| Ard Biesheuvel | 6ec4e251 | 2017-07-13 18:16:01 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 2012 Intel Corporation |
| 3 | * Copyright (C) 2017 Linaro Ltd. <ard.biesheuvel@linaro.org> |
| 4 | * |
| 5 | * This program is free software; you can redistribute it and/or |
| 6 | * modify it under the terms of the GNU General Public License |
| 7 | * as published by the Free Software Foundation; version 2 |
| 8 | * of the License. |
| 9 | */ |
| 10 | |
| 11 | #include <arm_neon.h> |
| 12 | |
| 13 | static const uint8x16_t x0f = { |
| 14 | 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, |
| 15 | 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, |
| 16 | }; |
| 17 | |
| 18 | #ifdef CONFIG_ARM |
| 19 | /* |
| 20 | * AArch32 does not provide this intrinsic natively because it does not |
| 21 | * implement the underlying instruction. AArch32 only provides a 64-bit |
| 22 | * wide vtbl.8 instruction, so use that instead. |
| 23 | */ |
| 24 | static uint8x16_t vqtbl1q_u8(uint8x16_t a, uint8x16_t b) |
| 25 | { |
| 26 | union { |
| 27 | uint8x16_t val; |
| 28 | uint8x8x2_t pair; |
| 29 | } __a = { a }; |
| 30 | |
| 31 | return vcombine_u8(vtbl2_u8(__a.pair, vget_low_u8(b)), |
| 32 | vtbl2_u8(__a.pair, vget_high_u8(b))); |
| 33 | } |
| 34 | #endif |
| 35 | |
| 36 | void __raid6_2data_recov_neon(int bytes, uint8_t *p, uint8_t *q, uint8_t *dp, |
| 37 | uint8_t *dq, const uint8_t *pbmul, |
| 38 | const uint8_t *qmul) |
| 39 | { |
| 40 | uint8x16_t pm0 = vld1q_u8(pbmul); |
| 41 | uint8x16_t pm1 = vld1q_u8(pbmul + 16); |
| 42 | uint8x16_t qm0 = vld1q_u8(qmul); |
| 43 | uint8x16_t qm1 = vld1q_u8(qmul + 16); |
| 44 | |
| 45 | /* |
| 46 | * while ( bytes-- ) { |
| 47 | * uint8_t px, qx, db; |
| 48 | * |
| 49 | * px = *p ^ *dp; |
| 50 | * qx = qmul[*q ^ *dq]; |
| 51 | * *dq++ = db = pbmul[px] ^ qx; |
| 52 | * *dp++ = db ^ px; |
| 53 | * p++; q++; |
| 54 | * } |
| 55 | */ |
| 56 | |
| 57 | while (bytes) { |
| 58 | uint8x16_t vx, vy, px, qx, db; |
| 59 | |
| 60 | px = veorq_u8(vld1q_u8(p), vld1q_u8(dp)); |
| 61 | vx = veorq_u8(vld1q_u8(q), vld1q_u8(dq)); |
| 62 | |
| 63 | vy = (uint8x16_t)vshrq_n_s16((int16x8_t)vx, 4); |
| 64 | vx = vqtbl1q_u8(qm0, vandq_u8(vx, x0f)); |
| 65 | vy = vqtbl1q_u8(qm1, vandq_u8(vy, x0f)); |
| 66 | qx = veorq_u8(vx, vy); |
| 67 | |
| 68 | vy = (uint8x16_t)vshrq_n_s16((int16x8_t)px, 4); |
| 69 | vx = vqtbl1q_u8(pm0, vandq_u8(px, x0f)); |
| 70 | vy = vqtbl1q_u8(pm1, vandq_u8(vy, x0f)); |
| 71 | vx = veorq_u8(vx, vy); |
| 72 | db = veorq_u8(vx, qx); |
| 73 | |
| 74 | vst1q_u8(dq, db); |
| 75 | vst1q_u8(dp, veorq_u8(db, px)); |
| 76 | |
| 77 | bytes -= 16; |
| 78 | p += 16; |
| 79 | q += 16; |
| 80 | dp += 16; |
| 81 | dq += 16; |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | void __raid6_datap_recov_neon(int bytes, uint8_t *p, uint8_t *q, uint8_t *dq, |
| 86 | const uint8_t *qmul) |
| 87 | { |
| 88 | uint8x16_t qm0 = vld1q_u8(qmul); |
| 89 | uint8x16_t qm1 = vld1q_u8(qmul + 16); |
| 90 | |
| 91 | /* |
| 92 | * while (bytes--) { |
| 93 | * *p++ ^= *dq = qmul[*q ^ *dq]; |
| 94 | * q++; dq++; |
| 95 | * } |
| 96 | */ |
| 97 | |
| 98 | while (bytes) { |
| 99 | uint8x16_t vx, vy; |
| 100 | |
| 101 | vx = veorq_u8(vld1q_u8(q), vld1q_u8(dq)); |
| 102 | |
| 103 | vy = (uint8x16_t)vshrq_n_s16((int16x8_t)vx, 4); |
| 104 | vx = vqtbl1q_u8(qm0, vandq_u8(vx, x0f)); |
| 105 | vy = vqtbl1q_u8(qm1, vandq_u8(vy, x0f)); |
| 106 | vx = veorq_u8(vx, vy); |
| 107 | vy = veorq_u8(vx, vld1q_u8(p)); |
| 108 | |
| 109 | vst1q_u8(dq, vx); |
| 110 | vst1q_u8(p, vy); |
| 111 | |
| 112 | bytes -= 16; |
| 113 | p += 16; |
| 114 | q += 16; |
| 115 | dq += 16; |
| 116 | } |
| 117 | } |