1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3 * This file and its contents are supplied under the terms of the
4 * Common Development and Distribution License ("CDDL"), version 1.0.
5 * You may only use this file in accordance with the terms of version
6 * 1.0 of the CDDL.
7 *
8 * A full copy of the text of the CDDL should have accompanied this
9 * source. A copy of the CDDL is also available via the Internet at
10 * https://opensource.org/license/CDDL-1.0.
11 */
12 /*
13 * Copyright (C) 2016 Gvozden Nešković. All rights reserved.
14 */
15
16 #include <sys/isa_defs.h>
17
18 #if defined(__x86_64) && HAVE_SIMD(SSE2)
19
20 #include <sys/types.h>
21 #include <sys/simd.h>
22 #include <sys/debug.h>
23
24 #ifdef __linux__
25 #define __asm __asm__ __volatile__
26 #endif
27
28 #define _REG_CNT(_0, _1, _2, _3, _4, _5, _6, _7, N, ...) N
29 #define REG_CNT(r...) _REG_CNT(r, 8, 7, 6, 5, 4, 3, 2, 1)
30
31 #define VR0_(REG, ...) "xmm"#REG
32 #define VR1_(_1, REG, ...) "xmm"#REG
33 #define VR2_(_1, _2, REG, ...) "xmm"#REG
34 #define VR3_(_1, _2, _3, REG, ...) "xmm"#REG
35 #define VR4_(_1, _2, _3, _4, REG, ...) "xmm"#REG
36 #define VR5_(_1, _2, _3, _4, _5, REG, ...) "xmm"#REG
37 #define VR6_(_1, _2, _3, _4, _5, _6, REG, ...) "xmm"#REG
38 #define VR7_(_1, _2, _3, _4, _5, _6, _7, REG, ...) "xmm"#REG
39
40 #define VR0(r...) VR0_(r, 1, 2, 3, 4, 5, 6)
41 #define VR1(r...) VR1_(r, 1, 2, 3, 4, 5, 6)
42 #define VR2(r...) VR2_(r, 1, 2, 3, 4, 5, 6)
43 #define VR3(r...) VR3_(r, 1, 2, 3, 4, 5, 6)
44 #define VR4(r...) VR4_(r, 1, 2, 3, 4, 5, 6)
45 #define VR5(r...) VR5_(r, 1, 2, 3, 4, 5, 6)
46 #define VR6(r...) VR6_(r, 1, 2, 3, 4, 5, 6)
47 #define VR7(r...) VR7_(r, 1, 2, 3, 4, 5, 6)
48
49 #define ELEM_SIZE 16
50
51 typedef struct v {
52 uint8_t b[ELEM_SIZE] __attribute__((aligned(ELEM_SIZE)));
53 } v_t;
54
55 #define XOR_ACC(src, r...) \
56 { \
57 switch (REG_CNT(r)) { \
58 case 4: \
59 __asm( \
60 "pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
61 "pxor 0x10(%[SRC]), %%" VR1(r) "\n" \
62 "pxor 0x20(%[SRC]), %%" VR2(r) "\n" \
63 "pxor 0x30(%[SRC]), %%" VR3(r) "\n" \
64 : : [SRC] "r" (src)); \
65 break; \
66 case 2: \
67 __asm( \
68 "pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
69 "pxor 0x10(%[SRC]), %%" VR1(r) "\n" \
70 : : [SRC] "r" (src)); \
71 break; \
72 case 1: \
73 __asm("pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
74 : : [SRC] "r" (src)); \
75 break; \
76 } \
77 }
78
79 #define XOR(r...) \
80 { \
81 switch (REG_CNT(r)) { \
82 case 8: \
83 __asm( \
84 "pxor %" VR0(r) ", %" VR4(r) "\n" \
85 "pxor %" VR1(r) ", %" VR5(r) "\n" \
86 "pxor %" VR2(r) ", %" VR6(r) "\n" \
87 "pxor %" VR3(r) ", %" VR7(r)); \
88 break; \
89 case 4: \
90 __asm( \
91 "pxor %" VR0(r) ", %" VR2(r) "\n" \
92 "pxor %" VR1(r) ", %" VR3(r)); \
93 break; \
94 case 2: \
95 __asm( \
96 "pxor %" VR0(r) ", %" VR1(r)); \
97 break; \
98 } \
99 }
100
101 #define ZERO(r...) XOR(r, r)
102
103 #define COPY(r...) \
104 { \
105 switch (REG_CNT(r)) { \
106 case 8: \
107 __asm( \
108 "movdqa %" VR0(r) ", %" VR4(r) "\n" \
109 "movdqa %" VR1(r) ", %" VR5(r) "\n" \
110 "movdqa %" VR2(r) ", %" VR6(r) "\n" \
111 "movdqa %" VR3(r) ", %" VR7(r)); \
112 break; \
113 case 4: \
114 __asm( \
115 "movdqa %" VR0(r) ", %" VR2(r) "\n" \
116 "movdqa %" VR1(r) ", %" VR3(r)); \
117 break; \
118 case 2: \
119 __asm( \
120 "movdqa %" VR0(r) ", %" VR1(r)); \
