lyra2: support for SM 2.1 cards (GTX 460)

also fix the build (scrypt) for this arch.

else, 318,26 kH/s on a GTX 460...

lyra2/cuda_lyra2.cu

@@ -6,11 +6,15 @@
 #include <stdio.h>
 #include <memory.h>
 
-#include "cuda_lyra2_vectors.h"
-
 #define TPB50 16
 #define TPB52 8
 
+#include "cuda_lyra2_sm2.cuh"
+
+#if !defined(__CUDA_ARCH__) ||  __CUDA_ARCH__ >= 500
+
+#include "cuda_lyra2_vectors.h"
+
 #define uint2x4 uint28
 #define memshift 3
 
@@ -238,6 +242,11 @@ void lyra2_gpu_hash_32(uint32_t threads, uint32_t startNounce, uint2 *g_hash)
 		g_hash[thread + threads*3] = ((uint2*)state)[3];
 	}
 }
+#else
+/* for unsupported SM arch */
+__device__ void* DMatrix;
+__global__ void lyra2_gpu_hash_32(uint32_t threads, uint32_t startNounce, uint2 *g_hash) {}
+#endif
 
 __host__
 void lyra2_cpu_init(int thr_id, uint32_t threads, uint64_t* d_matrix)
@@ -252,9 +261,14 @@ void lyra2_cpu_hash_32(int thr_id, uint32_t threads, uint32_t startNounce, uint6
 	int dev_id = device_map[thr_id % MAX_GPUS];
 	uint32_t tpb = TPB52;
 	if (device_sm[dev_id] == 500) tpb = TPB50;
+	if (device_sm[dev_id] <= 300) tpb = TPB30;
 
 	dim3 grid((threads + tpb - 1) / tpb);
 	dim3 block(tpb);
 
-	lyra2_gpu_hash_32 <<< grid, block >>> (threads, startNounce, (uint2*)d_hash);
+	if (device_sm[dev_id] >= 500)
+		lyra2_gpu_hash_32 <<< grid, block >>> (threads, startNounce, (uint2*)d_hash);
+	else
+		lyra2_gpu_hash_32_sm2 <<< grid, block >>> (threads, startNounce, d_hash);
+
 }

