Tweak v2.2 implementation

Unoptimized

crypto/asm/cnv2_double_main_loop_sandybridge.inc

@@ -161,18 +161,8 @@ main_loop_double_sandybridge:
 	mov	r9, QWORD PTR [rbx+8]
 
 	xor	edx, 16
-	mov	eax, edx
-	movdqu	xmm0, XMMWORD PTR [rdx+r13]
-	xor	edx, 48
-	paddq	xmm0, xmm7
-	movdqu	xmm1, XMMWORD PTR [rdx+r13]
-	movdqu	XMMWORD PTR [rdx+r13], xmm0
-	paddq	xmm1, xmm2
-	xor	edx, 16
-	movdqu	xmm0, XMMWORD PTR [rdx+r13]
-	movdqu	XMMWORD PTR [rdx+r13], xmm1
-	paddq	xmm0, xmm3
-	movdqu	XMMWORD PTR [rax+r13], xmm0
+	mov r8d, edx
+	mov r15d, edx
 
 	movq	rdx, xmm5
 	shl	rdx, 32
@@ -181,6 +171,25 @@ main_loop_double_sandybridge:
 	xor	r10, rdx
 	mov	rax, r10
 	mul	r11
+	movq xmm0, rax
+	movq xmm1, rdx
+	punpcklqdq xmm1, xmm0
+
+	movdqu	xmm0, XMMWORD PTR [r8+r13]
+	pxor xmm0, xmm1
+	xor	r8d, 48
+	paddq	xmm0, xmm7
+	movdqu	xmm1, XMMWORD PTR [r8+r13]
+	xor rdx, [r8+r13]
+	xor rax, [r8+r13+8]
+	movdqu	XMMWORD PTR [r8+r13], xmm0
+	paddq	xmm1, xmm2
+	xor	r8d, 16
+	movdqu	xmm0, XMMWORD PTR [r8+r13]
+	movdqu	XMMWORD PTR [r8+r13], xmm1
+	paddq	xmm0, xmm3
+	movdqu	XMMWORD PTR [r15+r13], xmm0
+
 	mov	r11, QWORD PTR [rsp+256]
 	add	r11, rdx
 	mov	rdx, QWORD PTR [rsp+264]
@@ -278,14 +287,20 @@ sqrt_fix_2_ret_sandybridge:
 
 	mov	rax, r10
 	mul	r9
+	movq xmm0, rax
+	movq xmm3, rdx
+	punpcklqdq xmm3, xmm0
 
 	mov r12d, ecx
 	mov r8d, ecx
 	xor r12d, 16
 	xor r8d, 32
 	xor ecx, 48
 	movdqu	xmm0, XMMWORD PTR [r12+rsi]
+	pxor xmm0, xmm3
 	movdqu	xmm1, XMMWORD PTR [r8+rsi]
+	xor rdx, [r8+rsi]
+	xor rax, [r8+rsi+8]
 	movdqu	xmm3, XMMWORD PTR [rcx+rsi]
 	add	rdi, rdx
 	add	rbp, rax

crypto/asm/cnv2_main_loop_ivybridge.inc

@@ -114,24 +114,30 @@ $sqrt_fixup_ivybridge_ret:
 	mov	 r9, r10
 	mov	 rax, rdi
 	mul	 rbp
+	movq xmm0, rax
+	movq xmm1, rdx
+	punpcklqdq xmm1, xmm0
 
 	xor	 r9, 16
 	mov	 rcx, r10
 	xor	 rcx, 32
 	xor	 r10, 48
-	add	 r8, rdx
-	add	 r11, rax
-	movdqu	 xmm0, XMMWORD PTR [r10+rbx]
 	movdqu	 xmm2, XMMWORD PTR [r9+rbx]
+	pxor xmm2, xmm1
+	movdqu	 xmm0, XMMWORD PTR [r10+rbx]
 	paddq	 xmm0, xmm5
 	movdqu	 xmm1, XMMWORD PTR [rcx+rbx]
+	xor rdx, [rcx+rbx]
+	xor rax, [rcx+rbx+8]
 	paddq	 xmm2, xmm4
 	paddq	 xmm1, xmm7
 	movdqa	 xmm5, xmm4
 	movdqu	 XMMWORD PTR [r9+rbx], xmm0
 	movdqa	 xmm4, xmm6
 	movdqu	 XMMWORD PTR [rcx+rbx], xmm2
 	movdqu	 XMMWORD PTR [r10+rbx], xmm1
+	add	 r8, rdx
+	add	 r11, rax
 	mov	 QWORD PTR [r14], r8
 	xor	 r8, rdi
 	mov	 r10, r8