121 break; \
122 default: \
123 VERIFY(0); \
124 } \
125 }
126
127 #define LOAD(src, r...) \
128 { \
129 switch (REG_CNT(r)) { \
130 case 4: \
131 __asm( \
132 "movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
133 "movdqa 0x10(%[SRC]), %%" VR1(r) "\n" \
134 "movdqa 0x20(%[SRC]), %%" VR2(r) "\n" \
135 "movdqa 0x30(%[SRC]), %%" VR3(r) "\n" \
136 : : [SRC] "r" (src)); \
137 break; \
138 case 2: \
139 __asm( \
140 "movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
141 "movdqa 0x10(%[SRC]), %%" VR1(r) "\n" \
142 : : [SRC] "r" (src)); \
143 break; \
144 case 1: \
145 __asm( \
146 "movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
147 : : [SRC] "r" (src)); \
148 break; \
149 } \
150 }
151
152 #define STORE(dst, r...) \
153 { \
154 switch (REG_CNT(r)) { \
155 case 4: \
156 __asm( \
157 "movdqa %%" VR0(r)", 0x00(%[DST])\n" \
158 "movdqa %%" VR1(r)", 0x10(%[DST])\n" \
159 "movdqa %%" VR2(r)", 0x20(%[DST])\n" \
160 "movdqa %%" VR3(r)", 0x30(%[DST])\n" \
161 : : [DST] "r" (dst)); \
162 break; \
163 case 2: \
164 __asm( \
165 "movdqa %%" VR0(r)", 0x00(%[DST])\n" \
166 "movdqa %%" VR1(r)", 0x10(%[DST])\n" \
167 : : [DST] "r" (dst)); \
168 break; \
169 case 1: \
170 __asm( \
171 "movdqa %%" VR0(r)", 0x00(%[DST])\n" \
172 : : [DST] "r" (dst)); \
173 break; \
174 default: \
175 VERIFY(0); \
176 } \
177 }
178
179 #define MUL2_SETUP() \
180 { \
181 __asm( \
182 "movd %[mask], %%xmm15\n" \
183 "pshufd $0x0, %%xmm15, %%xmm15\n" \
184 : : [mask] "r" (0x1d1d1d1d)); \
185 }
186
187 #define _MUL2_x1(a0) \
188 { \
189 __asm( \
190 "pxor %xmm14, %xmm14\n" \
191 "pcmpgtb %" a0", %xmm14\n" \
192 "pand %xmm15, %xmm14\n" \
193 "paddb %" a0", %" a0 "\n" \
194 "pxor %xmm14, %" a0); \
195 }
196
197 #define _MUL2_x2(a0, a1) \
198 { \
199 __asm( \
200 "pxor %xmm14, %xmm14\n" \
201 "pxor %xmm13, %xmm13\n" \
202 "pcmpgtb %" a0", %xmm14\n" \
203 "pcmpgtb %" a1", %xmm13\n" \
204 "pand %xmm15, %xmm14\n" \
205 "pand %xmm15, %xmm13\n" \
206 "paddb %" a0", %" a0 "\n" \
207 "paddb %" a1", %" a1 "\n" \
208 "pxor %xmm14, %" a0 "\n" \
209 "pxor %xmm13, %" a1); \
210 }
211
212 #define MUL2(r...) \
213 { \
214 switch (REG_CNT(r)) { \
215 case 4: \
216 _MUL2_x2(VR0(r), VR1(r)); \
217 _MUL2_x2(VR2(r), VR3(r)); \
218 break; \
219 case 2: \
220 _MUL2_x2(VR0(r), VR1(r)); \
221 break; \
222 case 1: \
223 _MUL2_x1(VR0(r)); \
224 break; \
225 } \
226 }
227
228 #define MUL4(r...) \
229 { \
230 MUL2(r); \
231 MUL2(r); \
232 }
233
234 /* General multiplication by adding powers of two */
235
236 #define _MUL_PARAM(x, in, acc) \
237 { \
238 if (x & 0x01) { COPY(in, acc); } else { ZERO(acc); } \
239 if (x & 0xfe) { MUL2(in); } \
240 if (x & 0x02) { XOR(in, acc); } \
241 if (x & 0xfc) { MUL2(in); } \
242 if (x & 0x04) { XOR(in, acc); } \
243 if (x & 0xf8) { MUL2(in); } \
244 if (x & 0x08) { XOR(in, acc); } \
245 if (x & 0xf0) { MUL2(in); } \
246 if (x & 0x10) { XOR(in, acc); } \
247 if (x & 0xe0) { MUL2(in); } \
248 if (x & 0x20) { XOR(in, acc); } \
249 if (x & 0xc0) { MUL2(in); } \
250 if (x & 0x40) { XOR(in, acc); } \
251 if (x & 0x80) { MUL2(in); XOR(in, acc); } \
252 }
253
254 #define _mul_x1_in 11
255 #define _mul_x1_acc 12
256
257 #define MUL_x1_DEFINE(x) \
258 static void \
259 mul_x1_ ## x(void) { _MUL_PARAM(x, _mul_x1_in, _mul_x1_acc); }
260
261 #define _mul_x2_in 9, 10
262 #define _mul_x2_acc 11, 12
263
264 #define MUL_x2_DEFINE(x) \
