Merge pull request #15 from JohnnyFFM/avx512

Avx512

blagominer.sln

@@ -13,12 +13,16 @@ Global
 		DebugAVX|x86 = DebugAVX|x86
 		DebugAVX2|x64 = DebugAVX2|x64
 		DebugAVX2|x86 = DebugAVX2|x86
+		DebugAVX512|x64 = DebugAVX512|x64
+		DebugAVX512|x86 = DebugAVX512|x86
 		Release|x64 = Release|x64
 		Release|x86 = Release|x86
 		ReleaseAVX|x64 = ReleaseAVX|x64
 		ReleaseAVX|x86 = ReleaseAVX|x86
 		ReleaseAVX2|x64 = ReleaseAVX2|x64
 		ReleaseAVX2|x86 = ReleaseAVX2|x86
+		ReleaseAVX512|x64 = ReleaseAVX512|x64
+		ReleaseAVX512|x86 = ReleaseAVX512|x86
 	EndGlobalSection
 	GlobalSection(ProjectConfigurationPlatforms) = postSolution
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.Debug|x64.ActiveCfg = Debug|x64
@@ -33,6 +37,10 @@ Global
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.DebugAVX2|x64.Build.0 = DebugAVX2|x64
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.DebugAVX2|x86.ActiveCfg = DebugAVX2|Win32
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.DebugAVX2|x86.Build.0 = DebugAVX2|Win32
+		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.DebugAVX512|x64.ActiveCfg = DebugAVX512|x64
+		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.DebugAVX512|x64.Build.0 = DebugAVX512|x64
+		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.DebugAVX512|x86.ActiveCfg = DebugAVX512|Win32
+		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.DebugAVX512|x86.Build.0 = DebugAVX512|Win32
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.Release|x64.ActiveCfg = Release|x64
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.Release|x64.Build.0 = Release|x64
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.Release|x86.ActiveCfg = Release|Win32
@@ -45,6 +53,10 @@ Global
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.ReleaseAVX2|x64.Build.0 = ReleaseAVX2|x64
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.ReleaseAVX2|x86.ActiveCfg = ReleaseAVX2|Win32
 		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.ReleaseAVX2|x86.Build.0 = ReleaseAVX2|Win32
+		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.ReleaseAVX512|x64.ActiveCfg = ReleaseAVX512|x64
+		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.ReleaseAVX512|x64.Build.0 = ReleaseAVX512|x64
+		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.ReleaseAVX512|x86.ActiveCfg = ReleaseAVX512|Win32
+		{7E2BCD41-DBC7-4565-8C4D-E92F039C2286}.ReleaseAVX512|x86.Build.0 = ReleaseAVX512|Win32
 	EndGlobalSection
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE

blagominer/blagominer.h

@@ -5,6 +5,9 @@
 #include "shabal.h"
 
 // blago version
+#ifdef __AVX512F__
+	char const *const version = "v1.170997_AVX512";
+#else
 #ifdef __AVX2__
 	char const *const version = "v1.170997_AVX2";
 #else
@@ -15,6 +18,7 @@
 		//	char const *const version = "v1.170997";
 	#endif
 #endif 
+#endif 
 
 extern HANDLE hHeap;							//heap
 

blagominer/blagominer.vcxproj.filters

@@ -110,6 +110,9 @@
     <ClCompile Include="mshabal_128.c">
       <Filter>Quelldateien</Filter>
     </ClCompile>
+    <ClCompile Include="mshabal_512.c">
+      <Filter>Quelldateien</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ResourceCompile Include="blagominer.rc">

blagominer/miner.conf

@@ -1,9 +1,9 @@
 {
    "Mode" : "pool",
-   "Server" : "50-50-pool.burst.cryptoguru.org",
+   "Server" : "pool.dev.burst-test.net",
    "Port": 8124,
 
-   "UpdaterAddr" : "50-50-pool.burst.cryptoguru.org",
+   "UpdaterAddr" : "pool.dev.burst-test.net",
    "UpdaterPort": "8124",
 
    "InfoAddr" : "127.0.0.1",
@@ -12,14 +12,14 @@
    "EnableProxy": false,
    "ProxyPort": 8126,
 
