#ifndef GLOBALS #define GLOBALS 1 #if !defined(VECTORS1) #define VECTORS1 #endif #ifdef VECTORS4 typedef uint4 vo; #elif defined(VECTORS2) typedef uint2 vo; #else typedef uint vo; #endif #ifndef NFACTOR #define NFACTOR 10 #endif #if NFACTOR == 1 # define __NF__ 2u #elif NFACTOR == 2 # define __NF__ 4u #elif NFACTOR == 3 # define __NF__ 8u #elif NFACTOR == 4 # define __NF__ 16u #elif NFACTOR == 5 # define __NF__ 32u #elif NFACTOR == 6 # define __NF__ 64u #elif NFACTOR == 7 # define __NF__ 128u #elif NFACTOR == 8 # define __NF__ 256u #elif NFACTOR == 9 # define __NF__ 512u #elif NFACTOR == 10 # define __NF__ 1024u #elif NFACTOR == 11 # define __NF__ 2048u #elif NFACTOR == 12 # define __NF__ 4096u #elif NFACTOR == 13 # define __NF__ 8192u #elif NFACTOR == 14 # define __NF__ 16384u #elif NFACTOR == 15 # define __NF__ 32768u #elif NFACTOR == 16 # define __NF__ 65536u #elif NFACTOR == 17 # define __NF__ 131072u #elif NFACTOR == 18 # define __NF__ 262144u #elif NFACTOR == 19 # define __NF__ 524288u #elif NFACTOR == 20 # define __NF__ 1048576u #else # define __NF__ 1024u #endif #define DEFNFACTOR(n) \ const uint n = __NF__; // const uint n = (__NF__/LOOKUP_GAP+(__NF__%LOOKUP_GAP>0)); #define E0 0x00FF00FFU #define E1 0xFF00FF00U #define DecAllKA uint K00 = 0x428a2f98U; \ uint K01 = 0x71374491U; \ uint K02 = 0xb5c0fbcfU; \ uint K03 = 0xe9b5dba5U; \ uint K04 = 0x3956c25bU; \ uint K05 = 0x59f111f1U; \ uint K06 = 0x923f82a4U; \ uint K07 = 0xab1c5ed5U; \ uint K08 = 0xd807aa98U; \ uint K09 = 0x12835b01U; \ uint K10 = 0x243185beU; \ uint K11 = 0x550c7dc3U; \ uint K12 = 0x72be5d74U; \ uint K13 = 0x80deb1feU; \ uint K14 = 0x9bdc06a7U; \ uint K76 = 0xc19bf174U; \ uint K15 = 0xe49b69c1U; #define DecAllKB uint K16 = 0xefbe4786U; \ uint K17 = 0x0fc19dc6U; \ uint K18 = 0x240ca1ccU; \ uint K19 = 0x2de92c6fU; \ uint K20 = 0x4a7484aaU; \ uint K21 = 0x5cb0a9dcU; \ uint K22 = 0x76f988daU; \ uint K23 = 0x983e5152U; \ uint K24 = 0xa831c66dU; \ uint K25 = 0xb00327c8U; \ uint K26 = 0xbf597fc7U; \ uint K27 = 0xc6e00bf3U; \ uint K28 = 0xd5a79147U; \ uint K29 = 0x06ca6351U; \ uint K30 = 0x14292967U; \ uint K31 = 0x27b70a85U; #define DecAllKC uint K32 = 0x2e1b2138U; \ uint K33 = 0x4d2c6dfcU; \ uint K34 = 0x53380d13U; \ uint K35 = 0x650a7354U; \ uint K36 = 0x766a0abbU; \ uint K37 = 0x81c2c92eU; \ uint K38 = 0x92722c85U; \ uint K39 = 0xa2bfe8a1U; \ uint K40 = 0xa81a664bU; \ uint K41 = 0xc24b8b70U; \ uint K42 = 0xc76c51a3U; \ uint K43 = 0xd192e819U; \ uint K44 = 0xd6990624U; \ uint K45 = 0xf40e3585U; \ uint K46 = 0x106aa070U; \ uint K47 = 0x19a4c116U; #define DecAllKD uint K48 = 