265 static void \
266 mul_x2_ ## x(void) { _MUL_PARAM(x, _mul_x2_in, _mul_x2_acc); }
267
268 MUL_x1_DEFINE(0); MUL_x1_DEFINE(1); MUL_x1_DEFINE(2); MUL_x1_DEFINE(3);
269 MUL_x1_DEFINE(4); MUL_x1_DEFINE(5); MUL_x1_DEFINE(6); MUL_x1_DEFINE(7);
270 MUL_x1_DEFINE(8); MUL_x1_DEFINE(9); MUL_x1_DEFINE(10); MUL_x1_DEFINE(11);
271 MUL_x1_DEFINE(12); MUL_x1_DEFINE(13); MUL_x1_DEFINE(14); MUL_x1_DEFINE(15);
272 MUL_x1_DEFINE(16); MUL_x1_DEFINE(17); MUL_x1_DEFINE(18); MUL_x1_DEFINE(19);
273 MUL_x1_DEFINE(20); MUL_x1_DEFINE(21); MUL_x1_DEFINE(22); MUL_x1_DEFINE(23);
274 MUL_x1_DEFINE(24); MUL_x1_DEFINE(25); MUL_x1_DEFINE(26); MUL_x1_DEFINE(27);
275 MUL_x1_DEFINE(28); MUL_x1_DEFINE(29); MUL_x1_DEFINE(30); MUL_x1_DEFINE(31);
276 MUL_x1_DEFINE(32); MUL_x1_DEFINE(33); MUL_x1_DEFINE(34); MUL_x1_DEFINE(35);
277 MUL_x1_DEFINE(36); MUL_x1_DEFINE(37); MUL_x1_DEFINE(38); MUL_x1_DEFINE(39);
278 MUL_x1_DEFINE(40); MUL_x1_DEFINE(41); MUL_x1_DEFINE(42); MUL_x1_DEFINE(43);
279 MUL_x1_DEFINE(44); MUL_x1_DEFINE(45); MUL_x1_DEFINE(46); MUL_x1_DEFINE(47);
280 MUL_x1_DEFINE(48); MUL_x1_DEFINE(49); MUL_x1_DEFINE(50); MUL_x1_DEFINE(51);
281 MUL_x1_DEFINE(52); MUL_x1_DEFINE(53); MUL_x1_DEFINE(54); MUL_x1_DEFINE(55);
282 MUL_x1_DEFINE(56); MUL_x1_DEFINE(57); MUL_x1_DEFINE(58); MUL_x1_DEFINE(59);
283 MUL_x1_DEFINE(60); MUL_x1_DEFINE(61); MUL_x1_DEFINE(62); MUL_x1_DEFINE(63);
284 MUL_x1_DEFINE(64); MUL_x1_DEFINE(65); MUL_x1_DEFINE(66); MUL_x1_DEFINE(67);
285 MUL_x1_DEFINE(68); MUL_x1_DEFINE(69); MUL_x1_DEFINE(70); MUL_x1_DEFINE(71);
286 MUL_x1_DEFINE(72); MUL_x1_DEFINE(73); MUL_x1_DEFINE(74); MUL_x1_DEFINE(75);
287 MUL_x1_DEFINE(76); MUL_x1_DEFINE(77); MUL_x1_DEFINE(78); MUL_x1_DEFINE(79);
288 MUL_x1_DEFINE(80); MUL_x1_DEFINE(81); MUL_x1_DEFINE(82); MUL_x1_DEFINE(83);
289 MUL_x1_DEFINE(84); MUL_x1_DEFINE(85); MUL_x1_DEFINE(86); MUL_x1_DEFINE(87);
290 MUL_x1_DEFINE(88); MUL_x1_DEFINE(89); MUL_x1_DEFINE(90); MUL_x1_DEFINE(91);
291 MUL_x1_DEFINE(92); MUL_x1_DEFINE(93); MUL_x1_DEFINE(94); MUL_x1_DEFINE(95);
292 MUL_x1_DEFINE(96); MUL_x1_DEFINE(97); MUL_x1_DEFINE(98); MUL_x1_DEFINE(99);
293 MUL_x1_DEFINE(100); MUL_x1_DEFINE(101); MUL_x1_DEFINE(102); MUL_x1_DEFINE(103);
294 MUL_x1_DEFINE(104); MUL_x1_DEFINE(105); MUL_x1_DEFINE(106); MUL_x1_DEFINE(107);
295 MUL_x1_DEFINE(108); MUL_x1_DEFINE(109); MUL_x1_DEFINE(110); MUL_x1_DEFINE(111);
296 MUL_x1_DEFINE(112); MUL_x1_DEFINE(113); MUL_x1_DEFINE(114); MUL_x1_DEFINE(115);
297 MUL_x1_DEFINE(116); MUL_x1_DEFINE(117); MUL_x1_DEFINE(118); MUL_x1_DEFINE(119);
298 MUL_x1_DEFINE(120); MUL_x1_DEFINE(121); MUL_x1_DEFINE(122); MUL_x1_DEFINE(123);
299 MUL_x1_DEFINE(124); MUL_x1_DEFINE(125); MUL_x1_DEFINE(126); MUL_x1_DEFINE(127);
300 MUL_x1_DEFINE(128); MUL_x1_DEFINE(129); MUL_x1_DEFINE(130); MUL_x1_DEFINE(131);
301 MUL_x1_DEFINE(132); MUL_x1_DEFINE(133); MUL_x1_DEFINE(134); MUL_x1_DEFINE(135);
302 MUL_x1_DEFINE(136); MUL_x1_DEFINE(137); MUL_x1_DEFINE(138); MUL_x1_DEFINE(139);
303 MUL_x1_DEFINE(140); MUL_x1_DEFINE(141); MUL_x1_DEFINE(142); MUL_x1_DEFINE(143);
304 MUL_x1_DEFINE(144); MUL_x1_DEFINE(145); MUL_x1_DEFINE(146); MUL_x1_DEFINE(147);
305 MUL_x1_DEFINE(148); MUL_x1_DEFINE(149); MUL_x1_DEFINE(150); MUL_x1_DEFINE(151);
306 MUL_x1_DEFINE(152); MUL_x1_DEFINE(153); MUL_x1_DEFINE(154); MUL_x1_DEFINE(155);
307 MUL_x1_DEFINE(156); MUL_x1_DEFINE(157); MUL_x1_DEFINE(158); MUL_x1_DEFINE(159);
