GOSTSec
/
ccminer-gostd-lite
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
Browse Source

whirlpool: upgrade with alexis variant (2x) 

to check on older cards... maybe breaks some
master
Tanguy Pruvot 8 years ago
parent
de41bc9d21
commit
18637503f2
6 changed files with 959 additions and 2521 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 3
      README.txt
  2. 13
      cuda_vectors.h
  3. 1
      sha256/cuda_sha256t.cu
  4. 190
      x15/cuda_whirlpool_tables.cuh
  5. 3232
      x15/cuda_x15_whirlpool.cu
  6. 25
      x15/whirlpool.cu

3
README.txt
Unescape View File

@ -275,11 +275,12 @@ features.
>>> RELEASE HISTORY <<< >>> RELEASE HISTORY <<<
Mar. 08th 2017 Mar. 08th 2017 v2.0
Handle cryptonight, wildkeccak and cryptonight-lite Handle cryptonight, wildkeccak and cryptonight-lite
Add a serie of new algos: timetravel, bastion, hmq1725, sha256t Add a serie of new algos: timetravel, bastion, hmq1725, sha256t
Import lyra2z from djm34 work... Import lyra2z from djm34 work...
Rework the common skein512 (used in most algos except skein ;) Rework the common skein512 (used in most algos except skein ;)
Upgrade whirlpool algo with alexis version (2x faster)
Store the share diff of second nonce(s) in most algos Store the share diff of second nonce(s) in most algos
Hardware monitoring thread to get more accurate power readings Hardware monitoring thread to get more accurate power readings
Small changes for the quiet mode & max-log-rate to reduce logs Small changes for the quiet mode & max-log-rate to reduce logs

13
cuda_vectors.h
Unescape View File

@ -32,3 +32,16 @@ static __device__ __forceinline__ uint2 operator+ (const uint2 a, const uint32_t
#endif #endif
} }
/* whirlpool ones */
#ifdef __CUDA_ARCH__
__device__ __forceinline__
uint2 ROL16(const uint2 a) {
uint2 result;
result.x = __byte_perm(a.x, a.y, 0x1076);
result.y = __byte_perm(a.y, a.x, 0x1076);
return result;
}
#else
#define ROL16(a) make_uint2(a.x, a.y) /* bad, just to define it */
#endif

1
sha256/cuda_sha256t.cu
Unescape View File

@ -444,6 +444,7 @@ void sha256t_gpu_hash_shared(const uint32_t threads, const uint32_t startNonce,
__host__ __host__
void sha256t_init(int thr_id) void sha256t_init(int thr_id)
{ {
cuda_get_arch(thr_id);
cudaMemcpyToSymbol(c_K, cpu_K, sizeof(cpu_K), 0, cudaMemcpyHostToDevice); cudaMemcpyToSymbol(c_K, cpu_K, sizeof(cpu_K), 0, cudaMemcpyHostToDevice);
CUDA_SAFE_CALL(cudaMalloc(&d_resNonces[thr_id], 2*sizeof(uint32_t))); CUDA_SAFE_CALL(cudaMalloc(&d_resNonces[thr_id], 2*sizeof(uint32_t)));
} }

