groestl: use sp bitslice enhancement, prepare SM 2.x variant

todo: simd512 SM 2.x variant (shfl op), and groestl/myriad functions
10 years ago · b521acb480
9 changed files with 532 additions and 468 deletions
--- a/Makefile.am
+++ b/Makefile.am
@ -71,8 +71,10 @@ ccminer_CPPFLAGS = @LIBCURL_CPPFLAGS@ @OPENMP_CFLAGS@ $(CPPFLAGS) $(PTHREAD_FLAG
				@@ -71,8 +71,10 @@ ccminer_CPPFLAGS = @LIBCURL_CPPFLAGS@ @OPENMP_CFLAGS@ $(CPPFLAGS) $(PTHREAD_FLAG

 nvcc_ARCH  = -gencode=arch=compute_50,code=\"sm_50,compute_50\"
 #nvcc_ARCH += -gencode=arch=compute_52,code=\"sm_52,compute_52\"
-#nvcc_ARCH += -gencode=arch=compute_35,code=\"sm_35,compute_35\"
-#nvcc_ARCH += -gencode=arch=compute_30,code=\"sm_30,compute_30\"
+
+#nvcc_ARCH = -gencode=arch=compute_35,code=\"sm_35,compute_35\"
+#nvcc_ARCH = -gencode=arch=compute_30,code=\"sm_30,compute_30\"
+#nvcc_ARCH = -gencode=arch=compute_20,code=\"sm_21,compute_20\"

 nvcc_FLAGS = $(nvcc_ARCH) @CUDA_INCLUDES@ -I. @CUDA_CFLAGS@
 nvcc_FLAGS += $(JANSSON_INCLUDES) --ptxas-options="-v"
--- a/bitslice_transformations_quad.cu
+++ b/bitslice_transformations_quad.cu
@ -1,442 +1,133 @@
				@@ -1,442 +1,133 @@
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 300
-/**
- * __shfl() returns the value of var held by the thread whose ID is given by srcLane.
- * If srcLane is outside the range 0..width-1, the thread's own value of var is returned.
- */
-#undef __shfl
-#define __shfl(var, srcLane, width) (uint32_t)(var)
-#endif
+/* File included in quark/groestl (quark/jha,nist5/X11+) and groest/myriad coins for SM 3+ */
+
+#define merge8(z,x,y)\
+	z=__byte_perm(x, y, 0x5140); \
+
+#define SWAP8(x,y)\
+	x=__byte_perm(x, y, 0x5410); \
+	y=__byte_perm(x, y, 0x7632);
+
+#define SWAP4(x,y)\
+	t = (y<<4); \
+	t = (x ^ t); \
+	t = 0xf0f0f0f0UL & t; \
+	x = (x ^ t); \
+	t=  t>>4;\
+	y=  y ^ t;
+
+#define SWAP2(x,y)\
+	t = (y<<2); \
+	t = (x ^ t); \
+	t = 0xccccccccUL & t; \
+	x = (x ^ t); \
+	t=  t>>2;\
+	y=  y ^ t;
+
+#define SWAP1(x,y)\
+	t = (y+y); \
+	t = (x ^ t); \
+	t = 0xaaaaaaaaUL & t; \
+	x = (x ^ t); \
+	t=  t>>1;\
+	y=  y ^ t;

