Asm optimized variant 1

CMakeLists.txt

@@ -189,7 +189,7 @@ target_link_libraries(xmr-stak-c ${LIBS})
 
 add_library(xmr-stak-asm
     STATIC
-    "crypto/asm/cnv2_main_loop.S"
+    "crypto/asm/cn_main_loop.S"
 )
 set_property(TARGET xmr-stak-asm PROPERTY LINKER_LANGUAGE C)
 

crypto/asm/cnv2_main_loop.S → crypto/asm/cn_main_loop.S

@@ -1,10 +1,19 @@
 #define ALIGN .align
 .intel_syntax noprefix
 .section .text
+.global cnv1_mainloop_sandybridge_asm
 .global cnv2_mainloop_ivybridge_asm
 .global cnv2_mainloop_ryzen_asm
 .global cnv2_double_mainloop_sandybridge_asm
 
+ALIGN 64
+cnv1_mainloop_sandybridge_asm:
+	sub rsp, 48
+	mov rcx, rdi
+	#include "cnv1_mainloop_sandybridge.inc"
+	add rsp, 48
+	ret 0
+
 ALIGN 64
 cnv2_mainloop_ivybridge_asm:
 	sub rsp, 48

crypto/asm/cnv2_main_loop.asm → crypto/asm/cn_main_loop.asm

@@ -1,8 +1,15 @@
 _TEXT_CNV2_MAINLOOP SEGMENT PAGE READ EXECUTE
+PUBLIC cnv1_mainloop_sandybridge_asm
 PUBLIC cnv2_mainloop_ivybridge_asm
 PUBLIC cnv2_mainloop_ryzen_asm
 PUBLIC cnv2_double_mainloop_sandybridge_asm
 
+ALIGN 64
+cnv1_mainloop_sandybridge_asm PROC
+	INCLUDE cnv1_mainloop_sandybridge.inc
+	ret 0
+cnv1_mainloop_sandybridge_asm ENDP
+
 ALIGN 64
 cnv2_mainloop_ivybridge_asm PROC
 	INCLUDE cnv2_main_loop_ivybridge.inc

crypto/asm/cnv1_mainloop_sandybridge.inc

@@ -0,0 +1,70 @@
+	mov	QWORD PTR [rsp+8], rbx
+	mov	QWORD PTR [rsp+16], rbp
+	mov	QWORD PTR [rsp+24], rsi
+	mov	QWORD PTR [rsp+32], rdi
+	push	r14
+	push	r15
+	mov	rax, QWORD PTR [rcx+48]
+	mov	ebp, 524288
+	xor	rax, QWORD PTR [rcx+16]
+	mov	rdx, QWORD PTR [rcx+56]
+	xor	rdx, QWORD PTR [rcx+24]
+	mov	r8, QWORD PTR [rcx+32]
+	xor	r8, QWORD PTR [rcx]
+	movq	xmm3, rax
+	mov	rax, QWORD PTR [rcx+256]
+	mov	rdi, QWORD PTR [rcx+40]
+	movq	xmm0, rdx
+	xor	rdi, QWORD PTR [rcx+8]
+	mov	rdx, r8
+	mov	r15, QWORD PTR [rcx+264]
+	and	edx, 2097136
+	mov	r14, QWORD PTR [rax+35]
+	xor	r14, QWORD PTR [rcx+192]
+	mov	rsi, QWORD PTR [rcx+224]
+	punpcklqdq xmm3, xmm0
+	movdqu	xmm2, XMMWORD PTR [rdx+rsi]
+
+	ALIGN 64
+cnv1_mainloop_sandybridge:
+	movq	xmm0, rdi
+	movq	xmm1, r8
+	punpcklqdq xmm1, xmm0
+	aesenc	xmm2, xmm1
+	movq r10, xmm2
+	mov r9d, r10d
+	and r9d, 2097136
+	add r9, rsi
+	movdqa	xmm0, xmm2
+	pxor	xmm0, xmm3
+	movdqa	xmm3, xmm2
+	movdqu	XMMWORD PTR [rdx+rsi], xmm0
+	psrldq	xmm0, 11
+	movq	rax, xmm0
+	movzx	eax, al
+	movzx	eax, BYTE PTR [rax+r15]
+	mov	BYTE PTR [rsi+rdx+11], al
+	mov	rbx, QWORD PTR [r9]
+	mov	r11, QWORD PTR [r9+8]
+	mov	rax, rbx
+	mul	r10
+	add	r8, rdx
+	mov	QWORD PTR [r9], r8
+	add	rdi, rax
+	mov	rax, r14
+	xor	rax, rdi
+	mov	QWORD PTR [r9+8], rax
+	xor	r8, rbx
+	mov	rdx, r8
+	and	edx, 2097136
+	movdqu	xmm2, XMMWORD PTR [rdx+rsi]
+	xor	rdi, r11
+	dec	ebp
+	jne	cnv1_mainloop_sandybridge
+
+	mov	rbx, QWORD PTR [rsp+24]
+	mov	rbp, QWORD PTR [rsp+32]
+	mov	rsi, QWORD PTR [rsp+40]
+	mov	rdi, QWORD PTR [rsp+48]
+	pop	r15
+	pop	r14

