Merge pull request #1512 from psychocrypt/topic-stelliteFork

support stellite v4 fork

2 files changed, 66 insertions, 24 deletions

diff --git a/xmrstak/backend/cpu/crypto/cryptonight_aesni.h b/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
index e2f4d88..da0a8af 100644
--- a/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
+++ b/xmrstak/backend/cpu/crypto/cryptonight_aesni.h
@@ -422,6 +422,7 @@ void cn_implode_scratchpad(const __m128i* input, __m128i* output)
 	_mm_store_si128(output + 11, xout7);
 }
 
+template<xmrstak_algo ALGO>
 inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
 {
 	mem_out[0] = _mm_cvtsi128_si64(tmp);
@@ -431,10 +432,21 @@ inline void cryptonight_monero_tweak(uint64_t* mem_out, __m128i tmp)
 
 	uint8_t x = static_cast<uint8_t>(vh >> 24);
 	static const uint16_t table = 0x7531;
-	const uint8_t index = (((x >> 3) & 6) | (x & 1)) << 1;
-	vh ^= ((table >> index) & 0x3) << 28;
+	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
+	{
+		const uint8_t index = (((x >> 3) & 6) | (x & 1)) << 1;
+		vh ^= ((table >> index) & 0x3) << 28;
+
+		mem_out[1] = vh;
+	}
+	else if(ALGO == cryptonight_stellite)
+	{
+		const uint8_t index = (((x >> 4) & 6) | (x & 1)) << 1;
+		vh ^= ((table >> index) & 0x3) << 28;
+
+		mem_out[1] = vh;
+	}
 
-	mem_out[1] = vh;
 }
 
 template<xmrstak_algo ALGO, bool SOFT_AES, bool PREFETCH>
@@ -444,7 +456,7 @@ void cryptonight_hash(const void* input, size_t len, void* output, cryptonight_c
 	constexpr size_t ITERATIONS = cn_select_iter<ALGO>();
 	constexpr size_t MEM = cn_select_memory<ALGO>();
 
-	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc) && len < 43)
+	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite) && len < 43)
 	{
 		memset(output, 0, 32);
 		return;
@@ -453,7 +465,7 @@ void cryptonight_hash(const void* input, size_t len, void* output, cryptonight_c
 	keccak((const uint8_t *)input, len, ctx0->hash_state, 200);
 
 	uint64_t monero_const;
-	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
+	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite)
 	{
 		monero_const  =  *reinterpret_cast<const uint64_t*>(reinterpret_cast<const uint8_t*>(input) + 35);
 		monero_const ^=  *(reinterpret_cast<const uint64_t*>(ctx0->hash_state) + 24);
@@ -482,8 +494,8 @@ void cryptonight_hash(const void* input, size_t len, void* output, cryptonight_c
 		else
 			cx = _mm_aesenc_si128(cx, _mm_set_epi64x(ah0, al0));
 
-		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
-			cryptonight_monero_tweak((uint64_t*)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
+		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite)
+			cryptonight_monero_tweak<ALGO>((uint64_t*)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
 		else
 			_mm_store_si128((__m128i *)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
 
@@ -506,7 +518,7 @@ void cryptonight_hash(const void* input, size_t len, void* output, cryptonight_c
 			_mm_prefetch((const char*)&l0[al0 & MASK], _MM_HINT_T0);
 		ah0 += lo;
 
-		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
+		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite)
 		{
 			if(ALGO == cryptonight_ipbc)
 				((uint64_t*)&l0[idx0 & MASK])[1] = ah0 ^ monero_const ^ ((uint64_t*)&l0[idx0 & MASK])[0];
@@ -549,7 +561,7 @@ void cryptonight_double_hash(const void* input, size_t len, void* output, crypto
 	constexpr size_t ITERATIONS = cn_select_iter<ALGO>();
 	constexpr size_t MEM = cn_select_memory<ALGO>();
 
-	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc) && len < 43)
+	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite) && len < 43)
 	{
 		memset(output, 0, 64);
 		return;
@@ -559,7 +571,7 @@ void cryptonight_double_hash(const void* input, size_t len, void* output, crypto
 	keccak((const uint8_t *)input+len, len, ctx[1]->hash_state, 200);
 
 	uint64_t monero_const_0, monero_const_1;
-	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
+	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite)
 	{
 		monero_const_0  =  *reinterpret_cast<const uint64_t*>(reinterpret_cast<const uint8_t*>(input) + 35);
 		monero_const_0 ^=  *(reinterpret_cast<const uint64_t*>(ctx[0]->hash_state) + 24);
@@ -597,8 +609,8 @@ void cryptonight_double_hash(const void* input, size_t len, void* output, crypto
 		else
 			cx = _mm_aesenc_si128(cx, _mm_set_epi64x(axh0, axl0));
 