308 MUL_x1_DEFINE(160); MUL_x1_DEFINE(161); MUL_x1_DEFINE(162); MUL_x1_DEFINE(163);
309 MUL_x1_DEFINE(164); MUL_x1_DEFINE(165); MUL_x1_DEFINE(166); MUL_x1_DEFINE(167);
310 MUL_x1_DEFINE(168); MUL_x1_DEFINE(169); MUL_x1_DEFINE(170); MUL_x1_DEFINE(171);
311 MUL_x1_DEFINE(172); MUL_x1_DEFINE(173); MUL_x1_DEFINE(174); MUL_x1_DEFINE(175);
312 MUL_x1_DEFINE(176); MUL_x1_DEFINE(177); MUL_x1_DEFINE(178); MUL_x1_DEFINE(179);
313 MUL_x1_DEFINE(180); MUL_x1_DEFINE(181); MUL_x1_DEFINE(182); MUL_x1_DEFINE(183);
314 MUL_x1_DEFINE(184); MUL_x1_DEFINE(185); MUL_x1_DEFINE(186); MUL_x1_DEFINE(187);
315 MUL_x1_DEFINE(188); MUL_x1_DEFINE(189); MUL_x1_DEFINE(190); MUL_x1_DEFINE(191);
316 MUL_x1_DEFINE(192); MUL_x1_DEFINE(193); MUL_x1_DEFINE(194); MUL_x1_DEFINE(195);
317 MUL_x1_DEFINE(196); MUL_x1_DEFINE(197); MUL_x1_DEFINE(198); MUL_x1_DEFINE(199);
318 MUL_x1_DEFINE(200); MUL_x1_DEFINE(201); MUL_x1_DEFINE(202); MUL_x1_DEFINE(203);
319 MUL_x1_DEFINE(204); MUL_x1_DEFINE(205); MUL_x1_DEFINE(206); MUL_x1_DEFINE(207);
320 MUL_x1_DEFINE(208); MUL_x1_DEFINE(209); MUL_x1_DEFINE(210); MUL_x1_DEFINE(211);
321 MUL_x1_DEFINE(212); MUL_x1_DEFINE(213); MUL_x1_DEFINE(214); MUL_x1_DEFINE(215);
322 MUL_x1_DEFINE(216); MUL_x1_DEFINE(217); MUL_x1_DEFINE(218); MUL_x1_DEFINE(219);
323 MUL_x1_DEFINE(220); MUL_x1_DEFINE(221); MUL_x1_DEFINE(222); MUL_x1_DEFINE(223);
324 MUL_x1_DEFINE(224); MUL_x1_DEFINE(225); MUL_x1_DEFINE(226); MUL_x1_DEFINE(227);
325 MUL_x1_DEFINE(228); MUL_x1_DEFINE(229); MUL_x1_DEFINE(230); MUL_x1_DEFINE(231);
326 MUL_x1_DEFINE(232); MUL_x1_DEFINE(233); MUL_x1_DEFINE(234); MUL_x1_DEFINE(235);
327 MUL_x1_DEFINE(236); MUL_x1_DEFINE(237); MUL_x1_DEFINE(238); MUL_x1_DEFINE(239);
328 MUL_x1_DEFINE(240); MUL_x1_DEFINE(241); MUL_x1_DEFINE(242); MUL_x1_DEFINE(243);
329 MUL_x1_DEFINE(244); MUL_x1_DEFINE(245); MUL_x1_DEFINE(246); MUL_x1_DEFINE(247);
330 MUL_x1_DEFINE(248); MUL_x1_DEFINE(249); MUL_x1_DEFINE(250); MUL_x1_DEFINE(251);
331 MUL_x1_DEFINE(252); MUL_x1_DEFINE(253); MUL_x1_DEFINE(254); MUL_x1_DEFINE(255);
332
333 MUL_x2_DEFINE(0); MUL_x2_DEFINE(1); MUL_x2_DEFINE(2); MUL_x2_DEFINE(3);
334 MUL_x2_DEFINE(4); MUL_x2_DEFINE(5); MUL_x2_DEFINE(6); MUL_x2_DEFINE(7);
335 MUL_x2_DEFINE(8); MUL_x2_DEFINE(9); MUL_x2_DEFINE(10); MUL_x2_DEFINE(11);
336 MUL_x2_DEFINE(12); MUL_x2_DEFINE(13); MUL_x2_DEFINE(14); MUL_x2_DEFINE(15);
337 MUL_x2_DEFINE(16); MUL_x2_DEFINE(17); MUL_x2_DEFINE(18); MUL_x2_DEFINE(19);
338 MUL_x2_DEFINE(20); MUL_x2_DEFINE(21); MUL_x2_DEFINE(22); MUL_x2_DEFINE(23);
339 MUL_x2_DEFINE(24); MUL_x2_DEFINE(25); MUL_x2_DEFINE(26); MUL_x2_DEFINE(27);
340 MUL_x2_DEFINE(28); MUL_x2_DEFINE(29); MUL_x2_DEFINE(30); MUL_x2_DEFINE(31);
341 MUL_x2_DEFINE(32); MUL_x2_DEFINE(33); MUL_x2_DEFINE(34); MUL_x2_DEFINE(35);
342 MUL_x2_DEFINE(36); MUL_x2_DEFINE(37); MUL_x2_DEFINE(38); MUL_x2_DEFINE(39);
343 MUL_x2_DEFINE(40); MUL_x2_DEFINE(41); MUL_x2_DEFINE(42); MUL_x2_DEFINE(43);
344 MUL_x2_DEFINE(44); MUL_x2_DEFINE(45); MUL_x2_DEFINE(46); MUL_x2_DEFINE(47);
345 MUL_x2_DEFINE(48); MUL_x2_DEFINE(49); MUL_x2_DEFINE(50); MUL_x2_DEFINE(51);
346 MUL_x2_DEFINE(52); MUL_x2_DEFINE(53); MUL_x2_DEFINE(54); MUL_x2_DEFINE(55);
347 MUL_x2_DEFINE(56); MUL_x2_DEFINE(57); MUL_x2_DEFINE(58); MUL_x2_DEFINE(59);
348 MUL_x2_DEFINE(60); MUL_x2_DEFINE(61); MUL_x2_DEFINE(62); MUL_x2_DEFINE(63);
349 MUL_x2_DEFINE(64); MUL_x2_DEFINE(65); MUL_x2_DEFINE(66); MUL_x2_DEFINE(67);
