void cryptonight_hash_ctx_aes_ni(void* output, const void* input, size_t len, struct cryptonight_ctx* ctx) {
hash_process(&ctx->state.hs, (const uint8_t*) input, len);
ctx->aes_ctx = (oaes_ctx*) oaes_alloc();
size_t i, j;
memcpy(ctx->text, ctx->state.init, INIT_SIZE_BYTE);
oaes_key_import_data(ctx->aes_ctx, ctx->state.hs.b, AES_KEY_SIZE);
for (i = 0; likely(i < MEMORY); i += INIT_SIZE_BYTE) {
fast_aesb_pseudo_round_mut(&ctx->text[AES_BLOCK_SIZE * 0], ctx->aes_ctx->key->exp_data);
fast_aesb_pseudo_round_mut(&ctx->text[AES_BLOCK_SIZE * 1], ctx->aes_ctx->key->exp_data);
fast_aesb_pseudo_round_mut(&ctx->text[AES_BLOCK_SIZE * 2], ctx->aes_ctx->key->exp_data);
fast_aesb_pseudo_round_mut(&ctx->text[AES_BLOCK_SIZE * 3], ctx->aes_ctx->key->exp_data);
fast_aesb_pseudo_round_mut(&ctx->text[AES_BLOCK_SIZE * 4], ctx->aes_ctx->key->exp_data);
fast_aesb_pseudo_round_mut(&ctx->text[AES_BLOCK_SIZE * 5], ctx->aes_ctx->key->exp_data);
fast_aesb_pseudo_round_mut(&ctx->text[AES_BLOCK_SIZE * 6], ctx->aes_ctx->key->exp_data);
fast_aesb_pseudo_round_mut(&ctx->text[AES_BLOCK_SIZE * 7], ctx->aes_ctx->key->exp_data);
memcpy(&ctx->long_state[i], ctx->text, INIT_SIZE_BYTE);
}
xor_blocks_dst(&ctx->state.k[0], &ctx->state.k[32], ctx->a);
xor_blocks_dst(&ctx->state.k[16], &ctx->state.k[48], ctx->b);
for (i = 0; likely(i < ITER / 4); ++i) {
/* Dependency chain: address -> read value ------+
* written value <-+ hard function (AES or MUL) <+
* next address <-+
*/
/* Iteration 1 */
j = e2i(ctx->a);
fast_aesb_single_round(&ctx->long_state[j], ctx->c, ctx->a);
xor_blocks_dst(ctx->c, ctx->b, &ctx->long_state[j]);
/* Iteration 2 */
mul_sum_xor_dst(ctx->c, ctx->a, &ctx->long_state[e2i(ctx->c)]);
/* Iteration 3 */
j = e2i(ctx->a);
fast_aesb_single_round(&ctx->long_state[j], ctx->b, ctx->a);
xor_blocks_dst(ctx->b, ctx->c, &ctx->long_state[j]);
/* Iteration 4 */
mul_sum_xor_dst(ctx->b, ctx->a, &ctx->long_state[e2i(ctx->b)]);
}
memcpy(ctx->text, ctx->state.init, INIT_SIZE_BYTE);
oaes_key_import_data(ctx->aes_ctx, &ctx->state.hs.b[32], AES_KEY_SIZE);
for (i = 0; likely(i < MEMORY); i += INIT_SIZE_BYTE) {
xor_blocks(&ctx->text[0 * AES_BLOCK_SIZE], &ctx->long_state[i + 0 * AES_BLOCK_SIZE]);
fast_aesb_pseudo_round_mut(&ctx->text[0 * AES_BLOCK_SIZE], ctx->aes_ctx->key->exp_data);
xor_blocks(&ctx->text[1 * AES_BLOCK_SIZE], &ctx->long_state[i + 1 * AES_BLOCK_SIZE]);
fast_aesb_pseudo_round_mut(&ctx->text[1 * AES_BLOCK_SIZE], ctx->aes_ctx->key->exp_data);
xor_blocks(&ctx->text[2 * AES_BLOCK_SIZE], &ctx->long_state[i + 2 * AES_BLOCK_SIZE]);
fast_aesb_pseudo_round_mut(&ctx->text[2 * AES_BLOCK_SIZE], ctx->aes_ctx->key->exp_data);
xor_blocks(&ctx->text[3 * AES_BLOCK_SIZE], &ctx->long_state[i + 3 * AES_BLOCK_SIZE]);
fast_aesb_pseudo_round_mut(&ctx->text[3 * AES_BLOCK_SIZE], ctx->aes_ctx->key->exp_data);
xor_blocks(&ctx->text[4 * AES_BLOCK_SIZE], &ctx->long_state[i + 4 * AES_BLOCK_SIZE]);
fast_aesb_pseudo_round_mut(&ctx->text[4 * AES_BLOCK_SIZE], ctx->aes_ctx->key->exp_data);
xor_blocks(&ctx->text[5 * AES_BLOCK_SIZE], &ctx->long_state[i + 5 * AES_BLOCK_SIZE]);
fast_aesb_pseudo_round_mut(&ctx->text[5 * AES_BLOCK_SIZE], ctx->aes_ctx->key->exp_data);
xor_blocks(&ctx->text[6 * AES_BLOCK_SIZE], &ctx->long_state[i + 6 * AES_BLOCK_SIZE]);
fast_aesb_pseudo_round_mut(&ctx->text[6 * AES_BLOCK_SIZE], ctx->aes_ctx->key->exp_data);
