1 /* -------------------------------------------
2 *
3 * neon.uc - RAID-6 syndrome calculation usi
4 *
5 * Copyright (C) 2012 Rob Herring
6 * Copyright (C) 2015 Linaro Ltd. <ard.bieshe
7 *
8 * Based on altivec.uc:
9 * Copyright 2002-2004 H. Peter Anvin - Al
10 *
11 * This program is free software; you can re
12 * it under the terms of the GNU General Pub
13 * the Free Software Foundation, Inc., 53 Te
14 * Boston MA 02111-1307, USA; either version
15 * (at your option) any later version; incor
16 *
17 * -------------------------------------------
18
19 /*
20 * neon$#.c
21 *
22 * $#-way unrolled NEON intrinsics math RAID-6
23 *
24 * This file is postprocessed using unroll.awk
25 */
26
27 #include <arm_neon.h>
28 #include "neon.h"
29
30 typedef uint8x16_t unative_t;
31
32 #define NSIZE sizeof(unative_t)
33
34 /*
35 * The SHLBYTE() operation shifts each byte le
36 * rolling over into the next byte
37 */
38 static inline unative_t SHLBYTE(unative_t v)
39 {
40 return vshlq_n_u8(v, 1);
41 }
42
43 /*
44 * The MASK() operation returns 0xFF in any by
45 * bit is 1, 0x00 for any byte for which the h
46 */
47 static inline unative_t MASK(unative_t v)
48 {
49 return (unative_t)vshrq_n_s8((int8x16_
50 }
51
52 static inline unative_t PMUL(unative_t v, unat
53 {
54 return (unative_t)vmulq_p8((poly8x16_t
55 }
56
57 void raid6_neon$#_gen_syndrome_real(int disks,
58 {
59 uint8_t **dptr = (uint8_t **)ptrs;
60 uint8_t *p, *q;
61 int d, z, z0;
62
63 register unative_t wd$$, wq$$, wp$$, w
64 const unative_t x1d = vdupq_n_u8(0x1d)
65
66 z0 = disks - 3; /* Highest dat
67 p = dptr[z0+1]; /* XOR parity
68 q = dptr[z0+2]; /* RS syndrome
69
70 for ( d = 0 ; d < bytes ; d += NSIZE*$
71 wq$$ = wp$$ = vld1q_u8(&dptr[z
72 for ( z = z0-1 ; z >= 0 ; z--
73 wd$$ = vld1q_u8(&dptr[
74 wp$$ = veorq_u8(wp$$,
75 w2$$ = MASK(wq$$);
76 w1$$ = SHLBYTE(wq$$);
77
78 w2$$ = vandq_u8(w2$$,
79 w1$$ = veorq_u8(w1$$,
80 wq$$ = veorq_u8(w1$$,
81 }
82 vst1q_u8(&p[d+NSIZE*$$], wp$$)
83 vst1q_u8(&q[d+NSIZE*$$], wq$$)
84 }
85 }
86
87 void raid6_neon$#_xor_syndrome_real(int disks,
88 unsigned l
89 {
90 uint8_t **dptr = (uint8_t **)ptrs;
91 uint8_t *p, *q;
92 int d, z, z0;
93
94 register unative_t wd$$, wq$$, wp$$, w
95 const unative_t x1d = vdupq_n_u8(0x1d)
96
97 z0 = stop; /* P/Q right s
98 p = dptr[disks-2]; /* XOR parity
99 q = dptr[disks-1]; /* RS syndrome
100
101 for ( d = 0 ; d < bytes ; d += NSIZE*$
102 wq$$ = vld1q_u8(&dptr[z0][d+$$
103 wp$$ = veorq_u8(vld1q_u8(&p[d+
104
105 /* P/Q data pages */
106 for ( z = z0-1 ; z >= start ;
107 wd$$ = vld1q_u8(&dptr[
108 wp$$ = veorq_u8(wp$$,
109 w2$$ = MASK(wq$$);
110 w1$$ = SHLBYTE(wq$$);
111
112 w2$$ = vandq_u8(w2$$,
113 w1$$ = veorq_u8(w1$$,
114 wq$$ = veorq_u8(w1$$,
115 }
116 /* P/Q left side optimization
117 for ( z = start-1 ; z >= 3 ; z
118 w2$$ = vshrq_n_u8(wq$$
119 w1$$ = vshlq_n_u8(wq$$
120
121 w2$$ = PMUL(w2$$, x1d)
122 wq$$ = veorq_u8(w1$$,
123 }
124
125 switch (z) {
126 case 2:
127 w2$$ = vshrq_n_u8(wq$$
128 w1$$ = vshlq_n_u8(wq$$
129
130 w2$$ = PMUL(w2$$, x1d)
131 wq$$ = veorq_u8(w1$$,
132 break;
133 case 1:
134 w2$$ = vshrq_n_u8(wq$$
135 w1$$ = vshlq_n_u8(wq$$
136
137 w2$$ = PMUL(w2$$, x1d)
138 wq$$ = veorq_u8(w1$$,
139 break;
140 case 0:
141 w2$$ = MASK(wq$$);
142 w1$$ = SHLBYTE(wq$$);
143
144 w2$$ = vandq_u8(w2$$,
145 wq$$ = veorq_u8(w1$$,
146 }
147 w1$$ = vld1q_u8(&q[d+NSIZE*$$]
148 wq$$ = veorq_u8(wq$$, w1$$);
149
150 vst1q_u8(&p[d+NSIZE*$$], wp$$)
151 vst1q_u8(&q[d+NSIZE*$$], wq$$)
152 }
153 }
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.