190
x15/cuda_whirlpool_tables.cuh
Unescape View File

@ -0,0 +1,190 @@
/* CUDA whirlpool huge constant tables */
static const uint64_t old1_T0[256] = {
0x78D8C07818281818, 0xAF2605AF23652323, 0xF9B87EF9C657C6C6, 0x6FFB136FE825E8E8,
0xA1CB4CA187948787, 0x6211A962B8D5B8B8, 0x0509080501030101, 0x6E0D426E4FD14F4F,
0xEE9BADEE365A3636, 0x04FF5904A6F7A6A6, 0xBD0CDEBDD26BD2D2, 0x060EFB06F502F5F5,
0x8096EF80798B7979, 0xCE305FCE6FB16F6F, 0xEF6DFCEF91AE9191, 0x07F8AA0752F65252,
0xFD4727FD60A06060, 0x76358976BCD9BCBC, 0xCD37ACCD9BB09B9B, 0x8C8A048C8E8F8E8E,
0x15D27115A3F8A3A3, 0x3C6C603C0C140C0C, 0x8A84FF8A7B8D7B7B, 0xE180B5E1355F3535,
0x69F5E8691D271D1D, 0x47B35347E03DE0E0, 0xAC21F6ACD764D7D7, 0xED9C5EEDC25BC2C2,
0x96436D962E722E2E, 0x7A29627A4BDD4B4B, 0x215DA321FE1FFEFE, 0x16D5821657F95757,
0x41BDA841153F1515, 0xB6E89FB677997777, 0xEB92A5EB37593737, 0x569E7B56E532E5E5,
0xD9138CD99FBC9F9F, 0x1723D317F00DF0F0, 0x7F206A7F4ADE4A4A, 0x95449E95DA73DADA,
0x25A2FA2558E85858, 0xCACF06CAC946C9C9, 0x8D7C558D297B2929, 0x225A50220A1E0A0A,
0x4F50E14FB1CEB1B1, 0x1AC9691AA0FDA0A0, 0xDA147FDA6BBD6B6B, 0xABD95CAB85928585,
0x733C8173BDDABDBD, 0x348FD2345DE75D5D, 0x5090805010301010, 0x0307F303F401F4F4,
0xC0DD16C0CB40CBCB, 0xC6D3EDC63E423E3E, 0x112D2811050F0505, 0xE6781FE667A96767,
0x53977353E431E4E4, 0xBB0225BB27692727, 0x5873325841C34141, 0x9DA72C9D8B808B8B,
0x01F65101A7F4A7A7, 0x94B2CF947D877D7D, 0xFB49DCFB95A29595, 0x9F568E9FD875D8D8,
0x30708B30FB10FBFB, 0x71CD2371EE2FEEEE, 0x91BBC7917C847C7C, 0xE37117E366AA6666,
0x8E7BA68EDD7ADDDD, 0x4BAFB84B17391717, 0x4645024647C94747, 0xDC1A84DC9EBF9E9E,
0xC5D41EC5CA43CACA, 0x995875992D772D2D, 0x792E9179BFDCBFBF, 0x1B3F381B07090707,
0x23AC0123ADEAADAD, 0x2FB0EA2F5AEE5A5A, 0xB5EF6CB583988383, 0xFFB685FF33553333,
0xF25C3FF263A56363, 0x0A12100A02060202, 0x38933938AAE3AAAA, 0xA8DEAFA871937171,
0xCFC60ECFC845C8C8, 0x7DD1C87D192B1919, 0x703B727049DB4949, 0x9A5F869AD976D9D9,
0x1D31C31DF20BF2F2, 0x48A84B48E338E3E3, 0x2AB9E22A5BED5B5B, 0x92BC349288858888,
0xC83EA4C89AB39A9A, 0xBE0B2DBE266A2626, 0xFABF8DFA32563232, 0x4A59E94AB0CDB0B0,
0x6AF21B6AE926E9E9, 0x337778330F110F0F, 0xA633E6A6D562D5D5, 0xBAF474BA809D8080,
0x7C27997CBEDFBEBE, 0xDEEB26DECD4ACDCD, 0xE489BDE4345C3434, 0x75327A7548D84848,
0x2454AB24FF1CFFFF, 0x8F8DF78F7A8E7A7A, 0xEA64F4EA90AD9090, 0x3E9DC23E5FE15F5F,
0xA03D1DA020602020, 0xD50F67D568B86868, 0x72CAD0721A2E1A1A, 0x2CB7192CAEEFAEAE,
