1 /* ------------------------------------------- 1 /* -----------------------------------------------------------------------
2 * 2 *
3 * neon.uc - RAID-6 syndrome calculation usi 3 * neon.uc - RAID-6 syndrome calculation using ARM NEON instructions
4 * 4 *
5 * Copyright (C) 2012 Rob Herring 5 * Copyright (C) 2012 Rob Herring
6 * Copyright (C) 2015 Linaro Ltd. <ard.bieshe 6 * Copyright (C) 2015 Linaro Ltd. <ard.biesheuvel@linaro.org>
7 * 7 *
8 * Based on altivec.uc: 8 * Based on altivec.uc:
9 * Copyright 2002-2004 H. Peter Anvin - Al 9 * Copyright 2002-2004 H. Peter Anvin - All Rights Reserved
10 * 10 *
11 * This program is free software; you can re 11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Pub 12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation, Inc., 53 Te 13 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
14 * Boston MA 02111-1307, USA; either version 14 * Boston MA 02111-1307, USA; either version 2 of the License, or
15 * (at your option) any later version; incor 15 * (at your option) any later version; incorporated herein by reference.
16 * 16 *
17 * ------------------------------------------- 17 * ----------------------------------------------------------------------- */
18 18
19 /* 19 /*
20 * neon$#.c 20 * neon$#.c
21 * 21 *
22 * $#-way unrolled NEON intrinsics math RAID-6 22 * $#-way unrolled NEON intrinsics math RAID-6 instruction set
23 * 23 *
24 * This file is postprocessed using unroll.awk 24 * This file is postprocessed using unroll.awk
25 */ 25 */
26 26
27 #include <arm_neon.h> 27 #include <arm_neon.h>
28 #include "neon.h" <<
29 28
30 typedef uint8x16_t unative_t; 29 typedef uint8x16_t unative_t;
31 30
32 #define NSIZE sizeof(unative_t) 31 #define NSIZE sizeof(unative_t)
33 32
34 /* 33 /*
35 * The SHLBYTE() operation shifts each byte le 34 * The SHLBYTE() operation shifts each byte left by 1, *not*
36 * rolling over into the next byte 35 * rolling over into the next byte
37 */ 36 */
38 static inline unative_t SHLBYTE(unative_t v) 37 static inline unative_t SHLBYTE(unative_t v)
39 { 38 {
40 return vshlq_n_u8(v, 1); 39 return vshlq_n_u8(v, 1);
41 } 40 }
42 41
43 /* 42 /*
44 * The MASK() operation returns 0xFF in any by 43 * The MASK() operation returns 0xFF in any byte for which the high
45 * bit is 1, 0x00 for any byte for which the h 44 * bit is 1, 0x00 for any byte for which the high bit is 0.
