Add memory to cryptonight_ctx.

This commit is contained in:
XMRig 2017-04-21 11:56:11 +03:00
parent f29d05bdde
commit 8ab4c1c8bd
13 changed files with 85 additions and 81 deletions

View file

@ -40,6 +40,6 @@ extern void (* const extra_hashes[4])(const void *, size_t, char *);
void cryptonight_init(int variant);
void cryptonight_hash(void* output, const void* input, size_t input_len);
int scanhash_cryptonight(int thr_id, uint32_t *hash, uint32_t *restrict pdata, const uint32_t *restrict ptarget, uint32_t max_nonce, unsigned long *restrict hashes_done, const char *memory, struct cryptonight_ctx *persistentctx);
int scanhash_cryptonight(int thr_id, uint32_t *hash, uint32_t *restrict pdata, const uint32_t *restrict ptarget, uint32_t max_nonce, unsigned long *restrict hashes_done, struct cryptonight_ctx *restrict ctx);
#endif /* __CRYPTONIGHT_H__ */

View file

@ -222,13 +222,13 @@ static inline void cn_implode_scratchpad(const __m128i* input, __m128i* output)
}
void cryptonight_av1_aesni(void *restrict output, const void *restrict input, char *restrict memory, struct cryptonight_ctx *restrict ctx)
void cryptonight_av1_aesni(void *restrict output, const void *restrict input, struct cryptonight_ctx *restrict ctx)
{
keccak((const uint8_t *) input, 76, ctx->state, 200);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) memory);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) ctx->memory);
const uint8_t* l0 = memory;
const uint8_t* l0 = ctx->memory;
uint64_t* h0 = (uint64_t*) ctx->state;
uint64_t al0 = h0[0] ^ h0[4];
@ -239,16 +239,16 @@ void cryptonight_av1_aesni(void *restrict output, const void *restrict input, ch
for (size_t i = 0; __builtin_expect(i < 0x80000, 1); i++) {
__m128i cx;
cx = _mm_load_si128((__m128i *)&l0[idx0 & 0x1FFFF0]);
cx = _mm_load_si128((__m128i *) &l0[idx0 & 0x1FFFF0]);
cx = _mm_aesenc_si128(cx, _mm_set_epi64x(ah0, al0));
_mm_store_si128((__m128i *)&l0[idx0 & 0x1FFFF0], _mm_xor_si128(bx0, cx));
_mm_store_si128((__m128i *) &l0[idx0 & 0x1FFFF0], _mm_xor_si128(bx0, cx));
idx0 = _mm_cvtsi128_si64(cx);
bx0 = cx;
uint64_t hi, lo, cl, ch;
cl = ((uint64_t*)&l0[idx0 & 0x1FFFF0])[0];
ch = ((uint64_t*)&l0[idx0 & 0x1FFFF0])[1];
cl = ((uint64_t*) &l0[idx0 & 0x1FFFF0])[0];
ch = ((uint64_t*) &l0[idx0 & 0x1FFFF0])[1];
lo = _umul128(idx0, cl, &hi);
al0 += hi;
@ -262,7 +262,7 @@ void cryptonight_av1_aesni(void *restrict output, const void *restrict input, ch
idx0 = al0;
}
cn_implode_scratchpad((__m128i*) memory, (__m128i*) ctx->state);
cn_implode_scratchpad((__m128i*) ctx->memory, (__m128i*) ctx->state);
keccakf(h0, 24);
extra_hashes[ctx->state[0] & 3](ctx->state, 200, output);

View file

@ -222,13 +222,13 @@ static inline void cn_implode_scratchpad(const __m128i* input, __m128i* output)
}
void cryptonight_av2_aesni_stak(void *restrict output, const void *restrict input, char *restrict memory, struct cryptonight_ctx *restrict ctx)
void cryptonight_av2_aesni_stak(void *restrict output, const void *restrict input, struct cryptonight_ctx *restrict ctx)
{
keccak((const uint8_t *) input, 76, ctx->state, 200);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) memory);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) ctx->memory);
const uint8_t* l0 = memory;
const uint8_t* l0 = ctx->memory;
uint64_t* h0 = (uint64_t*) ctx->state;
uint64_t al0 = h0[0] ^ h0[4];
@ -266,7 +266,7 @@ void cryptonight_av2_aesni_stak(void *restrict output, const void *restrict inpu
_mm_prefetch((const char*)&l0[idx0 & 0x1FFFF0], _MM_HINT_T0);
}
cn_implode_scratchpad((__m128i*) memory, (__m128i*) ctx->state);
cn_implode_scratchpad((__m128i*) ctx->memory, (__m128i*) ctx->state);
keccakf(h0, 24);
extra_hashes[ctx->state[0] & 3](ctx->state, 200, output);

View file

@ -222,13 +222,13 @@ static inline void cn_implode_scratchpad(const __m128i* input, __m128i* output)
}
void cryptonight_av3_aesni_bmi2(void *restrict output, const void *restrict input, char *restrict memory, struct cryptonight_ctx *restrict ctx)
void cryptonight_av3_aesni_bmi2(void *restrict output, const void *restrict input, struct cryptonight_ctx *restrict ctx)
{
keccak((const uint8_t *) input, 76, ctx->state, 200);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) memory);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) ctx->memory);
const uint8_t* l0 = memory;
const uint8_t* l0 = ctx->memory;
uint64_t* h0 = (uint64_t*) ctx->state;
uint64_t al0 = h0[0] ^ h0[4];
@ -262,7 +262,7 @@ void cryptonight_av3_aesni_bmi2(void *restrict output, const void *restrict inpu
idx0 = al0;
}
cn_implode_scratchpad((__m128i*) memory, (__m128i*) ctx->state);
cn_implode_scratchpad((__m128i*) ctx->memory, (__m128i*) ctx->state);
keccakf(h0, 24);
extra_hashes[ctx->state[0] & 3](ctx->state, 200, output);