-   "Paths":["\\\\.\\PHYSICALDRIVE4", "j:\\plot"],
+   "Paths":["d:\\plot"],
    "CacheSize" : 16384,
    "CacheSize2" : 512000,
 
    "Debug": true,
    "UseHDDWakeUp": true,
 
-   "TargetDeadline": 400000,
+   "TargetDeadline": 32000000,
 
    "SendInterval": 100,
    "UpdateInterval": 950,
@@ -31,5 +31,5 @@
    "WinSizeX": 76,
    "WinSizeY": 50,
 
-  "POC2StartBlock": 502000
+  "POC2StartBlock": 5020
 }

blagominer/mshabal.h

@@ -111,6 +111,8 @@ extern "C" {
 
 #define MSHABAL256_FACTOR 2
 
+#define MSHABAL512_FACTOR 4
+
   /*
   * The context structure for a Shabal computation. Contents are
   * private. Such a structure should be allocated and released by
@@ -151,13 +153,69 @@ extern "C" {
 #pragma pack()
 
 
+  /*
+  * The context structure for a Shabal computation. Contents are
+  * private. Such a structure should be allocated and released by
+  * the caller, in any memory area.
+  */
+#pragma pack(1)
+  typedef struct {
+	  unsigned char buf0[64];
+	  unsigned char buf1[64];
+	  unsigned char buf2[64];
+	  unsigned char buf3[64];
+	  unsigned char buf4[64];
+	  unsigned char buf5[64];
+	  unsigned char buf6[64];
+	  unsigned char buf7[64];
+	  unsigned char buf8[64];
+	  unsigned char buf9[64];
+	  unsigned char buf10[64];
+	  unsigned char buf11[64];
+	  unsigned char buf12[64];
+	  unsigned char buf13[64];
+	  unsigned char buf14[64];
+	  unsigned char buf15[64];
+	  unsigned char* xbuf0;
+	  unsigned char* xbuf1;
+	  unsigned char* xbuf2;
+	  unsigned char* xbuf3;
+	  unsigned char* xbuf4;
+	  unsigned char* xbuf5;
+	  unsigned char* xbuf6;
+	  unsigned char* xbuf7;
+	  unsigned char* xbuf8;
+	  unsigned char* xbuf9;
+	  unsigned char* xbuf10;
+	  unsigned char* xbuf11;
+	  unsigned char* xbuf12;
+	  unsigned char* xbuf13;
+	  unsigned char* xbuf14;
+	  unsigned char* xbuf15;
+	  size_t ptr;
+	  mshabal_u32 state[(12 + 16 + 16) * 4 * MSHABAL512_FACTOR];
+	  mshabal_u32 Whigh, Wlow;
+	  unsigned out_size;
+  } mshabal512_context;
+
+
+#pragma pack(1)
+  typedef struct {
+	  mshabal_u32 state[(12 + 16 + 16) * 4 * MSHABAL512_FACTOR];
+	  mshabal_u32 Whigh, Wlow;
+	  unsigned out_size;
+  } mshabal512_context_fast;
+
+#pragma pack()
+
   /*
    * Initialize a context structure. The output size must be a multiple
    * of 32, between 32 and 512 (inclusive). The output size is expressed
    * in bits.
    */
   void simd128_mshabal_init(mshabal_context *sc, unsigned out_size);
   void simd256_mshabal_init(mshabal256_context *sc, unsigned out_size);
+  void simd512_mshabal_init(mshabal512_context *sc, unsigned out_size);
 