46 */ 45 */
47 static inline unative_t MASK(unative_t v) 46 static inline unative_t MASK(unative_t v)
48 { 47 {
49 return (unative_t)vshrq_n_s8((int8x16_ 48 return (unative_t)vshrq_n_s8((int8x16_t)v, 7);
50 } 49 }
51 50
52 static inline unative_t PMUL(unative_t v, unat 51 static inline unative_t PMUL(unative_t v, unative_t u)
53 { 52 {
54 return (unative_t)vmulq_p8((poly8x16_t 53 return (unative_t)vmulq_p8((poly8x16_t)v, (poly8x16_t)u);
55 } 54 }
56 55
57 void raid6_neon$#_gen_syndrome_real(int disks, 56 void raid6_neon$#_gen_syndrome_real(int disks, unsigned long bytes, void **ptrs)
58 { 57 {
59 uint8_t **dptr = (uint8_t **)ptrs; 58 uint8_t **dptr = (uint8_t **)ptrs;
60 uint8_t *p, *q; 59 uint8_t *p, *q;
61 int d, z, z0; 60 int d, z, z0;
62 61
63 register unative_t wd$$, wq$$, wp$$, w 62 register unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
64 const unative_t x1d = vdupq_n_u8(0x1d) 63 const unative_t x1d = vdupq_n_u8(0x1d);
65 64
66 z0 = disks - 3; /* Highest dat 65 z0 = disks - 3; /* Highest data disk */
67 p = dptr[z0+1]; /* XOR parity 66 p = dptr[z0+1]; /* XOR parity */
68 q = dptr[z0+2]; /* RS syndrome 67 q = dptr[z0+2]; /* RS syndrome */
69 68
70 for ( d = 0 ; d < bytes ; d += NSIZE*$ 69 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
71 wq$$ = wp$$ = vld1q_u8(&dptr[z 70 wq$$ = wp$$ = vld1q_u8(&dptr[z0][d+$$*NSIZE]);
72 for ( z = z0-1 ; z >= 0 ; z-- 71 for ( z = z0-1 ; z >= 0 ; z-- ) {
73 wd$$ = vld1q_u8(&dptr[ 72 wd$$ = vld1q_u8(&dptr[z][d+$$*NSIZE]);
74 wp$$ = veorq_u8(wp$$, 73 wp$$ = veorq_u8(wp$$, wd$$);
75 w2$$ = MASK(wq$$); 74 w2$$ = MASK(wq$$);
76 w1$$ = SHLBYTE(wq$$); 75 w1$$ = SHLBYTE(wq$$);
77 76
78 w2$$ = vandq_u8(w2$$, 77 w2$$ = vandq_u8(w2$$, x1d);
79 w1$$ = veorq_u8(w1$$, 78 w1$$ = veorq_u8(w1$$, w2$$);
80 wq$$ = veorq_u8(w1$$, 79 wq$$ = veorq_u8(w1$$, wd$$);
81 } 80 }
82 vst1q_u8(&p[d+NSIZE*$$], wp$$) 81 vst1q_u8(&p[d+NSIZE*$$], wp$$);
83 vst1q_u8(&q[d+NSIZE*$$], wq$$) 82 vst1q_u8(&q[d+NSIZE*$$], wq$$);
84 } 83 }
85 } 84 }
86 85
87 void raid6_neon$#_xor_syndrome_real(int disks, 86 void raid6_neon$#_xor_syndrome_real(int disks, int start, int stop,
88 unsigned l 87 unsigned long bytes, void **ptrs)
89 { 88 {
90 uint8_t **dptr = (uint8_t **)ptrs; 89 uint8_t **dptr = (uint8_t **)ptrs;
91 uint8_t *p, *q; 90 uint8_t *p, *q;
92 int d, z, z0; 91 int d, z, z0;
93 92
94 register unative_t wd$$, wq$$, wp$$, w 93 register unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
95 const unative_t x1d = vdupq_n_u8(0x1d) 94 const unative_t x1d = vdupq_n_u8(0x1d);
96 95
97 z0 = stop; /* P/Q right s 96 z0 = stop; /* P/Q right side optimization */