-#ifdef __CUDA_ARCH__

 __device__ __forceinline__
-void to_bitslice_quad(uint32_t *input, uint32_t *output)
+void to_bitslice_quad(uint32_t *const __restrict__ input, uint32_t *const __restrict__ output)
 {
-    uint32_t other[8];
-    const int n = threadIdx.x % 4;
-
-    #pragma unroll
-    for (int i = 0; i < 8; i++) {
-        input[i] = __shfl((int)input[i], n ^ (3*(n >=1 && n <=2)), 4);
-        other[i] = __shfl((int)input[i], (threadIdx.x + 1) % 4, 4);
-        input[i] = __shfl((int)input[i], threadIdx.x & 2, 4);
-        other[i] = __shfl((int)other[i], threadIdx.x & 2, 4);
-        if (threadIdx.x & 1) {
-            input[i] = __byte_perm(input[i], 0, 0x1032);
-            other[i] = __byte_perm(other[i], 0, 0x1032);
-        }
-    }
-
-    output[ 0] =  (input[ 0] & 0x00000001);
-    output[ 0] |= ((other[ 0] & 0x00000001) << 1);
-    output[ 0] |= ((input[ 1] & 0x00000001) << 2);
-    output[ 0] |= ((other[ 1] & 0x00000001) << 3);
-    output[ 0] |= ((input[ 2] & 0x00000001) << 4);
-    output[ 0] |= ((other[ 2] & 0x00000001) << 5);
-    output[ 0] |= ((input[ 3] & 0x00000001) << 6);
-    output[ 0] |= ((other[ 3] & 0x00000001) << 7);
-    output[ 0] |= ((input[ 4] & 0x00000001) << 8);
-    output[ 0] |= ((other[ 4] & 0x00000001) << 9);
-    output[ 0] |= ((input[ 5] & 0x00000001) <<10);
-    output[ 0] |= ((other[ 5] & 0x00000001) <<11);
-    output[ 0] |= ((input[ 6] & 0x00000001) <<12);
-    output[ 0] |= ((other[ 6] & 0x00000001) <<13);
-    output[ 0] |= ((input[ 7] & 0x00000001) <<14);
-    output[ 0] |= ((other[ 7] & 0x00000001) <<15);
-    output[ 0] |= ((input[ 0] & 0x00000100) << 8);
-    output[ 0] |= ((other[ 0] & 0x00000100) << 9);
-    output[ 0] |= ((input[ 1] & 0x00000100) <<10);
-    output[ 0] |= ((other[ 1] & 0x00000100) <<11);
-    output[ 0] |= ((input[ 2] & 0x00000100) <<12);
-    output[ 0] |= ((other[ 2] & 0x00000100) <<13);
-    output[ 0] |= ((input[ 3] & 0x00000100) <<14);
-    output[ 0] |= ((other[ 3] & 0x00000100) <<15);
-    output[ 0] |= ((input[ 4] & 0x00000100) <<16);
-    output[ 0] |= ((other[ 4] & 0x00000100) <<17);
-    output[ 0] |= ((input[ 5] & 0x00000100) <<18);
-    output[ 0] |= ((other[ 5] & 0x00000100) <<19);
-    output[ 0] |= ((input[ 6] & 0x00000100) <<20);
-    output[ 0] |= ((other[ 6] & 0x00000100) <<21);
-    output[ 0] |= ((input[ 7] & 0x00000100) <<22);
-    output[ 0] |= ((other[ 7] & 0x00000100) <<23);
-
-    output[ 1] = ((input[ 0] & 0x00000002) >> 1);
-    output[ 1] |=  (other[ 0] & 0x00000002);
-    output[ 1] |= ((input[ 1] & 0x00000002) << 1);
-    output[ 1] |= ((other[ 1] & 0x00000002) << 2);
-    output[ 1] |= ((input[ 2] & 0x00000002) << 3);
-    output[ 1] |= ((other[ 2] & 0x00000002) << 4);
-    output[ 1] |= ((input[ 3] & 0x00000002) << 5);
-    output[ 1] |= ((other[ 3] & 0x00000002) << 6);
-    output[ 1] |= ((input[ 4] & 0x00000002) << 7);
-    output[ 1] |= ((other[ 4] & 0x00000002) << 8);
-    output[ 1] |= ((input[ 5] & 0x00000002) << 9);
-    output[ 1] |= ((other[ 5] & 0x00000002) <<10);
-    output[ 1] |= ((input[ 6] & 0x00000002) <<11);
-    output[ 1] |= ((other[ 6] & 0x00000002) <<12);
-    output[ 1] |= ((input[ 7] & 0x00000002) <<13);
-    output[ 1] |= ((other[ 7] & 0x00000002) <<14);
-    output[ 1] |= ((input[ 0] & 0x00000200) << 7);
-    output[ 1] |= ((other[ 0] & 0x00000200) << 8);
-    output[ 1] |= ((input[ 1] & 0x00000200) << 9);
-    output[ 1] |= ((other[ 1] & 0x00000200) <<10);
-    output[ 1] |= ((input[ 2] & 0x00000200) <<11);
-    output[ 1] |= ((other[ 2] & 0x00000200) <<12);
-    output[ 1] |= ((input[ 3] & 0x00000200) <<13);
-    output[ 1] |= ((other[ 3] & 0x00000200) <<14);
-    output[ 1] |= ((input[ 4] & 0x00000200) <<15);
-    output[ 1] |= ((other[ 4] & 0x00000200) <<16);
-    output[ 1] |= ((input[ 5] & 0x00000200) <<17);
-    output[ 1] |= ((other[ 5] & 0x00000200) <<18);
-    output[ 1] |= ((input[ 6] & 0x00000200) <<19);
-    output[ 1] |= ((other[ 6] & 0x00000200) <<20);
-    output[ 1] |= ((input[ 7] & 0x00000200) <<21);
-    output[ 1] |= ((other[ 7] & 0x00000200) <<22);
-
-    output[ 2] = ((input[ 0] & 0x00000004) >> 2);
-    output[ 2] |= ((other[ 0] & 0x00000004) >> 1);
-    output[ 2] |=  (input[ 1] & 0x00000004);
-    output[ 2] |= ((other[ 1] & 0x00000004) << 1);
-    output[ 2] |= ((input[ 2] & 0x00000004) << 2);
-    output[ 2] |= ((other[ 2] & 0x00000004) << 3);
-    output[ 2] |= ((input[ 3] & 0x00000004) << 4);
-    output[ 2] |= ((other[ 3] & 0x00000004) << 5);
-    output[ 2] |= ((input[ 4] & 0x00000004) << 6);
-    output[ 2] |= ((other[ 4] & 0x00000004) << 7);
-    output[ 2] |= ((input[ 5] & 0x00000004) << 8);
-    output[ 2] |= ((other[ 5] & 0x00000004) << 9);
-    output[ 2] |= ((input[ 6] & 0x00000004) <<10);
-    output[ 2] |= ((other[ 6] & 0x00000004) <<11);
-    output[ 2] |= ((input[ 7] & 0x00000004) <<12);
-    output[ 2] |= ((other[ 7] & 0x00000004) <<13);
-    output[ 2] |= ((input[ 0] & 0x00000400) << 6);
-    output[ 2] |= ((other[ 0] & 0x00000400) << 7);
-    output[ 2] |= ((input[ 1] & 0x00000400) << 8);
-    output[ 2] |= ((other[ 1] & 0x00000400) << 9);
-    output[ 2] |= ((input[ 2] & 0x00000400) <<10);
-    output[ 2] |= ((other[ 2] & 0x00000400) <<11);
-    output[ 2] |= ((input[ 3] & 0x00000400) <<12);
-    output[ 2] |= ((other[ 3] & 0x00000400) <<13);
-    output[ 2] |= ((input[ 4] & 0x00000400) <<14);
-    output[ 2] |= ((other[ 4] & 0x00000400) <<15);
-    output[ 2] |= ((input[ 5] & 0x00000400) <<16);
-    output[ 2] |= ((other[ 5] & 0x00000400) <<17);
-    output[ 2] |= ((input[ 6] & 0x00000400) <<18);
-    output[ 2] |= ((other[ 6] & 0x00000400) <<19);
-    output[ 2] |= ((input[ 7] & 0x00000400) <<20);
-    output[ 2] |= ((other[ 7] & 0x00000400) <<21);
-
-    output[ 3] = ((input[ 0] & 0x00000008) >> 3);
-    output[ 3] |= ((other[ 0] & 0x00000008) >> 2);
-    output[ 3] |= ((input[ 1] & 0x00000008) >> 1);
-    output[ 3] |=  (other[ 1] & 0x00000008);
-    output[ 3] |= ((input[ 2] & 0x00000008) << 1);
-    output[ 3] |= ((other[ 2] & 0x00000008) << 2);
-    output[ 3] |= ((input[ 3] & 0x00000008) << 3);
-    output[ 3] |= ((other[ 3] & 0x00000008) << 4);
-    output[ 3] |= ((input[ 4] & 0x00000008) << 5);
-    output[ 3] |= ((other[ 4] & 0x00000008) << 6);
-    output[ 3] |= ((input[ 5] & 0x00000008) << 7);
-    output[ 3] |= ((other[ 5] & 0x00000008) << 8);
-    output[ 3] |= ((input[ 6] & 0x00000008) << 9);
-    output[ 3] |= ((other[ 6] & 0x00000008) <<10);
-    output[ 3] |= ((input[ 7] & 0x00000008) <<11);
-    output[ 3] |= ((other[ 7] & 0x00000008) <<12);
-    output[ 3] |= ((input[ 0] & 0x00000800) << 5);
-    output[ 3] |= ((other[ 0] & 0x00000800) << 6);
-    output[ 3] |= ((input[ 1] & 0x00000800) << 7);
-    output[ 3] |= ((other[ 1] & 0x00000800) << 8);
-    output[ 3] |= ((input[ 2] & 0x00000800) << 9);
-    output[ 3] |= ((other[ 2] & 0x00000800) <<10);
-    output[ 3] |= ((input[ 3] & 0x00000800) <<11);
-    output[ 3] |= ((other[ 3] & 0x00000800) <<12);
-    output[ 3] |= ((input[ 4] & 0x00000800) <<13);
-    output[ 3] |= ((other[ 4] & 0x00000800) <<14);
-    output[ 3] |= ((input[ 5] & 0x00000800) <<15);
-    output[ 3] |= ((other[ 5] & 0x00000800) <<16);
-    output[ 3] |= ((input[ 6] & 0x00000800) <<17);
-    output[ 3] |= ((other[ 6] & 0x00000800) <<18);
-    output[ 3] |= ((input[ 7] & 0x00000800) <<19);
-    output[ 3] |= ((other[ 7] & 0x00000800) <<20);
-
-    output[ 4] = ((input[ 0] & 0x00000010) >> 4);
-    output[ 4] |= ((other[ 0] & 0x00000010) >> 3);
-    output[ 4] |= ((input[ 1] & 0x00000010) >> 2);
-    output[ 4] |= ((other[ 1] & 0x00000010) >> 1);
-    output[ 4] |=  (input[ 2] & 0x00000010);