-		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
-			cryptonight_monero_tweak((uint64_t*)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
+		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite)
+			cryptonight_monero_tweak<ALGO>((uint64_t*)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
 		else
 			_mm_store_si128((__m128i *)&l0[idx0 & MASK], _mm_xor_si128(bx0, cx));
 
@@ -615,8 +627,8 @@ void cryptonight_double_hash(const void* input, size_t len, void* output, crypto
 		else
 			cx = _mm_aesenc_si128(cx, _mm_set_epi64x(axh1, axl1));
 
-		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
-			cryptonight_monero_tweak((uint64_t*)&l1[idx1 & MASK], _mm_xor_si128(bx1, cx));
+		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite)
+			cryptonight_monero_tweak<ALGO>((uint64_t*)&l1[idx1 & MASK], _mm_xor_si128(bx1, cx));
 		else
 			_mm_store_si128((__m128i *)&l1[idx1 & MASK], _mm_xor_si128(bx1, cx));
 
@@ -636,7 +648,7 @@ void cryptonight_double_hash(const void* input, size_t len, void* output, crypto
 		axh0 += lo;
 		((uint64_t*)&l0[idx0 & MASK])[0] = axl0;
 
-		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
+		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite)
 		{
 			if(ALGO == cryptonight_ipbc)
 				((uint64_t*)&l0[idx0 & MASK])[1] = axh0 ^ monero_const_0 ^ ((uint64_t*)&l0[idx0 & MASK])[0];
@@ -672,7 +684,7 @@ void cryptonight_double_hash(const void* input, size_t len, void* output, crypto
 		axh1 += lo;
 		((uint64_t*)&l1[idx1 & MASK])[0] = axl1;
 
-		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc)
+		if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite)
 		{
 			if(ALGO == cryptonight_ipbc)
 				((uint64_t*)&l1[idx1 & MASK])[1] = axh1 ^ monero_const_1 ^ ((uint64_t*)&l1[idx1 & MASK])[0];
@@ -724,8 +736,8 @@ void cryptonight_double_hash(const void* input, size_t len, void* output, crypto
 	else							\
 		c = _mm_aesenc_si128(c, a);			\
 	b = _mm_xor_si128(b, c);				\
-	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc) \
-		cryptonight_monero_tweak((uint64_t*)ptr, b); \
+	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite) \
+		cryptonight_monero_tweak<ALGO>((uint64_t*)ptr, b); \
 	else \
 		_mm_store_si128(ptr, b);\
 
@@ -739,7 +751,7 @@ void cryptonight_double_hash(const void* input, size_t len, void* output, crypto
 #define CN_STEP4(a, b, c, l, mc, ptr, idx)				\
 	lo = _umul128(idx, _mm_cvtsi128_si64(b), &hi);		\
 	a = _mm_add_epi64(a, _mm_set_epi64x(lo, hi));		\
-	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc) \
+	if(ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite) \
 	{ \
 		_mm_store_si128(ptr, _mm_xor_si128(a, mc)); \
 		if (ALGO == cryptonight_ipbc) \
@@ -770,7 +782,7 @@ void cryptonight_triple_hash(const void* input, size_t len, void* output, crypto
 	constexpr size_t ITERATIONS = cn_select_iter<ALGO>();
 	constexpr size_t MEM = cn_select_memory<ALGO>();
 
-	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc) && len < 43)
+	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite) && len < 43)
 	{
 		memset(output, 0, 32 * 3);
 		return;
@@ -864,7 +876,7 @@ void cryptonight_quad_hash(const void* input, size_t len, void* output, cryptoni
 	constexpr size_t ITERATIONS = cn_select_iter<ALGO>();
 	constexpr size_t MEM = cn_select_memory<ALGO>();
 
-	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc) && len < 43)
+	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite) && len < 43)
 	{
 		memset(output, 0, 32 * 4);
 		return;
@@ -973,7 +985,7 @@ void cryptonight_penta_hash(const void* input, size_t len, void* output, crypton
 	constexpr size_t ITERATIONS = cn_select_iter<ALGO>();
 	constexpr size_t MEM = cn_select_memory<ALGO>();
 
