xref: /freebsd/sys/libkern/x86/crc32_sse42.c (revision 25ecdc7d52770caf1c9b44b5ec11f468f6b636f3)
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 __FBSDID("$FreeBSD$");
28 
29 /*
30  * This file is compiled in userspace in order to run ATF unit tests.
31  */
32 #ifndef _KERNEL
33 #include <stdint.h>
34 #include <stdlib.h>
35 #else
36 #include <sys/param.h>
37 #include <sys/kernel.h>
38 #endif
39 #include <sys/gsb_crc32.h>
40 
41 static __inline uint32_t
42 _mm_crc32_u8(uint32_t x, uint8_t y)
43 {
44 	/*
45 	 * clang (at least 3.9.[0-1]) pessimizes "rm" (y) and "m" (y)
46 	 * significantly and "r" (y) a lot by copying y to a different
47 	 * local variable (on the stack or in a register), so only use
48 	 * the latter.  This costs a register and an instruction but
49 	 * not a uop.
50 	 */
51 	__asm("crc32b %1,%0" : "+r" (x) : "r" (y));
52 	return (x);
53 }
54 
55 #ifdef __amd64__
56 static __inline uint64_t
57 _mm_crc32_u64(uint64_t x, uint64_t y)
58 {
59 	__asm("crc32q %1,%0" : "+r" (x) : "r" (y));
60 	return (x);
61 }
62 #else
63 static __inline uint32_t
64 _mm_crc32_u32(uint32_t x, uint32_t y)
65 {
66 	__asm("crc32l %1,%0" : "+r" (x) : "r" (y));
67 	return (x);
68 }
69 #endif
70 
71 /* CRC-32C (iSCSI) polynomial in reversed bit order. */
72 #define POLY	0x82f63b78
73 
74 /*
75  * Block sizes for three-way parallel crc computation.  LONG and SHORT must
76  * both be powers of two.
77  */
78 #define LONG	128
79 #define SHORT	64
80 
81 /*
82  * Tables for updating a crc for LONG, 2 * LONG, SHORT and 2 * SHORT bytes
83  * of value 0 later in the input stream, in the same way that the hardware
84  * would, but in software without calculating intermediate steps.
85  */
86 static uint32_t crc32c_long[4][256];
87 static uint32_t crc32c_2long[4][256];
88 static uint32_t crc32c_short[4][256];
89 static uint32_t crc32c_2short[4][256];
90 
91 /*
92  * Multiply a matrix times a vector over the Galois field of two elements,
93  * GF(2).  Each element is a bit in an unsigned integer.  mat must have at
94  * least as many entries as the power of two for most significant one bit in
95  * vec.
96  */
97 static inline uint32_t
98 gf2_matrix_times(uint32_t *mat, uint32_t vec)
99 {
100 	uint32_t sum;
101 
102 	sum = 0;
103 	while (vec) {
104 		if (vec & 1)
105 			sum ^= *mat;
106 		vec >>= 1;
107 		mat++;
108 	}
109 	return (sum);
110 }
111 
112 /*
113  * Multiply a matrix by itself over GF(2).  Both mat and square must have 32
114  * rows.
115  */
116 static inline void
117 gf2_matrix_square(uint32_t *square, uint32_t *mat)
118 {
119 	int n;
120 
121 	for (n = 0; n < 32; n++)
122 		square[n] = gf2_matrix_times(mat, mat[n]);
123 }
124 
125 /*
126  * Construct an operator to apply len zeros to a crc.  len must be a power of
127  * two.  If len is not a power of two, then the result is the same as for the
128  * largest power of two less than len.  The result for len == 0 is the same as
129  * for len == 1.  A version of this routine could be easily written for any
130  * len, but that is not needed for this application.
131  */
132 static void
133 crc32c_zeros_op(uint32_t *even, size_t len)
134 {
135 	uint32_t odd[32];       /* odd-power-of-two zeros operator */
136 	uint32_t row;
137 	int n;
138 
139 	/* put operator for one zero bit in odd */
140 	odd[0] = POLY;              /* CRC-32C polynomial */
141 	row = 1;
142 	for (n = 1; n < 32; n++) {
143 		odd[n] = row;
144 		row <<= 1;
145 	}
146 
147 	/* put operator for two zero bits in even */
148 	gf2_matrix_square(even, odd);
149 
150 	/* put operator for four zero bits in odd */
151 	gf2_matrix_square(odd, even);
152 
153 	/*
154 	 * first square will put the operator for one zero byte (eight zero
155 	 * bits), in even -- next square puts operator for two zero bytes in
156 	 * odd, and so on, until len has been rotated down to zero
157 	 */
158 	do {
159 		gf2_matrix_square(even, odd);
160 		len >>= 1;
161 		if (len == 0)
162 			return;
163 		gf2_matrix_square(odd, even);
164 		len >>= 1;
165 	} while (len);
166 
167 	/* answer ended up in odd -- copy to even */
168 	for (n = 0; n < 32; n++)
169 		even[n] = odd[n];
170 }
171 
172 /*
173  * Take a length and build four lookup tables for applying the zeros operator
174  * for that length, byte-by-byte on the operand.