350 MUL_x2_DEFINE(68); MUL_x2_DEFINE(69); MUL_x2_DEFINE(70); MUL_x2_DEFINE(71);
351 MUL_x2_DEFINE(72); MUL_x2_DEFINE(73); MUL_x2_DEFINE(74); MUL_x2_DEFINE(75);
352 MUL_x2_DEFINE(76); MUL_x2_DEFINE(77); MUL_x2_DEFINE(78); MUL_x2_DEFINE(79);
353 MUL_x2_DEFINE(80); MUL_x2_DEFINE(81); MUL_x2_DEFINE(82); MUL_x2_DEFINE(83);
354 MUL_x2_DEFINE(84); MUL_x2_DEFINE(85); MUL_x2_DEFINE(86); MUL_x2_DEFINE(87);
355 MUL_x2_DEFINE(88); MUL_x2_DEFINE(89); MUL_x2_DEFINE(90); MUL_x2_DEFINE(91);
356 MUL_x2_DEFINE(92); MUL_x2_DEFINE(93); MUL_x2_DEFINE(94); MUL_x2_DEFINE(95);
357 MUL_x2_DEFINE(96); MUL_x2_DEFINE(97); MUL_x2_DEFINE(98); MUL_x2_DEFINE(99);
358 MUL_x2_DEFINE(100); MUL_x2_DEFINE(101); MUL_x2_DEFINE(102); MUL_x2_DEFINE(103);
359 MUL_x2_DEFINE(104); MUL_x2_DEFINE(105); MUL_x2_DEFINE(106); MUL_x2_DEFINE(107);
360 MUL_x2_DEFINE(108); MUL_x2_DEFINE(109); MUL_x2_DEFINE(110); MUL_x2_DEFINE(111);
361 MUL_x2_DEFINE(112); MUL_x2_DEFINE(113); MUL_x2_DEFINE(114); MUL_x2_DEFINE(115);
362 MUL_x2_DEFINE(116); MUL_x2_DEFINE(117); MUL_x2_DEFINE(118); MUL_x2_DEFINE(119);
363 MUL_x2_DEFINE(120); MUL_x2_DEFINE(121); MUL_x2_DEFINE(122); MUL_x2_DEFINE(123);
364 MUL_x2_DEFINE(124); MUL_x2_DEFINE(125); MUL_x2_DEFINE(126); MUL_x2_DEFINE(127);
365 MUL_x2_DEFINE(128); MUL_x2_DEFINE(129); MUL_x2_DEFINE(130); MUL_x2_DEFINE(131);
366 MUL_x2_DEFINE(132); MUL_x2_DEFINE(133); MUL_x2_DEFINE(134); MUL_x2_DEFINE(135);
367 MUL_x2_DEFINE(136); MUL_x2_DEFINE(137); MUL_x2_DEFINE(138); MUL_x2_DEFINE(139);
368 MUL_x2_DEFINE(140); MUL_x2_DEFINE(141); MUL_x2_DEFINE(142); MUL_x2_DEFINE(143);
369 MUL_x2_DEFINE(144); MUL_x2_DEFINE(145); MUL_x2_DEFINE(146); MUL_x2_DEFINE(147);
370 MUL_x2_DEFINE(148); MUL_x2_DEFINE(149); MUL_x2_DEFINE(150); MUL_x2_DEFINE(151);
371 MUL_x2_DEFINE(152); MUL_x2_DEFINE(153); MUL_x2_DEFINE(154); MUL_x2_DEFINE(155);
372 MUL_x2_DEFINE(156); MUL_x2_DEFINE(157); MUL_x2_DEFINE(158); MUL_x2_DEFINE(159);
373 MUL_x2_DEFINE(160); MUL_x2_DEFINE(161); MUL_x2_DEFINE(162); MUL_x2_DEFINE(163);
374 MUL_x2_DEFINE(164); MUL_x2_DEFINE(165); MUL_x2_DEFINE(166); MUL_x2_DEFINE(167);
375 MUL_x2_DEFINE(168); MUL_x2_DEFINE(169); MUL_x2_DEFINE(170); MUL_x2_DEFINE(171);
376 MUL_x2_DEFINE(172); MUL_x2_DEFINE(173); MUL_x2_DEFINE(174); MUL_x2_DEFINE(175);
377 MUL_x2_DEFINE(176); MUL_x2_DEFINE(177); MUL_x2_DEFINE(178); MUL_x2_DEFINE(179);
378 MUL_x2_DEFINE(180); MUL_x2_DEFINE(181); MUL_x2_DEFINE(182); MUL_x2_DEFINE(183);
379 MUL_x2_DEFINE(184); MUL_x2_DEFINE(185); MUL_x2_DEFINE(186); MUL_x2_DEFINE(187);
380 MUL_x2_DEFINE(188); MUL_x2_DEFINE(189); MUL_x2_DEFINE(190); MUL_x2_DEFINE(191);
381 MUL_x2_DEFINE(192); MUL_x2_DEFINE(193); MUL_x2_DEFINE(194); MUL_x2_DEFINE(195);
382 MUL_x2_DEFINE(196); MUL_x2_DEFINE(197); MUL_x2_DEFINE(198); MUL_x2_DEFINE(199);
383 MUL_x2_DEFINE(200); MUL_x2_DEFINE(201); MUL_x2_DEFINE(202); MUL_x2_DEFINE(203);
384 MUL_x2_DEFINE(204); MUL_x2_DEFINE(205); MUL_x2_DEFINE(206); MUL_x2_DEFINE(207);
385 MUL_x2_DEFINE(208); MUL_x2_DEFINE(209); MUL_x2_DEFINE(210); MUL_x2_DEFINE(211);
386 MUL_x2_DEFINE(212); MUL_x2_DEFINE(213); MUL_x2_DEFINE(214); MUL_x2_DEFINE(215);
387 MUL_x2_DEFINE(216); MUL_x2_DEFINE(217); MUL_x2_DEFINE(218); MUL_x2_DEFINE(219);
388 MUL_x2_DEFINE(220); MUL_x2_DEFINE(221); MUL_x2_DEFINE(222); MUL_x2_DEFINE(223);
389 MUL_x2_DEFINE(224); MUL_x2_DEFINE(225); MUL_x2_DEFINE(226); MUL_x2_DEFINE(227);
