diff --git a/xmrstak/backend/cpu/crypto/cryptonight_aesni.h b/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
index 89c508990f82a28f4d14fcb9aedbb83bdf4e52d4..27347609625dfb2f873fe578ad9dd9a1adeda3c4 100644
--- a/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
+++ b/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
@@ -19,6 +19,7 @@
#include "xmrstak/backend/cryptonight.hpp"
#include <memory.h>
#include <stdio.h>
+#include <cfenv>
#ifdef __GNUC__
#include <x86intrin.h>
@@ -422,6 +423,27 @@ void cn_implode_scratchpad(const __m128i* input, __m128i* output)
_mm_store_si128(output + 11, xout7);
}
+inline __m128i int_sqrt33_1_double_precision(const uint64_t n0)
+{
+ __m128d x = _mm_castsi128_pd(_mm_add_epi64(_mm_cvtsi64_si128(n0 >> 12), _mm_set_epi64x(0, 1023ULL << 52)));
+ x = _mm_sqrt_sd(_mm_setzero_pd(), x);
+ uint64_t r = static_cast<uint64_t>(_mm_cvtsi128_si64(_mm_castpd_si128(x)));
+
+ const uint64_t s = r >> 20;
+ r >>= 19;
+
+ uint64_t x2 = (s - (1022ULL << 32)) * (r - s - (1022ULL << 32) + 1);
+
+#if defined _MSC_VER || (__GNUC__ >= 7)
+ _addcarry_u64(_subborrow_u64(0, x2, n0, (unsigned long long int*)&x2), r, 0, (unsigned long long int*)&r);
+#else
+ // GCC versions prior to 7 don't generate correct assembly for _subborrow_u64 -> _addcarry_u64 sequence
+ // Fallback to simpler code
+ if (x2 < n0) ++r;
+#endif
+ return _mm_cvtsi64_si128(r);
+}
+
inline __m128i aes_round_bittube2(const __m128i& val, const __m128i& key)
{
alignas(16) uint32_t k[4];
@@ -467,6 +489,51 @@ inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
}
+inline void set_float_rounding_mode()
+{
+#ifdef _MSC_VER
+ _control87(RC_DOWN, MCW_RC);
+#else
+ std::fesetround(FE_DOWNWARD);
+#endif
+}
+
+#define CN_MONERO_V8_SHUFFLE(n, l0, idx0, ax0, bx0, bx1) \
+ /* Shuffle the other 3x16 byte chunks in the current 64-byte cache line */ \
+ if(ALGO == cryptonight_monero_v8) \
+ { \
+ const uint64_t idx1 = idx0 & MASK; \
+ const __m128i chunk1 = _mm_load_si128((__m128i *)&l0[idx1 ^ 0x10]); \
+ const __m128i chunk2 = _mm_load_si128((__m128i *)&l0[idx1 ^ 0x20]); \
+ const __m128i chunk3 = _mm_load_si128((__m128i *)&l0[idx1 ^ 0x30]); \
+ _mm_store_si128((__m128i *)&l0[idx1 ^ 0x10], _mm_add_epi64(chunk3, bx1)); \
+ _mm_store_si128((__m128i *)&l0[idx1 ^ 0x20], _mm_add_epi64(chunk1, bx0)); \
+ _mm_store_si128((__m128i *)&l0[idx1 ^ 0x30], _mm_add_epi64(chunk2, ax0)); \
+ }
+
+#define CN_MONERO_V8_DIV(n, cx, sqrt_result_xmm, division_result_xmm, cl) \
+ if(ALGO == cryptonight_monero_v8) \
+ { \
+ const uint64_t sqrt_result = static_cast<uint64_t>(_mm_cvtsi128_si64(sqrt_result_xmm)); \
+ /* Use division and square root results from the _previous_ iteration to hide the latency */ \
+ const uint64_t cx_64 = _mm_cvtsi128_si64(cx); \
+ cl ^= static_cast<uint64_t>(_mm_cvtsi128_si64(division_result_xmm)) ^ (sqrt_result << 32); \
+ const uint32_t d = (cx_64 + (sqrt_result << 1)) | 0x80000001UL; \
+ /* Most and least significant bits in the divisor are set to 1 \
+ * to make sure we don't divide by a small or even number, \
+ * so there are no shortcuts for such cases \
+ * \
+ * Quotient may be as large as (2^64 - 1)/(2^31 + 1) = 8589934588 = 2^33 - 4 \
+ * We drop the highest bit to fit both quotient and remainder in 32 bits \
+ */ \
+ /* Compiler will optimize it to a single div instruction */ \