crypto/asm/cnv2_main_loop_ryzen.inc

@@ -109,15 +109,21 @@ $main_loop_ryzen:
 $sqrt_fixup_ryzen_ret:
 	mov	rax, rsi
 	mul	r14
+	movq xmm1, rax
+	movq xmm0, rdx
+	punpcklqdq xmm0, xmm1
 
 	mov	r9d, r10d
 	mov	ecx, r10d
 	xor	r9d, 16
 	xor	ecx, 32
 	xor	r10d, 48
-	movdqa	xmm0, XMMWORD PTR [r10+rbx]
-	movdqa	xmm2, XMMWORD PTR [r9+rbx]
 	movdqa	xmm1, XMMWORD PTR [rcx+rbx]
+	xor rdx, [rcx+rbx]
+	xor rax, [rcx+rbx+8]
+	movdqa	xmm2, XMMWORD PTR [r9+rbx]
+	pxor xmm2, xmm0
+	movdqa	xmm0, XMMWORD PTR [r10+rbx]
 	paddq	xmm0, xmm4
 	paddq	xmm2, xmm3
 	paddq	xmm1, xmm6

crypto/cryptonight_aesni.h

@@ -446,11 +446,10 @@ void cryptonight_hash(const void* input, size_t len, void* output, cryptonight_c
 		// Shuffle the other 3x16 byte chunks in the current 64-byte cache line
 		if (SHUFFLE)
 		{
-			// Shuffle constants here were chosen carefully
-			// to maximize permutation cycle length
-			// and have no 2-byte elements stay in their places
-			const __m128i chunk1 = _mm_load_si128((__m128i *)&l0[idx1 ^ 0x10]);
+			const __m128i chunk1 = _mm_xor_si128(_mm_load_si128((__m128i *)&l0[idx1 ^ 0x10]), _mm_set_epi64x(lo, hi));
 			const __m128i chunk2 = _mm_load_si128((__m128i *)&l0[idx1 ^ 0x20]);
+			hi ^= ((uint64_t*)&l0[idx1 ^ 0x20])[0];
+			lo ^= ((uint64_t*)&l0[idx1 ^ 0x20])[1];
 			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));
@@ -637,25 +636,27 @@ void cryptonight_double_hash(const void* input1, size_t len1, void* output1, con
 		cl = ((uint64_t*)&l0[idx01])[0];
 		ch = ((uint64_t*)&l0[idx01])[1];
 
+		if (INT_MATH)
+		{
+			const uint64_t sqrt_result0 = _mm_cvtsi128_si64(sqrt_result_xmm);
+			cl ^= static_cast<uint64_t>(_mm_cvtsi128_si64(division_result_xmm)) ^ (sqrt_result0 << 32);
+		}
+
+		lo = _umul128(idx00, cl, &hi);
+
 		if (SHUFFLE)
 		{
 			uint32_t k = idx01 ^ 0x10;
-			const __m128i chunk1 = _mm_load_si128((__m128i *)&l0[k]); k ^= 0x30;
+			const __m128i chunk1 = _mm_xor_si128(_mm_load_si128((__m128i *)&l0[k]), _mm_set_epi64x(lo, hi)); k ^= 0x30;
 			const __m128i chunk2 = _mm_load_si128((__m128i *)&l0[k]);
+			hi ^= ((uint64_t*)&l0[k])[0];
+			lo ^= ((uint64_t*)&l0[k])[1];
 			_mm_store_si128((__m128i *)&l0[k], _mm_add_epi64(chunk1, bx00)); k ^= 0x10;
 			const __m128i chunk3 = _mm_load_si128((__m128i *)&l0[k]);
 			_mm_store_si128((__m128i *)&l0[k], _mm_add_epi64(chunk2, ax0)); k ^= 0x20;
 			_mm_store_si128((__m128i *)&l0[k], _mm_add_epi64(chunk3, bx01));
 		}
 
