1/*
2 * Argon2 source code package
3 *
4 * Written by Daniel Dinu and Dmitry Khovratovich, 2015
5 *
6 * This work is licensed under a Creative Commons CC0 1.0 License/Waiver.
7 *
8 * You should have received a copy of the CC0 Public Domain Dedication along
9 * with
10 * this software. If not, see
11 * <http://creativecommons.org/publicdomain/zero/1.0/>.
12 */
13
14#include <stdint.h>
15#include <stdlib.h>
16#include <string.h>
17
18#include "argon2-core.h"
19#include "argon2.h"
20#include "private/common.h"
21#include "private/sse2_64_32.h"
22
23#if defined(HAVE_AVX512FINTRIN_H) && defined(HAVE_AVX2INTRIN_H) && \
24    defined(HAVE_EMMINTRIN_H) &&  defined(HAVE_TMMINTRIN_H) && defined(HAVE_SMMINTRIN_H)
25
26# ifdef __GNUC__
27#  pragma GCC target("sse2")
28#  pragma GCC target("ssse3")
29#  pragma GCC target("sse4.1")
30#  pragma GCC target("avx2")
31#  pragma GCC target("avx512f")
32# endif
33
34# ifdef _MSC_VER
35#  include <intrin.h> /* for _mm_set_epi64x */
36# endif
37#include <emmintrin.h>
38#include <immintrin.h>
39#include <smmintrin.h>
40#include <tmmintrin.h>
41
42# include "blamka-round-avx512f.h"
43
44static void
45fill_block(__m512i *state, const uint8_t *ref_block, uint8_t *next_block)
46{
47    __m512i  block_XY[ARGON2_512BIT_WORDS_IN_BLOCK];
48    uint32_t i;
49
50    for (i = 0; i < ARGON2_512BIT_WORDS_IN_BLOCK; i++) {
51        block_XY[i] = state[i] = _mm512_xor_si512(
52            state[i], _mm512_loadu_si512((__m512i const *) (&ref_block[64 * i])));
53    }
54
55    for (i = 0; i < 2; ++i) {
56        BLAKE2_ROUND_1(
57            state[8 * i + 0], state[8 * i + 1], state[8 * i + 2], state[8 * i + 3],
58            state[8 * i + 4], state[8 * i + 5], state[8 * i + 6], state[8 * i + 7]);
59    }
60
61    for (i = 0; i < 2; ++i) {
62        BLAKE2_ROUND_2(
63            state[2 * 0 + i], state[2 * 1 + i], state[2 * 2 + i], state[2 * 3 + i],
64            state[2 * 4 + i], state[2 * 5 + i], state[2 * 6 + i], state[2 * 7 + i]);
65    }
66
67    for (i = 0; i < ARGON2_512BIT_WORDS_IN_BLOCK; i++) {
68        state[i] = _mm512_xor_si512(state[i], block_XY[i]);
69        _mm512_storeu_si512((__m512i *) (&next_block[64 * i]), state[i]);
70    }
71}
72
73static void
74fill_block_with_xor(__m512i *state, const uint8_t *ref_block,
75                    uint8_t *next_block)
76{
77    __m512i  block_XY[ARGON2_512BIT_WORDS_IN_BLOCK];
78    uint32_t i;
79
80    for (i = 0; i < ARGON2_512BIT_WORDS_IN_BLOCK; i++) {
81        state[i] = _mm512_xor_si512(
82            state[i], _mm512_loadu_si512((__m512i const *) (&ref_block[64 * i])));
83        block_XY[i] = _mm512_xor_si512(
84            state[i], _mm512_loadu_si512((__m512i const *) (&next_block[64 * i])));
85    }
86
87    for (i = 0; i < 2; ++i) {
88        BLAKE2_ROUND_1(
89            state[8 * i + 0], state[8 * i + 1], state[8 * i + 2], state[8 * i + 3],
90            state[8 * i + 4], state[8 * i + 5], state[8 * i + 6], state[8 * i + 7]);
91    }
92
93    for (i = 0; i < 2; ++i) {
94        BLAKE2_ROUND_2(
95            state[2 * 0 + i], state[2 * 1 + i], state[2 * 2 + i], state[2 * 3 + i],
96            state[2 * 4 + i], state[2 * 5 + i], state[2 * 6 + i], state[2 * 7 + i]);
97    }
98
99    for (i = 0; i < ARGON2_512BIT_WORDS_IN_BLOCK; i++) {
100        state[i] = _mm512_xor_si512(state[i], block_XY[i]);
101        _mm512_storeu_si512((__m512i *) (&next_block[64 * i]), state[i]);
102    }
103}
104
105static void
106generate_addresses(const argon2_instance_t *instance,
107                   const argon2_position_t *position, uint64_t *pseudo_rands)
108{
109    block    address_block, input_block, tmp_block;
110    uint32_t i;
111
112    init_block_value(&address_block, 0);
113    init_block_value(&input_block, 0);
114
115    if (instance != NULL && position != NULL) {
116        input_block.v[0] = position->pass;
117        input_block.v[1] = position->lane;
118        input_block.v[2] = position->slice;
119        input_block.v[3] = instance->memory_blocks;
120        input_block.v[4] = instance->passes;
121        input_block.v[5] = instance->type;
122
123        for (i = 0; i < instance->segment_length; ++i) {
124            if (i % ARGON2_ADDRESSES_IN_BLOCK == 0) {
125                /* Temporary zero-initialized blocks */
126                __m512i zero_block[ARGON2_512BIT_WORDS_IN_BLOCK];
127                __m512i zero2_block[ARGON2_512BIT_WORDS_IN_BLOCK];
128
