1 /* -----------------------------------------------------------------------
3 * neon.uc - RAID-6 syndrome calculation using ARM NEON instructions
5 * Copyright (C) 2012 Rob Herring
6 * Copyright (C) 2015 Linaro Ltd. <ard.biesheuvel@linaro.org>
9 * Copyright 2002-2004 H. Peter Anvin - All Rights Reserved
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
14 * Boston MA 02111-1307, USA; either version 2 of the License, or
15 * (at your option) any later version; incorporated herein by reference.
17 * ----------------------------------------------------------------------- */
22 * $#-way unrolled NEON intrinsics math RAID-6 instruction set
24 * This file is postprocessed using unroll.awk
29 typedef uint8x16_t unative_t;
31 #define NSIZE sizeof(unative_t)
34 * The SHLBYTE() operation shifts each byte left by 1, *not*
35 * rolling over into the next byte
37 static inline unative_t SHLBYTE(unative_t v)
39 return vshlq_n_u8(v, 1);
43 * The MASK() operation returns 0xFF in any byte for which the high
44 * bit is 1, 0x00 for any byte for which the high bit is 0.
46 static inline unative_t MASK(unative_t v)
48 return (unative_t)vshrq_n_s8((int8x16_t)v, 7);
51 static inline unative_t PMUL(unative_t v, unative_t u)
53 return (unative_t)vmulq_p8((poly8x16_t)v, (poly8x16_t)u);
56 void raid6_neon$#_gen_syndrome_real(int disks, unsigned long bytes, void **ptrs)
58 uint8_t **dptr = (uint8_t **)ptrs;
62 register unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
63 const unative_t x1d = vdupq_n_u8(0x1d);
65 z0 = disks - 3; /* Highest data disk */
66 p = dptr[z0+1]; /* XOR parity */
67 q = dptr[z0+2]; /* RS syndrome */
69 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
70 wq$$ = wp$$ = vld1q_u8(&dptr[z0][d+$$*NSIZE]);
71 for ( z = z0-1 ; z >= 0 ; z-- ) {
72 wd$$ = vld1q_u8(&dptr[z][d+$$*NSIZE]);
73 wp$$ = veorq_u8(wp$$, wd$$);
77 w2$$ = vandq_u8(w2$$, x1d);
78 w1$$ = veorq_u8(w1$$, w2$$);
79 wq$$ = veorq_u8(w1$$, wd$$);
81 vst1q_u8(&p[d+NSIZE*$$], wp$$);
82 vst1q_u8(&q[d+NSIZE*$$], wq$$);
86 void raid6_neon$#_xor_syndrome_real(int disks, int start, int stop,
87 unsigned long bytes, void **ptrs)
89 uint8_t **dptr = (uint8_t **)ptrs;
93 register unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
94 const unative_t x1d = vdupq_n_u8(0x1d);
96 z0 = stop; /* P/Q right side optimization */
97 p = dptr[disks-2]; /* XOR parity */
98 q = dptr[disks-1]; /* RS syndrome */
100 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
101 wq$$ = vld1q_u8(&dptr[z0][d+$$*NSIZE]);
102 wp$$ = veorq_u8(vld1q_u8(&p[d+$$*NSIZE]), wq$$);
105 for ( z = z0-1 ; z >= start ; z-- ) {
106 wd$$ = vld1q_u8(&dptr[z][d+$$*NSIZE]);
107 wp$$ = veorq_u8(wp$$, wd$$);
109 w1$$ = SHLBYTE(wq$$);
111 w2$$ = vandq_u8(w2$$, x1d);
112 w1$$ = veorq_u8(w1$$, w2$$);
113 wq$$ = veorq_u8(w1$$, wd$$);
115 /* P/Q left side optimization */
116 for ( z = start-1 ; z >= 3 ; z -= 4 ) {
117 w2$$ = vshrq_n_u8(wq$$, 4);
118 w1$$ = vshlq_n_u8(wq$$, 4);
120 w2$$ = PMUL(w2$$, x1d);
121 wq$$ = veorq_u8(w1$$, w2$$);
126 w2$$ = vshrq_n_u8(wq$$, 5);
127 w1$$ = vshlq_n_u8(wq$$, 3);
129 w2$$ = PMUL(w2$$, x1d);
130 wq$$ = veorq_u8(w1$$, w2$$);
133 w2$$ = vshrq_n_u8(wq$$, 6);
134 w1$$ = vshlq_n_u8(wq$$, 2);
136 w2$$ = PMUL(w2$$, x1d);
137 wq$$ = veorq_u8(w1$$, w2$$);
141 w1$$ = SHLBYTE(wq$$);
143 w2$$ = vandq_u8(w2$$, x1d);
144 wq$$ = veorq_u8(w1$$, w2$$);
146 w1$$ = vld1q_u8(&q[d+NSIZE*$$]);
147 wq$$ = veorq_u8(wq$$, w1$$);
149 vst1q_u8(&p[d+NSIZE*$$], wp$$);
150 vst1q_u8(&q[d+NSIZE*$$], wq$$);