175  */
176 static void
177 crc32c_zeros(uint32_t zeros[][256], size_t len)
178 {
179 	uint32_t op[32];
180 	uint32_t n;
181 
182 	crc32c_zeros_op(op, len);
183 	for (n = 0; n < 256; n++) {
184 		zeros[0][n] = gf2_matrix_times(op, n);
185 		zeros[1][n] = gf2_matrix_times(op, n << 8);
186 		zeros[2][n] = gf2_matrix_times(op, n << 16);
187 		zeros[3][n] = gf2_matrix_times(op, n << 24);
188 	}
189 }
190 
191 /* Apply the zeros operator table to crc. */
192 static inline uint32_t
193 crc32c_shift(uint32_t zeros[][256], uint32_t crc)
194 {
195 
196 	return (zeros[0][crc & 0xff] ^ zeros[1][(crc >> 8) & 0xff] ^
197 	    zeros[2][(crc >> 16) & 0xff] ^ zeros[3][crc >> 24]);
198 }
199 
200 /* Initialize tables for shifting crcs. */
201 static void
202 #ifndef _KERNEL
203 __attribute__((__constructor__))
204 #endif
205 crc32c_init_hw(void)
206 {
207 	crc32c_zeros(crc32c_long, LONG);
208 	crc32c_zeros(crc32c_2long, 2 * LONG);
209 	crc32c_zeros(crc32c_short, SHORT);
210 	crc32c_zeros(crc32c_2short, 2 * SHORT);
211 }
212 #ifdef _KERNEL
213 SYSINIT(crc32c_sse42, SI_SUB_LOCK, SI_ORDER_ANY, crc32c_init_hw, NULL);
214 #endif
215 
216 /* Compute CRC-32C using the Intel hardware instruction. */
217 uint32_t
218 sse42_crc32c(uint32_t crc, const unsigned char *buf, unsigned len)
219 {
220 #ifdef __amd64__
221 	const size_t align = 8;
222 #else
223 	const size_t align = 4;
224 #endif
225 	const unsigned char *next, *end;
226 #ifdef __amd64__
227 	uint64_t crc0, crc1, crc2;
228 #else
229 	uint32_t crc0, crc1, crc2;
230 #endif
231 
232 	next = buf;
233 	crc0 = crc;
234 
235 	/* Compute the crc to bring the data pointer to an aligned boundary. */
236 	while (len && ((uintptr_t)next & (align - 1)) != 0) {
237 		crc0 = _mm_crc32_u8(crc0, *next);
238 		next++;
239 		len--;
240 	}
241 
242 #if LONG > SHORT
243 	/*
244 	 * Compute the crc on sets of LONG*3 bytes, executing three independent
245 	 * crc instructions, each on LONG bytes -- this is optimized for the
246 	 * Nehalem, Westmere, Sandy Bridge, and Ivy Bridge architectures, which
247 	 * have a throughput of one crc per cycle, but a latency of three
248 	 * cycles.
249 	 */
250 	crc = 0;
251 	while (len >= LONG * 3) {
252 		crc1 = 0;
253 		crc2 = 0;
254 		end = next + LONG;
255 		do {
256 #ifdef __amd64__
257 			crc0 = _mm_crc32_u64(crc0, *(const uint64_t *)next);
258 			crc1 = _mm_crc32_u64(crc1,
259 			    *(const uint64_t *)(next + LONG));
260 			crc2 = _mm_crc32_u64(crc2,
261 			    *(const uint64_t *)(next + (LONG * 2)));
262 #else
263 			crc0 = _mm_crc32_u32(crc0, *(const uint32_t *)next);
264 			crc1 = _mm_crc32_u32(crc1,
265 			    *(const uint32_t *)(next + LONG));
266 			crc2 = _mm_crc32_u32(crc2,
267 			    *(const uint32_t *)(next + (LONG * 2)));
268 #endif
269 			next += align;
270 		} while (next < end);
271 		/*-
272 		 * Update the crc.  Try to do it in parallel with the inner
273 		 * loop.  'crc' is used to accumulate crc0 and crc1
274 		 * produced by the inner loop so that the next iteration
275 		 * of the loop doesn't depend on anything except crc2.