View file

@ -203,13 +203,13 @@ static inline void cn_implode_scratchpad(const __m128i* input, __m128i* output)
}
void cryptonight_av4_softaes(void *restrict output, const void *restrict input, char *restrict memory, struct cryptonight_ctx *restrict ctx)
void cryptonight_av4_softaes(void *restrict output, const void *restrict input, struct cryptonight_ctx *restrict ctx)
{
keccak((const uint8_t *) input, 76, ctx->state, 200);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) memory);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) ctx->memory);
const uint8_t* l0 = memory;
const uint8_t* l0 = ctx->memory;
uint64_t* h0 = (uint64_t*) ctx->state;
uint64_t al0 = h0[0] ^ h0[4];
@ -243,7 +243,7 @@ void cryptonight_av4_softaes(void *restrict output, const void *restrict input,
idx0 = al0;
}
cn_implode_scratchpad((__m128i*) memory, (__m128i*) ctx->state);
cn_implode_scratchpad((__m128i*) ctx->memory, (__m128i*) ctx->state);
keccakf(h0, 24);
extra_hashes[ctx->state[0] & 3](ctx->state, 200, output);

View file

@ -213,8 +213,10 @@ static inline void cn_implode_scratchpad(const __m128i* input, __m128i* output)
}
void cryptonight_av5_aesni_experimental(void *restrict output, const void *restrict input, char *restrict memory, struct cryptonight_ctx *restrict ctx)
void cryptonight_av5_aesni_experimental(void *restrict output, const void *restrict input, struct cryptonight_ctx *restrict ctx)
{
const uint8_t* memory = ctx->memory;
keccak((const uint8_t *) input, 76, ctx->state, 200);
cn_explode_scratchpad((__m128i*) ctx->state, (__m128i*) memory);

View file

@ -37,18 +37,18 @@
#if defined(__x86_64__)
void cryptonight_av1_aesni(void* output, const void* input, const char *memory, struct cryptonight_ctx* ctx);
void cryptonight_av2_aesni_stak(void* output, const void* input, const char *memory, struct cryptonight_ctx* ctx);
void cryptonight_av3_aesni_bmi2(void* output, const void* input, const char *memory, struct cryptonight_ctx* ctx);
void cryptonight_av4_softaes(void* output, const void* input, const char *memory, struct cryptonight_ctx* ctx);
void cryptonight_av5_aesni_experimental(void* output, const void* input, const char *memory, struct cryptonight_ctx* ctx);
void cryptonight_av1_aesni(void* output, const void* input, struct cryptonight_ctx* ctx);
void cryptonight_av2_aesni_stak(void* output, const void* input, struct cryptonight_ctx* ctx);
void cryptonight_av3_aesni_bmi2(void* output, const void* input, struct cryptonight_ctx* ctx);
void cryptonight_av4_softaes(void* output, const void* input, struct cryptonight_ctx* ctx);
void cryptonight_av5_aesni_experimental(void* output, const void* input, struct cryptonight_ctx* ctx);
#elif defined(__i386__)
void cryptonight_av1_aesni32(void* output, const void* input, const char *memory, struct cryptonight_ctx* ctx);
#endif
void cryptonight_av4_softaes(void* output, const void* input, const char *memory, struct cryptonight_ctx* ctx);
void cryptonight_av4_softaes(void* output, const void* input, struct cryptonight_ctx* ctx);
void (*cryptonight_hash_ctx)(void* output, const void* input, const char *memory, struct cryptonight_ctx* ctx) = NULL;
void (*cryptonight_hash_ctx)(void* output, const void* input, struct cryptonight_ctx* ctx) = NULL;
void cryptonight_init(int variant)
@ -111,26 +111,15 @@ static inline void do_skein_hash(const void* input, size_t len, char* output) {
void (* const extra_hashes[4])(const void *, size_t, char *) = {do_blake_hash, do_groestl_hash, do_jh_hash, do_skein_hash};
void cryptonight_hash(void* output, const void* input, size_t len) {
uint8_t *memory __attribute((aligned(16))) = (uint8_t *) malloc(MEMORY);
struct cryptonight_ctx *ctx = (struct cryptonight_ctx*)malloc(sizeof(struct cryptonight_ctx));
cryptonight_hash_ctx(output, input, memory, ctx);
free(memory);
free(ctx);
}
#ifndef BUILD_TEST
int scanhash_cryptonight(int thr_id, uint32_t *hash, uint32_t *restrict pdata, const uint32_t *restrict ptarget, uint32_t max_nonce, unsigned long *restrict hashes_done, const char *restrict memory, struct cryptonight_ctx *persistentctx) {
int scanhash_cryptonight(int thr_id, uint32_t *hash, uint32_t *restrict pdata, const uint32_t *restrict ptarget, uint32_t max_nonce, unsigned long *restrict hashes_done, struct cryptonight_ctx *restrict ctx) {
uint32_t *nonceptr = (uint32_t*) (((char*)pdata) + 39);
uint32_t n = *nonceptr - 1;
const uint32_t first_nonce = n + 1;
do {
*nonceptr = ++n;
cryptonight_hash_ctx(hash, pdata, memory, persistentctx);
cryptonight_hash_ctx(hash, pdata, ctx);
if (unlikely(hash[7] < ptarget[7])) {
*hashes_done = n - first_nonce + 1;