xref: /freebsd/sys/libkern/x86/crc32_sse42.c (revision 5e3190f700637fcfc1a52daeaa4a031fdd2557c7)
1 /*
2  * Derived from crc32c.c version 1.1 by Mark Adler.
3  *
4  * Copyright (C) 2013 Mark Adler
5  *
6  * This software is provided 'as-is', without any express or implied warranty.
7  * In no event will the author be held liable for any damages arising from the
8  * use of this software.
9  *
10  * Permission is granted to anyone to use this software for any purpose,
11  * including commercial applications, and to alter it and redistribute it
12  * freely, subject to the following restrictions:
13  *
14  * 1. The origin of this software must not be misrepresented; you must not
15  *    claim that you wrote the original software. If you use this software
16  *    in a product, an acknowledgment in the product documentation would be
17  *    appreciated but is not required.
18  * 2. Altered source versions must be plainly marked as such, and must not be
19  *    misrepresented as being the original software.
20  * 3. This notice may not be removed or altered from any source distribution.
21  *
22  * Mark Adler
23  * madler@alumni.caltech.edu
24  */
25 
26 #include <sys/cdefs.h>
27 /*
28  * This file is compiled in userspace in order to run ATF unit tests.
29  */
30 #ifndef _KERNEL
31 #include <stdint.h>
32 #include <stdlib.h>
33 #else
34 #include <sys/param.h>
35 #include <sys/kernel.h>
36 #endif
37 #include <sys/gsb_crc32.h>
38 
39 static __inline uint32_t
40 _mm_crc32_u8(uint32_t x, uint8_t y)
41 {
42 	/*
43 	 * clang (at least 3.9.[0-1]) pessimizes "rm" (y) and "m" (y)
44 	 * significantly and "r" (y) a lot by copying y to a different
45 	 * local variable (on the stack or in a register), so only use
46 	 * the latter.  This costs a register and an instruction but
47 	 * not a uop.
48 	 */
49 	__asm("crc32b %1,%0" : "+r" (x) : "r" (y));
50 	return (x);
51 }
52 
53 #ifdef __amd64__
54 static __inline uint64_t
55 _mm_crc32_u64(uint64_t x, uint64_t y)
56 {
57 	__asm("crc32q %1,%0" : "+r" (x) : "r" (y));
58 	return (x);
59 }
60 #else
61 static __inline uint32_t
62 _mm_crc32_u32(uint32_t x, uint32_t y)
63 {
64 	__asm("crc32l %1,%0" : "+r" (x) : "r" (y));
65 	return (x);
66 }
67 #endif
68 
69 /* CRC-32C (iSCSI) polynomial in reversed bit order. */
70 #define POLY	0x82f63b78
71 
72 /*
73  * Block sizes for three-way parallel crc computation.  LONG and SHORT must
74  * both be powers of two.
75  */
76 #define LONG	128
77 #define SHORT	64
78 
79 /*
80  * Tables for updating a crc for LONG, 2 * LONG, SHORT and 2 * SHORT bytes
81  * of value 0 later in the input stream, in the same way that the hardware
82  * would, but in software without calculating intermediate steps.
83  */
84 static uint32_t crc32c_long[4][256];
85 static uint32_t crc32c_2long[4][256];
86 static uint32_t crc32c_short[4][256];
87 static uint32_t crc32c_2short[4][256];
88 
89 /*
90  * Multiply a matrix times a vector over the Galois field of two elements,
91  * GF(2).  Each element is a bit in an unsigned integer.  mat must have at
92  * least as many entries as the power of two for most significant one bit in
93  * vec.
94  */
95 static inline uint32_t
96 gf2_matrix_times(uint32_t *mat, uint32_t vec)
97 {
98 	uint32_t sum;
99 
100 	sum = 0;
101 	while (vec) {
102 		if (vec & 1)
103 			sum ^= *mat;
104 		vec >>= 1;
105 		mat++;
106 	}
107 	return (sum);
108 }
109 
110 /*
111  * Multiply a matrix by itself over GF(2).  Both mat and square must have 32
112  * rows.
113  */
114 static inline void
115 gf2_matrix_square(uint32_t *square, uint32_t *mat)
116 {
117 	int n;
118 
119 	for (n = 0; n < 32; n++)
120 		square[n] = gf2_matrix_times(mat, mat[n]);
121 }
122 
123 /*
124  * Construct an operator to apply len zeros to a crc.  len must be a power of
125  * two.  If len is not a power of two, then the result is the same as for the
126  * largest power of two less than len.  The result for len == 0 is the same as
127  * for len == 1.  A version of this routine could be easily written for any
128  * len, but that is not needed for this application.