-		if (INT_MATH)
-		{
-			const uint64_t sqrt_result0 = _mm_cvtsi128_si64(sqrt_result_xmm);
-			cl ^= static_cast<uint64_t>(_mm_cvtsi128_si64(division_result_xmm)) ^ (sqrt_result0 << 32);
-		}
-
-		lo = _umul128(idx00, cl, &hi);
-
 		axl0 += hi;
 		axh0 += lo;
 		((uint64_t*)&l0[idx01])[0] = axl0;
@@ -683,8 +684,10 @@ void cryptonight_double_hash(const void* input1, size_t len1, void* output1, con
 		if (SHUFFLE)
 		{
 			uint32_t k = idx11 ^ 0x10;
-			const __m128i chunk1 = _mm_load_si128((__m128i *)&l1[k]); k ^= 0x30;
+			const __m128i chunk1 = _mm_xor_si128(_mm_load_si128((__m128i *)&l1[k]), _mm_set_epi64x(lo, hi)); k ^= 0x30;
 			const __m128i chunk2 = _mm_load_si128((__m128i *)&l1[k]);
+			hi ^= ((uint64_t*)&l1[k])[0];
+			lo ^= ((uint64_t*)&l1[k])[1];
 			_mm_store_si128((__m128i *)&l1[k], _mm_add_epi64(chunk1, bx10)); k ^= 0x10;
 			const __m128i chunk3 = _mm_load_si128((__m128i *)&l1[k]);
 			_mm_store_si128((__m128i *)&l1[k], _mm_add_epi64(chunk2, ax1)); k ^= 0x20;

minethd.cpp

@@ -218,6 +218,16 @@ cryptonight_ctx* minethd_alloc_ctx()
 	return nullptr; //Should never happen
 }
 
+static void print_hash(const char* input, const char* hash)
+{
+	printf("HASH(\"%s\") = ", input);
+	for (int k = 0; k < 32; ++k)
+	{
+		printf("%02x", static_cast<uint8_t>(hash[k]));
+	}
+	printf("\n");
+}
+
 bool minethd::self_test()
 {
 	alloc_msg msg = { 0 };
@@ -327,6 +337,7 @@ bool minethd::self_test()
 
 			if (memcmp(hash, reference_hash[i], HASH_SIZE) != 0)
 			{
+				print_hash(input.c_str(), hash);
 				printer::inst()->print_msg(L0, "Cryptonight hash self-test (variant %d) failed.", i);
 				return false;
 			}
@@ -335,6 +346,8 @@ bool minethd::self_test()
 			{
 				if (memcmp(hash_dbl, reference_hash_dbl[i], HASH_SIZE * 2) != 0)
 				{
+					print_hash(prev_input.c_str(), hash_dbl);
+					print_hash(input.c_str(), hash_dbl + HASH_SIZE);
 					printer::inst()->print_msg(L0, "Cryptonight double hash self-test (variant %d) failed.", i);
 					return false;
 				}
@@ -356,13 +369,16 @@ bool minethd::self_test()
 
 					if (memcmp(hash, reference_hash[i], HASH_SIZE) != 0)
 					{
+						print_hash(input.c_str(), hash);
 						printer::inst()->print_msg(L0, "Cryptonight hash self-test (variant 2, asm version %d) failed.", j);
 						return false;
 					}
 					if (!prev_input.empty())
 					{
 						if (memcmp(hash_dbl, reference_hash_dbl[i], HASH_SIZE * 2) != 0)
 						{
+							print_hash(prev_input.c_str(), hash_dbl);
+							print_hash(input.c_str(), hash_dbl + HASH_SIZE);
 							printer::inst()->print_msg(L0, "Cryptonight double hash self-test (variant 2, asm version %d) failed.", j);
 							return false;
 						}

