1/*
2 * version 20110505
3 * D. J. Bernstein
4 * Public domain.
5 *
6 * Based on crypto_core/salsa208/armneon/core.c from SUPERCOP 20130419
7 */
8
9#define ROUNDS 8
10static void
11salsa20_8_intrinsic(void * input)
12{
13  int i;
14
15  const uint32x4_t abab = {-1,0,-1,0};
16
17  /*
18   * This is modified since we only have one argument. Usually you'd rearrange
19   * the constant, key, and input bytes, but we just have one linear array to
20   * rearrange which is a bit easier.
21   */
22
23  /*
24   * Change the input to be diagonals as if it's a 4x4 matrix of 32-bit values.
25   */
26  uint32x4_t x0x5x10x15;
27  uint32x4_t x12x1x6x11;
28  uint32x4_t x8x13x2x7;
29  uint32x4_t x4x9x14x3;
30
31  uint32x4_t x0x1x10x11;
32  uint32x4_t x12x13x6x7;
33  uint32x4_t x8x9x2x3;
34  uint32x4_t x4x5x14x15;
35
36  uint32x4_t x0x1x2x3;
37  uint32x4_t x4x5x6x7;
38  uint32x4_t x8x9x10x11;
39  uint32x4_t x12x13x14x15;
40
41  x0x1x2x3 = vld1q_u8((uint8_t *) input);
42  x4x5x6x7 = vld1q_u8(16 + (uint8_t *) input);
43  x8x9x10x11 = vld1q_u8(32 + (uint8_t *) input);
44  x12x13x14x15 = vld1q_u8(48 + (uint8_t *) input);
45
46  x0x1x10x11 = vcombine_u32(vget_low_u32(x0x1x2x3), vget_high_u32(x8x9x10x11));
47  x4x5x14x15 = vcombine_u32(vget_low_u32(x4x5x6x7), vget_high_u32(x12x13x14x15));
48  x8x9x2x3 = vcombine_u32(vget_low_u32(x8x9x10x11), vget_high_u32(x0x1x2x3));
49  x12x13x6x7 = vcombine_u32(vget_low_u32(x12x13x14x15), vget_high_u32(x4x5x6x7));
50
51  x0x5x10x15 = vbslq_u32(abab,x0x1x10x11,x4x5x14x15);
52  x8x13x2x7 = vbslq_u32(abab,x8x9x2x3,x12x13x6x7);
53  x4x9x14x3 = vbslq_u32(abab,x4x5x14x15,x8x9x2x3);
54  x12x1x6x11 = vbslq_u32(abab,x12x13x6x7,x0x1x10x11);
55
56  uint32x4_t start0 = x0x5x10x15;
57  uint32x4_t start1 = x12x1x6x11;
58  uint32x4_t start3 = x4x9x14x3;
59  uint32x4_t start2 = x8x13x2x7;
60
61  /* From here on this should be the same as the SUPERCOP version. */
62
63  uint32x4_t diag0 = start0;
64  uint32x4_t diag1 = start1;
65  uint32x4_t diag2 = start2;
66  uint32x4_t diag3 = start3;
67
68  uint32x4_t a0;
69  uint32x4_t a1;
70  uint32x4_t a2;
71  uint32x4_t a3;
72
73  for (i = ROUNDS;i > 0;i -= 2) {
74    a0 = diag1 + diag0;
75    diag3 ^= vsriq_n_u32(vshlq_n_u32(a0,7),a0,25);
76    a1 = diag0 + diag3;
77    diag2 ^= vsriq_n_u32(vshlq_n_u32(a1,9),a1,23);
78    a2 = diag3 + diag2;
79    diag1 ^= vsriq_n_u32(vshlq_n_u32(a2,13),a2,19);
80    a3 = diag2 + diag1;
81    diag0 ^= vsriq_n_u32(vshlq_n_u32(a3,18),a3,14);
82
83    diag3 = vextq_u32(diag3,diag3,3);
84    diag2 = vextq_u32(diag2,diag2,2);
85    diag1 = vextq_u32(diag1,diag1,1);
86
87    a0 = diag3 + diag0;
88    diag1 ^= vsriq_n_u32(vshlq_n_u32(a0,7),a0,25);
89    a1 = diag0 + diag1;
90    diag2 ^= vsriq_n_u32(vshlq_n_u32(a1,9),a1,23);
91    a2 = diag1 + diag2;
92    diag3 ^= vsriq_n_u32(vshlq_n_u32(a2,13),a2,19);
93    a3 = diag2 + diag3;
94    diag0 ^= vsriq_n_u32(vshlq_n_u32(a3,18),a3,14);
95
96    diag1 = vextq_u32(diag1,diag1,3);
97    diag2 = vextq_u32(diag2,diag2,2);
98    diag3 = vextq_u32(diag3,diag3,1);
99  }
100
101  x0x5x10x15 = diag0 + start0;
102  x12x1x6x11 = diag1 + start1;
103  x8x13x2x7 = diag2 + start2;
104  x4x9x14x3 = diag3 + start3;
105
106  x0x1x10x11 = vbslq_u32(abab,x0x5x10x15,x12x1x6x11);
107  x12x13x6x7 = vbslq_u32(abab,x12x1x6x11,x8x13x2x7);
108  x8x9x2x3 = vbslq_u32(abab,x8x13x2x7,x4x9x14x3);
109  x4x5x14x15 = vbslq_u32(abab,x4x9x14x3,x0x5x10x15);
110
111  x0x1x2x3 = vcombine_u32(vget_low_u32(x0x1x10x11),vget_high_u32(x8x9x2x3));
112  x4x5x6x7 = vcombine_u32(vget_low_u32(x4x5x14x15),vget_high_u32(x12x13x6x7));
113  x8x9x10x11 = vcombine_u32(vget_low_u32(x8x9x2x3),vget_high_u32(x0x1x10x11));
114  x12x13x14x15 = vcombine_u32(vget_low_u32(x12x13x6x7),vget_high_u32(x4x5x14x15));
115
116  vst1q_u8((uint8_t *) input,(uint8x16_t) x0x1x2x3);
117  vst1q_u8(16 + (uint8_t *) input,(uint8x16_t) x4x5x6x7);
118  vst1q_u8(32 + (uint8_t *) input,(uint8x16_t) x8x9x10x11);
119  vst1q_u8(48 + (uint8_t *) input,(uint8x16_t) x12x13x14x15);
120}
121