Coverage Report

Created: 2024-11-29 06:10

/src/botan/src/lib/stream/chacha/chacha_avx2/chacha_avx2.cpp
Line
Count
Source (jump to first uncovered line)
1
/*
2
* (C) 2018 Jack Lloyd
3
*
4
* Botan is released under the Simplified BSD License (see license.txt)
5
*/
6
7
#include <botan/internal/chacha.h>
8
9
#include <botan/internal/simd_avx2.h>
10
11
namespace Botan {
12
13
//static
14
BOTAN_AVX2_FN
15
18.9k
void ChaCha::chacha_avx2_x8(uint8_t output[64 * 8], uint32_t state[16], size_t rounds) {
16
18.9k
   SIMD_8x32::reset_registers();
17
18
18.9k
   BOTAN_ASSERT(rounds % 2 == 0, "Valid rounds");
19
18.9k
   const SIMD_8x32 CTR0 = SIMD_8x32(0, 1, 2, 3, 4, 5, 6, 7);
20
21
18.9k
   const uint32_t C = 0xFFFFFFFF - state[12];
22
18.9k
   const SIMD_8x32 CTR1 = SIMD_8x32(0, C < 1, C < 2, C < 3, C < 4, C < 5, C < 6, C < 7);
23
24
18.9k
   SIMD_8x32 R00 = SIMD_8x32::splat(state[0]);
25
18.9k
   SIMD_8x32 R01 = SIMD_8x32::splat(state[1]);
26
18.9k
   SIMD_8x32 R02 = SIMD_8x32::splat(state[2]);
27
18.9k
   SIMD_8x32 R03 = SIMD_8x32::splat(state[3]);
28
18.9k
   SIMD_8x32 R04 = SIMD_8x32::splat(state[4]);
29
18.9k
   SIMD_8x32 R05 = SIMD_8x32::splat(state[5]);
30
18.9k
   SIMD_8x32 R06 = SIMD_8x32::splat(state[6]);
31
18.9k
   SIMD_8x32 R07 = SIMD_8x32::splat(state[7]);
32
18.9k
   SIMD_8x32 R08 = SIMD_8x32::splat(state[8]);
33
18.9k
   SIMD_8x32 R09 = SIMD_8x32::splat(state[9]);
34
18.9k
   SIMD_8x32 R10 = SIMD_8x32::splat(state[10]);
35
18.9k
   SIMD_8x32 R11 = SIMD_8x32::splat(state[11]);
36
18.9k
   SIMD_8x32 R12 = SIMD_8x32::splat(state[12]) + CTR0;
37
18.9k
   SIMD_8x32 R13 = SIMD_8x32::splat(state[13]) + CTR1;
38
18.9k
   SIMD_8x32 R14 = SIMD_8x32::splat(state[14]);
39
18.9k
   SIMD_8x32 R15 = SIMD_8x32::splat(state[15]);
40
41
208k
   for(size_t r = 0; r != rounds / 2; ++r) {
42
189k
      R00 += R04;
43
189k
      R01 += R05;
44
189k
      R02 += R06;
45
189k
      R03 += R07;
46
47
189k
      R12 ^= R00;
48
189k
      R13 ^= R01;
49
189k
      R14 ^= R02;
50
189k
      R15 ^= R03;
51
52
189k
      R12 = R12.rotl<16>();
53
189k
      R13 = R13.rotl<16>();
54
189k
      R14 = R14.rotl<16>();
55
189k
      R15 = R15.rotl<16>();
56
57
189k
      R08 += R12;
58
189k
      R09 += R13;
59
189k
      R10 += R14;
60
189k
      R11 += R15;
61
62
189k
      R04 ^= R08;
63
189k
      R05 ^= R09;
64
189k
      R06 ^= R10;
65
189k
      R07 ^= R11;
66
67
189k
      R04 = R04.rotl<12>();
68
189k
      R05 = R05.rotl<12>();
69
189k
      R06 = R06.rotl<12>();
70
189k
      R07 = R07.rotl<12>();
71
72
189k
      R00 += R04;
73
189k
      R01 += R05;
74
189k
      R02 += R06;
75
189k
      R03 += R07;
76
77
189k
      R12 ^= R00;
78
189k
      R13 ^= R01;
79
189k
      R14 ^= R02;
80
189k
      R15 ^= R03;
81
82
189k
      R12 = R12.rotl<8>();
83
189k
      R13 = R13.rotl<8>();
84
189k
      R14 = R14.rotl<8>();
85
189k
      R15 = R15.rotl<8>();
86
87
189k
      R08 += R12;
88
189k
      R09 += R13;
89
189k
      R10 += R14;
90
189k
      R11 += R15;
91
92
189k
      R04 ^= R08;
93
189k
      R05 ^= R09;
94
189k
      R06 ^= R10;
95
189k
      R07 ^= R11;
96
97
189k
      R04 = R04.rotl<7>();
98
189k
      R05 = R05.rotl<7>();
99
189k
      R06 = R06.rotl<7>();
100
189k
      R07 = R07.rotl<7>();
101
102
189k
      R00 += R05;
103
189k
      R01 += R06;