129  */
130 static void
131 crc32c_zeros_op(uint32_t *even, size_t len)
132 {
133 	uint32_t odd[32];       /* odd-power-of-two zeros operator */
134 	uint32_t row;
135 	int n;
136 
137 	/* put operator for one zero bit in odd */
138 	odd[0] = POLY;              /* CRC-32C polynomial */
139 	row = 1;
140 	for (n = 1; n < 32; n++) {
141 		odd[n] = row;
142 		row <<= 1;
143 	}
144 
145 	/* put operator for two zero bits in even */
146 	gf2_matrix_square(even, odd);
147 
148 	/* put operator for four zero bits in odd */
149 	gf2_matrix_square(odd, even);
150 
151 	/*
152 	 * first square will put the operator for one zero byte (eight zero
153 	 * bits), in even -- next square puts operator for two zero bytes in
154 	 * odd, and so on, until len has been rotated down to zero
155 	 */
156 	do {
157 		gf2_matrix_square(even, odd);
158 		len >>= 1;
159 		if (len == 0)
160 			return;
161 		gf2_matrix_square(odd, even);
162 		len >>= 1;
163 	} while (len);
164 
165 	/* answer ended up in odd -- copy to even */
166 	for (n = 0; n < 32; n++)
167 		even[n] = odd[n];
168 }
169 
170 /*
171  * Take a length and build four lookup tables for applying the zeros operator
172  * for that length, byte-by-byte on the operand.
173  */
174 static void
175 crc32c_zeros(uint32_t zeros[][256], size_t len)
176 {
177 	uint32_t op[32];
178 	uint32_t n;
179 
180 	crc32c_zeros_op(op, len);
181 	for (n = 0; n < 256; n++) {
182 		zeros[0][n] = gf2_matrix_times(op, n);
183 		zeros[1][n] = gf2_matrix_times(op, n << 8);
184 		zeros[2][n] = gf2_matrix_times(op, n << 16);
185 		zeros[3][n] = gf2_matrix_times(op, n << 24);
186 	}
187 }
188 
189 /* Apply the zeros operator table to crc. */
190 static inline uint32_t
191 crc32c_shift(uint32_t zeros[][256], uint32_t crc)
192 {
193 
194 	return (zeros[0][crc & 0xff] ^ zeros[1][(crc >> 8) & 0xff] ^
195 	    zeros[2][(crc >> 16) & 0xff] ^ zeros[3][crc >> 24]);
196 }
197 
198 /* Initialize tables for shifting crcs. */
199 static void
200 #ifndef _KERNEL
201 __attribute__((__constructor__))
202 #endif
203 crc32c_init_hw(void)
204 {
205 	crc32c_zeros(crc32c_long, LONG);
206 	crc32c_zeros(crc32c_2long, 2 * LONG);
207 	crc32c_zeros(crc32c_short, SHORT);
208 	crc32c_zeros(crc32c_2short, 2 * SHORT);
209 }
210 #ifdef _KERNEL
211 SYSINIT(crc32c_sse42, SI_SUB_LOCK, SI_ORDER_ANY, crc32c_init_hw, NULL);
212 #endif
213 
214 /* Compute CRC-32C using the Intel hardware instruction. */
215 uint32_t
216 sse42_crc32c(uint32_t crc, const unsigned char *buf, unsigned len)
217 {
218 #ifdef __amd64__
219 	const size_t align = 8;
220 #else
221 	const size_t align = 4;
222 #endif
223 	const unsigned char *next, *end;
224 #ifdef __amd64__
225 	uint64_t crc0, crc1, crc2;
226 #else
227 	uint32_t crc0, crc1, crc2;
228 #endif
229 
230 	next = buf;
231 	crc0 = crc;
232 
233 	/* Compute the crc to bring the data pointer to an aligned boundary. */
234 	while (len && ((uintptr_t)next & (align - 1)) != 0) {
235 		crc0 = _mm_crc32_u8(crc0, *next);
236 		next++;
237 		len--;
238 	}
239 
240 #if LONG > SHORT
241 	/*
242 	 * Compute the crc on sets of LONG*3 bytes, executing three independent
243 	 * crc instructions, each on LONG bytes -- this is optimized for the
244 	 * Nehalem, Westmere, Sandy Bridge, and Ivy Bridge architectures, which
245 	 * have a throughput of one crc per cycle, but a latency of three
246 	 * cycles.
247 	 */
248 	crc = 0;
249 	while (len >= LONG * 3) {
250 		crc1 = 0;
251 		crc2 = 0;
252 		end = next + LONG;
253 		do {
254 #ifdef __amd64__
255 			crc0 = _mm_crc32_u64(crc0, *(const uint64_t *)next);
256 			crc1 = _mm_crc32_u64(crc1,
257 			    *(const uint64_t *)(next + LONG));
258 			crc2 = _mm_crc32_u64(crc2,
259 			    *(const uint64_t *)(next + (LONG * 2)));
260 #else
261 			crc0 = _mm_crc32_u32(crc0, *(const uint32_t *)next);
262 			crc1 = _mm_crc32_u32(crc1,
263 			    *(const uint32_t *)(next + LONG));
264 			crc2 = _mm_crc32_u32(crc2,
265 			    *(const uint32_t *)(next + (LONG * 2)));
266 #endif
267 			next += align;
268 		} while (next < end);
269 		/*-
270 		 * Update the crc.  Try to do it in parallel with the inner
271 		 * loop.  'crc' is used to accumulate crc0 and crc1
272 		 * produced by the inner loop so that the next iteration
273 		 * of the loop doesn't depend on anything except crc2.
