1*4882a593Smuzhiyun // SPDX-License-Identifier: GPL-2.0-only
2*4882a593Smuzhiyun /*
3*4882a593Smuzhiyun * Copyright (C) 2012 Intel Corporation
4*4882a593Smuzhiyun * Copyright (C) 2017 Linaro Ltd. <ard.biesheuvel@linaro.org>
5*4882a593Smuzhiyun */
6*4882a593Smuzhiyun
7*4882a593Smuzhiyun #include <arm_neon.h>
8*4882a593Smuzhiyun
9*4882a593Smuzhiyun #ifdef CONFIG_ARM
10*4882a593Smuzhiyun /*
11*4882a593Smuzhiyun * AArch32 does not provide this intrinsic natively because it does not
12*4882a593Smuzhiyun * implement the underlying instruction. AArch32 only provides a 64-bit
13*4882a593Smuzhiyun * wide vtbl.8 instruction, so use that instead.
14*4882a593Smuzhiyun */
vqtbl1q_u8(uint8x16_t a,uint8x16_t b)15*4882a593Smuzhiyun static uint8x16_t vqtbl1q_u8(uint8x16_t a, uint8x16_t b)
16*4882a593Smuzhiyun {
17*4882a593Smuzhiyun union {
18*4882a593Smuzhiyun uint8x16_t val;
19*4882a593Smuzhiyun uint8x8x2_t pair;
20*4882a593Smuzhiyun } __a = { a };
21*4882a593Smuzhiyun
22*4882a593Smuzhiyun return vcombine_u8(vtbl2_u8(__a.pair, vget_low_u8(b)),
23*4882a593Smuzhiyun vtbl2_u8(__a.pair, vget_high_u8(b)));
24*4882a593Smuzhiyun }
25*4882a593Smuzhiyun #endif
26*4882a593Smuzhiyun
__raid6_2data_recov_neon(int bytes,uint8_t * p,uint8_t * q,uint8_t * dp,uint8_t * dq,const uint8_t * pbmul,const uint8_t * qmul)27*4882a593Smuzhiyun void __raid6_2data_recov_neon(int bytes, uint8_t *p, uint8_t *q, uint8_t *dp,
28*4882a593Smuzhiyun uint8_t *dq, const uint8_t *pbmul,
29*4882a593Smuzhiyun const uint8_t *qmul)
30*4882a593Smuzhiyun {
31*4882a593Smuzhiyun uint8x16_t pm0 = vld1q_u8(pbmul);
32*4882a593Smuzhiyun uint8x16_t pm1 = vld1q_u8(pbmul + 16);
33*4882a593Smuzhiyun uint8x16_t qm0 = vld1q_u8(qmul);
34*4882a593Smuzhiyun uint8x16_t qm1 = vld1q_u8(qmul + 16);
35*4882a593Smuzhiyun uint8x16_t x0f = vdupq_n_u8(0x0f);
36*4882a593Smuzhiyun
37*4882a593Smuzhiyun /*
38*4882a593Smuzhiyun * while ( bytes-- ) {
39*4882a593Smuzhiyun * uint8_t px, qx, db;
40*4882a593Smuzhiyun *
41*4882a593Smuzhiyun * px = *p ^ *dp;
42*4882a593Smuzhiyun * qx = qmul[*q ^ *dq];
43*4882a593Smuzhiyun * *dq++ = db = pbmul[px] ^ qx;
44*4882a593Smuzhiyun * *dp++ = db ^ px;
45*4882a593Smuzhiyun * p++; q++;
46*4882a593Smuzhiyun * }
47*4882a593Smuzhiyun */
48*4882a593Smuzhiyun
49*4882a593Smuzhiyun while (bytes) {
50*4882a593Smuzhiyun uint8x16_t vx, vy, px, qx, db;
51*4882a593Smuzhiyun