129                memset(zero_block, 0, sizeof(zero_block));
130                memset(zero2_block, 0, sizeof(zero2_block));
131                init_block_value(&address_block, 0);
132                init_block_value(&tmp_block, 0);
133                /* Increasing index counter */
134                input_block.v[6]++;
135                /* First iteration of G */
136                fill_block_with_xor(zero_block, (uint8_t *) &input_block.v,
137                                    (uint8_t *) &tmp_block.v);
138                /* Second iteration of G */
139                fill_block_with_xor(zero2_block, (uint8_t *) &tmp_block.v,
140                                    (uint8_t *) &address_block.v);
141            }
142
143            pseudo_rands[i] = address_block.v[i % ARGON2_ADDRESSES_IN_BLOCK];
144        }
145    }
146}
147
148void
149fill_segment_avx512f(const argon2_instance_t *instance,
150                     argon2_position_t        position)
151{
152    block    *ref_block = NULL, *curr_block = NULL;
153    uint64_t  pseudo_rand, ref_index, ref_lane;
154    uint32_t  prev_offset, curr_offset;
155    uint32_t  starting_index, i;
156    __m512i   state[ARGON2_512BIT_WORDS_IN_BLOCK];
157    int       data_independent_addressing = 1;
158
159    /* Pseudo-random values that determine the reference block position */
160    uint64_t *pseudo_rands = NULL;
161
162    if (instance == NULL) {
163        return;
164    }
165
166    if (instance->type == Argon2_id &&
167        (position.pass != 0 || position.slice >= ARGON2_SYNC_POINTS / 2)) {
168        data_independent_addressing = 0;
169    }
170
171    pseudo_rands = instance->pseudo_rands;
172
173    if (data_independent_addressing) {
174        generate_addresses(instance, &position, pseudo_rands);
175    }
176
177    starting_index = 0;
178
179    if ((0 == position.pass) && (0 == position.slice)) {
180        starting_index = 2; /* we have already generated the first two blocks */
181    }
182
183    /* Offset of the current block */
184    curr_offset = position.lane * instance->lane_length +
185                  position.slice * instance->segment_length + starting_index;
186
187    if (0 == curr_offset % instance->lane_length) {
188        /* Last block in this lane */
189        prev_offset = curr_offset + instance->lane_length - 1;
190    } else {
191        /* Previous block */
192        prev_offset = curr_offset - 1;
193    }
194
195    memcpy(state, ((instance->region->memory + prev_offset)->v),
196           ARGON2_BLOCK_SIZE);
197
198    for (i = starting_index; i < instance->segment_length;
199         ++i, ++curr_offset, ++prev_offset) {
200        /*1.1 Rotating prev_offset if needed */
201        if (curr_offset % instance->lane_length == 1) {
202            prev_offset = curr_offset - 1;
203        }
204
205        /* 1.2 Computing the index of the reference block */
206        /* 1.2.1 Taking pseudo-random value from the previous block */
207        if (data_independent_addressing) {
208#pragma warning(push)
209#pragma warning(disable : 6385)
210            pseudo_rand = pseudo_rands[i];
211#pragma warning(pop)
212        } else {
213            pseudo_rand = instance->region->memory[prev_offset].v[0];
214        }
215
216        /* 1.2.2 Computing the lane of the reference block */
217        ref_lane = ((pseudo_rand >> 32)) % instance->lanes;
218
219        if ((position.pass == 0) && (position.slice == 0)) {
220            /* Can not reference other lanes yet */
221            ref_lane = position.lane;
222        }
223
224        /* 1.2.3 Computing the number of possible reference block within the
225         * lane.
226         */
227        position.index = i;
228        ref_index = index_alpha(instance, &position, pseudo_rand & 0xFFFFFFFF,
229                                ref_lane == position.lane);
230
231        /* 2 Creating a new block */
232        ref_block = instance->region->memory +
233                    instance->lane_length * ref_lane + ref_index;
234        curr_block = instance->region->memory + curr_offset;
235        if (position.pass != 0) {
236            fill_block_with_xor(state, (uint8_t *) ref_block->v,
237                                (uint8_t *) curr_block->v);
238        } else {
239            fill_block(state, (uint8_t *) ref_block->v,
240                       (uint8_t *) curr_block->v);
241        }
242    }
243}
244#endif
245