+ const uint64_t cx_s = _mm_cvtsi128_si64(_mm_srli_si128(cx, 8)); \
+ const uint64_t division_result = static_cast<uint32_t>(cx_s / d) + ((cx_s % d) << 32); \
+ division_result_xmm = _mm_cvtsi64_si128(static_cast<int64_t>(division_result)); \
+ /* Use division_result as an input for the square root to prevent parallel implementation in hardware */ \
+ sqrt_result_xmm = int_sqrt33_1_double_precision(cx_64 + division_result); \
+ }
+
#define CN_INIT_SINGLE \
if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite || ALGO == cryptonight_masari || ALGO == cryptonight_bittube2) && len < 43) \
{ \
@@ -474,7 +541,7 @@ inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
return; \
}
-#define CN_INIT(n, monero_const, l0, ax0, bx0, idx0, ptr0) \
+#define CN_INIT(n, monero_const, l0, ax0, bx0, idx0, ptr0, bx1, sqrt_result_xmm, division_result_xmm) \
keccak((const uint8_t *)input + len * n, len, ctx[n]->hash_state, 200); \
uint64_t monero_const; \
if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite || ALGO == cryptonight_masari || ALGO == cryptonight_bittube2) \
@@ -489,16 +556,27 @@ inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
uint64_t idx0; \
__m128i bx0; \
uint8_t* l0 = ctx[n]->long_state; \
+ /* BEGIN cryptonight_monero_v8 variables */ \
+ __m128i bx1; \
+ __m128i division_result_xmm; \
+ __m128i sqrt_result_xmm; \
+ /* END cryptonight_monero_v8 variables */ \
{ \
uint64_t* h0 = (uint64_t*)ctx[n]->hash_state; \
idx0 = h0[0] ^ h0[4]; \
ax0 = _mm_set_epi64x(h0[1] ^ h0[5], idx0); \
bx0 = _mm_set_epi64x(h0[3] ^ h0[7], h0[2] ^ h0[6]); \
+ if(ALGO == cryptonight_monero_v8) \
+ { \
+ bx1 = _mm_set_epi64x(h0[9] ^ h0[11], h0[8] ^ h0[10]); \
+ division_result_xmm = _mm_cvtsi64_si128(h0[12]); \
+ sqrt_result_xmm = _mm_cvtsi64_si128(h0[13]); \
+ set_float_rounding_mode(); \
+ } \
} \
__m128i *ptr0
-
-#define CN_STEP1(n, monero_const, l0, ax0, bx0, idx0, ptr0, cx) \
+#define CN_STEP1(n, monero_const, l0, ax0, bx0, idx0, ptr0, cx, bx1) \
__m128i cx; \
ptr0 = (__m128i *)&l0[idx0 & MASK]; \
cx = _mm_load_si128(ptr0); \
@@ -512,7 +590,8 @@ inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
cx = soft_aesenc(cx, ax0); \
else \
cx = _mm_aesenc_si128(cx, ax0); \
- }
+ } \
+ CN_MONERO_V8_SHUFFLE(n, l0, idx0, ax0, bx0, bx1)
#define CN_STEP2(n, monero_const, l0, ax0, bx0, idx0, ptr0, cx) \
if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite || ALGO == cryptonight_masari || ALGO == cryptonight_bittube2) \
@@ -524,15 +603,22 @@ inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
ptr0 = (__m128i *)&l0[idx0 & MASK]; \
if(PREFETCH) \
_mm_prefetch((const char*)ptr0, _MM_HINT_T0); \
- bx0 = cx; \
+ if(ALGO != cryptonight_monero_v8) \
+ bx0 = cx
-#define CN_STEP3(n, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0) \
+#define CN_STEP3(n, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0, cx, bx1, sqrt_result_xmm, division_result_xmm) \
uint64_t lo, cl, ch; \
uint64_t al0 = _mm_cvtsi128_si64(ax0); \
uint64_t ah0 = ((uint64_t*)&ax0)[1]; \
cl = ((uint64_t*)ptr0)[0]; \
ch = ((uint64_t*)ptr0)[1]; \
- \
+ CN_MONERO_V8_DIV(n, cx, sqrt_result_xmm, division_result_xmm, cl); \
+ CN_MONERO_V8_SHUFFLE(n, l0, idx0, ax0, bx0, bx1); \
+ if(ALGO == cryptonight_monero_v8) \
+ { \
+ bx1 = bx0; \
+ bx0 = cx; \
+ } \
{ \
uint64_t hi; \
lo = _umul128(idx0, cl, &hi); \