52*4882a593Smuzhiyun px = veorq_u8(vld1q_u8(p), vld1q_u8(dp));
53*4882a593Smuzhiyun vx = veorq_u8(vld1q_u8(q), vld1q_u8(dq));
54*4882a593Smuzhiyun
55*4882a593Smuzhiyun vy = vshrq_n_u8(vx, 4);
56*4882a593Smuzhiyun vx = vqtbl1q_u8(qm0, vandq_u8(vx, x0f));
57*4882a593Smuzhiyun vy = vqtbl1q_u8(qm1, vy);
58*4882a593Smuzhiyun qx = veorq_u8(vx, vy);
59*4882a593Smuzhiyun
60*4882a593Smuzhiyun vy = vshrq_n_u8(px, 4);
61*4882a593Smuzhiyun vx = vqtbl1q_u8(pm0, vandq_u8(px, x0f));
62*4882a593Smuzhiyun vy = vqtbl1q_u8(pm1, vy);
63*4882a593Smuzhiyun vx = veorq_u8(vx, vy);
64*4882a593Smuzhiyun db = veorq_u8(vx, qx);
65*4882a593Smuzhiyun
66*4882a593Smuzhiyun vst1q_u8(dq, db);
67*4882a593Smuzhiyun vst1q_u8(dp, veorq_u8(db, px));
68*4882a593Smuzhiyun
69*4882a593Smuzhiyun bytes -= 16;
70*4882a593Smuzhiyun p += 16;
71*4882a593Smuzhiyun q += 16;
72*4882a593Smuzhiyun dp += 16;
73*4882a593Smuzhiyun dq += 16;
74*4882a593Smuzhiyun }
75*4882a593Smuzhiyun }
76*4882a593Smuzhiyun
__raid6_datap_recov_neon(int bytes,uint8_t * p,uint8_t * q,uint8_t * dq,const uint8_t * qmul)77*4882a593Smuzhiyun void __raid6_datap_recov_neon(int bytes, uint8_t *p, uint8_t *q, uint8_t *dq,
78*4882a593Smuzhiyun const uint8_t *qmul)
79*4882a593Smuzhiyun {
80*4882a593Smuzhiyun uint8x16_t qm0 = vld1q_u8(qmul);
81*4882a593Smuzhiyun uint8x16_t qm1 = vld1q_u8(qmul + 16);
82*4882a593Smuzhiyun uint8x16_t x0f = vdupq_n_u8(0x0f);
83*4882a593Smuzhiyun
84*4882a593Smuzhiyun /*
85*4882a593Smuzhiyun * while (bytes--) {
86*4882a593Smuzhiyun * *p++ ^= *dq = qmul[*q ^ *dq];
87*4882a593Smuzhiyun * q++; dq++;
88*4882a593Smuzhiyun * }
89*4882a593Smuzhiyun */
90*4882a593Smuzhiyun
91*4882a593Smuzhiyun while (bytes) {
92*4882a593Smuzhiyun uint8x16_t vx, vy;
93*4882a593Smuzhiyun
94*4882a593Smuzhiyun vx = veorq_u8(vld1q_u8(q), vld1q_u8(dq));
95*4882a593Smuzhiyun
96*4882a593Smuzhiyun vy = vshrq_n_u8(vx, 4);
97*4882a593Smuzhiyun vx = vqtbl1q_u8(qm0, vandq_u8(vx, x0f));
98*4882a593Smuzhiyun vy = vqtbl1q_u8(qm1, vy);
99*4882a593Smuzhiyun vx = veorq_u8(vx, vy);
100*4882a593Smuzhiyun vy = veorq_u8(vx, vld1q_u8(p));
101*4882a593Smuzhiyun
102*4882a593Smuzhiyun vst1q_u8(dq, vx);
103*4882a593Smuzhiyun vst1q_u8(p, vy);
104*4882a593Smuzhiyun
105*4882a593Smuzhiyun bytes -= 16;
106*4882a593Smuzhiyun p += 16;
107*4882a593Smuzhiyun q += 16;
108*4882a593Smuzhiyun dq += 16;
109*4882a593Smuzhiyun }
110*4882a593Smuzhiyun }
111