98 p = dptr[disks-2]; /* XOR parity 97 p = dptr[disks-2]; /* XOR parity */
99 q = dptr[disks-1]; /* RS syndrome 98 q = dptr[disks-1]; /* RS syndrome */
100 99
101 for ( d = 0 ; d < bytes ; d += NSIZE*$ 100 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
102 wq$$ = vld1q_u8(&dptr[z0][d+$$ 101 wq$$ = vld1q_u8(&dptr[z0][d+$$*NSIZE]);
103 wp$$ = veorq_u8(vld1q_u8(&p[d+ 102 wp$$ = veorq_u8(vld1q_u8(&p[d+$$*NSIZE]), wq$$);
104 103
105 /* P/Q data pages */ 104 /* P/Q data pages */
106 for ( z = z0-1 ; z >= start ; 105 for ( z = z0-1 ; z >= start ; z-- ) {
107 wd$$ = vld1q_u8(&dptr[ 106 wd$$ = vld1q_u8(&dptr[z][d+$$*NSIZE]);
108 wp$$ = veorq_u8(wp$$, 107 wp$$ = veorq_u8(wp$$, wd$$);
109 w2$$ = MASK(wq$$); 108 w2$$ = MASK(wq$$);
110 w1$$ = SHLBYTE(wq$$); 109 w1$$ = SHLBYTE(wq$$);
111 110
112 w2$$ = vandq_u8(w2$$, 111 w2$$ = vandq_u8(w2$$, x1d);
113 w1$$ = veorq_u8(w1$$, 112 w1$$ = veorq_u8(w1$$, w2$$);
114 wq$$ = veorq_u8(w1$$, 113 wq$$ = veorq_u8(w1$$, wd$$);
115 } 114 }
116 /* P/Q left side optimization 115 /* P/Q left side optimization */
117 for ( z = start-1 ; z >= 3 ; z 116 for ( z = start-1 ; z >= 3 ; z -= 4 ) {
118 w2$$ = vshrq_n_u8(wq$$ 117 w2$$ = vshrq_n_u8(wq$$, 4);
119 w1$$ = vshlq_n_u8(wq$$ 118 w1$$ = vshlq_n_u8(wq$$, 4);
120 119
121 w2$$ = PMUL(w2$$, x1d) 120 w2$$ = PMUL(w2$$, x1d);
122 wq$$ = veorq_u8(w1$$, 121 wq$$ = veorq_u8(w1$$, w2$$);
123 } 122 }
124 123
125 switch (z) { 124 switch (z) {
126 case 2: 125 case 2:
127 w2$$ = vshrq_n_u8(wq$$ 126 w2$$ = vshrq_n_u8(wq$$, 5);
128 w1$$ = vshlq_n_u8(wq$$ 127 w1$$ = vshlq_n_u8(wq$$, 3);
129 128
130 w2$$ = PMUL(w2$$, x1d) 129 w2$$ = PMUL(w2$$, x1d);
131 wq$$ = veorq_u8(w1$$, 130 wq$$ = veorq_u8(w1$$, w2$$);
132 break; 131 break;
133 case 1: 132 case 1:
134 w2$$ = vshrq_n_u8(wq$$ 133 w2$$ = vshrq_n_u8(wq$$, 6);
135 w1$$ = vshlq_n_u8(wq$$ 134 w1$$ = vshlq_n_u8(wq$$, 2);
136 135
137 w2$$ = PMUL(w2$$, x1d) 136 w2$$ = PMUL(w2$$, x1d);
138 wq$$ = veorq_u8(w1$$, 137 wq$$ = veorq_u8(w1$$, w2$$);
139 break; 138 break;
140 case 0: 139 case 0:
141 w2$$ = MASK(wq$$); 140 w2$$ = MASK(wq$$);
142 w1$$ = SHLBYTE(wq$$); 141 w1$$ = SHLBYTE(wq$$);
143 142
144 w2$$ = vandq_u8(w2$$, 143 w2$$ = vandq_u8(w2$$, x1d);
145 wq$$ = veorq_u8(w1$$, 144 wq$$ = veorq_u8(w1$$, w2$$);
146 } 145 }
147 w1$$ = vld1q_u8(&q[d+NSIZE*$$] 146 w1$$ = vld1q_u8(&q[d+NSIZE*$$]);
148 wq$$ = veorq_u8(wq$$, w1$$); 147 wq$$ = veorq_u8(wq$$, w1$$);
149 148
150 vst1q_u8(&p[d+NSIZE*$$], wp$$) 149 vst1q_u8(&p[d+NSIZE*$$], wp$$);
151 vst1q_u8(&q[d+NSIZE*$$], wq$$) 150 vst1q_u8(&q[d+NSIZE*$$], wq$$);
152 } 151 }
153 } 152 }
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.