-	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc) && len < 43)
+	if((ALGO == cryptonight_monero || ALGO == cryptonight_aeon || ALGO == cryptonight_ipbc || ALGO == cryptonight_stellite) && len < 43)
 	{
 		memset(output, 0, 32 * 5);
 		return;
diff --git a/xmrstak/backend/cpu/minethd.cpp b/xmrstak/backend/cpu/minethd.cpp
index b53eac7..31ef926 100644
--- a/xmrstak/backend/cpu/minethd.cpp
+++ b/xmrstak/backend/cpu/minethd.cpp
@@ -289,6 +289,9 @@ bool minethd::self_test()
 	else if(::jconf::inst()->GetCurrentCoinSelection().GetDescription(1).GetMiningAlgo() == cryptonight_ipbc)
 	{
 	}
+	else if(::jconf::inst()->GetCurrentCoinSelection().GetDescription(1).GetMiningAlgo() == cryptonight_stellite)
+	{
+	}
 	for (int i = 0; i < MAX_N; i++)
 		cryptonight_free_ctx(ctx[i]);
 
@@ -378,6 +381,9 @@ minethd::cn_hash_fun minethd::func_selector(bool bHaveAes, bool bNoPrefetch, xmr
 	case cryptonight_ipbc:
 		algv = 5;
 		break;
+	case cryptonight_stellite:
+		algv = 6;
+		break;
 	default:
 		algv = 2;
 		break;
@@ -407,7 +413,11 @@ minethd::cn_hash_fun minethd::func_selector(bool bHaveAes, bool bNoPrefetch, xmr
 		cryptonight_hash<cryptonight_ipbc, false, false>,
 		cryptonight_hash<cryptonight_ipbc, true, false>,
 		cryptonight_hash<cryptonight_ipbc, false, true>,
-		cryptonight_hash<cryptonight_ipbc, true, true>
+		cryptonight_hash<cryptonight_ipbc, true, true>,
+		cryptonight_hash<cryptonight_stellite, false, false>,
+		cryptonight_hash<cryptonight_stellite, true, false>,
+		cryptonight_hash<cryptonight_stellite, false, true>,
+		cryptonight_hash<cryptonight_stellite, true, true>
 	};
 
 	std::bitset<2> digit;
@@ -547,6 +557,9 @@ minethd::cn_hash_fun_multi minethd::func_multi_selector(size_t N, bool bHaveAes,
 	case cryptonight_ipbc:
 		algv = 5;
 		break;
+	case cryptonight_stellite:
+		algv = 6;
+		break;
 	default:
 		algv = 2;
 		break;
@@ -653,7 +666,24 @@ minethd::cn_hash_fun_multi minethd::func_multi_selector(size_t N, bool bHaveAes,
 		cryptonight_penta_hash<cryptonight_ipbc, false, false>,
 		cryptonight_penta_hash<cryptonight_ipbc, true, false>,
 		cryptonight_penta_hash<cryptonight_ipbc, false, true>,
-		cryptonight_penta_hash<cryptonight_ipbc, true, true>
+		cryptonight_penta_hash<cryptonight_ipbc, true, true>,
+
+		cryptonight_double_hash<cryptonight_stellite, false, false>,
+		cryptonight_double_hash<cryptonight_stellite, true, false>,
+		cryptonight_double_hash<cryptonight_stellite, false, true>,
+		cryptonight_double_hash<cryptonight_stellite, true, true>,
+		cryptonight_triple_hash<cryptonight_stellite, false, false>,
+		cryptonight_triple_hash<cryptonight_stellite, true, false>,
+		cryptonight_triple_hash<cryptonight_stellite, false, true>,
+		cryptonight_triple_hash<cryptonight_stellite, true, true>,
+		cryptonight_quad_hash<cryptonight_stellite, false, false>,
+		cryptonight_quad_hash<cryptonight_stellite, true, false>,
+		cryptonight_quad_hash<cryptonight_stellite, false, true>,
+		cryptonight_quad_hash<cryptonight_stellite, true, true>,
+		cryptonight_penta_hash<cryptonight_stellite, false, false>,
+		cryptonight_penta_hash<cryptonight_stellite, true, false>,
+		cryptonight_penta_hash<cryptonight_stellite, false, true>,
+		cryptonight_penta_hash<cryptonight_stellite, true, true>,
 	};
 
 	std::bitset<2> digit;