-    output[ 4] |= ((other[ 2] & 0x00000010) << 1);
-    output[ 4] |= ((input[ 3] & 0x00000010) << 2);
-    output[ 4] |= ((other[ 3] & 0x00000010) << 3);
-    output[ 4] |= ((input[ 4] & 0x00000010) << 4);
-    output[ 4] |= ((other[ 4] & 0x00000010) << 5);
-    output[ 4] |= ((input[ 5] & 0x00000010) << 6);
-    output[ 4] |= ((other[ 5] & 0x00000010) << 7);
-    output[ 4] |= ((input[ 6] & 0x00000010) << 8);
-    output[ 4] |= ((other[ 6] & 0x00000010) << 9);
-    output[ 4] |= ((input[ 7] & 0x00000010) <<10);
-    output[ 4] |= ((other[ 7] & 0x00000010) <<11);
-    output[ 4] |= ((input[ 0] & 0x00001000) << 4);
-    output[ 4] |= ((other[ 0] & 0x00001000) << 5);
-    output[ 4] |= ((input[ 1] & 0x00001000) << 6);
-    output[ 4] |= ((other[ 1] & 0x00001000) << 7);
-    output[ 4] |= ((input[ 2] & 0x00001000) << 8);
-    output[ 4] |= ((other[ 2] & 0x00001000) << 9);
-    output[ 4] |= ((input[ 3] & 0x00001000) <<10);
-    output[ 4] |= ((other[ 3] & 0x00001000) <<11);
-    output[ 4] |= ((input[ 4] & 0x00001000) <<12);
-    output[ 4] |= ((other[ 4] & 0x00001000) <<13);
-    output[ 4] |= ((input[ 5] & 0x00001000) <<14);
-    output[ 4] |= ((other[ 5] & 0x00001000) <<15);
-    output[ 4] |= ((input[ 6] & 0x00001000) <<16);
-    output[ 4] |= ((other[ 6] & 0x00001000) <<17);
-    output[ 4] |= ((input[ 7] & 0x00001000) <<18);
-    output[ 4] |= ((other[ 7] & 0x00001000) <<19);
-
-    output[ 5] = ((input[ 0] & 0x00000020) >> 5);
-    output[ 5] |= ((other[ 0] & 0x00000020) >> 4);
-    output[ 5] |= ((input[ 1] & 0x00000020) >> 3);
-    output[ 5] |= ((other[ 1] & 0x00000020) >> 2);
-    output[ 5] |= ((input[ 2] & 0x00000020) >> 1);
-    output[ 5] |=  (other[ 2] & 0x00000020);
-    output[ 5] |= ((input[ 3] & 0x00000020) << 1);
-    output[ 5] |= ((other[ 3] & 0x00000020) << 2);
-    output[ 5] |= ((input[ 4] & 0x00000020) << 3);
-    output[ 5] |= ((other[ 4] & 0x00000020) << 4);
-    output[ 5] |= ((input[ 5] & 0x00000020) << 5);
-    output[ 5] |= ((other[ 5] & 0x00000020) << 6);
-    output[ 5] |= ((input[ 6] & 0x00000020) << 7);
-    output[ 5] |= ((other[ 6] & 0x00000020) << 8);
-    output[ 5] |= ((input[ 7] & 0x00000020) << 9);
-    output[ 5] |= ((other[ 7] & 0x00000020) <<10);
-    output[ 5] |= ((input[ 0] & 0x00002000) << 3);
-    output[ 5] |= ((other[ 0] & 0x00002000) << 4);
-    output[ 5] |= ((input[ 1] & 0x00002000) << 5);
-    output[ 5] |= ((other[ 1] & 0x00002000) << 6);
-    output[ 5] |= ((input[ 2] & 0x00002000) << 7);
-    output[ 5] |= ((other[ 2] & 0x00002000) << 8);
-    output[ 5] |= ((input[ 3] & 0x00002000) << 9);
-    output[ 5] |= ((other[ 3] & 0x00002000) <<10);
-    output[ 5] |= ((input[ 4] & 0x00002000) <<11);
-    output[ 5] |= ((other[ 4] & 0x00002000) <<12);
-    output[ 5] |= ((input[ 5] & 0x00002000) <<13);
-    output[ 5] |= ((other[ 5] & 0x00002000) <<14);
-    output[ 5] |= ((input[ 6] & 0x00002000) <<15);
-    output[ 5] |= ((other[ 6] & 0x00002000) <<16);
-    output[ 5] |= ((input[ 7] & 0x00002000) <<17);
-    output[ 5] |= ((other[ 7] & 0x00002000) <<18);
-
-    output[ 6] = ((input[ 0] & 0x00000040) >> 6);
-    output[ 6] |= ((other[ 0] & 0x00000040) >> 5);
-    output[ 6] |= ((input[ 1] & 0x00000040) >> 4);
-    output[ 6] |= ((other[ 1] & 0x00000040) >> 3);
-    output[ 6] |= ((input[ 2] & 0x00000040) >> 2);
-    output[ 6] |= ((other[ 2] & 0x00000040) >> 1);
-    output[ 6] |=  (input[ 3] & 0x00000040);
-    output[ 6] |= ((other[ 3] & 0x00000040) << 1);
-    output[ 6] |= ((input[ 4] & 0x00000040) << 2);
-    output[ 6] |= ((other[ 4] & 0x00000040) << 3);
-    output[ 6] |= ((input[ 5] & 0x00000040) << 4);
-    output[ 6] |= ((other[ 5] & 0x00000040) << 5);
-    output[ 6] |= ((input[ 6] & 0x00000040) << 6);
-    output[ 6] |= ((other[ 6] & 0x00000040) << 7);
-    output[ 6] |= ((input[ 7] & 0x00000040) << 8);
-    output[ 6] |= ((other[ 7] & 0x00000040) << 9);
-    output[ 6] |= ((input[ 0] & 0x00004000) << 2);
-    output[ 6] |= ((other[ 0] & 0x00004000) << 3);
-    output[ 6] |= ((input[ 1] & 0x00004000) << 4);
-    output[ 6] |= ((other[ 1] & 0x00004000) << 5);
-    output[ 6] |= ((input[ 2] & 0x00004000) << 6);
-    output[ 6] |= ((other[ 2] & 0x00004000) << 7);
-    output[ 6] |= ((input[ 3] & 0x00004000) << 8);
-    output[ 6] |= ((other[ 3] & 0x00004000) << 9);
-    output[ 6] |= ((input[ 4] & 0x00004000) <<10);
-    output[ 6] |= ((other[ 4] & 0x00004000) <<11);
-    output[ 6] |= ((input[ 5] & 0x00004000) <<12);
-    output[ 6] |= ((other[ 5] & 0x00004000) <<13);
-    output[ 6] |= ((input[ 6] & 0x00004000) <<14);
-    output[ 6] |= ((other[ 6] & 0x00004000) <<15);
-    output[ 6] |= ((input[ 7] & 0x00004000) <<16);
-    output[ 6] |= ((other[ 7] & 0x00004000) <<17);
-
-    output[ 7] = ((input[ 0] & 0x00000080) >> 7);
-    output[ 7] |= ((other[ 0] & 0x00000080) >> 6);
-    output[ 7] |= ((input[ 1] & 0x00000080) >> 5);
-    output[ 7] |= ((other[ 1] & 0x00000080) >> 4);
-    output[ 7] |= ((input[ 2] & 0x00000080) >> 3);
-    output[ 7] |= ((other[ 2] & 0x00000080) >> 2);
-    output[ 7] |= ((input[ 3] & 0x00000080) >> 1);
-    output[ 7] |=  (other[ 3] & 0x00000080);
-    output[ 7] |= ((input[ 4] & 0x00000080) << 1);
-    output[ 7] |= ((other[ 4] & 0x00000080) << 2);
-    output[ 7] |= ((input[ 5] & 0x00000080) << 3);
-    output[ 7] |= ((other[ 5] & 0x00000080) << 4);
-    output[ 7] |= ((input[ 6] & 0x00000080) << 5);
-    output[ 7] |= ((other[ 6] & 0x00000080) << 6);
-    output[ 7] |= ((input[ 7] & 0x00000080) << 7);
-    output[ 7] |= ((other[ 7] & 0x00000080) << 8);
-    output[ 7] |= ((input[ 0] & 0x00008000) << 1);
-    output[ 7] |= ((other[ 0] & 0x00008000) << 2);
-    output[ 7] |= ((input[ 1] & 0x00008000) << 3);
-    output[ 7] |= ((other[ 1] & 0x00008000) << 4);
-    output[ 7] |= ((input[ 2] & 0x00008000) << 5);
-    output[ 7] |= ((other[ 2] & 0x00008000) << 6);
-    output[ 7] |= ((input[ 3] & 0x00008000) << 7);
-    output[ 7] |= ((other[ 3] & 0x00008000) << 8);
-    output[ 7] |= ((input[ 4] & 0x00008000) << 9);
-    output[ 7] |= ((other[ 4] & 0x00008000) <<10);
-    output[ 7] |= ((input[ 5] & 0x00008000) <<11);
-    output[ 7] |= ((other[ 5] & 0x00008000) <<12);
-    output[ 7] |= ((input[ 6] & 0x00008000) <<13);
-    output[ 7] |= ((other[ 6] & 0x00008000) <<14);
-    output[ 7] |= ((input[ 7] & 0x00008000) <<15);
-    output[ 7] |= ((other[ 7] & 0x00008000) <<16);
+	uint32_t other[8];
+	uint32_t d[8];
+	uint32_t t;
+	const unsigned int n = threadIdx.x & 3;
+
+	#pragma unroll
+	for (int i = 0; i < 8; i++) {
+		input[i] = __shfl((int)input[i], n ^ (3*(n >=1 && n <=2)), 4);
+		other[i] = __shfl((int)input[i], (threadIdx.x + 1) & 3, 4);
+		input[i] = __shfl((int)input[i], threadIdx.x & 2, 4);
+		other[i] = __shfl((int)other[i], threadIdx.x & 2, 4);
+		if (threadIdx.x & 1) {
+			input[i] = __byte_perm(input[i], 0, 0x1032);
+			other[i] = __byte_perm(other[i], 0, 0x1032);
+		}
+	}
+
+	merge8(d[0], input[0], input[4]);
+	merge8(d[1], other[0], other[4]);
+	merge8(d[2], input[1], input[5]);
+	merge8(d[3], other[1], other[5]);
+	merge8(d[4], input[2], input[6]);
+	merge8(d[5], other[2], other[6]);
+	merge8(d[6], input[3], input[7]);
+	merge8(d[7], other[3], other[7]);
+
+	SWAP1(d[0], d[1]);
+	SWAP1(d[2], d[3]);
+	SWAP1(d[4], d[5]);
+	SWAP1(d[6], d[7]);
+
+	SWAP2(d[0], d[2]);
+	SWAP2(d[1], d[3]);
+	SWAP2(d[4], d[6]);
+	SWAP2(d[5], d[7]);
+
+	SWAP4(d[0], d[4]);
+	SWAP4(d[1], d[5]);
+	SWAP4(d[2], d[6]);
+	SWAP4(d[3], d[7]);
+
+	output[0] = d[0];
+	output[1] = d[1];
+	output[2] = d[2];
+	output[3] = d[3];
+	output[4] = d[4];
+	output[5] = d[5];
+	output[6] = d[6];
+	output[7] = d[7];
 }