   /*
    * Process some more data bytes; four chunks of data, pointed to by
@@ -176,6 +234,12 @@ extern "C" {
 	  void *data0, void *data1, void *data2, void *data3,
 	  void *data4, void *data5, void *data6, void *data7,
 	  size_t len);
+  void simd512_mshabal(mshabal512_context *sc,
+	  void *data0, void *data1, void *data2, void *data3,
+	  void *data4, void *data5, void *data6, void *data7, 
+	  void *data8, void *data9, void *data10, void *data11,
+	  void *data12, void *data13, void *data14, void *data15,
+	  size_t len);
   /*
    * Terminate the Shabal computation incarnated by the provided context
    * structure. "n" shall be a value between 0 and 7 (inclusive): this is
@@ -205,7 +269,16 @@ extern "C" {
 	  unsigned n,
 	  void *dst0, void *dst1, void *dst2, void *dst3,
 	  void *dst4, void *dst5, void *dst6, void *dst7);
-
+  void simd512_mshabal_close(mshabal512_context *sc,
+	  unsigned ub0, unsigned ub1, unsigned ub2, unsigned ub3,
+	  unsigned ub4, unsigned ub5, unsigned ub6, unsigned ub7,
+	  unsigned ub8, unsigned ub9, unsigned ub10, unsigned ub11,
+	  unsigned ub12, unsigned ub13, unsigned ub14, unsigned ub15,
+	  unsigned n,
+	  void *dst0, void *dst1, void *dst2, void *dst3,
+	  void *dst4, void *dst5, void *dst6, void *dst7,
+	  void *dst8, void *dst9, void *dst10, void *dst11,
+	  void *dst12, void *dst13, void *dst14, void *dst15);
   /*
    * Combined open and close routines
    */
@@ -220,6 +293,14 @@ extern "C" {
 		  void *u1, void *u2,
 		  void *dst0, void *dst1, void *dst2, void *dst3, void *dst4, void *dst5, void *dst6, void *dst7,
 		  unsigned n);
+  void
+	  simd512_mshabal_openclose_fast(mshabal512_context_fast *sc,
+		  void *u1, void *u2,
+		  void *dst0, void *dst1, void *dst2, void *dst3, void *dst4, void *dst5, void *dst6, void *dst7,
+		  void *dst8, void *dst9, void *dst10, void *dst11, void *dst12, void *dst13, void *dst14, void *dst15,
+		  unsigned n);
+
+
 #ifdef  __cplusplus
 }
 #endif

blagominer/mshabal_128.c

@@ -634,10 +634,10 @@ extern "C" {
 	  }
 
 	  //transfer results to ram
-	  for (j = 0; j < 12; j++)
-		  _mm_storeu_si128((__m128i *)sc->state + j, A[j]);
-	  for (j = 0; j < 16; j++) {
-		  _mm_storeu_si128((__m128i *)sc->state + j + 12, B[j]);
+	  //for (j = 0; j < 12; j++)
+		 // _mm_storeu_si128((__m128i *)sc->state + j, A[j]);
+	  for (j = 8; j < 10; j++) {
+		//  _mm_storeu_si128((__m128i *)sc->state + j + 12, B[j]);
 		  _mm_storeu_si128((__m128i *)sc->state + j + 28, C[j]);
 	  }
   }
@@ -655,7 +655,7 @@ extern "C" {
 	  //extract results
 	  out_size_w32 = sc->out_size >> 5;
 	  off = 4 * (28 + (16 - out_size_w32));
-	  for (z = 0; z < out_size_w32; z++) {
+	  for (z = 0; z < 2; z++) {
 		  unsigned y = off + (z << 2);
 		  ((u32 *)dst0)[z] = sc->state[y + 0];
 		  ((u32 *)dst1)[z] = sc->state[y + 1];

blagominer/mshabal_256.c

@@ -698,10 +698,10 @@ extern "C" {
 	  }
 
 	  //transfer results to ram
-	  for (j = 0; j < 12; j++)
-		  _mm256_storeu_si256((__m256i *)sc->state + j, A[j]);
-	  for (j = 0; j < 16; j++) {
-		  _mm256_storeu_si256((__m256i *)sc->state + j + 12, B[j]);
+	 // for (j = 0; j < 12; j++)
+		 // _mm256_storeu_si256((__m256i *)sc->state + j, A[j]);
+	  for (j = 8; j < 10; j++) {
+		  //_mm256_storeu_si256((__m256i *)sc->state + j + 12, B[j]);
 		  _mm256_storeu_si256((__m256i *)sc->state + j + 28, C[j]);
 	  }
   }
@@ -724,7 +724,7 @@ extern "C" {
 	  //extract results
 	  out_size_w32 = sc->out_size >> 5;
 	  off = MSHABAL256_FACTOR * 4 * (28 + (16 - out_size_w32));
-	  for (z = 0; z < out_size_w32; z++) {
+	  for (z = 0; z < 2; z++) {
 		  unsigned y = off + MSHABAL256_FACTOR * (z << 2);
 		  ((u32 *)dst0)[z] = sc->state[y + 0];
 		  ((u32 *)dst1)[z] = sc->state[y + 1];