0001 /* -----------------------------------------------------------------------
0002 *
0003 * neon.uc - RAID-6 syndrome calculation using ARM NEON instructions
0004 *
0005 * Copyright (C) 2012 Rob Herring
0006 * Copyright (C) 2015 Linaro Ltd. <ard.biesheuvel@linaro.org>
0007 *
0008 * Based on altivec.uc:
0009 * Copyright 2002-2004 H. Peter Anvin - All Rights Reserved
0010 *
0011 * This program is free software; you can redistribute it and/or modify
0012 * it under the terms of the GNU General Public License as published by
0013 * the Free Software Foundation, Inc., 53 Temple Place Ste 330,
0014 * Boston MA 02111-1307, USA; either version 2 of the License, or
0015 * (at your option) any later version; incorporated herein by reference.
0016 *
0017 * ----------------------------------------------------------------------- */
0018
0019 /*
0020 * neon$#.c
0021 *
0022 * $#-way unrolled NEON intrinsics math RAID-6 instruction set
0023 *
0024 * This file is postprocessed using unroll.awk
0025 */
0026
0027 #include <arm_neon.h>
0028
0029 typedef uint8x16_t unative_t;
0030
0031 #define NSIZE sizeof(unative_t)
0032
0033 /*
0034 * The SHLBYTE() operation shifts each byte left by 1, *not*
0035 * rolling over into the next byte
0036 */
0037 static inline unative_t SHLBYTE(unative_t v)
0038 {
0039 return vshlq_n_u8(v, 1);
0040 }
0041
0042 /*
0043 * The MASK() operation returns 0xFF in any byte for which the high
0044 * bit is 1, 0x00 for any byte for which the high bit is 0.
0045 */
0046 static inline unative_t MASK(unative_t v)
0047 {
0048 return (unative_t)vshrq_n_s8((int8x16_t)v, 7);
0049 }
0050
0051 static inline unative_t PMUL(unative_t v, unative_t u)
0052 {
0053 return (unative_t)vmulq_p8((poly8x16_t)v, (poly8x16_t)u);
0054 }
0055
0056 void raid6_neon$#_gen_syndrome_real(int disks, unsigned long bytes, void **ptrs)
0057 {
0058 uint8_t **dptr = (uint8_t **)ptrs;
0059 uint8_t *p, *q;
0060 int d, z, z0;
0061
0062 register unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
0063 const unative_t x1d = vdupq_n_u8(0x1d);
0064
0065 z0 = disks - 3; /* Highest data disk */
0066 p = dptr[z0+1]; /* XOR parity */
0067 q = dptr[z0+2]; /* RS syndrome */
0068
0069 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
0070 wq$$ = wp$$ = vld1q_u8(&dptr[z0][d+$$*NSIZE]);
0071 for ( z = z0-1 ; z >= 0 ; z-- ) {
0072 wd$$ = vld1q_u8(&dptr[z][d+$$*NSIZE]);
0073 wp$$ = veorq_u8(wp$$, wd$$);
0074 w2$$ = MASK(wq$$);
0075 w1$$ = SHLBYTE(wq$$);
0076
0077 w2$$ = vandq_u8(w2$$, x1d);
0078 w1$$ = veorq_u8(w1$$, w2$$);
0079 wq$$ = veorq_u8(w1$$, wd$$);
0080 }
0081 vst1q_u8(&p[d+NSIZE*$$], wp$$);
0082 vst1q_u8(&q[d+NSIZE*$$], wq$$);
0083 }
0084 }
0085
0086 void raid6_neon$#_xor_syndrome_real(int disks, int start, int stop,
0087 unsigned long bytes, void **ptrs)
0088 {
0089 uint8_t **dptr = (uint8_t **)ptrs;
0090 uint8_t *p, *q;
0091 int d, z, z0;
0092
0093 register unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
0094 const unative_t x1d = vdupq_n_u8(0x1d);
0095
0096 z0 = stop; /* P/Q right side optimization */
0097 p = dptr[disks-2]; /* XOR parity */
0098 q = dptr[disks-1]; /* RS syndrome */
0099
0100 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
0101 wq$$ = vld1q_u8(&dptr[z0][d+$$*NSIZE]);
0102 wp$$ = veorq_u8(vld1q_u8(&p[d+$$*NSIZE]), wq$$);
0103
0104 /* P/Q data pages */
0105 for ( z = z0-1 ; z >= start ; z-- ) {
0106 wd$$ = vld1q_u8(&dptr[z][d+$$*NSIZE]);
0107 wp$$ = veorq_u8(wp$$, wd$$);
0108 w2$$ = MASK(wq$$);
0109 w1$$ = SHLBYTE(wq$$);
0110
0111 w2$$ = vandq_u8(w2$$, x1d);
0112 w1$$ = veorq_u8(w1$$, w2$$);
0113 wq$$ = veorq_u8(w1$$, wd$$);
0114 }
0115 /* P/Q left side optimization */
0116 for ( z = start-1 ; z >= 3 ; z -= 4 ) {
0117 w2$$ = vshrq_n_u8(wq$$, 4);
0118 w1$$ = vshlq_n_u8(wq$$, 4);
0119
0120 w2$$ = PMUL(w2$$, x1d);
0121 wq$$ = veorq_u8(w1$$, w2$$);
0122 }
0123
0124 switch (z) {
0125 case 2:
0126 w2$$ = vshrq_n_u8(wq$$, 5);
0127 w1$$ = vshlq_n_u8(wq$$, 3);
0128
0129 w2$$ = PMUL(w2$$, x1d);
0130 wq$$ = veorq_u8(w1$$, w2$$);
0131 break;
0132 case 1:
0133 w2$$ = vshrq_n_u8(wq$$, 6);
0134 w1$$ = vshlq_n_u8(wq$$, 2);
0135
0136 w2$$ = PMUL(w2$$, x1d);
0137 wq$$ = veorq_u8(w1$$, w2$$);
0138 break;
0139 case 0:
0140 w2$$ = MASK(wq$$);
0141 w1$$ = SHLBYTE(wq$$);
0142
0143 w2$$ = vandq_u8(w2$$, x1d);
0144 wq$$ = veorq_u8(w1$$, w2$$);
0145 }
0146 w1$$ = vld1q_u8(&q[d+NSIZE*$$]);
0147 wq$$ = veorq_u8(wq$$, w1$$);
0148
0149 vst1q_u8(&p[d+NSIZE*$$], wp$$);
0150 vst1q_u8(&q[d+NSIZE*$$], wq$$);
0151 }
0152 }