crypto/cryptonight.h

@@ -14,6 +14,8 @@ typedef struct {
 	uint8_t hash_state[224]; // Need only 200, explicit align
 	uint8_t* long_state;
 	uint8_t ctx_info[24]; //Use some of the extra memory for flags
+	const void* input;
+	uint8_t* variant1_table;
 } cryptonight_ctx;
 
 typedef struct {

minethd.cpp

@@ -344,12 +344,12 @@ bool minethd::self_test()
 				}
 			}
 
-			if (jconf::inst()->HaveHardwareAes() && (i == 2))
+			if (jconf::inst()->HaveHardwareAes() && (i > 0))
 			{
 				for (int j = 1; j <= 2; ++j)
 				{
 					char hash[32];
-					cn_hash_fun hash_fun = func_selector(true, 2, j);
+					cn_hash_fun hash_fun = func_selector(true, i, j);
 					cn_hash_fun_dbl hash_fun_dbl = func_dbl_selector(jconf::inst()->HaveHardwareAes(), i, j);
 
 					hash_fun(input.c_str(), input.length(), hash, ctx0);
@@ -361,7 +361,7 @@ 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 2, asm version %d) failed.", j);
+						printer::inst()->print_msg(L0, "Cryptonight hash self-test (variant %d, asm version %d) failed.", i, j);
 						return false;
 					}
 					if (!prev_input.empty())
@@ -370,7 +370,7 @@ bool minethd::self_test()
 						{
 							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);
+							printer::inst()->print_msg(L0, "Cryptonight double hash self-test (variant %d, asm version %d) failed.", i, j);
 							return false;
 						}
 					}
@@ -413,12 +413,15 @@ int minethd::pgo_instrument()
 		}
 	}
 
-	for (int i = 1; i <= 2; ++i)
+	for (int variant = 0; variant <= 2; ++variant)
 	{
-		hash_fun = func_selector(true, 2, i);
-		hash_fun_dbl = func_dbl_selector(true, 2, i);
-		hash_fun(input, sizeof(input), hash, ctx0);
-		hash_fun_dbl(input, sizeof(input), hash, input, sizeof(input), hash + 32, ctx0, ctx1);
+		for (int i = 1; i <= 2; ++i)
+		{
+			hash_fun = func_selector(true, variant, i);
+			hash_fun_dbl = func_dbl_selector(true, variant, i);
+			hash_fun(input, sizeof(input), hash, ctx0);
+			hash_fun_dbl(input, sizeof(input), hash, input, sizeof(input), hash + 32, ctx0, ctx1);
+		}
 	}
 
 	cryptonight_free_ctx(ctx0);
@@ -479,6 +482,7 @@ void minethd::consume_work()
 	iConsumeCnt++;
 }
 
+extern "C" void cnv1_mainloop_sandybridge_asm(cryptonight_ctx* ctx0);
 extern "C" void cnv2_mainloop_ivybridge_asm(cryptonight_ctx* ctx0);
 extern "C" void cnv2_mainloop_ryzen_asm(cryptonight_ctx* ctx0);
 extern "C" void cnv2_double_mainloop_sandybridge_asm(cryptonight_ctx* ctx0, cryptonight_ctx* ctx1);
@@ -490,6 +494,26 @@ uint64_t min_cycles = uint64_t(-1);
 