390 MUL_x2_DEFINE(228); MUL_x2_DEFINE(229); MUL_x2_DEFINE(230); MUL_x2_DEFINE(231);
391 MUL_x2_DEFINE(232); MUL_x2_DEFINE(233); MUL_x2_DEFINE(234); MUL_x2_DEFINE(235);
392 MUL_x2_DEFINE(236); MUL_x2_DEFINE(237); MUL_x2_DEFINE(238); MUL_x2_DEFINE(239);
393 MUL_x2_DEFINE(240); MUL_x2_DEFINE(241); MUL_x2_DEFINE(242); MUL_x2_DEFINE(243);
394 MUL_x2_DEFINE(244); MUL_x2_DEFINE(245); MUL_x2_DEFINE(246); MUL_x2_DEFINE(247);
395 MUL_x2_DEFINE(248); MUL_x2_DEFINE(249); MUL_x2_DEFINE(250); MUL_x2_DEFINE(251);
396 MUL_x2_DEFINE(252); MUL_x2_DEFINE(253); MUL_x2_DEFINE(254); MUL_x2_DEFINE(255);
397
398
399
400 typedef void (*mul_fn_ptr_t)(void);
401
402 static const mul_fn_ptr_t __attribute__((aligned(256)))
403 gf_x1_mul_fns[256] = {
404 mul_x1_0, mul_x1_1, mul_x1_2, mul_x1_3, mul_x1_4, mul_x1_5,
405 mul_x1_6, mul_x1_7, mul_x1_8, mul_x1_9, mul_x1_10, mul_x1_11,
406 mul_x1_12, mul_x1_13, mul_x1_14, mul_x1_15, mul_x1_16, mul_x1_17,
407 mul_x1_18, mul_x1_19, mul_x1_20, mul_x1_21, mul_x1_22, mul_x1_23,
408 mul_x1_24, mul_x1_25, mul_x1_26, mul_x1_27, mul_x1_28, mul_x1_29,
409 mul_x1_30, mul_x1_31, mul_x1_32, mul_x1_33, mul_x1_34, mul_x1_35,
410 mul_x1_36, mul_x1_37, mul_x1_38, mul_x1_39, mul_x1_40, mul_x1_41,
411 mul_x1_42, mul_x1_43, mul_x1_44, mul_x1_45, mul_x1_46, mul_x1_47,
412 mul_x1_48, mul_x1_49, mul_x1_50, mul_x1_51, mul_x1_52, mul_x1_53,
413 mul_x1_54, mul_x1_55, mul_x1_56, mul_x1_57, mul_x1_58, mul_x1_59,
414 mul_x1_60, mul_x1_61, mul_x1_62, mul_x1_63, mul_x1_64, mul_x1_65,
415 mul_x1_66, mul_x1_67, mul_x1_68, mul_x1_69, mul_x1_70, mul_x1_71,
416 mul_x1_72, mul_x1_73, mul_x1_74, mul_x1_75, mul_x1_76, mul_x1_77,
417 mul_x1_78, mul_x1_79, mul_x1_80, mul_x1_81, mul_x1_82, mul_x1_83,
418 mul_x1_84, mul_x1_85, mul_x1_86, mul_x1_87, mul_x1_88, mul_x1_89,
419 mul_x1_90, mul_x1_91, mul_x1_92, mul_x1_93, mul_x1_94, mul_x1_95,
420 mul_x1_96, mul_x1_97, mul_x1_98, mul_x1_99, mul_x1_100, mul_x1_101,
421 mul_x1_102, mul_x1_103, mul_x1_104, mul_x1_105, mul_x1_106, mul_x1_107,
422 mul_x1_108, mul_x1_109, mul_x1_110, mul_x1_111, mul_x1_112, mul_x1_113,
423 mul_x1_114, mul_x1_115, mul_x1_116, mul_x1_117, mul_x1_118, mul_x1_119,
424 mul_x1_120, mul_x1_121, mul_x1_122, mul_x1_123, mul_x1_124, mul_x1_125,
425 mul_x1_126, mul_x1_127, mul_x1_128, mul_x1_129, mul_x1_130, mul_x1_131,
426 mul_x1_132, mul_x1_133, mul_x1_134, mul_x1_135, mul_x1_136, mul_x1_137,
427 mul_x1_138, mul_x1_139, mul_x1_140, mul_x1_141, mul_x1_142, mul_x1_143,
428 mul_x1_144, mul_x1_145, mul_x1_146, mul_x1_147, mul_x1_148, mul_x1_149,
429 mul_x1_150, mul_x1_151, mul_x1_152, mul_x1_153, mul_x1_154, mul_x1_155,
430 mul_x1_156, mul_x1_157, mul_x1_158, mul_x1_159, mul_x1_160, mul_x1_161,
431 mul_x1_162, mul_x1_163, mul_x1_164, mul_x1_165, mul_x1_166, mul_x1_167,
432 mul_x1_168, mul_x1_169, mul_x1_170, mul_x1_171, mul_x1_172, mul_x1_173,
433 mul_x1_174, mul_x1_175, mul_x1_176, mul_x1_177, mul_x1_178, mul_x1_179,
434 mul_x1_180, mul_x1_181, mul_x1_182, mul_x1_183, mul_x1_184, mul_x1_185,
435 mul_x1_186, mul_x1_187, mul_x1_188, mul_x1_189, mul_x1_190, mul_x1_191,
436 mul_x1_192, mul_x1_193, mul_x1_194, mul_x1_195, mul_x1_196, mul_x1_197,
437 mul_x1_198, mul_x1_199, mul_x1_200, mul_x1_201, mul_x1_202, mul_x1_203,
438 mul_x1_204, mul_x1_205, mul_x1_206, mul_x1_207, mul_x1_208, mul_x1_209,
439 mul_x1_210, mul_x1_211, mul_x1_212, mul_x1_213, mul_x1_214, mul_x1_215,
