1/* -----------------------------------------------------------------------
2 *
3 *   neon.uc - RAID-6 syndrome calculation using ARM NEON instructions
4 *
5 *   Copyright (C) 2012 Rob Herring
6 *   Copyright (C) 2015 Linaro Ltd. <ard.biesheuvel@linaro.org>
7 *
8 *   Based on altivec.uc:
9 *     Copyright 2002-2004 H. Peter Anvin - All Rights Reserved
10 *
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.
16 *
17 * ----------------------------------------------------------------------- */
18
19/*
20 * neon$#.c
21 *
22 * $#-way unrolled NEON intrinsics math RAID-6 instruction set
23 *
24 * This file is postprocessed using unroll.awk
25 */
26
27#include <arm_neon.h>
28
29typedef uint8x16_t unative_t;
30
31#define NBYTES(x) ((unative_t){x,x,x,x, x,x,x,x, x,x,x,x, x,x,x,x})
32#define NSIZE	sizeof(unative_t)
33
34/*
35 * The SHLBYTE() operation shifts each byte left by 1, *not*
36 * rolling over into the next byte
37 */
38static inline unative_t SHLBYTE(unative_t v)
39{
40	return vshlq_n_u8(v, 1);
41}
42
43/*
44 * The MASK() operation returns 0xFF in any byte for which the high
45 * bit is 1, 0x00 for any byte for which the high bit is 0.
46 */
47static inline unative_t MASK(unative_t v)
48{
49	const uint8x16_t temp = NBYTES(0);
50	return (unative_t)vcltq_s8((int8x16_t)v, (int8x16_t)temp);
51}
52
53void raid6_neon$#_gen_syndrome_real(int disks, unsigned long bytes, void **ptrs)
54{
55	uint8_t **dptr = (uint8_t **)ptrs;
56	uint8_t *p, *q;
57	int d, z, z0;
58
59	register unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
60	const unative_t x1d = NBYTES(0x1d);
61
62	z0 = disks - 3;		/* Highest data disk */
63	p = dptr[z0+1];		/* XOR parity */
64	q = dptr[z0+2];		/* RS syndrome */
65
66	for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
67		wq$$ = wp$$ = vld1q_u8(&dptr[z0][d+$$*NSIZE]);
68		for ( z = z0-1 ; z >= 0 ; z-- ) {
69			wd$$ = vld1q_u8(&dptr[z][d+$$*NSIZE]);
70			wp$$ = veorq_u8(wp$$, wd$$);
71			w2$$ = MASK(wq$$);
72			w1$$ = SHLBYTE(wq$$);
73
74			w2$$ = vandq_u8(w2$$, x1d);
75			w1$$ = veorq_u8(w1$$, w2$$);
76			wq$$ = veorq_u8(w1$$, wd$$);
77		}
78		vst1q_u8(&p[d+NSIZE*$$], wp$$);
79		vst1q_u8(&q[d+NSIZE*$$], wq$$);
80	}
81}
82
83void raid6_neon$#_xor_syndrome_real(int disks, int start, int stop,
84				    unsigned long bytes, void **ptrs)
85{
86	uint8_t **dptr = (uint8_t **)ptrs;
87	uint8_t *p, *q;
88	int d, z, z0;
89
90	register unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
91	const unative_t x1d = NBYTES(0x1d);
92
93	z0 = stop;		/* P/Q right side optimization */
94	p = dptr[disks-2];	/* XOR parity */
95	q = dptr[disks-1];	/* RS syndrome */
96
97	for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
98		wq$$ = vld1q_u8(&dptr[z0][d+$$*NSIZE]);
99		wp$$ = veorq_u8(vld1q_u8(&p[d+$$*NSIZE]), wq$$);
100
101		/* P/Q data pages */
102		for ( z = z0-1 ; z >= start ; z-- ) {
103			wd$$ = vld1q_u8(&dptr[z][d+$$*NSIZE]);
104			wp$$ = veorq_u8(wp$$, wd$$);
105			w2$$ = MASK(wq$$);
106			w1$$ = SHLBYTE(wq$$);
107
108			w2$$ = vandq_u8(w2$$, x1d);
109			w1$$ = veorq_u8(w1$$, w2$$);
110			wq$$ = veorq_u8(w1$$, wd$$);
111		}
112		/* P/Q left side optimization */
113		for ( z = start-1 ; z >= 0 ; z-- ) {
114			w2$$ = MASK(wq$$);
115			w1$$ = SHLBYTE(wq$$);
116
117			w2$$ = vandq_u8(w2$$, x1d);
118			wq$$ = veorq_u8(w1$$, w2$$);
119		}
120		w1$$ = vld1q_u8(&q[d+NSIZE*$$]);
121		wq$$ = veorq_u8(wq$$, w1$$);
122
123		vst1q_u8(&p[d+NSIZE*$$], wp$$);
124		vst1q_u8(&q[d+NSIZE*$$], wq$$);
125	}
126}
127