 ALIGN(64) uint8_t variant1_table[256];
 
+void cryptonight_hash_v1_asm(const void* input, size_t len, void* output, cryptonight_ctx* ctx0)
+{
+	keccak((const uint8_t *)input, len, ctx0->hash_state, 200);
+	cn_explode_scratchpad<MEMORY, false>((__m128i*)ctx0->hash_state, (__m128i*)ctx0->long_state);
+
+#ifdef PERFORMANCE_TUNING
+	t1 = __rdtsc();
+#endif
+	ctx0->input = input;
+	ctx0->variant1_table = variant1_table;
+	cnv1_mainloop_sandybridge_asm(ctx0);
+#ifdef PERFORMANCE_TUNING
+	t2 = __rdtsc();
+#endif
+
+	cn_implode_scratchpad<MEMORY, false>((__m128i*)ctx0->long_state, (__m128i*)ctx0->hash_state);
+	keccakf((uint64_t*)ctx0->hash_state, 24);
+	extra_hashes[ctx0->hash_state[0] & 3](ctx0->hash_state, 200, (char*)output);
+}
+
 template<int asm_version>
 void cryptonight_hash_v2_asm(const void* input, size_t len, void* output, cryptonight_ctx* ctx0)
 {
@@ -548,16 +572,23 @@ minethd::cn_hash_fun minethd::func_selector(bool bHaveAes, int variant, int asm_
 	// function as a two digit binary
 	// Digit order SOFT_AES, NO_PREFETCH, SHUFFLE, INT_MATH
 
-	if (bHaveAes && (variant == 2) && (asm_version > 0))
+	if (bHaveAes && (asm_version > 0))
 	{
-		switch (asm_version)
+		if (variant == 1)
+		{
+			return cryptonight_hash_v1_asm;
+		}
+		else if (variant == 2)
 		{
-		case 1:
-			// Intel Ivy Bridge (Xeon v2, Core i7/i5/i3 3xxx, Pentium G2xxx, Celeron G1xxx)
-			return cryptonight_hash_v2_asm<1>;
-		case 2:
-			// AMD Ryzen (1xxx and 2xxx series)
-			return cryptonight_hash_v2_asm<2>;
+			switch (asm_version)
+			{
+			case 1:
+				// Intel Ivy Bridge (Xeon v2, Core i7/i5/i3 3xxx, Pentium G2xxx, Celeron G1xxx)
+				return cryptonight_hash_v2_asm<1>;
+			case 2:
+				// AMD Ryzen (1xxx and 2xxx series)
+				return cryptonight_hash_v2_asm<2>;
+			}
 		}
 	}
 

xmr-stak-cpu.vcxproj

@@ -211,7 +211,7 @@
     <ClInclude Include="webdesign.h" />
   </ItemGroup>
   <ItemGroup>
-    <MASM Include="crypto\asm\cnv2_main_loop.asm" />
+    <MASM Include="crypto\asm\cn_main_loop.asm" />
   </ItemGroup>
   <ItemGroup>
     <None Include="crypto\asm\cnv2_double_main_loop_sandybridge.inc" />

xmr-stak-cpu.vcxproj.filters

@@ -254,11 +254,6 @@
       <Filter>Header Files</Filter>
     </ClInclude>
   </ItemGroup>
-  <ItemGroup>
-    <MASM Include="crypto\asm\cnv2_main_loop.asm">
-      <Filter>Source Files\asm</Filter>
-    </MASM>
-  </ItemGroup>
   <ItemGroup>
     <None Include="crypto\asm\cnv2_main_loop_ivybridge.inc">
       <Filter>Source Files\asm</Filter>
@@ -270,4 +265,9 @@
       <Filter>Source Files\asm</Filter>
     </None>
   </ItemGroup>
+  <ItemGroup>
+    <MASM Include="crypto\asm\cn_main_loop.asm">
+      <Filter>Source Files\asm</Filter>
+    </MASM>
+  </ItemGroup>
 </Project>