 __device__ __forceinline__
-void from_bitslice_quad(uint32_t *input, uint32_t *output)
+void from_bitslice_quad(const uint32_t *const __restrict__ input, uint32_t *const __restrict__ output)
 {
-    output[ 0] = ((input[ 0] & 0x00000100) >> 8);
-    output[ 0] |= ((input[ 1] & 0x00000100) >> 7);
-    output[ 0] |= ((input[ 2] & 0x00000100) >> 6);
-    output[ 0] |= ((input[ 3] & 0x00000100) >> 5);
-    output[ 0] |= ((input[ 4] & 0x00000100) >> 4);
-    output[ 0] |= ((input[ 5] & 0x00000100) >> 3);
-    output[ 0] |= ((input[ 6] & 0x00000100) >> 2);
-    output[ 0] |= ((input[ 7] & 0x00000100) >> 1);
-    output[ 0] |= ((input[ 0] & 0x01000000) >>16);
-    output[ 0] |= ((input[ 1] & 0x01000000) >>15);
-    output[ 0] |= ((input[ 2] & 0x01000000) >>14);
-    output[ 0] |= ((input[ 3] & 0x01000000) >>13);
-    output[ 0] |= ((input[ 4] & 0x01000000) >>12);
-    output[ 0] |= ((input[ 5] & 0x01000000) >>11);
-    output[ 0] |= ((input[ 6] & 0x01000000) >>10);
-    output[ 0] |= ((input[ 7] & 0x01000000) >> 9);
-    output[ 2] = ((input[ 0] & 0x00000200) >> 9);
-    output[ 2] |= ((input[ 1] & 0x00000200) >> 8);
-    output[ 2] |= ((input[ 2] & 0x00000200) >> 7);
-    output[ 2] |= ((input[ 3] & 0x00000200) >> 6);
-    output[ 2] |= ((input[ 4] & 0x00000200) >> 5);
-    output[ 2] |= ((input[ 5] & 0x00000200) >> 4);
-    output[ 2] |= ((input[ 6] & 0x00000200) >> 3);
-    output[ 2] |= ((input[ 7] & 0x00000200) >> 2);
-    output[ 2] |= ((input[ 0] & 0x02000000) >>17);
-    output[ 2] |= ((input[ 1] & 0x02000000) >>16);
-    output[ 2] |= ((input[ 2] & 0x02000000) >>15);
-    output[ 2] |= ((input[ 3] & 0x02000000) >>14);
-    output[ 2] |= ((input[ 4] & 0x02000000) >>13);
-    output[ 2] |= ((input[ 5] & 0x02000000) >>12);
-    output[ 2] |= ((input[ 6] & 0x02000000) >>11);
-    output[ 2] |= ((input[ 7] & 0x02000000) >>10);
-    output[ 4] = ((input[ 0] & 0x00000400) >>10);
-    output[ 4] |= ((input[ 1] & 0x00000400) >> 9);
-    output[ 4] |= ((input[ 2] & 0x00000400) >> 8);
-    output[ 4] |= ((input[ 3] & 0x00000400) >> 7);
-    output[ 4] |= ((input[ 4] & 0x00000400) >> 6);
-    output[ 4] |= ((input[ 5] & 0x00000400) >> 5);
-    output[ 4] |= ((input[ 6] & 0x00000400) >> 4);
-    output[ 4] |= ((input[ 7] & 0x00000400) >> 3);
-    output[ 4] |= ((input[ 0] & 0x04000000) >>18);
-    output[ 4] |= ((input[ 1] & 0x04000000) >>17);
-    output[ 4] |= ((input[ 2] & 0x04000000) >>16);
-    output[ 4] |= ((input[ 3] & 0x04000000) >>15);
-    output[ 4] |= ((input[ 4] & 0x04000000) >>14);
-    output[ 4] |= ((input[ 5] & 0x04000000) >>13);
-    output[ 4] |= ((input[ 6] & 0x04000000) >>12);
-    output[ 4] |= ((input[ 7] & 0x04000000) >>11);
-    output[ 6] = ((input[ 0] & 0x00000800) >>11);
-    output[ 6] |= ((input[ 1] & 0x00000800) >>10);
-    output[ 6] |= ((input[ 2] & 0x00000800) >> 9);
-    output[ 6] |= ((input[ 3] & 0x00000800) >> 8);
-    output[ 6] |= ((input[ 4] & 0x00000800) >> 7);
-    output[ 6] |= ((input[ 5] & 0x00000800) >> 6);
-    output[ 6] |= ((input[ 6] & 0x00000800) >> 5);
-    output[ 6] |= ((input[ 7] & 0x00000800) >> 4);
-    output[ 6] |= ((input[ 0] & 0x08000000) >>19);
-    output[ 6] |= ((input[ 1] & 0x08000000) >>18);
-    output[ 6] |= ((input[ 2] & 0x08000000) >>17);
-    output[ 6] |= ((input[ 3] & 0x08000000) >>16);
-    output[ 6] |= ((input[ 4] & 0x08000000) >>15);
-    output[ 6] |= ((input[ 5] & 0x08000000) >>14);
-    output[ 6] |= ((input[ 6] & 0x08000000) >>13);
-    output[ 6] |= ((input[ 7] & 0x08000000) >>12);
-    output[ 8] = ((input[ 0] & 0x00001000) >>12);
-    output[ 8] |= ((input[ 1] & 0x00001000) >>11);
-    output[ 8] |= ((input[ 2] & 0x00001000) >>10);
-    output[ 8] |= ((input[ 3] & 0x00001000) >> 9);
-    output[ 8] |= ((input[ 4] & 0x00001000) >> 8);
-    output[ 8] |= ((input[ 5] & 0x00001000) >> 7);
-    output[ 8] |= ((input[ 6] & 0x00001000) >> 6);
-    output[ 8] |= ((input[ 7] & 0x00001000) >> 5);
-    output[ 8] |= ((input[ 0] & 0x10000000) >>20);
-    output[ 8] |= ((input[ 1] & 0x10000000) >>19);
-    output[ 8] |= ((input[ 2] & 0x10000000) >>18);
-    output[ 8] |= ((input[ 3] & 0x10000000) >>17);
-    output[ 8] |= ((input[ 4] & 0x10000000) >>16);
-    output[ 8] |= ((input[ 5] & 0x10000000) >>15);
-    output[ 8] |= ((input[ 6] & 0x10000000) >>14);
-    output[ 8] |= ((input[ 7] & 0x10000000) >>13);
-    output[10] = ((input[ 0] & 0x00002000) >>13);
-    output[10] |= ((input[ 1] & 0x00002000) >>12);
-    output[10] |= ((input[ 2] & 0x00002000) >>11);
-    output[10] |= ((input[ 3] & 0x00002000) >>10);
-    output[10] |= ((input[ 4] & 0x00002000) >> 9);
-    output[10] |= ((input[ 5] & 0x00002000) >> 8);
-    output[10] |= ((input[ 6] & 0x00002000) >> 7);
-    output[10] |= ((input[ 7] & 0x00002000) >> 6);
-    output[10] |= ((input[ 0] & 0x20000000) >>21);
-    output[10] |= ((input[ 1] & 0x20000000) >>20);
-    output[10] |= ((input[ 2] & 0x20000000) >>19);
-    output[10] |= ((input[ 3] & 0x20000000) >>18);
-    output[10] |= ((input[ 4] & 0x20000000) >>17);
-    output[10] |= ((input[ 5] & 0x20000000) >>16);
-    output[10] |= ((input[ 6] & 0x20000000) >>15);
-    output[10] |= ((input[ 7] & 0x20000000) >>14);
-    output[12] = ((input[ 0] & 0x00004000) >>14);
-    output[12] |= ((input[ 1] & 0x00004000) >>13);
-    output[12] |= ((input[ 2] & 0x00004000) >>12);
-    output[12] |= ((input[ 3] & 0x00004000) >>11);
-    output[12] |= ((input[ 4] & 0x00004000) >>10);
-    output[12] |= ((input[ 5] & 0x00004000) >> 9);
-    output[12] |= ((input[ 6] & 0x00004000) >> 8);
-    output[12] |= ((input[ 7] & 0x00004000) >> 7);
-    output[12] |= ((input[ 0] & 0x40000000) >>22);
-    output[12] |= ((input[ 1] & 0x40000000) >>21);
-    output[12] |= ((input[ 2] & 0x40000000) >>20);
-    output[12] |= ((input[ 3] & 0x40000000) >>19);
-    output[12] |= ((input[ 4] & 0x40000000) >>18);
-    output[12] |= ((input[ 5] & 0x40000000) >>17);
-    output[12] |= ((input[ 6] & 0x40000000) >>16);
-    output[12] |= ((input[ 7] & 0x40000000) >>15);
-    output[14] = ((input[ 0] & 0x00008000) >>15);
-    output[14] |= ((input[ 1] & 0x00008000) >>14);
-    output[14] |= ((input[ 2] & 0x00008000) >>13);
-    output[14] |= ((input[ 3] & 0x00008000) >>12);
-    output[14] |= ((input[ 4] & 0x00008000) >>11);
-    output[14] |= ((input[ 5] & 0x00008000) >>10);
-    output[14] |= ((input[ 6] & 0x00008000) >> 9);
-    output[14] |= ((input[ 7] & 0x00008000) >> 8);
-    output[14] |= ((input[ 0] & 0x80000000) >>23);
-    output[14] |= ((input[ 1] & 0x80000000) >>22);
-    output[14] |= ((input[ 2] & 0x80000000) >>21);
-    output[14] |= ((input[ 3] & 0x80000000) >>20);
-    output[14] |= ((input[ 4] & 0x80000000) >>19);
-    output[14] |= ((input[ 5] & 0x80000000) >>18);
-    output[14] |= ((input[ 6] & 0x80000000) >>17);
-    output[14] |= ((input[ 7] & 0x80000000) >>16);
-
-#pragma unroll 8
-    for (int i = 0; i < 16; i+=2) {
-        if (threadIdx.x & 1) output[i] = __byte_perm(output[i], 0, 0x1032);
-        output[i] = __byte_perm(output[i], __shfl((int)output[i], (threadIdx.x+1)&3, 4), 0x7610);
-        output[i+1] = __shfl((int)output[i], (threadIdx.x+2)&3, 4);
-        if (threadIdx.x & 3) output[i] = output[i+1] = 0;
-    }
+	uint32_t d[8];
+	uint32_t t;
+
+	d[0] = __byte_perm(input[0], input[4], 0x7531);
+	d[1] = __byte_perm(input[1], input[5], 0x7531);
+	d[2] = __byte_perm(input[2], input[6], 0x7531);
+	d[3] = __byte_perm(input[3], input[7], 0x7531);
+
+	SWAP1(d[0], d[1]);
+	SWAP1(d[2], d[3]);
+
+	SWAP2(d[0], d[2]);
+	SWAP2(d[1], d[3]);
+
+	t = __byte_perm(d[0], d[2], 0x5410);
+	d[2] = __byte_perm(d[0], d[2], 0x7632);
+	d[0] = t;
+
+	t = __byte_perm(d[1], d[3], 0x5410);
+	d[3] = __byte_perm(d[1], d[3], 0x7632);
+	d[1] = t;
+
+	SWAP4(d[0], d[2]);
+	SWAP4(d[1], d[3]);
+
+	output[0] = d[0];
+	output[2] = d[1];
+	output[4] = d[0] >> 16;
+	output[6] = d[1] >> 16;
+	output[8] = d[2];
+	output[10] = d[3];
+	output[12] = d[2] >> 16;
+	output[14] = d[3] >> 16;
+
+	#pragma unroll 8
+	for (int i = 0; i < 16; i+=2) {
+		if (threadIdx.x & 1) output[i] = __byte_perm(output[i], 0, 0x1032);
+		output[i] = __byte_perm(output[i], __shfl((int)output[i], (threadIdx.x+1)&3, 4), 0x7610);
+		output[i+1] = __shfl((int)output[i], (threadIdx.x+2)&3, 4);
+		if (threadIdx.x & 3) output[i] = output[i+1] = 0;
+	}
 }
-
-#else
-
-/* host "fake" functions */
-#define from_bitslice_quad(st, out)
-#define to_bitslice_quad(in, msg) in[0] = (uint32_t) in[0];
-
-#endif /* device only code */
--- a/configure.ac
+++ b/configure.ac
@ -1,4 +1,4 @@
				@@ -1,4 +1,4 @@
-AC_INIT([ccminer], [1.5.2-git])
+AC_INIT([ccminer], [1.5.2])