104
189k
      R02 += R07;
105
189k
      R03 += R04;
106
107
189k
      R15 ^= R00;
108
189k
      R12 ^= R01;
109
189k
      R13 ^= R02;
110
189k
      R14 ^= R03;
111
112
189k
      R15 = R15.rotl<16>();
113
189k
      R12 = R12.rotl<16>();
114
189k
      R13 = R13.rotl<16>();
115
189k
      R14 = R14.rotl<16>();
116
117
189k
      R10 += R15;
118
189k
      R11 += R12;
119
189k
      R08 += R13;
120
189k
      R09 += R14;
121
122
189k
      R05 ^= R10;
123
189k
      R06 ^= R11;
124
189k
      R07 ^= R08;
125
189k
      R04 ^= R09;
126
127
189k
      R05 = R05.rotl<12>();
128
189k
      R06 = R06.rotl<12>();
129
189k
      R07 = R07.rotl<12>();
130
189k
      R04 = R04.rotl<12>();
131
132
189k
      R00 += R05;
133
189k
      R01 += R06;
134
189k
      R02 += R07;
135
189k
      R03 += R04;
136
137
189k
      R15 ^= R00;
138
189k
      R12 ^= R01;
139
189k
      R13 ^= R02;
140
189k
      R14 ^= R03;
141
142
189k
      R15 = R15.rotl<8>();
143
189k
      R12 = R12.rotl<8>();
144
189k
      R13 = R13.rotl<8>();
145
189k
      R14 = R14.rotl<8>();
146
147
189k
      R10 += R15;
148
189k
      R11 += R12;
149
189k
      R08 += R13;
150
189k
      R09 += R14;
151
152
189k
      R05 ^= R10;
153
189k
      R06 ^= R11;
154
189k
      R07 ^= R08;
155
189k
      R04 ^= R09;
156
157
189k
      R05 = R05.rotl<7>();
158
189k
      R06 = R06.rotl<7>();
159
189k
      R07 = R07.rotl<7>();
160
189k
      R04 = R04.rotl<7>();
161
189k
   }
162
163
18.9k
   R00 += SIMD_8x32::splat(state[0]);
164
18.9k
   R01 += SIMD_8x32::splat(state[1]);
165
18.9k
   R02 += SIMD_8x32::splat(state[2]);
166
18.9k
   R03 += SIMD_8x32::splat(state[3]);
167
18.9k
   R04 += SIMD_8x32::splat(state[4]);
168
18.9k
   R05 += SIMD_8x32::splat(state[5]);
169
18.9k
   R06 += SIMD_8x32::splat(state[6]);
170
18.9k
   R07 += SIMD_8x32::splat(state[7]);
171
18.9k
   R08 += SIMD_8x32::splat(state[8]);
172
18.9k
   R09 += SIMD_8x32::splat(state[9]);
173
18.9k
   R10 += SIMD_8x32::splat(state[10]);
174
18.9k
   R11 += SIMD_8x32::splat(state[11]);
175
18.9k
   R12 += SIMD_8x32::splat(state[12]) + CTR0;
176
18.9k
   R13 += SIMD_8x32::splat(state[13]) + CTR1;
177
18.9k
   R14 += SIMD_8x32::splat(state[14]);
178
18.9k
   R15 += SIMD_8x32::splat(state[15]);
179
180
18.9k
   SIMD_8x32::transpose(R00, R01, R02, R03, R04, R05, R06, R07);
181
18.9k
   SIMD_8x32::transpose(R08, R09, R10, R11, R12, R13, R14, R15);
182
183
18.9k
   R00.store_le(output);
184
18.9k
   R08.store_le(output + 32 * 1);
185
18.9k
   R01.store_le(output + 32 * 2);
186
18.9k
   R09.store_le(output + 32 * 3);
187
18.9k
   R02.store_le(output + 32 * 4);
188
18.9k
   R10.store_le(output + 32 * 5);
189
18.9k
   R03.store_le(output + 32 * 6);
190
18.9k
   R11.store_le(output + 32 * 7);
191
18.9k
   R04.store_le(output + 32 * 8);
192
18.9k
   R12.store_le(output + 32 * 9);
193
18.9k
   R05.store_le(output + 32 * 10);
194
18.9k
   R13.store_le(output + 32 * 11);
195
18.9k
   R06.store_le(output + 32 * 12);
196
18.9k
   R14.store_le(output + 32 * 13);
197
18.9k
   R07.store_le(output + 32 * 14);
198
18.9k
   R15.store_le(output + 32 * 15);
199
200
18.9k
   SIMD_8x32::zero_registers();
201
202
18.9k
   state[12] += 8;
203
18.9k
   if(state[12] < 8) {
204
0
      state[13]++;
205
0
   }
206
18.9k
}
207
}  // namespace Botan