@@ -542,7 +628,6 @@ inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
((uint64_t*)ptr0)[0] = al0; \
if(PREFETCH) \
_mm_prefetch((const char*)ptr0, _MM_HINT_T0)
-
#define CN_STEP4(n, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0) \
if (ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite || ALGO == cryptonight_masari || ALGO == cryptonight_bittube2) \
@@ -622,6 +707,9 @@ inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
#define CN_ENUM_10(n, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10) n, x1 ## n, x2 ## n, x3 ## n, x4 ## n, x5 ## n, x6 ## n, x7 ## n, x8 ## n, x9 ## n, x10 ## n
#define CN_ENUM_11(n, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11) n, x1 ## n, x2 ## n, x3 ## n, x4 ## n, x5 ## n, x6 ## n, x7 ## n, x8 ## n, x9 ## n, x10 ## n, x11 ## n
#define CN_ENUM_12(n, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12) n, x1 ## n, x2 ## n, x3 ## n, x4 ## n, x5 ## n, x6 ## n, x7 ## n, x8 ## n, x9 ## n, x10 ## n, x11 ## n, x12 ## n
+#define CN_ENUM_13(n, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13) n, x1 ## n, x2 ## n, x3 ## n, x4 ## n, x5 ## n, x6 ## n, x7 ## n, x8 ## n, x9 ## n, x10 ## n, x11 ## n, x12 ## n, x13 ## n
+#define CN_ENUM_14(n, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14) n, x1 ## n, x2 ## n, x3 ## n, x4 ## n, x5 ## n, x6 ## n, x7 ## n, x8 ## n, x9 ## n, x10 ## n, x11 ## n, x12 ## n, x13 ## n, x14 ## n
+#define CN_ENUM_15(n, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15) n, x1 ## n, x2 ## n, x3 ## n, x4 ## n, x5 ## n, x6 ## n, x7 ## n, x8 ## n, x9 ## n, x10 ## n, x11 ## n, x12 ## n, x13 ## n, x14 ## n, x15 ## n
/** repeat a macro call multiple times
*
@@ -657,15 +745,14 @@ struct Cryptonight_hash<1>
constexpr size_t MEM = cn_select_memory<ALGO>();
CN_INIT_SINGLE;
- REPEAT_1(6, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0);
+ REPEAT_1(9, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0, bx1, sqrt_result_xmm, division_result_xmm);
// Optim - 90% time boundary
for(size_t i = 0; i < ITERATIONS; i++)
{
-
- REPEAT_1(7, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
+ REPEAT_1(8, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx, bx1);
REPEAT_1(7, CN_STEP2, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
- REPEAT_1(11, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
+ REPEAT_1(15, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0, cx, bx1, sqrt_result_xmm, division_result_xmm);
REPEAT_1(11, CN_STEP4, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
REPEAT_1(6, CN_STEP5, monero_const, l0, ax0, bx0, idx0, ptr0);
}
@@ -687,14 +774,14 @@ struct Cryptonight_hash<2>
constexpr size_t MEM = cn_select_memory<ALGO>();
CN_INIT_SINGLE;
- REPEAT_2(6, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0);
+ REPEAT_2(9, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0, bx1, sqrt_result_xmm, division_result_xmm);
// Optim - 90% time boundary
for(size_t i = 0; i < ITERATIONS; i++)
{
- REPEAT_2(7, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
+ REPEAT_2(8, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx, bx1);
REPEAT_2(7, CN_STEP2, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
- REPEAT_2(11, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
+ REPEAT_2(15, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0, cx, bx1, sqrt_result_xmm, division_result_xmm);