 AC_PREREQ([2.59c])
 AC_CANONICAL_SYSTEM
--- a/cpuminer-config.h
+++ b/cpuminer-config.h
@ -156,7 +156,7 @@
				@@ -156,7 +156,7 @@
 #define PACKAGE_NAME "ccminer"

 /* Define to the full name and version of this package. */
-#define PACKAGE_STRING "ccminer 1.5.2-git"
+#define PACKAGE_STRING "ccminer 1.5.2"

 /* Define to the one symbol short name of this package. */
 #define PACKAGE_TARNAME "ccminer"
@ -165,7 +165,7 @@
				@@ -165,7 +165,7 @@
 #define PACKAGE_URL ""

 /* Define to the version of this package. */
-#define PACKAGE_VERSION "1.5.2-git"
+#define PACKAGE_VERSION "1.5.2"

 /* If using the C implementation of alloca, define if you know the
   direction of stack growth for your system; otherwise it will be
@ -188,7 +188,7 @@
				@@ -188,7 +188,7 @@
 #define USE_XOP 1

 /* Version number of package */
-#define VERSION "1.5.2-git"
+#define VERSION "1.5.2"

 /* Define curl_free() as free() if our version of curl lacks curl_free. */
 /* #undef curl_free */
--- a/cuda_groestlcoin.cu
+++ b/cuda_groestlcoin.cu
@ -12,15 +12,20 @@ extern uint32_t *d_resultNonce[8];
				@@ -12,15 +12,20 @@ extern uint32_t *d_resultNonce[8];

 __constant__ uint32_t groestlcoin_gpu_msg[32];

-// 64 Register Variante für Compute 3.0
+#if __CUDA_ARCH__ >= 300
+
+// 64 Registers Variant for Compute 3.0
 #include "groestl_functions_quad.cu"
+
 #include "bitslice_transformations_quad.cu"
+#endif

 #define SWAB32(x) cuda_swab32(x)

 __global__ __launch_bounds__(256, 4)
 void groestlcoin_gpu_hash_quad(int threads, uint32_t startNounce, uint32_t *resNounce)
 {
+#if __CUDA_ARCH__ >= 300
    // durch 4 dividieren, weil jeweils 4 Threads zusammen ein Hash berechnen
    int thread = (blockDim.x * blockIdx.x + threadIdx.x) / 4;
    if (thread < threads)
@ -86,6 +91,7 @@ void groestlcoin_gpu_hash_quad(int threads, uint32_t startNounce, uint32_t *resN
				@@ -86,6 +91,7 @@ void groestlcoin_gpu_hash_quad(int threads, uint32_t startNounce, uint32_t *resN
                    resNounce[0] = nounce;
        }
    }
+#endif
 }

 // Setup-Funktionen
@ -139,6 +145,11 @@ __host__ void groestlcoin_cpu_hash(int thr_id, int threads, uint32_t startNounce
				@@ -139,6 +145,11 @@ __host__ void groestlcoin_cpu_hash(int thr_id, int threads, uint32_t startNounce
    // Größe des dynamischen Shared Memory Bereichs
    size_t shared_size = 0;