0x5E7DC95EB4C1B4B4, 0x19CE9A1954FC5454, 0xE57FECE593A89393, 0xAA2F0DAA22662222,
0xE96307E964AC6464, 0x122ADB12F10EF1F1, 0xA2CCBFA273957373, 0x5A82905A12361212,
0x5D7A3A5D40C04040, 0x2848402808180808, 0xE89556E8C358C3C3, 0x7BDF337BEC29ECEC,
0x904D9690DB70DBDB, 0x1FC0611FA1FEA1A1, 0x83911C838D8A8D8D, 0xC9C8F5C93D473D3D,
0xF15BCCF197A49797, 0x0000000000000000, 0xD4F936D4CF4CCFCF, 0x876E45872B7D2B2B,
0xB3E197B3769A7676, 0xB0E664B0829B8282, 0xA928FEA9D667D6D6, 0x77C3D8771B2D1B1B,
0x5B74C15BB5C2B5B5, 0x29BE1129AFECAFAF, 0xDF1D77DF6ABE6A6A, 0x0DEABA0D50F05050,
0x4C57124C45CF4545, 0x1838CB18F308F3F3, 0xF0AD9DF030503030, 0x74C42B74EF2CEFEF,
0xC3DAE5C33F413F3F, 0x1CC7921C55FF5555, 0x10DB7910A2FBA2A2, 0x65E90365EA23EAEA,
0xEC6A0FEC65AF6565, 0x6803B968BAD3BABA, 0x934A65932F712F2F, 0xE78E4EE7C05DC0C0,
0x8160BE81DE7FDEDE, 0x6CFCE06C1C241C1C, 0x2E46BB2EFD1AFDFD, 0x641F52644DD74D4D,
0xE076E4E092AB9292, 0xBCFA8FBC759F7575, 0x1E36301E060A0606, 0x98AE24988A838A8A,
0x404BF940B2CBB2B2, 0x59856359E637E6E6, 0x367E70360E120E0E, 0x63E7F8631F211F1F,
0xF75537F762A66262, 0xA33AEEA3D461D4D4, 0x32812932A8E5A8A8, 0xF452C4F496A79696,
0x3A629B3AF916F9F9, 0xF6A366F6C552C5C5, 0xB11035B1256F2525, 0x20ABF22059EB5959,
0xAED054AE84918484, 0xA7C5B7A772967272, 0xDDECD5DD394B3939, 0x61165A614CD44C4C,
0x3B94CA3B5EE25E5E, 0x859FE78578887878, 0xD8E5DDD838483838, 0x869814868C898C8C,
0xB217C6B2D16ED1D1, 0x0BE4410BA5F2A5A5, 0x4DA1434DE23BE2E2, 0xF84E2FF861A36161,
0x4542F145B3C8B3B3, 0xA53415A521632121, 0xD60894D69CB99C9C, 0x66EEF0661E221E1E,
0x5261225243C54343, 0xFCB176FCC754C7C7, 0x2B4FB32BFC19FCFC, 0x14242014040C0404,
0x08E3B20851F35151, 0xC725BCC799B69999, 0xC4224FC46DB76D6D, 0x396568390D170D0D,
0x35798335FA13FAFA, 0x8469B684DF7CDFDF, 0x9BA9D79B7E827E7E, 0xB4193DB4246C2424,
0xD7FEC5D73B4D3B3B, 0x3D9A313DABE0ABAB, 0xD1F03ED1CE4FCECE, 0x5599885511331111,
0x89830C898F8C8F8F, 0x6B044A6B4ED24E4E, 0x5166D151B7C4B7B7, 0x60E00B60EB20EBEB,
0xCCC1FDCC3C443C3C, 0xBFFD7CBF819E8181, 0xFE40D4FE94A19494, 0x0C1CEB0CF704F7F7,
0x6718A167B9D6B9B9, 0x5F8B985F13351313, 0x9C517D9C2C742C2C, 0xB805D6B8D368D3D3,
0x5C8C6B5CE734E7E7, 0xCB3957CB6EB26E6E, 0xF3AA6EF3C451C4C4, 0x0F1B180F03050303,
0x13DC8A1356FA5656, 0x495E1A4944CC4444, 0x9EA0DF9E7F817F7F, 0x37882137A9E6A9A9,
0x82674D822A7E2A2A, 0x6D0AB16DBBD0BBBB, 0xE28746E2C15EC1C1, 0x02F1A20253F55353,