276 		 *
277 		 * The full expression for the update is:
278 		 *     crc = S*S*S*crc + S*S*crc0 + S*crc1
279 		 * where the terms are polynomials modulo the CRC polynomial.
280 		 * We regroup this subtly as:
281 		 *     crc = S*S * (S*crc + crc0) + S*crc1.
282 		 * This has an extra dependency which reduces possible
283 		 * parallelism for the expression, but it turns out to be
284 		 * best to intentionally delay evaluation of this expression
285 		 * so that it competes less with the inner loop.
286 		 *
287 		 * We also intentionally reduce parallelism by feedng back
288 		 * crc2 to the inner loop as crc0 instead of accumulating
289 		 * it in crc.  This synchronizes the loop with crc update.
290 		 * CPU and/or compiler schedulers produced bad order without
291 		 * this.
292 		 *
293 		 * Shifts take about 12 cycles each, so 3 here with 2
294 		 * parallelizable take about 24 cycles and the crc update
295 		 * takes slightly longer.  8 dependent crc32 instructions
296 		 * can run in 24 cycles, so the 3-way blocking is worse
297 		 * than useless for sizes less than 8 * <word size> = 64
298 		 * on amd64.  In practice, SHORT = 32 confirms these
299 		 * timing calculations by giving a small improvement
300 		 * starting at size 96.  Then the inner loop takes about
301 		 * 12 cycles and the crc update about 24, but these are
302 		 * partly in parallel so the total time is less than the
303 		 * 36 cycles that 12 dependent crc32 instructions would
304 		 * take.
305 		 *
306 		 * To have a chance of completely hiding the overhead for
307 		 * the crc update, the inner loop must take considerably
308 		 * longer than 24 cycles.  LONG = 64 makes the inner loop
309 		 * take about 24 cycles, so is not quite large enough.
310 		 * LONG = 128 works OK.  Unhideable overheads are about
311 		 * 12 cycles per inner loop.  All assuming timing like
312 		 * Haswell.
313 		 */
314 		crc = crc32c_shift(crc32c_long, crc) ^ crc0;
315 		crc1 = crc32c_shift(crc32c_long, crc1);
316 		crc = crc32c_shift(crc32c_2long, crc) ^ crc1;
317 		crc0 = crc2;
318 		next += LONG * 2;
319 		len -= LONG * 3;
320 	}
321 	crc0 ^= crc;
322 #endif /* LONG > SHORT */
323 
324 	/*
325 	 * Do the same thing, but now on SHORT*3 blocks for the remaining data
326 	 * less than a LONG*3 block
327 	 */
328 	crc = 0;
329 	while (len >= SHORT * 3) {
330 		crc1 = 0;
331 		crc2 = 0;
332 		end = next + SHORT;
333 		do {
334 #ifdef __amd64__
335 			crc0 = _mm_crc32_u64(crc0, *(const uint64_t *)next);
336 			crc1 = _mm_crc32_u64(crc1,
337 			    *(const uint64_t *)(next + SHORT));
338 			crc2 = _mm_crc32_u64(crc2,
339 			    *(const uint64_t *)(next + (SHORT * 2)));
340 #else
341 			crc0 = _mm_crc32_u32(crc0, *(const uint32_t *)next);
342 			crc1 = _mm_crc32_u32(crc1,
343 			    *(const uint32_t *)(next + SHORT));
344 			crc2 = _mm_crc32_u32(crc2,
345 			    *(const uint32_t *)(next + (SHORT * 2)));
346 #endif
347 			next += align;
348 		} while (next < end);
349 		crc = crc32c_shift(crc32c_short, crc) ^ crc0;
350 		crc1 = crc32c_shift(crc32c_short, crc1);
351 		crc = crc32c_shift(crc32c_2short, crc) ^ crc1;
352 		crc0 = crc2;
353 		next += SHORT * 2;
354 		len -= SHORT * 3;
355 	}
356 	crc0 ^= crc;
357 
358 	/* Compute the crc on the remaining bytes at native word size. */
359 	end = next + (len - (len & (align - 1)));
360 	while (next < end) {
361 #ifdef __amd64__
362 		crc0 = _mm_crc32_u64(crc0, *(const uint64_t *)next);
363 #else
364 		crc0 = _mm_crc32_u32(crc0, *(const uint32_t *)next);
365 #endif
366 		next += align;
367 	}
368 	len &= (align - 1);
369 
370 	/* Compute the crc for any trailing bytes. */
371 	while (len) {
372 		crc0 = _mm_crc32_u8(crc0, *next);
373 		next++;
374 		len--;
375 	}
376 
377 	return ((uint32_t)crc0);
378 }
379