+    if (device_sm[device_map[thr_id]] < 300) {
+        printf("Sorry, This algo is not supported by this GPU arch (SM 3.0 required)");
+        return;
+    }
+
    cudaMemset(d_resultNonce[thr_id], 0xFF, sizeof(uint32_t));
    groestlcoin_gpu_hash_quad<<<grid, block, shared_size>>>(threads, startNounce, d_resultNonce[thr_id]);

--- a/cuda_myriadgroestl.cu
+++ b/cuda_myriadgroestl.cu
@ -5,6 +5,14 @@
				@@ -5,6 +5,14 @@

 #include "cuda_helper.h"

+#if __CUDA_ARCH__ >= 300
+
+// 64 Registers Variant for Compute 3.0
+#include "groestl_functions_quad.cu"
+
+#include "bitslice_transformations_quad.cu"
+#endif
+
 // globaler Speicher für alle HeftyHashes aller Threads
 __constant__ uint32_t pTarget[8]; // Single GPU
 uint32_t *d_outputHashes[8];
@ -40,13 +48,9 @@ uint32_t myr_sha256_cpu_w2Table[] = {
				@@ -40,13 +48,9 @@ uint32_t myr_sha256_cpu_w2Table[] = {
    0x69bc7ac4, 0xbd11375b, 0xe3ba71e5, 0x3b209ff2, 0x18feee17, 0xe25ad9e7, 0x13375046, 0x0515089d,
    0x4f0d0f04, 0x2627484e, 0x310128d2, 0xc668b434, 0x420841cc, 0x62d311b8, 0xe59ba771, 0x85a7a484 };

-// 64 Register Variante für Compute 3.0
-#include "groestl_functions_quad.cu"
-#include "bitslice_transformations_quad.cu"
-
 #define SWAB32(x)        ( ((x & 0x000000FF) << 24) | ((x & 0x0000FF00) << 8) | ((x & 0x00FF0000) >> 8) | ((x & 0xFF000000) >> 24) )

-#if __CUDA_ARCH__ < 350 
+#if __CUDA_ARCH__ < 320
    // Kepler (Compute 3.0)
    #define ROTR32(x, n) (((x) >> (n)) | ((x) << (32 - (n))))
 #else
@ -220,6 +224,7 @@ __device__ void myriadgroestl_gpu_sha256(uint32_t *message)
				@@ -220,6 +224,7 @@ __device__ void myriadgroestl_gpu_sha256(uint32_t *message)
 __global__ void __launch_bounds__(256, 4)
 myriadgroestl_gpu_hash_quad(int threads, uint32_t startNounce, uint32_t *hashBuffer)
 {
+#if __CUDA_ARCH__ >= 300
    // durch 4 dividieren, weil jeweils 4 Threads zusammen ein Hash berechnen
    int thread = (blockDim.x * blockIdx.x + threadIdx.x) / 4;
    if (thread < threads)
@ -250,11 +255,13 @@ __global__ void __launch_bounds__(256, 4)
				@@ -250,11 +255,13 @@ __global__ void __launch_bounds__(256, 4)
            for(int k=0;k<16;k++) outpHash[k] = out_state[k];
        }
    }
+#endif
 }

 __global__ void
 myriadgroestl_gpu_hash_quad2(int threads, uint32_t startNounce, uint32_t *resNounce, uint32_t *hashBuffer)
 {
+#if __CUDA_ARCH__ >= 300
    int thread = (blockDim.x * blockIdx.x + threadIdx.x);
    if (thread < threads)
    {
@ -291,6 +298,7 @@ __global__ void
				@@ -291,6 +298,7 @@ __global__ void
            if(resNounce[0] > nounce)
                resNounce[0] = nounce;
    }
+#endif
 }

 // Setup-Funktionen
@ -365,6 +373,11 @@ __host__ void myriadgroestl_cpu_hash(int thr_id, int threads, uint32_t startNoun
				@@ -365,6 +373,11 @@ __host__ void myriadgroestl_cpu_hash(int thr_id, int threads, uint32_t startNoun
    dim3 grid(factor*((threads + threadsperblock-1)/threadsperblock));
    dim3 block(threadsperblock);

+    if (device_sm[device_map[thr_id]] < 300) {
+        printf("Sorry, This algo is not supported by this GPU arch (SM 3.0 required)");
+        return;
+    }
+
    myriadgroestl_gpu_hash_quad<<<grid, block, shared_size>>>(threads, startNounce, d_outputHashes[thr_id]);
    dim3 grid2((threads + threadsperblock-1)/threadsperblock);
    myriadgroestl_gpu_hash_quad2<<<grid2, block, shared_size>>>(threads, startNounce, d_resultNonce[thr_id], d_outputHashes[thr_id]);
--- a/quark/cuda_quark_groestl512.cu
+++ b/quark/cuda_quark_groestl512.cu
@ -8,13 +8,17 @@
				@@ -8,13 +8,17 @@
 #define TPB 256
 #define THF 4

-// 64 Register Variante für Compute 3.0
+#if __CUDA_ARCH__ >= 300
 #include "groestl_functions_quad.cu"
 #include "bitslice_transformations_quad.cu"
+#endif
+
+#include "quark/cuda_quark_groestl512_sm20.cu"

 __global__ __launch_bounds__(TPB, THF)
 void quark_groestl512_gpu_hash_64_quad(int threads, uint32_t startNounce, uint32_t * __restrict g_hash, uint32_t * __restrict g_nonceVector)
 {
+#if __CUDA_ARCH__ >= 300
    // durch 4 dividieren, weil jeweils 4 Threads zusammen ein Hash berechnen
    int thread = (blockDim.x * blockIdx.x + threadIdx.x) >> 2;
    if (thread < threads)
@ -54,11 +58,13 @@ void quark_groestl512_gpu_hash_64_quad(int threads, uint32_t startNounce, uint32
				@@ -54,11 +58,13 @@ void quark_groestl512_gpu_hash_64_quad(int threads, uint32_t startNounce, uint32
            for(int k=0;k<16;k++) outpHash[k] = hash[k];
        }
    }
+#endif
 }

 __global__ void __launch_bounds__(TPB, THF)
 quark_doublegroestl512_gpu_hash_64_quad(int threads, uint32_t startNounce, uint32_t *g_hash, uint32_t *g_nonceVector)
 {
+#if __CUDA_ARCH__ >= 300
    int thread = (blockDim.x * blockIdx.x + threadIdx.x)>>2;
    if (thread < threads)
    {
@ -113,11 +119,15 @@ __global__ void __launch_bounds__(TPB, THF)
				@@ -113,11 +119,15 @@ __global__ void __launch_bounds__(TPB, THF)
            for(int k=0;k<16;k++) outpHash[k] = hash[k];
        }
    }
+#endif
 }

-// Setup-Funktionen
+
+
 __host__ void quark_groestl512_cpu_init(int thr_id, int threads)
 {
+    if (device_sm[device_map[thr_id]] < 300)
+        quark_groestl512_sm20_init(thr_id, threads);
 }

 __host__ void quark_groestl512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order)
@ -135,7 +145,10 @@ __host__ void quark_groestl512_cpu_hash_64(int thr_id, int threads, uint32_t sta
				@@ -135,7 +145,10 @@ __host__ void quark_groestl512_cpu_hash_64(int thr_id, int threads, uint32_t sta
    // Größe des dynamischen Shared Memory Bereichs
    size_t shared_size = 0;

-    quark_groestl512_gpu_hash_64_quad<<<grid, block, shared_size>>>(threads, startNounce, d_hash, d_nonceVector);
+    if (device_sm[device_map[thr_id]] >= 300)
+        quark_groestl512_gpu_hash_64_quad<<<grid, block, shared_size>>>(threads, startNounce, d_hash, d_nonceVector);
+    else
+        quark_groestl512_sm20_hash_64(thr_id, threads, startNounce, d_nonceVector, d_hash, order);

    // Strategisches Sleep Kommando zur Senkung der CPU Last
    MyStreamSynchronize(NULL, order, thr_id);
@ -143,21 +156,18 @@ __host__ void quark_groestl512_cpu_hash_64(int thr_id, int threads, uint32_t sta
				@@ -143,21 +156,18 @@ __host__ void quark_groestl512_cpu_hash_64(int thr_id, int threads, uint32_t sta