0x8B72AE8BDC79DCDC, 0x275358270B1D0B0B, 0xD3019CD39DBA9D9D, 0xC12B47C16CB46C6C,
0xF5A495F531533131, 0xB9F387B9749C7474, 0x0915E309F607F6F6, 0x434C0A4346CA4646,
0x26A50926ACE9ACAC, 0x97B53C9789868989, 0x44B4A044143C1414, 0x42BA5B42E13EE1E1,
0x4EA6B04E163A1616, 0xD2F7CDD23A4E3A3A, 0xD0066FD069BB6969, 0x2D41482D091B0909,
0xADD7A7AD70907070, 0x546FD954B6C7B6B6, 0xB71ECEB7D06DD0D0, 0x7ED63B7EED2AEDED,
0xDBE22EDBCC49CCCC, 0x57682A5742C64242, 0xC22CB4C298B59898, 0x0EED490EA4F1A4A4,
0x88755D8828782828, 0x3186DA315CE45C5C, 0x3F6B933FF815F8F8, 0xA4C244A486978686
};
static const uint64_t old1_RC[10] = {
0x4F01B887E8C62318, 0x52916F79F5D2A636, 0x357B0CA38E9BBC60, 0x57FE4B2EC2D7E01D, 0xDA4AF09FE5377715,
0x856BA0B10A29C958, 0x67053ECBF4105DBD, 0xD8957DA78B4127E4, 0x9E4717DD667CEEFB, 0x33835AAD07BF2DCA
};
static const uint64_t old1_precomputed_round_key_64[72] = {
0xECE7FCC7F808AB3B, 0x44E9CB45024545CB, 0xB289A43CA4A489FE, 0xC5E1F3E1E1C5A9A0,
0xFCD4FCFCAC5C48AC, 0x418F8F0E90F70E8F, 0x7979078514077946, 0xF8B87868B8F8D8F8,
0xE4B6012A17B9C1EF, 0xAFF6AD2866D6C68D, 0xB0C745991504968F, 0x74E93F120FE2E675,
0xC1D216198D2A495B, 0x35B6A53DF6716FD7, 0xD6983228CCDC354F, 0xC3B8BFCB242AB159,
0x4AC0340260B548D4, 0x3B3582BB4F9BB769, 0xD8683C4AF17C46F8, 0x9EE05220D8214E61,
0xFED2431F5098E0E5, 0x387AF50F0776E24F, 0x7FFD300A74DE5AE1, 0xB753981921AEB24B,
0x296F5919978BA0BD, 0x52CC9DD7031E59AC, 0xE5E646C9A5165AE3, 0x4F848BDF70532817,
0xA01ADD556B693C51, 0xC7268B83DDB75E97, 0xCB908F9316E2C076, 0x8FC39EE0D946E9C5,
0x5B256B0FE7937D86, 0xF77C51751A22932C, 0x2582A83CD34108C4, 0xE218CDE28A9C790D,
0xC93A572E592E3594, 0xEA70CF71BC4E55AC, 0xC135C8214155BC85, 0x50989643A6B8456E,
0x60E970DFD74C71E6, 0xC0B5A520AB7C88D8, 0x3A66AA761D5B1400, 0x0502AB3087300DE6,
0xC06203EDED483DEA, 0xC2620EDF55C1CB74, 0xF36A22CF9AA452F1, 0xB3502DC83BA2660F,
0xCF249B6DB82273C3, 0xC74DADAB026388F2, 0xB8B77B3275AFCDE8, 0xC9947382C6D6A3C0,
0x84938F2258A6BC21, 0x1E51E15A3B99CDF7, 0xC812F9AC41F5CC05, 0x21BFEC61E9B9393E,
0xF576066160243540, 0x3A62D1CB6404180D, 0x8807A55C2AC7AFE2, 0x804237B54859503E,
0x1619B3612106744B, 0xC1ECB5643D81C76F, 0xBA7CBB8C13214C6C, 0xD241AEAD7622701E,
0xDD900A1B66BF748C, 0xCACCF665EC2391FE, 0xF9BED90100B89447, 0x4CF5D284E56B7A0F,
0x003EB289B6993F96, 0xE9DB01146199245D, 0x97701270F3F41CCB, 0x9C8CA117E01E4B49
};
//----------------------------------------------------------------------------------------------------------
