsigned long * __restrict p2,
197 	const unsigned long * __restrict p3)
198 {
199 	uint64_t *dp1 = (uint64_t *)p1;
200 	uint64_t *dp2 = (uint64_t *)p2;
201 	uint64_t *dp3 = (uint64_t *)p3;
202 
203 	register uint64x2_t v0, v1, v2, v3;
204 	long lines = bytes / (sizeof(uint64x2_t) * 4);
205 
206 	do {
207 		/* p1 ^= p2 ^ p3 */
208 		v0 = eor3(vld1q_u64(dp1 + 0), vld1q_u64(dp2 + 0),
209 			  vld1q_u64(dp3 + 0));
210 		v1 = eor3(vld1q_u64(dp1 + 2), vld1q_u64(dp2 + 2),
211 			  vld1q_u64(dp3 + 2));
212 		v2 = eor3(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4),
213 			  vld1q_u64(dp3 + 4));
214 		v3 = eor3(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6),
215 			  vld1q_u64(dp3 + 6));
216 
217 		/* store */
218 		vst1q_u64(dp1 + 0, v0);
219 		vst1q_u64(dp1 + 2, v1);
220 		vst1q_u64(dp1 + 4, v2);
221 		vst1q_u64(dp1 + 6, v3);
222 
223 		dp1 += 8;
224 		dp2 += 8;
225 		dp3 += 8;
226 	} while (--lines > 0);
227 }
228 
229 static void xor_arm64_eor3_4(unsigned long bytes,
230 	unsigned long * __restrict p1,
231 	const unsigned long * __restrict p2,
232 	const unsigned long * __restrict p3,
233 	const unsigned long * __restrict p4)
234 {
235 	uint64_t *dp1 = (uint64_t *)p1;
236 	uint64_t *dp2 = (uint64_t *)p2;
237 	uint64_t *dp3 = (uint64_t *)p3;
238 	uint64_t *dp4 = (uint64_t *)p4;
239 
240 	register uint64x2_t v0, v1, v2, v3;
241 	long lines = bytes / (sizeof(uint64x2_t) * 4);
242 
243 	do {
244 		/* p1 ^= p2 ^ p3 */
245 		v0 = eor3(vld1q_u64(dp1 + 0), vld1q_u64(dp2 + 0),
246 			  vld1q_u64(dp3 + 0));
247 		v1 = eor3(vld1q_u64(dp1 + 2), vld1q_u64(dp2 + 2),
248 			  vld1q_u64(dp3 + 2));
249 		v2 = eor3(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4),
250 			  vld1q_u64(dp3 + 4));
251 		v3 = eor3(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6),
252 			  vld1q_u64(dp3 + 6));
253 
254 		/* p1 ^= p4 */
255 		v0 = veorq_u64(v0, vld1q_u64(dp4 + 0));
256 		v1 = veorq_u64(v1, vld1q_u64(dp4 + 2));
257 		v2 = veorq_u64(v2, vld1q_u64(dp4 + 4));
258 		v3 = veorq_u64(v3, vld1q_u64(dp4 + 6));
259 
260 		/* store */
261 		vst1q_u64(dp1 + 0, v0);
262 		vst1q_u64(dp1 + 2, v1);
263 		vst1q_u64(dp1 + 4, v2);
264 		vst1q_u64(dp1 + 6, v3);
265 
266 		dp1 += 8;
267 		dp2 += 8;
268 		dp3 += 8;
269 		dp4 += 8;
270 	} while (--lines > 0);
271 }
272 
273 static void xor_arm64_eor3_5(unsigned long bytes,
274 	unsigned long * __restrict p1,
275 	const unsigned long * __restrict p2,
276 	const unsigned long * __restrict p3,
277 	const unsigned long * __restrict p4,
278 	const unsigned long * __restrict p5)
279 {
280 	uint64_t *dp1 = (uint64_t *)p1;
281 	uint64_t *dp2 = (uint64_t *)p2;
282 	uint64_t *dp3 = (uint64_t *)p3;
283 	uint64_t *dp4 = (uint64_t *)p4;
284 	uint64_t *dp5 = (uint64_t *)p5;
285 
286 	register uint64x2_t v0, v1, v2, v3;
287 	long lines = bytes / (sizeof(uint64x2_t) * 4);
288 
289 	do {
290 		/* p1 ^= p2 ^ p3 */
291 		v0 = eor3(vld1q_u64(dp1 + 0), vld1q_u64(dp2 + 0),
292 			  vld1q_u64(dp3 + 0));
293 		v1 = eor3(vld1q_u64(dp1 + 2), vld1q_u64(dp2 + 2),
294 			  vld1q_u64(dp3 + 2));
295 		v2 = eor3(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4),
296 			  vld1q_u64(dp3 + 4));
297 		v3 = eor3(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6),
298 			  vld1q_u64(dp3 + 6));
299 
300 		/* p1 ^= p4 ^ p5 */
301 		v0 = eor3(v0, vld1q_u64(dp4 + 0), vld1q_u64(dp5 + 0));
302 		v1 = eor3(v1, vld1q_u64(dp4 + 2), vld1q_u64(dp5 + 2));
303 		v2 = eor3(v2, vld1q_u64(dp4 + 4), vld1q_u64(dp5 + 4));
304 		v3 = eor3(v3, vld1q_u64(dp4 + 6), vld1q_u64(dp5 + 6));
305 
306 		/* store */
307 		vst1q_u64(dp1 + 0, v0);
308 		vst1q_u64(dp1 + 2, v1);
309 		vst1q_u64(dp1 + 4, v2);
310 		vst1q_u64(dp1 + 6, v3);
311 
312 		dp1 += 8;
313 		dp2 += 8;
314 		dp3 += 8;
315 		dp4 += 8;
316 		dp5 += 8;
317 	} while (--lines > 0);
318 }
319 
320 static int __init xor_neon_init(void)
321 {
322 	if (IS_ENABLED(CONFIG_AS_HAS_SHA3) && cpu_have_named_feature(SHA3)) {
323 		xor_block_inner_neon.do_3 = xor_arm64_eor3_3;
324 		xor_block_inner_neon.do_4 = xor_arm64_eor3_4;
325 		xor_block_inner_neon.do_5 = xor_arm64_eor3_5;
326 	}
327 	return 0;
328 }
329 module_init(xor_neon_init);
330 
331 static void __exit xor_neon_exit(void)
332 {
333 }
334 module_exit(xor_neon_exit);
335 
336 MODULE_AUTHOR("Jackie Liu <liuyun01@kylinos.cn>");
337 MODULE_DESCRIPTION("ARMv8 XOR Extensions");
338 MODULE_LICENSE("GPL");
339