xor_blocks(&ctx->text[7 * AES_BLOCK_SIZE], &ctx->long_state[i + 7 * AES_BLOCK_SIZE]);
fast_aesb_pseudo_round_mut(&ctx->text[7 * AES_BLOCK_SIZE], ctx->aes_ctx->key->exp_data);
}
memcpy(ctx->state.init, ctx->text, INIT_SIZE_BYTE);
hash_permutation(&ctx->state.hs);
/*memcpy(hash, &state, 32);*/
extra_hashes[ctx->state.hs.b[0] & 3](&ctx->state, 200, output);
oaes_free((OAES_CTX **) &ctx->aes_ctx);
}
void cryptonight_hash(const char* input, char* output, uint32_t len, int variant, uint64_t height) {
struct cryptonight_ctx *ctx = alloca(sizeof(struct cryptonight_ctx));
hash_process(&ctx->state.hs, (const uint8_t*) input, len);
memcpy(ctx->text, ctx->state.init, INIT_SIZE_BYTE);
memcpy(ctx->aes_key, ctx->state.hs.b, AES_KEY_SIZE);
ctx->aes_ctx = (oaes_ctx*) oaes_alloc();
size_t i, j;
VARIANT1_INIT();
VARIANT2_INIT(ctx->b, ctx->state);
VARIANT4_RANDOM_MATH_INIT(ctx->state);
oaes_key_import_data(ctx->aes_ctx, ctx->aes_key, AES_KEY_SIZE);
for (i = 0; i < MEMORY / INIT_SIZE_BYTE; i++) {
for (j = 0; j < INIT_SIZE_BLK; j++) {
aesb_pseudo_round(&ctx->text[AES_BLOCK_SIZE * j],
&ctx->text[AES_BLOCK_SIZE * j],
ctx->aes_ctx->key->exp_data);
}
memcpy(&ctx->long_state[i * INIT_SIZE_BYTE], ctx->text, INIT_SIZE_BYTE);
}
for (i = 0; i < 16; i++) {
ctx->a[i] = ctx->state.k[i] ^ ctx->state.k[32 + i];
ctx->b[i] = ctx->state.k[16 + i] ^ ctx->state.k[48 + i];
}
for (i = 0; i < ITER / 2; i++) {
/* Dependency chain: address -> read value ------+
* written value <-+ hard function (AES or MUL) <+
* next address <-+
*/
/* Iteration 1 */
j = e2i(ctx->a);
aesb_single_round(&ctx->long_state[j * AES_BLOCK_SIZE], ctx->c, ctx->a);
VARIANT2_SHUFFLE_ADD(ctx->long_state, j * AES_BLOCK_SIZE, ctx->a, ctx->b, ctx->c);
xor_blocks_dst(ctx->c, ctx->b, &ctx->long_state[j * AES_BLOCK_SIZE]);
VARIANT1_1((uint8_t*)&ctx->long_state[j * AES_BLOCK_SIZE]);
/* Iteration 2 */
j = e2i(ctx->c);
uint64_t* dst = (uint64_t*)&ctx->long_state[j * AES_BLOCK_SIZE];
uint64_t t[2];
t[0] = dst[0];
t[1] = dst[1];
VARIANT2_INTEGER_MATH(t, ctx->c);
copy_block(ctx->a1, ctx->a);
VARIANT4_RANDOM_MATH(ctx->a, t, r, ctx->b, ctx->b + AES_BLOCK_SIZE);
uint64_t hi;
uint64_t lo = mul128(((uint64_t*)ctx->c)[0], t[0], &hi);
VARIANT2_2();
VARIANT2_SHUFFLE_ADD(ctx->long_state, j * AES_BLOCK_SIZE, ctx->a1, ctx->b, ctx->c);
((uint64_t*)ctx->a)[0] += hi;
((uint64_t*)ctx->a)[1] += lo;
dst[0] = ((uint64_t*)ctx->a)[0];
dst[1] = ((uint64_t*)ctx->a)[1];
((uint64_t*)ctx->a)[0] ^= t[0];
((uint64_t*)ctx->a)[1] ^= t[1];
VARIANT1_2((uint8_t*)&ctx->long_state[j * AES_BLOCK_SIZE]);
copy_block(ctx->b + AES_BLOCK_SIZE, ctx->b);
copy_block(ctx->b, ctx->c);
}
memcpy(ctx->text, ctx->state.init, INIT_SIZE_BYTE);
oaes_key_import_data(ctx->aes_ctx, &ctx->state.hs.b[32], AES_KEY_SIZE);
for (i = 0; i < MEMORY / INIT_SIZE_BYTE; i++) {
for (j = 0; j < INIT_SIZE_BLK; j++) {
xor_blocks(&ctx->text[j * AES_BLOCK_SIZE],
&ctx->long_state[i * INIT_SIZE_BYTE + j * AES_BLOCK_SIZE]);
aesb_pseudo_round(&ctx->text[j * AES_BLOCK_SIZE],
&ctx->text[j * AES_BLOCK_SIZE],
ctx->aes_ctx->key->exp_data);
}
}
memcpy(ctx->state.init, ctx->text, INIT_SIZE_BYTE);
hash_permutation(&ctx->state.hs);
/*memcpy(hash, &state, 32);*/
extra_hashes[ctx->state.hs.b[0] & 3](&ctx->state, 200, output);
oaes_free((OAES_CTX **) &ctx->aes_ctx);
}