static const uint64_t plain_T0[256] = {
0xD83078C018601818, 0x2646AF05238C2323, 0xB891F97EC63FC6C6, 0xFBCD6F13E887E8E8,
0xCB13A14C87268787, 0x116D62A9B8DAB8B8, 0x0902050801040101, 0x0D9E6E424F214F4F,
0x9B6CEEAD36D83636, 0xFF510459A6A2A6A6, 0x0CB9BDDED26FD2D2, 0x0EF706FBF5F3F5F5,
0x96F280EF79F97979, 0x30DECE5F6FA16F6F, 0x6D3FEFFC917E9191, 0xF8A407AA52555252,
0x47C0FD27609D6060, 0x35657689BCCABCBC, 0x372BCDAC9B569B9B, 0x8A018C048E028E8E,
0xD25B1571A3B6A3A3, 0x6C183C600C300C0C, 0x84F68AFF7BF17B7B, 0x806AE1B535D43535,
0xF53A69E81D741D1D, 0xB3DD4753E0A7E0E0, 0x21B3ACF6D77BD7D7, 0x9C99ED5EC22FC2C2,
0x435C966D2EB82E2E, 0x29967A624B314B4B, 0x5DE121A3FEDFFEFE, 0xD5AE168257415757,
0xBD2A41A815541515, 0xE8EEB69F77C17777, 0x926EEBA537DC3737, 0x9ED7567BE5B3E5E5,
0x1323D98C9F469F9F, 0x23FD17D3F0E7F0F0, 0x20947F6A4A354A4A, 0x44A9959EDA4FDADA,
0xA2B025FA587D5858, 0xCF8FCA06C903C9C9, 0x7C528D5529A42929, 0x5A1422500A280A0A,
0x507F4FE1B1FEB1B1, 0xC95D1A69A0BAA0A0, 0x14D6DA7F6BB16B6B, 0xD917AB5C852E8585,
0x3C677381BDCEBDBD, 0x8FBA34D25D695D5D, 0x9020508010401010, 0x07F503F3F4F7F4F4,
0xDD8BC016CB0BCBCB, 0xD37CC6ED3EF83E3E, 0x2D0A112805140505, 0x78CEE61F67816767,
0x97D55373E4B7E4E4, 0x024EBB25279C2727, 0x7382583241194141, 0xA70B9D2C8B168B8B,
0xF6530151A7A6A7A7, 0xB2FA94CF7DE97D7D, 0x4937FBDC956E9595, 0x56AD9F8ED847D8D8,
0x70EB308BFBCBFBFB, 0xCDC17123EE9FEEEE, 0xBBF891C77CED7C7C, 0x71CCE31766856666,
0x7BA78EA6DD53DDDD, 0xAF2E4BB8175C1717, 0x458E460247014747, 0x1A21DC849E429E9E,
0xD489C51ECA0FCACA, 0x585A99752DB42D2D, 0x2E637991BFC6BFBF, 0x3F0E1B38071C0707,
0xAC472301AD8EADAD, 0xB0B42FEA5A755A5A, 0xEF1BB56C83368383, 0xB666FF8533CC3333,
0x5CC6F23F63916363, 0x12040A1002080202, 0x93493839AA92AAAA, 0xDEE2A8AF71D97171,
0xC68DCF0EC807C8C8, 0xD1327DC819641919, 0x3B92707249394949, 0x5FAF9A86D943D9D9,
0x31F91DC3F2EFF2F2, 0xA8DB484BE3ABE3E3, 0xB9B62AE25B715B5B, 0xBC0D9234881A8888,
0x3E29C8A49A529A9A, 0x0B4CBE2D26982626, 0xBF64FA8D32C83232, 0x597D4AE9B0FAB0B0,
0xF2CF6A1BE983E9E9, 0x771E33780F3C0F0F, 0x33B7A6E6D573D5D5, 0xF41DBA74803A8080,
0x27617C99BEC2BEBE, 0xEB87DE26CD13CDCD, 0x8968E4BD34D03434, 0x3290757A483D4848,
0x54E324ABFFDBFFFF, 0x8DF48FF77AF57A7A, 0x643DEAF4907A9090, 0x9DBE3EC25F615F5F,
0x3D40A01D20802020, 0x0FD0D56768BD6868, 0xCA3472D01A681A1A, 0xB7412C19AE82AEAE,
0x7D755EC9B4EAB4B4, 0xCEA8199A544D5454, 0x7F3BE5EC93769393, 0x2F44AA0D22882222,