 __host__ void quark_doublegroestl512_cpu_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order)
 {
-    int threadsperblock = TPB;
-
-    // Compute 3.0 benutzt die registeroptimierte Quad Variante mit Warp Shuffle
-    // mit den Quad Funktionen brauchen wir jetzt 4 threads pro Hash, daher Faktor 4 bei der Blockzahl
    const int factor = THF;
+    int threadsperblock = TPB;

-    // berechne wie viele Thread Blocks wir brauchen
    dim3 grid(factor*((threads + threadsperblock-1)/threadsperblock));
    dim3 block(threadsperblock);

-    // Größe des dynamischen Shared Memory Bereichs
    size_t shared_size = 0;

-    quark_doublegroestl512_gpu_hash_64_quad<<<grid, block, shared_size>>>(threads, startNounce, d_hash, d_nonceVector);
+    if (device_sm[device_map[thr_id]] >= 300)
+        quark_doublegroestl512_gpu_hash_64_quad<<<grid, block, shared_size>>>(threads, startNounce, d_hash, d_nonceVector);
+    else
+        quark_doublegroestl512_sm20_hash_64(thr_id, threads, startNounce, d_nonceVector, d_hash, order);

-    // Strategisches Sleep Kommando zur Senkung der CPU Last
    MyStreamSynchronize(NULL, order, thr_id);
 }
--- a/quark/cuda_quark_groestl512_sm20.cu
+++ b/quark/cuda_quark_groestl512_sm20.cu
@ -0,0 +1,330 @@
				@@ -0,0 +1,330 @@
+// SM 2.1 variant
+
+// #include "cuda_helper.h"
+
+#define MAXWELL_OR_FERMI 0
+#define USE_SHARED 1
+
+#define SPH_C32(x)    ((uint32_t)(x ## U))
+#define SPH_T32(x)    ((x) & SPH_C32(0xFFFFFFFF))
+
+#define PC32up(j, r)   ((uint32_t)((j) + (r)))
+#define PC32dn(j, r)   0
+#define QC32up(j, r)   0xFFFFFFFF
+#define QC32dn(j, r)   (((uint32_t)(r) << 24) ^ SPH_T32(~((uint32_t)(j) << 24)))
+
+#define B32_0(x)    __byte_perm(x, 0, 0x4440)
+//((x) & 0xFF)
+#define B32_1(x)    __byte_perm(x, 0, 0x4441)
+//(((x) >> 8) & 0xFF)
+#define B32_2(x)    __byte_perm(x, 0, 0x4442)
+//(((x) >> 16) & 0xFF)
+#define B32_3(x)    __byte_perm(x, 0, 0x4443)
+//((x) >> 24)
+
+// a healthy mix between shared and textured access provides the highest speed on Compute 3.0 and 3.5!
+#define T0up(x) (*((uint32_t*)mixtabs + (    (x))))
+#define T0dn(x) tex1Dfetch(t0dn1, x)
+#define T1up(x) tex1Dfetch(t1up1, x)
+#define T1dn(x) (*((uint32_t*)mixtabs + (768+(x))))
+#define T2up(x) tex1Dfetch(t2up1, x)
+#define T2dn(x) (*((uint32_t*)mixtabs + (1280+(x))))
+#define T3up(x) (*((uint32_t*)mixtabs + (1536+(x))))
+#define T3dn(x) tex1Dfetch(t3dn1, x)
+
+texture<unsigned int, 1, cudaReadModeElementType> t0up1;
+texture<unsigned int, 1, cudaReadModeElementType> t0dn1;
+texture<unsigned int, 1, cudaReadModeElementType> t1up1;
+texture<unsigned int, 1, cudaReadModeElementType> t1dn1;
+texture<unsigned int, 1, cudaReadModeElementType> t2up1;
+texture<unsigned int, 1, cudaReadModeElementType> t2dn1;
+texture<unsigned int, 1, cudaReadModeElementType> t3up1;
+texture<unsigned int, 1, cudaReadModeElementType> t3dn1;
+
+extern uint32_t T0up_cpu[];
+extern uint32_t T0dn_cpu[];
+extern uint32_t T1up_cpu[];
+extern uint32_t T1dn_cpu[];
+extern uint32_t T2up_cpu[];
+extern uint32_t T2dn_cpu[];
+extern uint32_t T3up_cpu[];
+extern uint32_t T3dn_cpu[];
+
+#if __CUDA_ARCH__ < 300
+
+__device__ __forceinline__
+void quark_groestl512_perm_P(uint32_t *a, char *mixtabs)
+{
+	uint32_t t[32];
+
+	for(int r=0; r<14; r++)
+	{
+		switch(r) {
+		case 0:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 0); break;
+		case 1:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 1); break;
+		case 2:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 2); break;
+		case 3:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 3); break;
+		case 4:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 4); break;
+		case 5:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 5); break;
+		case 6:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 6); break;
+		case 7:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 7); break;
+		case 8:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 8); break;
+		case 9:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 9); break;
+		case 10:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 10); break;
+		case 11:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 11); break;
+		case 12:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 12); break;
+		case 13:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) a[(k*2)+0] ^= PC32up(k<< 4, 13); break;
+		}
+
+		// RBTT
+		#pragma unroll 16
+		for(int k=0;k<32;k+=2) {
+			uint32_t t0_0 = B32_0(a[(k     ) & 0x1f]), t9_0  = B32_0(a[(k +  9) & 0x1f]);
+			uint32_t t2_1 = B32_1(a[(k +  2) & 0x1f]), t11_1 = B32_1(a[(k + 11) & 0x1f]);
+			uint32_t t4_2 = B32_2(a[(k +  4) & 0x1f]), t13_2 = B32_2(a[(k + 13) & 0x1f]);
+			uint32_t t6_3 = B32_3(a[(k +  6) & 0x1f]), t23_3 = B32_3(a[(k + 23) & 0x1f]);
+
+			t[k + 0] =  T0up( t0_0 ) ^ T1up(  t2_1 ) ^ T2up(  t4_2 ) ^ T3up(  t6_3 ) ^
+						T0dn( t9_0 ) ^ T1dn( t11_1 ) ^ T2dn( t13_2 ) ^ T3dn( t23_3 );
+
+			t[k + 1] =  T0dn( t0_0 ) ^ T1dn(  t2_1 ) ^ T2dn(  t4_2 ) ^ T3dn(  t6_3 ) ^
+						T0up( t9_0 ) ^ T1up( t11_1 ) ^ T2up( t13_2 ) ^ T3up( t23_3 );
+		}
+
+		#pragma unroll 32
+		for(int k=0; k<32; k++) {
+			a[k] = t[k];
+		}
+	}
+}
+
+__device__ __forceinline__
+void quark_groestl512_perm_Q(uint32_t *a, char *mixtabs)
+{
+	for(int r=0; r<14; r++)
+	{
+		uint32_t t[32];
+
+		switch(r) {
+		case 0:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 0); a[(k*2)+1] ^= QC32dn(k<< 4, 0);} break;
+		case 1:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 1); a[(k*2)+1] ^= QC32dn(k<< 4, 1);} break;
+		case 2:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 2); a[(k*2)+1] ^= QC32dn(k<< 4, 2);} break;
+		case 3:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 3); a[(k*2)+1] ^= QC32dn(k<< 4, 3);} break;
+		case 4:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 4); a[(k*2)+1] ^= QC32dn(k<< 4, 4);} break;
+		case 5:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 5); a[(k*2)+1] ^= QC32dn(k<< 4, 5);} break;
+		case 6:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 6); a[(k*2)+1] ^= QC32dn(k<< 4, 6);} break;