440 mul_x1_216, mul_x1_217, mul_x1_218, mul_x1_219, mul_x1_220, mul_x1_221,
441 mul_x1_222, mul_x1_223, mul_x1_224, mul_x1_225, mul_x1_226, mul_x1_227,
442 mul_x1_228, mul_x1_229, mul_x1_230, mul_x1_231, mul_x1_232, mul_x1_233,
443 mul_x1_234, mul_x1_235, mul_x1_236, mul_x1_237, mul_x1_238, mul_x1_239,
444 mul_x1_240, mul_x1_241, mul_x1_242, mul_x1_243, mul_x1_244, mul_x1_245,
445 mul_x1_246, mul_x1_247, mul_x1_248, mul_x1_249, mul_x1_250, mul_x1_251,
446 mul_x1_252, mul_x1_253, mul_x1_254, mul_x1_255
447 };
448
449 static const mul_fn_ptr_t __attribute__((aligned(256)))
450 gf_x2_mul_fns[256] = {
451 mul_x2_0, mul_x2_1, mul_x2_2, mul_x2_3, mul_x2_4, mul_x2_5,
452 mul_x2_6, mul_x2_7, mul_x2_8, mul_x2_9, mul_x2_10, mul_x2_11,
453 mul_x2_12, mul_x2_13, mul_x2_14, mul_x2_15, mul_x2_16, mul_x2_17,
454 mul_x2_18, mul_x2_19, mul_x2_20, mul_x2_21, mul_x2_22, mul_x2_23,
455 mul_x2_24, mul_x2_25, mul_x2_26, mul_x2_27, mul_x2_28, mul_x2_29,
456 mul_x2_30, mul_x2_31, mul_x2_32, mul_x2_33, mul_x2_34, mul_x2_35,
457 mul_x2_36, mul_x2_37, mul_x2_38, mul_x2_39, mul_x2_40, mul_x2_41,
458 mul_x2_42, mul_x2_43, mul_x2_44, mul_x2_45, mul_x2_46, mul_x2_47,
459 mul_x2_48, mul_x2_49, mul_x2_50, mul_x2_51, mul_x2_52, mul_x2_53,
460 mul_x2_54, mul_x2_55, mul_x2_56, mul_x2_57, mul_x2_58, mul_x2_59,
461 mul_x2_60, mul_x2_61, mul_x2_62, mul_x2_63, mul_x2_64, mul_x2_65,
462 mul_x2_66, mul_x2_67, mul_x2_68, mul_x2_69, mul_x2_70, mul_x2_71,
463 mul_x2_72, mul_x2_73, mul_x2_74, mul_x2_75, mul_x2_76, mul_x2_77,
464 mul_x2_78, mul_x2_79, mul_x2_80, mul_x2_81, mul_x2_82, mul_x2_83,
465 mul_x2_84, mul_x2_85, mul_x2_86, mul_x2_87, mul_x2_88, mul_x2_89,
466 mul_x2_90, mul_x2_91, mul_x2_92, mul_x2_93, mul_x2_94, mul_x2_95,
467 mul_x2_96, mul_x2_97, mul_x2_98, mul_x2_99, mul_x2_100, mul_x2_101,
468 mul_x2_102, mul_x2_103, mul_x2_104, mul_x2_105, mul_x2_106, mul_x2_107,
469 mul_x2_108, mul_x2_109, mul_x2_110, mul_x2_111, mul_x2_112, mul_x2_113,
470 mul_x2_114, mul_x2_115, mul_x2_116, mul_x2_117, mul_x2_118, mul_x2_119,
471 mul_x2_120, mul_x2_121, mul_x2_122, mul_x2_123, mul_x2_124, mul_x2_125,
472 mul_x2_126, mul_x2_127, mul_x2_128, mul_x2_129, mul_x2_130, mul_x2_131,
473 mul_x2_132, mul_x2_133, mul_x2_134, mul_x2_135, mul_x2_136, mul_x2_137,
474 mul_x2_138, mul_x2_139, mul_x2_140, mul_x2_141, mul_x2_142, mul_x2_143,
475 mul_x2_144, mul_x2_145, mul_x2_146, mul_x2_147, mul_x2_148, mul_x2_149,
476 mul_x2_150, mul_x2_151, mul_x2_152, mul_x2_153, mul_x2_154, mul_x2_155,
477 mul_x2_156, mul_x2_157, mul_x2_158, mul_x2_159, mul_x2_160, mul_x2_161,
478 mul_x2_162, mul_x2_163, mul_x2_164, mul_x2_165, mul_x2_166, mul_x2_167,
479 mul_x2_168, mul_x2_169, mul_x2_170, mul_x2_171, mul_x2_172, mul_x2_173,
480 mul_x2_174, mul_x2_175, mul_x2_176, mul_x2_177, mul_x2_178, mul_x2_179,
481 mul_x2_180, mul_x2_181, mul_x2_182, mul_x2_183, mul_x2_184, mul_x2_185,
482 mul_x2_186, mul_x2_187, mul_x2_188, mul_x2_189, mul_x2_190, mul_x2_191,
483 mul_x2_192, mul_x2_193, mul_x2_194, mul_x2_195, mul_x2_196, mul_x2_197,
484 mul_x2_198, mul_x2_199, mul_x2_200, mul_x2_201, mul_x2_202, mul_x2_203,
485 mul_x2_204, mul_x2_205, mul_x2_206, mul_x2_207, mul_x2_208, mul_x2_209,
486 mul_x2_210, mul_x2_211, mul_x2_212, mul_x2_213, mul_x2_214, mul_x2_215,
487 mul_x2_216, mul_x2_217, mul_x2_218, mul_x2_219, mul_x2_220, mul_x2_221,
488 mul_x2_222, mul_x2_223, mul_x2_224, mul_x2_225, mul_x2_226, mul_x2_227,