REPEAT_2(11, CN_STEP4, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
REPEAT_2(6, CN_STEP5, monero_const, l0, ax0, bx0, idx0, ptr0);
}
@@ -716,14 +803,14 @@ struct Cryptonight_hash<3>
constexpr size_t MEM = cn_select_memory<ALGO>();
CN_INIT_SINGLE;
- REPEAT_3(6, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0);
+ REPEAT_3(9, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0, bx1, sqrt_result_xmm, division_result_xmm);
// Optim - 90% time boundary
for(size_t i = 0; i < ITERATIONS; i++)
{
- REPEAT_3(7, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
+ REPEAT_3(8, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx, bx1);
REPEAT_3(7, CN_STEP2, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
- REPEAT_3(11, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
+ REPEAT_3(15, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0, cx, bx1, sqrt_result_xmm, division_result_xmm);
REPEAT_3(11, CN_STEP4, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
REPEAT_3(6, CN_STEP5, monero_const, l0, ax0, bx0, idx0, ptr0);
}
@@ -745,14 +832,14 @@ struct Cryptonight_hash<4>
constexpr size_t MEM = cn_select_memory<ALGO>();
CN_INIT_SINGLE;
- REPEAT_4(6, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0);
+ REPEAT_4(9, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0, bx1, sqrt_result_xmm, division_result_xmm);
// Optim - 90% time boundary
for(size_t i = 0; i < ITERATIONS; i++)
{
- REPEAT_4(7, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
+ REPEAT_4(8, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx, bx1);
REPEAT_4(7, CN_STEP2, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
- REPEAT_4(11, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
+ REPEAT_4(15, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0, cx, bx1, sqrt_result_xmm, division_result_xmm);
REPEAT_4(11, CN_STEP4, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
REPEAT_4(6, CN_STEP5, monero_const, l0, ax0, bx0, idx0, ptr0);
}
@@ -774,14 +861,14 @@ struct Cryptonight_hash<5>
constexpr size_t MEM = cn_select_memory<ALGO>();
CN_INIT_SINGLE;
- REPEAT_5(6, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0);
+ REPEAT_5(9, CN_INIT, monero_const, l0, ax0, bx0, idx0, ptr0, bx1, sqrt_result_xmm, division_result_xmm);
// Optim - 90% time boundary
for(size_t i = 0; i < ITERATIONS; i++)
{
- REPEAT_5(7, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
+ REPEAT_5(8, CN_STEP1, monero_const, l0, ax0, bx0, idx0, ptr0, cx, bx1);
REPEAT_5(7, CN_STEP2, monero_const, l0, ax0, bx0, idx0, ptr0, cx);
- REPEAT_5(11, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
+ REPEAT_5(15, CN_STEP3, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0, cx, bx1, sqrt_result_xmm, division_result_xmm);
REPEAT_5(11, CN_STEP4, monero_const, l0, ax0, bx0, idx0, ptr0, lo, cl, ch, al0, ah0);
REPEAT_5(6, CN_STEP5, monero_const, l0, ax0, bx0, idx0, ptr0);
}
diff --git a/xmrstak/backend/cpu/minethd.cpp b/xmrstak/backend/cpu/minethd.cpp
index e11c82009b4ce17675097fd01d7249dff48278bf..87f4d3285f77e7ee74da31f312e20cbbf894afe9 100644
--- a/xmrstak/backend/cpu/minethd.cpp
+++ b/xmrstak/backend/cpu/minethd.cpp
@@ -489,6 +489,9 @@ minethd::cn_hash_fun minethd::func_multi_selector(bool bHaveAes, bool bNoPrefetc
case cryptonight_bittube2:
algv = 9;
break;
+ case cryptonight_monero_v8:
+ algv = 10;
+ break;
default:
algv = 2;
break;