0x63C8E907648D6464, 0x2AFF12DBF1E3F1F1, 0xCCE6A2BF73D17373, 0x82245A9012481212,
0x7A805D3A401D4040, 0x4810284008200808, 0x959BE856C32BC3C3, 0xDFC57B33EC97ECEC,
0x4DAB9096DB4BDBDB, 0xC05F1F61A1BEA1A1, 0x9107831C8D0E8D8D, 0xC87AC9F53DF43D3D,
0x5B33F1CC97669797, 0x0000000000000000, 0xF983D436CF1BCFCF, 0x6E5687452BAC2B2B,
0xE1ECB39776C57676, 0xE619B06482328282, 0x28B1A9FED67FD6D6, 0xC33677D81B6C1B1B,
0x74775BC1B5EEB5B5, 0xBE432911AF86AFAF, 0x1DD4DF776AB56A6A, 0xEAA00DBA505D5050,
0x578A4C1245094545, 0x38FB18CBF3EBF3F3, 0xAD60F09D30C03030, 0xC4C3742BEF9BEFEF,
0xDA7EC3E53FFC3F3F, 0xC7AA1C9255495555, 0xDB591079A2B2A2A2, 0xE9C96503EA8FEAEA,
0x6ACAEC0F65896565, 0x036968B9BAD2BABA, 0x4A5E93652FBC2F2F, 0x8E9DE74EC027C0C0,
0x60A181BEDE5FDEDE, 0xFC386CE01C701C1C, 0x46E72EBBFDD3FDFD, 0x1F9A64524D294D4D,
0x7639E0E492729292, 0xFAEABC8F75C97575, 0x360C1E3006180606, 0xAE0998248A128A8A,
0x4B7940F9B2F2B2B2, 0x85D15963E6BFE6E6, 0x7E1C36700E380E0E, 0xE73E63F81F7C1F1F,
0x55C4F73762956262, 0x3AB5A3EED477D4D4, 0x814D3229A89AA8A8, 0x5231F4C496629696,
0x62EF3A9BF9C3F9F9, 0xA397F666C533C5C5, 0x104AB13525942525, 0xABB220F259795959,
0xD015AE54842A8484, 0xC5E4A7B772D57272, 0xEC72DDD539E43939, 0x1698615A4C2D4C4C,
0x94BC3BCA5E655E5E, 0x9FF085E778FD7878, 0xE570D8DD38E03838, 0x980586148C0A8C8C,
0x17BFB2C6D163D1D1, 0xE4570B41A5AEA5A5, 0xA1D94D43E2AFE2E2, 0x4EC2F82F61996161,
0x427B45F1B3F6B3B3, 0x3442A51521842121, 0x0825D6949C4A9C9C, 0xEE3C66F01E781E1E,
0x6186522243114343, 0xB193FC76C73BC7C7, 0x4FE52BB3FCD7FCFC, 0x2408142004100404,
0xE3A208B251595151, 0x252FC7BC995E9999, 0x22DAC44F6DA96D6D, 0x651A39680D340D0D,
0x79E93583FACFFAFA, 0x69A384B6DF5BDFDF, 0xA9FC9BD77EE57E7E, 0x1948B43D24902424,
0xFE76D7C53BEC3B3B, 0x9A4B3D31AB96ABAB, 0xF081D13ECE1FCECE, 0x9922558811441111,
0x8303890C8F068F8F, 0x049C6B4A4E254E4E, 0x667351D1B7E6B7B7, 0xE0CB600BEB8BEBEB,
0xC178CCFD3CF03C3C, 0xFD1FBF7C813E8181, 0x4035FED4946A9494, 0x1CF30CEBF7FBF7F7,
0x186F67A1B9DEB9B9, 0x8B265F98134C1313, 0x51589C7D2CB02C2C, 0x05BBB8D6D36BD3D3,
0x8CD35C6BE7BBE7E7, 0x39DCCB576EA56E6E, 0xAA95F36EC437C4C4, 0x1B060F18030C0303,
0xDCAC138A56455656, 0x5E88491A440D4444, 0xA0FE9EDF7FE17F7F, 0x884F3721A99EA9A9,
0x6754824D2AA82A2A, 0x0A6B6DB1BBD6BBBB, 0x879FE246C123C1C1, 0xF1A602A253515353,
0x72A58BAEDC57DCDC, 0x531627580B2C0B0B, 0x0127D39C9D4E9D9D, 0x2BD8C1476CAD6C6C,