489 mul_x2_228, mul_x2_229, mul_x2_230, mul_x2_231, mul_x2_232, mul_x2_233,
490 mul_x2_234, mul_x2_235, mul_x2_236, mul_x2_237, mul_x2_238, mul_x2_239,
491 mul_x2_240, mul_x2_241, mul_x2_242, mul_x2_243, mul_x2_244, mul_x2_245,
492 mul_x2_246, mul_x2_247, mul_x2_248, mul_x2_249, mul_x2_250, mul_x2_251,
493 mul_x2_252, mul_x2_253, mul_x2_254, mul_x2_255
494 };
495
496 #define MUL(c, r...) \
497 { \
498 switch (REG_CNT(r)) { \
499 case 2: \
500 COPY(r, _mul_x2_in); \
501 gf_x2_mul_fns[c](); \
502 COPY(_mul_x2_acc, r); \
503 break; \
504 case 1: \
505 COPY(r, _mul_x1_in); \
506 gf_x1_mul_fns[c](); \
507 COPY(_mul_x1_acc, r); \
508 break; \
509 default: \
510 VERIFY(0); \
511 } \
512 }
513
514
515 #define raidz_math_begin() kfpu_begin()
516 #define raidz_math_end() kfpu_end()
517
518 #define SYN_STRIDE 4
519
520 #define ZERO_STRIDE 4
521 #define ZERO_DEFINE() {}
522 #define ZERO_D 0, 1, 2, 3
523
524 #define COPY_STRIDE 4
525 #define COPY_DEFINE() {}
526 #define COPY_D 0, 1, 2, 3
527
528 #define ADD_STRIDE 4
529 #define ADD_DEFINE() {}
530 #define ADD_D 0, 1, 2, 3
531
532 #define MUL_STRIDE 2
533 #define MUL_DEFINE() MUL2_SETUP()
534 #define MUL_D 0, 1
535
536 #define GEN_P_STRIDE 4
537 #define GEN_P_DEFINE() {}
538 #define GEN_P_P 0, 1, 2, 3
539
540 #define GEN_PQ_STRIDE 4
541 #define GEN_PQ_DEFINE() {}
542 #define GEN_PQ_D 0, 1, 2, 3
543 #define GEN_PQ_C 4, 5, 6, 7
544
545 #define GEN_PQR_STRIDE 4
546 #define GEN_PQR_DEFINE() {}
547 #define GEN_PQR_D 0, 1, 2, 3
548 #define GEN_PQR_C 4, 5, 6, 7
549
550 #define SYN_Q_DEFINE() {}
551 #define SYN_Q_D 0, 1, 2, 3
552 #define SYN_Q_X 4, 5, 6, 7
553
554 #define SYN_R_DEFINE() {}
555 #define SYN_R_D 0, 1, 2, 3
556 #define SYN_R_X 4, 5, 6, 7
557
558 #define SYN_PQ_DEFINE() {}
559 #define SYN_PQ_D 0, 1, 2, 3
560 #define SYN_PQ_X 4, 5, 6, 7
561
562 #define REC_PQ_STRIDE 2
563 #define REC_PQ_DEFINE() MUL2_SETUP()
564 #define REC_PQ_X 0, 1
565 #define REC_PQ_Y 2, 3
566 #define REC_PQ_T 4, 5
567
568 #define SYN_PR_DEFINE() {}
569 #define SYN_PR_D 0, 1, 2, 3
570 #define SYN_PR_X 4, 5, 6, 7
571
572 #define REC_PR_STRIDE 2
573 #define REC_PR_DEFINE() MUL2_SETUP()
574 #define REC_PR_X 0, 1
575 #define REC_PR_Y 2, 3
576 #define REC_PR_T 4, 5
577
578 #define SYN_QR_DEFINE() {}
579 #define SYN_QR_D 0, 1, 2, 3
580 #define SYN_QR_X 4, 5, 6, 7
581
582 #define REC_QR_STRIDE 2
583 #define REC_QR_DEFINE() MUL2_SETUP()
584 #define REC_QR_X 0, 1
585 #define REC_QR_Y 2, 3
586 #define REC_QR_T 4, 5
587
588 #define SYN_PQR_DEFINE() {}
589 #define SYN_PQR_D 0, 1, 2, 3
590 #define SYN_PQR_X 4, 5, 6, 7
591
592 #define REC_PQR_STRIDE 1
593 #define REC_PQR_DEFINE() MUL2_SETUP()
594 #define REC_PQR_X 0
595 #define REC_PQR_Y 1
596 #define REC_PQR_Z 2
597 #define REC_PQR_XS 3
598 #define REC_PQR_YS 4
599
600
601 #include <sys/vdev_raidz_impl.h>
602 #include "vdev_raidz_math_impl.h"
603
604 DEFINE_GEN_METHODS(sse2);
605 DEFINE_REC_METHODS(sse2);
606
607 static boolean_t
raidz_will_sse2_work(void)608 raidz_will_sse2_work(void)
609 {
610 return (kfpu_allowed() && zfs_sse_available() && zfs_sse2_available());
611 }
612
613 const raidz_impl_ops_t vdev_raidz_sse2_impl = {
614 .init = NULL,
615 .fini = NULL,
616 .gen = RAIDZ_GEN_METHODS(sse2),
617 .rec = RAIDZ_REC_METHODS(sse2),
618 .is_supported = &raidz_will_sse2_work,
619 .name = "sse2"
620 };
621
622 #endif /* defined(__x86_64) && HAVE_SIMD(SSE2) */
623