+		case 7:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 7); a[(k*2)+1] ^= QC32dn(k<< 4, 7);} break;
+		case 8:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 8); a[(k*2)+1] ^= QC32dn(k<< 4, 8);} break;
+		case 9:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 9); a[(k*2)+1] ^= QC32dn(k<< 4, 9);} break;
+		case 10:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 10); a[(k*2)+1] ^= QC32dn(k<< 4, 10);} break;
+		case 11:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 11); a[(k*2)+1] ^= QC32dn(k<< 4, 11);} break;
+		case 12:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 12); a[(k*2)+1] ^= QC32dn(k<< 4, 12);} break;
+		case 13:
+			#pragma unroll 16
+			for(int k=0;k<16;k++) { a[(k*2)+0] ^= QC32up(k<< 4, 13); a[(k*2)+1] ^= QC32dn(k<< 4, 13);} break;
+		}
+
+		// RBTT
+		#pragma unroll 16
+		for(int k=0;k<32;k+=2)
+		{
+			uint32_t t2_0  = B32_0(a[(k +  2) & 0x1f]), t1_0  = B32_0(a[(k +  1) & 0x1f]);
+			uint32_t t6_1  = B32_1(a[(k +  6) & 0x1f]), t5_1  = B32_1(a[(k +  5) & 0x1f]);
+			uint32_t t10_2 = B32_2(a[(k + 10) & 0x1f]), t9_2  = B32_2(a[(k +  9) & 0x1f]);
+			uint32_t t22_3 = B32_3(a[(k + 22) & 0x1f]), t13_3 = B32_3(a[(k + 13) & 0x1f]);
+
+			t[k + 0] =  T0up( t2_0 ) ^ T1up( t6_1 ) ^ T2up( t10_2 ) ^ T3up( t22_3 ) ^
+						T0dn( t1_0 ) ^ T1dn( t5_1 ) ^ T2dn(  t9_2 ) ^ T3dn( t13_3 );
+
+			t[k + 1] =  T0dn( t2_0 ) ^ T1dn( t6_1 ) ^ T2dn( t10_2 ) ^ T3dn( t22_3 ) ^
+						T0up( t1_0 ) ^ T1up( t5_1 ) ^ T2up(  t9_2 ) ^ T3up( t13_3 );
+		}
+		#pragma unroll 32
+		for(int k=0;k<32;k++)
+			a[k] = t[k];
+	}
+}
+
+#endif
+
+__global__
+void quark_groestl512_gpu_hash_64(int threads, uint32_t startNounce, uint32_t *g_hash, uint32_t *g_nonceVector)
+{
+#if __CUDA_ARCH__ < 300
+	extern __shared__ char mixtabs[];
+
+	if (threadIdx.x < 256)
+	{
+		*((uint32_t*)mixtabs + (    threadIdx.x)) = tex1Dfetch(t0up1, threadIdx.x);
+		*((uint32_t*)mixtabs + (256+threadIdx.x)) = tex1Dfetch(t0dn1, threadIdx.x);
+		*((uint32_t*)mixtabs + (512+threadIdx.x)) = tex1Dfetch(t1up1, threadIdx.x);
+		*((uint32_t*)mixtabs + (768+threadIdx.x)) = tex1Dfetch(t1dn1, threadIdx.x);
+		*((uint32_t*)mixtabs + (1024+threadIdx.x)) = tex1Dfetch(t2up1, threadIdx.x);
+		*((uint32_t*)mixtabs + (1280+threadIdx.x)) = tex1Dfetch(t2dn1, threadIdx.x);
+		*((uint32_t*)mixtabs + (1536+threadIdx.x)) = tex1Dfetch(t3up1, threadIdx.x);
+		*((uint32_t*)mixtabs + (1792+threadIdx.x)) = tex1Dfetch(t3dn1, threadIdx.x);
+	}
+
+	__syncthreads();
+
+	int thread = (blockDim.x * blockIdx.x + threadIdx.x);
+	if (thread < threads)
+	{
+		// GROESTL
+		uint32_t message[32];
+		uint32_t state[32];
+
+		uint32_t nounce = (g_nonceVector != NULL) ? g_nonceVector[thread] : (startNounce + thread);
+
+		int hashPosition = nounce - startNounce;
+		uint32_t *inpHash = &g_hash[16 * hashPosition];
+
+		#pragma unroll 16
+		for(int k=0; k<16; k++)
+			message[k] = inpHash[k];
+
+		#pragma unroll 14
+		for(int k=1; k<15; k++)
+			message[k+16] = 0;
+
+		message[16] = 0x80;
+		message[31] = 0x01000000;
+
+		#pragma unroll 32
+		for(int u=0; u<32; u++)
+			state[u] = message[u];
+		state[31] ^= 0x20000;
+
+		// Perm
+		quark_groestl512_perm_P(state, mixtabs);
+		state[31] ^= 0x20000;
+		quark_groestl512_perm_Q(message, mixtabs);
+
+		#pragma unroll 32
+		for(int u=0;u<32;u++) state[u] ^= message[u];
+
+		#pragma unroll 32
+		for(int u=0;u<32;u++) message[u] = state[u];
+
+		quark_groestl512_perm_P(message, mixtabs);
+
+		#pragma unroll 32
+		for(int u=0;u<32;u++) state[u] ^= message[u];
+
+		// Erzeugten Hash rausschreiben
+		uint32_t *outpHash = &g_hash[16 * hashPosition];
+
+		#pragma unroll 16
+		for(int k=0;k<16;k++) outpHash[k] = state[k+16];
+	}
+#endif
+}
+
+#define texDef(texname, texmem, texsource, texsize) \
+	unsigned int *texmem; \
+	cudaMalloc(&texmem, texsize); \
+	cudaMemcpy(texmem, texsource, texsize, cudaMemcpyHostToDevice); \
+	texname.normalized = 0; \
+	texname.filterMode = cudaFilterModePoint; \
+	texname.addressMode[0] = cudaAddressModeClamp; \
+	{ cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<unsigned int>(); \
+	  cudaBindTexture(NULL, &texname, texmem, &channelDesc, texsize ); } \
+
+__host__
+void quark_groestl512_sm20_init(int thr_id, int threads)
+{
+	// Texturen mit obigem Makro initialisieren
+	texDef(t0up1, d_T0up, T0up_cpu, sizeof(uint32_t)*256);
+	texDef(t0dn1, d_T0dn, T0dn_cpu, sizeof(uint32_t)*256);
+	texDef(t1up1, d_T1up, T1up_cpu, sizeof(uint32_t)*256);
+	texDef(t1dn1, d_T1dn, T1dn_cpu, sizeof(uint32_t)*256);
+	texDef(t2up1, d_T2up, T2up_cpu, sizeof(uint32_t)*256);
+	texDef(t2dn1, d_T2dn, T2dn_cpu, sizeof(uint32_t)*256);
+	texDef(t3up1, d_T3up, T3up_cpu, sizeof(uint32_t)*256);
+	texDef(t3dn1, d_T3dn, T3dn_cpu, sizeof(uint32_t)*256);
+}
+
+__host__
+void quark_groestl512_sm20_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order)
+{
+	int threadsperblock = 512;
+
+	dim3 grid((threads + threadsperblock-1)/threadsperblock);
+	dim3 block(threadsperblock);
+
+	size_t shared_size = 8 * 256 * sizeof(uint32_t);
+
+	quark_groestl512_gpu_hash_64<<<grid, block, shared_size>>>(threads, startNounce, d_hash, d_nonceVector);
+
+	// MyStreamSynchronize(NULL, order, thr_id);
+}
+
+__host__
+void quark_doublegroestl512_sm20_hash_64(int thr_id, int threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order)
+{
+	int threadsperblock = 512;
+
+	dim3 grid((threads + threadsperblock-1)/threadsperblock);
+	dim3 block(threadsperblock);
+
+	size_t shared_size = 8 * 256 * sizeof(uint32_t);
+
+	quark_groestl512_gpu_hash_64<<<grid, block, shared_size>>>(threads, startNounce, d_hash, d_nonceVector);
+	quark_groestl512_gpu_hash_64<<<grid, block, shared_size>>>(threads, startNounce, d_hash, d_nonceVector);
+
+	// MyStreamSynchronize(NULL, order, thr_id);
+}
--- a/x11/cuda_x11_simd512.cu
+++ b/x11/cuda_x11_simd512.cu
@ -181,6 +181,13 @@ do { \
				@@ -181,6 +181,13 @@ do { \
 #undef BUTTERFLY
 }

+#if defined(__CUDA_ARCH__)
+#if __CUDA_ARCH__ < 300
+  #define __shfl(var, srcLane, width) (uint32_t)(var)
+  // #error __shfl() not supported by SM 2.x
+#endif
+#endif
+
 /**
 * FFT_16 using w=2 as 16th root of unity
 * Unrolled decimation in frequency (DIF) radix-2 NTT.