0x1e376c08U; \ uint K49 = 0x2748774cU; \ uint K50 = 0x34b0bcb5U; \ uint K51 = 0x391c0cb3U; \ uint K52 = 0x4ed8aa4aU; \ uint K53 = 0x5b9cca4fU; \ uint K54 = 0x682e6ff3U; \ uint K55 = 0x748f82eeU; \ uint K56 = 0x78a5636fU; \ uint K57 = 0x84c87814U; \ uint K58 = 0x8cc70208U; \ uint K59 = 0x90befffaU; \ uint K60 = 0xa4506cebU; \ uint K61 = 0xbef9a3f7U; \ uint K62 = 0xc67178f2U; #define DecAllKE uint K63 = 0x98c7e2a2U; \ uint K64 = 0xfc08884dU; \ uint K65 = 0xcd2a11aeU; \ uint K66 = 0x510e527fU; \ uint K67 = 0x9b05688cU; \ uint K68 = 0xC3910C8EU; \ uint K69 = 0xfb6feee7U; \ uint K70 = 0x2a01a605U; \ uint K71 = 0x0c2e12e0U; \ uint K72 = 0x4498517BU; \ uint K73 = 0x6a09e667U; \ uint K74 = 0xa4ce148bU; \ uint K75 = 0x95F61999U; \ uint K77 = 0xBB67AE85U; \ uint K78 = 0x3C6EF372U; \ uint K79 = 0xA54FF53AU; \ uint K80 = 0x1F83D9ABU; \ uint K81 = 0x5BE0CD19U; //Search constants #define DecAllSK uint SK00 = 0x5C5C5C5CU; \ uint SK01 = 0x36363636U; \ uint SK02 = 0x80000000U; \ uint SK03 = 0x00000280U; \ uint SK04 = 0x000004a0U; \ uint SK05 = 0x00000300U; __constant uint K[64] = { 0x428a2f98U, 0x71374491U, 0xb5c0fbcfU, 0xe9b5dba5U, 0x3956c25bU, 0x59f111f1U, 0x923f82a4U, 0xab1c5ed5U, 0xd807aa98U, 0x12835b01U, 0x243185beU, // 10 0x550c7dc3U, 0x72be5d74U, 0x80deb1feU, 0x9bdc06a7U, 0xc19bf174U, 0xe49b69c1U, 0xefbe4786U, 0x0fc19dc6U, 0x240ca1ccU, 0x2de92c6fU, // 20 0x4a7484aaU, // 0x5cb0a9dcU, 0x76f988daU, 0x983e5152U, 0xa831c66dU, 0xb00327c8U, 0xbf597fc7U, 0xc6e00bf3U, 0xd5a79147U, 0x06ca6351U, // 30 0x14292967U, // 0x27b70a85U, 0x2e1b2138U, 0x4d2c6dfcU, 0x53380d13U, 0x650a7354U, 0x766a0abbU, 0x81c2c92eU, 0x92722c85U, 0xa2bfe8a1U, // 40 0xa81a664bU, // 0xc24b8b70U, 0xc76c51a3U, 0xd192e819U, 0xd6990624U, 0xf40e3585U, 0x106aa070U, 0x19a4c116U, 0x1e376c08U, 0x2748774cU, // 50 0x34b0bcb5U, // 0x391c0cb3U, 0x4ed8aa4aU, 0x5b9cca4fU, 0x682e6ff3U, 0x748f82eeU, 0x78a5636fU, 0x84c87814U, 0x8cc70208U, 0x90befffaU, // 60 0xa4506cebU, // 0xbef9a3f7U, 0xc67178f2U }; __constant uint freshK[18] = { 0x98c7e2a2U, //63 0xfc08884dU, 0xcd2a11aeU, 0x510e527fU, 0x9b05688cU, 0xC3910C8EU, 0xfb6feee7U, 0x2a01a605U, // 70 0x0c2e12e0U, 0x4498517BU, 0x6a09e667U, // 10 0xa4ce148bU, 0x95F61999U, // 0xc19bf174U, //76 removed 0xBB67AE85U, 0x3C6EF372U, 0xA54FF53AU, 0x1F83D9ABU, // 80 0x5BE0CD19U }; /* 0x5C5C5C5CU, 0x36363636U, 0x80000000U, // 0x000003FFU, //never used 0x00000280U, 0x000004a0U, 0x00000300U }; */ #define FOURdeclare(myvar, v1, v2, v3, v4) uint myvar ## 1 = v1; \ uint myvar ## 2 = v2; \ uint myvar ## 3 = v3; \ uint myvar ## 4 = v4; #define FOURassign(myvar, v1, v2, v3, v4) myvar ## 1 = v1; \ myvar ## 2 = v2; \ myvar ## 3 = v3; \ myvar ## 4 = v4; #define FOURcopy(var1, var2) var1 ## 1 = var2 ## 1; \ var1 ## 2 = var2 ## 2; \ var1 ## 3 = var2 ## 3; \ var1 ## 4 = var2 ## 4; #define FOURtovec(var1, var2) var1.x = var2 ## 1; \ var1.y = var2 ## 2; \ var1.z = var2 ## 3; \ var1.w = var2 ## 4; #define FOURfromvec(var1, var2) var1 ## 1 = var2.x; \ var1 ## 2 = var2.y; \ var1 ## 3 = var2.z; \ var1 ## 4 = var2.w; #define UNROLL_FACTOR 2 #define rotl(x,y) rotate(x,y) #define Ch(x,y,z) bitselect(z,y,x) #define Maj(x,y,z) Ch((x^z),y,z) #define EndianSwapa(n) (Ch(E0, rotl(n, 8U), rotl(n, 24U))) #define EndianSwapb(n) (rotl(n & E0, 24U)|rotl(n & E1, 8U)) #define Tr2(x) (rotl(x, 30U) ^ rotl(x, 19U) ^ rotl(x, 10U)) #define Tr1(x) (rotl(x, 26U) ^ rotl(x, 21U) ^ rotl(x, 7U)) #define Wr2(x) (rotl(x, 25U) ^ rotl(x, 14U) ^ (x>>3U)) #define Wr1(x) (rotl(x, 15U) ^ rotl(x, 13U) ^ (x>>10U)) #define RND(a, b, c, d, e, f, g, h, k) \ h += Tr1(e); \ h += Ch(e, f, g); \ h += k; \ d += h; \ h += Tr2(a); \ h += Maj(a, b, c); //#define zSIZE 8 //#define Coord(x,y,z) z+(x<<3)+((uint)(y*CONCURRENT_THREADS)<<3) //#define Coord(x,y,z) (x<<3)+((y<<3)*CONCURRENT_THREADS)+z //#define Coord(x,y,z) x*(z ## SIZE)+y*(CONCURRENT_THREADS)*(z ## SIZE)+z #define Coord(x,y,z) x*(z ## SIZE)+y*(x ## SIZE)*(z ## SIZE)+z #define CO Coord(x,y,z) //#define Coord(x,y,z) x+y*(x ## SIZE)+z*(y ## SIZE)*(x ## SIZE) //#define CO Coord(z,x,y) /* volatile uint fixedWa[8] = {0x428a2f99,0xd807aa98,0xf59b89c2,0xb707775c,0xad87a3ea,0xc91b1417,0xe64fb6a2,0xe0a1adbe}; volatile uint fixedWb[8] = {0xf1374491,0x12835b01,0x73924787,0x0468c23f,0xbcb1d3a3,0xc359dce1,0xe84d923a,0x7c728e11}; volatile uint fixedWc[8] = {0xb5c0fbcf,0x243185be,0x23c6886e,0xe7e72b4c,0x7b993186,0xa83253a7,0xe93a5730,0x511c78e4}; volatile uint fixedWd[8] = {0xe9b5dba5,0x550c7dc3,0xa42ca65c,0x49e1f1a2,0x562b9420,0x3b13c12d,0x09837686,0x315b45bd}; volatile uint fixedWe[8] = {0x3956c25b,0x72be5d74,0x15ed3627,0x4b99c816,0xbff3ca0c,0x9d3d725d,0x078ff753,0xfca71413}; volatile uint fixedWf[8] = {0x59f111f1,0x80deb1fe,0x4d6edcbf,0x926d1570,0xda4b0c23,0xd9031a84,0x29833341,0xea28f96a}; volatile uint fixedWg[8] = {0x923f82a4,0x9bdc06a7,0xe28217fc,0xaa0fc072,0x6cd8711a,0xb1a03340,0xd5de0b7e,0x79703128}; volatile uint fixedWh[8] = {0xab1c5ed5,0xc19bf794,0xef02488f,0xadb36e2c,0x8f337caa,0x16f58012,0x6948ccf4,0x4e1ef848}; */ /* __constant uint sK[6] = { 0x5C5C5C5CU, //82 0x36363636U, 0x80000000U, // 0x000003FFU, //never used 0x00000280U, 0x000004a0U, 0x00000300U }; */ #define FOUND (0xFF) #define SETFOUND(Xnonce) output[output[FOUND]++] = Xnonce #endif