tests.txt

@@ -1,40 +1,40 @@
 This is a test This is a test This is a test
 74d15836e33d14e164c2494648996eb5ed71a3ec2c72c2be225eda1b8a857aba
 0157c5ee188bbec8975285a3064ee92065217672fd69a1aebd0766c7b56ee0bd
-4cf1ff9ca46eb433b36cd9f70e02b14cc06bfd18ca77fa9ccaafd1fd96c674b0
+353fdc068fd47b03c04b9431e005e00b68c2168a3cc7335c8b9b308156591a4f
 Lorem ipsum dolor sit amet, consectetur adipiscing
 22ec483997cba20105378af3ec647ee5d20401d7df21c0bf4bf866bc55383e92
 755d58e48e53f795a0ed6b27c794018372922e5d1a256cdbf9fc442f59f284c9
-7d292e43f4751714ec07dbcb0e4bbffe2a7afb6066420960684ff57d7474c871
+72f134fc50880c330fe65a2cb7896d59b2e708a0221c6a9da3f69b3a702d8682
 elit, sed do eiusmod tempor incididunt ut labore
 c5efc04bf88b450e86537dc046339b16d35133c4d905ec7fa16bd28a67c4f2fe
 7158c9c0d5082df7f2ee236b994f385bd96fd09eda30e21643cb7351fd7301ce
-335563425256edebf1d92dc342369c2f4770ebb4112ba975659bd8a0f210abd0
+410919660ec540fc49d8695ff01f974226a2a28dbbac82949c12f541b9a62d2f
 et dolore magna aliqua. Ut enim ad minim veniam,
 628c400e4712cecb44d88572e9e8bb9be9a1221da1cb52ff8eefaf4adcc172eb
 7329cde3fbf98bec02578fcdcfeaf2cf11e2a1f105324f89c36470708bd6db16
-47758e86d2f57210366cec36fff26f9464d89efd116fe6ef28b718b5da120801
+4472fecfeb371e8b7942ce0378c0ba5e6d0c6361b669c587807365c787ae652d
 quis nostrud exercitation ullamco laboris nisi
 a0351d7aa54c2e7c774695af86f8bbb859a0ef9b0d4f0031dd1df5ea7ccc752d
 05066660ea3bc0568269cd95c212ad2bf2f2ced4e4cdb1f2bc5f766e88e4862b
-48787b48d5c68f0c1dd825c32580af741cc0ee314f08133135c1e86d87a24a95
+577568395203f1f1225f2982b637f7d5e61b47a0f546ba16d46020b471b74076
 ut aliquip ex ea commodo consequat. Duis aute
 677b3a14c1875eda0ca0c3d6c340413848b1ab0bf9d448dddd5714cbc6d170b9
 edc9f99dfd626ddc5604f8b387c7a88cc6fcb17cef46a3b917c2f8ffbd449982
-93bdf47495854f7cfaaca1af8c0f39ef4a3024c10eb0dea23726b0e06ef29e84
+f6fd7efe95a5c6c4bb46d9b429e3faf65b1ce439e116742d42b928e61de52385
 irure dolor in reprehenderit in voluptate velit
 8a73c33ebfd11d78db984486a298149d034051c61cdaf6ff7e783e46a6763edf
 44df1cbd33439b82f901bcad232f3908331330edad0c9b9af35d62f524fd92b4
-a375a71d0541057ccc96719150dfe10b6e6f486b19cf4a0835e19605413a8417
+422f8cfe8060cf6c3d9fd66f68e3c9977adb683aea2788029308bbe9bc50d728
 esse cillum dolore eu fugiat nulla pariatur.
 021007fa46b46110e7dd6c7f1bb392499d7461950efd884e6bb4260d57906b6f
 0fa9723e149c0772d16ae95b744186f419b48adcbfe685c99b53f6db44ba2668
-163478a76f8f1432533fbdd1284d65c89f37479e54f20841c6ce4eba56c73854
+512e62c8c8c833cfbd9d361442cb00d63c0a3fd8964cfd2fedc17c7c25ec2d4b
 Excepteur sint occaecat cupidatat non proident,
 d61f8a0722e9d38c691fe22613ef68c83a498dd24e3c382ee1abfa665d632371
 90c71412c2ca0c2e5789a98fb7ce36179d3c7f8b164f9aa07df56d44c9e9e96d
-356b0470c6eea75cad7a108179e232905b23bdaf03c2824c6e619d503ee93677
+12a794c1aa13d561c9c6111cee631ca9d0a321718d67d3416add9de1693ba41e
 sunt in culpa qui officia deserunt mollit anim id est laborum.
 75a105029f6b8c00429c427ffc7a64d84dbcdf2728ce0d2df9133cef91c9f8d3
 5944b5b0480e84dc233bcc37101c23077542433c868c67325e9c501cfd1b8151
-a47e2b007dc25bb279e197a1b91f67ecebe2ddd8791cd32dd2cb76dd21ed943f
+2659ff95fc74b6215c1dc741e85b7a9710101b30620212f80eb59c3c55993f9d

xmr-stak-cpu.vcxproj

@@ -214,6 +214,7 @@
     <MASM Include="crypto\asm\cnv2_main_loop.asm" />
   </ItemGroup>
   <ItemGroup>
+    <None Include="crypto\asm\cnv2_double_main_loop_sandybridge.inc" />
     <None Include="crypto\asm\cnv2_main_loop_ivybridge.inc" />
     <None Include="crypto\asm\cnv2_main_loop_ryzen.inc" />
   </ItemGroup>

xmr-stak-cpu.vcxproj.filters

@@ -266,5 +266,8 @@
     <None Include="crypto\asm\cnv2_main_loop_ryzen.inc">
       <Filter>Source Files\asm</Filter>
     </None>
+    <None Include="crypto\asm\cnv2_double_main_loop_sandybridge.inc">
+      <Filter>Source Files\asm</Filter>
+    </None>
   </ItemGroup>
 </Project>