lyra2/cuda_lyra2_sm2.cuh

@@ -0,0 +1,225 @@
+#include <memory.h>
+
+#ifdef __INTELLISENSE__
+/* just for vstudio code colors */
+#undef __CUDA_ARCH__
+#define __CUDA_ARCH__ 300
+#endif
+
+#include "cuda_helper.h"
+
+#define TPB30 160
+
+#if __CUDA_ARCH__ >= 200 && __CUDA_ARCH__ <= 350
+
+static __constant__ uint2 blake2b_IV[8] = {
+	{ 0xf3bcc908, 0x6a09e667 },
+	{ 0x84caa73b, 0xbb67ae85 },
+	{ 0xfe94f82b, 0x3c6ef372 },
+	{ 0x5f1d36f1, 0xa54ff53a },
+	{ 0xade682d1, 0x510e527f },
+	{ 0x2b3e6c1f, 0x9b05688c },
+	{ 0xfb41bd6b, 0x1f83d9ab },
+	{ 0x137e2179, 0x5be0cd19 }
+};
+
+#define reduceDuplexRow(rowIn, rowInOut, rowOut) { \
+	for (int i = 0; i < 8; i++) { \
+		for (int j = 0; j < 12; j++) \
+			state[j] ^= Matrix[12 * i + j][rowIn] + Matrix[12 * i + j][rowInOut]; \
+		round_lyra(state); \
+		for (int j = 0; j < 12; j++) \
+			Matrix[j + 12 * i][rowOut] ^= state[j]; \
+		Matrix[0 + 12 * i][rowInOut] ^= state[11]; \
+		Matrix[1 + 12 * i][rowInOut] ^= state[0]; \
+		Matrix[2 + 12 * i][rowInOut] ^= state[1]; \
+		Matrix[3 + 12 * i][rowInOut] ^= state[2]; \
+		Matrix[4 + 12 * i][rowInOut] ^= state[3]; \
+		Matrix[5 + 12 * i][rowInOut] ^= state[4]; \
+		Matrix[6 + 12 * i][rowInOut] ^= state[5]; \
+		Matrix[7 + 12 * i][rowInOut] ^= state[6]; \
+		Matrix[8 + 12 * i][rowInOut] ^= state[7]; \
+		Matrix[9 + 12 * i][rowInOut] ^= state[8]; \
+		Matrix[10+ 12 * i][rowInOut] ^= state[9]; \
+		Matrix[11+ 12 * i][rowInOut] ^= state[10]; \
+	} \
+  }
+
+#define absorbblock(in)  { \
+	state[0] ^= Matrix[0][in]; \
+	state[1] ^= Matrix[1][in]; \
+	state[2] ^= Matrix[2][in]; \
+	state[3] ^= Matrix[3][in]; \
+	state[4] ^= Matrix[4][in]; \
+	state[5] ^= Matrix[5][in]; \
+	state[6] ^= Matrix[6][in]; \
+	state[7] ^= Matrix[7][in]; \
+	state[8] ^= Matrix[8][in]; \
+	state[9] ^= Matrix[9][in]; \
+	state[10] ^= Matrix[10][in]; \
+	state[11] ^= Matrix[11][in]; \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+	round_lyra(state); \
+  }
+
+static __device__ __forceinline__
+void Gfunc(uint2 & a, uint2 &b, uint2 &c, uint2 &d)
+{
+	a += b; d ^= a; d = SWAPUINT2(d);
+	c += d; b ^= c; b = ROR2(b, 24);
+	a += b; d ^= a; d = ROR2(d, 16);
+	c += d; b ^= c; b = ROR2(b, 63);
+}
+
+__device__ __forceinline__
+static void round_lyra(uint2 *s)
+{
+	Gfunc(s[0], s[4], s[8],  s[12]);
+	Gfunc(s[1], s[5], s[9],  s[13]);
+	Gfunc(s[2], s[6], s[10], s[14]);
+	Gfunc(s[3], s[7], s[11], s[15]);
+	Gfunc(s[0], s[5], s[10], s[15]);
+	Gfunc(s[1], s[6], s[11], s[12]);
+	Gfunc(s[2], s[7], s[8],  s[13]);
+	Gfunc(s[3], s[4], s[9],  s[14]);
+}
+
+__device__ __forceinline__
+void reduceDuplexRowSetup(const int rowIn, const int rowInOut, const int rowOut, uint2 state[16], uint2 Matrix[96][8])
+{
+#if __CUDA_ARCH__ > 500
+	#pragma unroll
+#endif
+	for (int i = 0; i < 8; i++)
+	{
+		#pragma unroll
+		for (int j = 0; j < 12; j++)
+			state[j] ^= Matrix[12 * i + j][rowIn] + Matrix[12 * i + j][rowInOut];
+
+		round_lyra(state);
+
+		#pragma unroll
+		for (int j = 0; j < 12; j++)
+			Matrix[j + 84 - 12 * i][rowOut] = Matrix[12 * i + j][rowIn] ^ state[j];
+
+		Matrix[0 + 12 * i][rowInOut] ^= state[11];
+		Matrix[1 + 12 * i][rowInOut] ^= state[0];
+		Matrix[2 + 12 * i][rowInOut] ^= state[1];
+		Matrix[3 + 12 * i][rowInOut] ^= state[2];
+		Matrix[4 + 12 * i][rowInOut] ^= state[3];
+		Matrix[5 + 12 * i][rowInOut] ^= state[4];
+		Matrix[6 + 12 * i][rowInOut] ^= state[5];
+		Matrix[7 + 12 * i][rowInOut] ^= state[6];
+		Matrix[8 + 12 * i][rowInOut] ^= state[7];
+		Matrix[9 + 12 * i][rowInOut] ^= state[8];
+		Matrix[10 + 12 * i][rowInOut] ^= state[9];
+		Matrix[11 + 12 * i][rowInOut] ^= state[10];
+	}
+}
+
+__global__ __launch_bounds__(TPB30, 1)
+void lyra2_gpu_hash_32_sm2(uint32_t threads, uint32_t startNounce, uint64_t *g_hash)
+{
+	uint32_t thread = (blockDim.x * blockIdx.x + threadIdx.x);
+	if (thread < threads)
+	{
+		uint2 state[16];
+
+		#pragma unroll
+		for (int i = 0; i<4; i++) {
+			LOHI(state[i].x, state[i].y, g_hash[threads*i + thread]);
+		} //password
+
+		#pragma unroll
+		for (int i = 0; i<4; i++) {
+			state[i + 4] = state[i];
+		} //salt
+
+		#pragma unroll
+		for (int i = 0; i<8; i++) {
+			state[i + 8] = blake2b_IV[i];
+		}
+
+		// blake2blyra x2
+		//#pragma unroll 24
+		for (int i = 0; i<24; i++) {
+			round_lyra(state);
+		} //because 12 is not enough
+
+		uint2 Matrix[96][8]; // not cool
+
+		// reducedSqueezeRow0
+		#pragma unroll 8
+		for (int i = 0; i < 8; i++)
+		{
+			#pragma unroll 12
+			for (int j = 0; j<12; j++) {
+				Matrix[j + 84 - 12 * i][0] = state[j];
+			}
+			round_lyra(state);
+		}
+
+		// reducedSqueezeRow1
+		#pragma unroll 8
+		for (int i = 0; i < 8; i++)
+		{
+			#pragma unroll 12
+			for (int j = 0; j<12; j++) {
+				state[j] ^= Matrix[j + 12 * i][0];
+			}
+			round_lyra(state);
+			#pragma unroll 12
+			for (int j = 0; j<12; j++) {
+				Matrix[j + 84 - 12 * i][1] = Matrix[j + 12 * i][0] ^ state[j];
+			}
+		}
+
+		reduceDuplexRowSetup(1, 0, 2, state, Matrix);
+		reduceDuplexRowSetup(2, 1, 3, state, Matrix);
+		reduceDuplexRowSetup(3, 0, 4, state, Matrix);
+		reduceDuplexRowSetup(4, 3, 5, state, Matrix);
+		reduceDuplexRowSetup(5, 2, 6, state, Matrix);
+		reduceDuplexRowSetup(6, 1, 7, state, Matrix);
+
+		uint32_t rowa;
+		rowa = state[0].x & 7;
+		reduceDuplexRow(7, rowa, 0);
+		rowa = state[0].x & 7;
+		reduceDuplexRow(0, rowa, 3);
+		rowa = state[0].x & 7;
+		reduceDuplexRow(3, rowa, 6);
+		rowa = state[0].x & 7;
+		reduceDuplexRow(6, rowa, 1);
+		rowa = state[0].x & 7;
+		reduceDuplexRow(1, rowa, 4);
+		rowa = state[0].x & 7;
+		reduceDuplexRow(4, rowa, 7);
+		rowa = state[0].x & 7;
+		reduceDuplexRow(7, rowa, 2);
+		rowa = state[0].x & 7;
+		reduceDuplexRow(2, rowa, 5);
+
+		absorbblock(rowa);
+
+		#pragma unroll
+		for (int i = 0; i<4; i++) {
+			g_hash[threads*i + thread] = devectorize(state[i]);
+		}
+
+	} //thread
+}
+
+#else
+/* if __CUDA_ARCH__ < 200 .. host */
+__global__ void lyra2_gpu_hash_32_sm2(uint32_t threads, uint32_t startNounce, uint64_t *g_hash) {}
+#endif