274 		 *
275 		 * The full expression for the update is:
276 		 *     crc = S*S*S*crc + S*S*crc0 + S*crc1
277 		 * where the terms are polynomials modulo the CRC polynomial.
278 		 * We regroup this subtly as:
279 		 *     crc = S*S * (S*crc + crc0) + S*crc1.
280 		 * This has an extra dependency which reduces possible
281 		 * parallelism for the expression, but it turns out to be
282 		 * best to intentionally delay evaluation of this expression
283 		 * so that it competes less with the inner loop.
284 		 *
285 		 * We also intentionally reduce parallelism by feedng back
286 		 * crc2 to the inner loop as crc0 instead of accumulating
287 		 * it in crc.  This synchronizes the loop with crc update.
288 		 * CPU and/or compiler schedulers produced bad order without
289 		 * this.
290 		 *
291 		 * Shifts take about 12 cycles each, so 3 here with 2
292 		 * parallelizable take about 24 cycles and the crc update
293 		 * takes slightly longer.  8 dependent crc32 instructions
294 		 * can run in 24 cycles, so the 3-way blocking is worse
295 		 * than useless for sizes less than 8 * <word size> = 64
296 		 * on amd64.  In practice, SHORT = 32 confirms these
297 		 * timing calculations by giving a small improvement
298 		 * starting at size 96.  Then the inner loop takes about
299 		 * 12 cycles and the crc update about 24, but these are
300 		 * partly in parallel so the total time is less than the
301 		 * 36 cycles that 12 dependent crc32 instructions would
302 		 * take.
303 		 *
304 		 * To have a chance of completely hiding the overhead for
305 		 * the crc update, the inner loop must take considerably
306 		 * longer than 24 cycles.  LONG = 64 makes the inner loop
307 		 * take about 24 cycles, so is not quite large enough.
308 		 * LONG = 128 works OK.  Unhideable overheads are about
309 		 * 12 cycles per inner loop.  All assuming timing like
310 		 * Haswell.
311 		 */
312 		crc = crc32c_shift(crc32c_long, crc) ^ crc0;
313 		crc1 = crc32c_shift(crc32c_long, crc1);
314 		crc = crc32c_shift(crc32c_2long, crc) ^ crc1;
315 		crc0 = crc2;
316 		next += LONG * 2;
317 		len -= LONG * 3;
318 	}
319 	crc0 ^= crc;
320 #endif /* LONG > SHORT */
321 
322 	/*
323 	 * Do the same thing, but now on SHORT*3 blocks for the remaining data
324 	 * less than a LONG*3 block
325 	 */
326 	crc = 0;
327 	while (len >= SHORT * 3) {
328 		crc1 = 0;
329 		crc2 = 0;
330 		end = next + SHORT;
331 		do {
332 #ifdef __amd64__
333 			crc0 = _mm_crc32_u64(crc0, *(const uint64_t *)next);
334 			crc1 = _mm_crc32_u64(crc1,
335 			    *(const uint64_t *)(next + SHORT));
336 			crc2 = _mm_crc32_u64(crc2,
337 			    *(const uint64_t *)(next + (SHORT * 2)));
338 #else
339 			crc0 = _mm_crc32_u32(crc0, *(const uint32_t *)next);
340 			crc1 = _mm_crc32_u32(crc1,
341 			    *(const uint32_t *)(next + SHORT));
342 			crc2 = _mm_crc32_u32(crc2,
343 			    *(const uint32_t *)(next + (SHORT * 2)));
344 #endif
345 			next += align;
346 		} while (next < end);
347 		crc = crc32c_shift(crc32c_short, crc) ^ crc0;
348 		crc1 = crc32c_shift(crc32c_short, crc1);
349 		crc = crc32c_shift(crc32c_2short, crc) ^ crc1;
350 		crc0 = crc2;
351 		next += SHORT * 2;
352 		len -= SHORT * 3;
353 	}
354 	crc0 ^= crc;
355 
356 	/* Compute the crc on the remaining bytes at native word size. */
357 	end = next + (len - (len & (align - 1)));
358 	while (next < end) {
359 #ifdef __amd64__
360 		crc0 = _mm_crc32_u64(crc0, *(const uint64_t *)next);
361 #else
362 		crc0 = _mm_crc32_u32(crc0, *(const uint32_t *)next);
363 #endif
364 		next += align;
365 	}
366 	len &= (align - 1);
367 
368 	/* Compute the crc for any trailing bytes. */
369 	while (len) {
370 		crc0 = _mm_crc32_u8(crc0, *next);
371 		next++;
372 		len--;
373 	}
374 
375 	return ((uint32_t)crc0);
376 }
377