lyra2/cuda_lyra2v2.cu

@@ -1,4 +1,5 @@
 #include <stdio.h>
+#include <stdint.h>
 #include <memory.h>
 
 #ifdef __INTELLISENSE__
@@ -9,16 +10,12 @@
 #define TPB52 10
 #define TPB50 16
 
-#include "cuda_lyra2_vectors.h"
-
 #include "cuda_lyra2v2_sm3.cuh"
 
-#ifndef __CUDA_ARCH__
-__device__ void *DMatrix;
-#endif
-
 #if __CUDA_ARCH__ >= 500
 
+#include "cuda_lyra2_vectors.h"
+
 #define Nrow 4
 #define Ncol 4
 #define u64type uint2
@@ -346,6 +343,8 @@ void lyra2v2_gpu_hash_32(uint32_t threads, uint32_t startNounce, uint2 *outputHa
 	}
 }
 #else
+#include "cuda_helper.h"
+__device__ void* DMatrix;
 __global__ void lyra2v2_gpu_hash_32(uint32_t threads, uint32_t startNounce, uint2 *outputHash) {}
 #endif
 

lyra2/lyra2RE.cu

@@ -15,6 +15,7 @@ static uint64_t* d_matrix[MAX_GPUS];
 extern void blake256_cpu_init(int thr_id, uint32_t threads);
 extern void blake256_cpu_hash_80(const int thr_id, const uint32_t threads, const uint32_t startNonce, uint64_t *Hash, int order);
 extern void blake256_cpu_setBlock_80(uint32_t *pdata);
+
 extern void keccak256_cpu_hash_32(int thr_id, uint32_t threads, uint32_t startNonce, uint64_t *d_outputHash, int order);
 extern void keccak256_cpu_init(int thr_id, uint32_t threads);
 extern void keccak256_cpu_free(int thr_id);
@@ -110,7 +111,7 @@ extern "C" int scanhash_lyra2(int thr_id, struct work* work, uint32_t max_nonce,
 		init[thr_id] = true;
 	}
 
-	uint32_t endiandata[20];
+	uint32_t _ALIGN(128) endiandata[20];
 	for (int k=0; k < 20; k++)
 		be32enc(&endiandata[k], pdata[k]);
 

scrypt/nv_kernel2.cu

@@ -24,6 +24,8 @@
 	#define __ldg(x) (*(x))
 #endif
 
+#if !defined(__CUDA_ARCH__) ||  __CUDA_ARCH__ >= 300
+
 // grab lane ID
 static __device__ __inline__ unsigned int __laneId() { unsigned int laneId; asm( "mov.u32 %0, %%laneid;" : "=r"( laneId ) ); return laneId; }
 
@@ -635,3 +637,6 @@ template <int ALGO> __global__ void nv2_scrypt_core_kernelB_LG(uint32_t *g_odata
 
 	__transposed_write_BC(B, C, (uint4*)(g_odata), 1);
 }
+
+#endif /* prevent SM 2 */
+

scrypt/titan_kernel.cu

@@ -28,6 +28,8 @@ typedef enum
 	#define __ldg(x) (*(x))
 #endif
 
+#if !defined(__CUDA_ARCH__) ||  __CUDA_ARCH__ >= 300
+
 // scratchbuf constants (pointers to scratch buffer for each warp, i.e. 32 hashes)
 __constant__ uint32_t* c_V[TOTAL_WARP_LIMIT];
 
@@ -736,3 +738,5 @@ bool TitanKernel::run_kernel(dim3 grid, dim3 threads, int WARPS_PER_BLOCK, int t
 
 	return success;
 }
+
+#endif /* prevent SM 2 */