0xA462F59531C43131, 0xF3E8B98774CD7474, 0x15F109E3F6FFF6F6, 0x4C8C430A46054646,
0xA5452609AC8AACAC, 0xB50F973C891E8989, 0xB42844A014501414, 0xBADF425BE1A3E1E1,
0xA62C4EB016581616, 0xF774D2CD3AE83A3A, 0x06D2D06F69B96969, 0x41122D4809240909,
0xD7E0ADA770DD7070, 0x6F7154D9B6E2B6B6, 0x1EBDB7CED067D0D0, 0xD6C77E3BED93EDED,
0xE285DB2ECC17CCCC, 0x6884572A42154242, 0x2C2DC2B4985A9898, 0xED550E49A4AAA4A4,
0x7550885D28A02828, 0x86B831DA5C6D5C5C, 0x6BED3F93F8C7F8F8, 0xC211A44486228686
};
static const uint64_t plain_RC[10] = {
0x4F01B887E8C62318, 0x52916F79F5D2A636, 0x357B0CA38E9BBC60, 0x57FE4B2EC2D7E01D, 0xDA4AF09FE5377715,
0x856BA0B10A29C958, 0x67053ECBF4105DBD, 0xD8957DA78B4127E4, 0x9E4717DD667CEEFB, 0x33835AAD07BF2DCA
};
static const uint64_t plain_precomputed_round_key_64[72] = {
0x24AED1EAF889AB3B, 0xAFCBE94566454544, 0x89B2A4C5A4A4FE70, 0xA0E1CCE1E1A9FAC5,
0xFCB8FCFC5CC0AC48, 0x698F8F90260EF78F, 0x797985D707147996, 0xF878C8B868F8A8F8,
0x58704630DBBF19D3, 0xDB37CFAFD1235B29, 0x98AC958BC28A2C01, 0xA706B2C0B19E6381,
0xDB09B2B07A605E44, 0x71BC8CBCCF2C5B73, 0xD3DDEDEF240967DC, 0x197D3BD7F03B8D7B,
0x866511DEC1AABE38, 0x7F33874AD0F37C68, 0x57F0AD98DBFA37F3, 0xBC8D35EE5842E2C5,
0x7E246E99E8F00911, 0x0134B010EDD6C501, 0xD3EC287BF152C9FB, 0x4027F1C70CDC5632,
0x14CF9B9420A525AF, 0x4D53C4E3A92636C1, 0xE1F94077867D0FE6, 0x29066AE2BBE65D91,
0x8D5EFE4CCC545A96, 0xA63A3262CB31E9BE, 0x476A849618597BB1, 0x31AF592736C9F0D4,
0xB00B3725C0B5F9E2, 0xA5948416A2CB2B39, 0x148C34FACEF88A60, 0x19928C416437A57A,
0x893F83FAA146F3B3, 0x7CCF0278483F4997, 0x238F001EBAE8ADDC, 0x3D32B0ED494F7792,
0x2FFF4D7782634175, 0x00460355D038FAFF, 0x61F3983E49027DBF, 0x0BCEE59AC260A8F4,
0x279D5DEE445ADFC8, 0xA4007504555AF423, 0x8CE2F902121016B0, 0x1D33336829CD30AC,
0x89AD846882F16B03, 0x637146D862C64099, 0x10C2194B173E434C, 0xC586FF4CD3CF9CE2,
0x5326DF42A011FF21, 0x134BE46CCB008E1B, 0xCEB747A3F73B12A6, 0xCA33283B0E9018D9,
0xF92C9A0A7A671CD0, 0xB2B6634A532F942A, 0xB4A8ACFE46224288, 0x5935583DC75C4A47,
0xA16F5CA55D92A674, 0x395C73C48CE61777, 0xC61AEC530B3B2A08, 0x62E74D81EB58F62A,
0x3ABCEE01B6489548, 0x818EED6BC66B0DA5, 0x755A2688CF3DCEE0, 0xE99CF6C0DB4A8CC2,
0x1385717FD59CB754, 0x7B0B7D978A4B4143, 0x7A15F6DBBB351963, 0x27820137F64E7A6A
};

3232
x15/cuda_x15_whirlpool.cu
View File

File diff suppressed because it is too large Load Diff

25
x15/whirlpool.cu
Unescape View File

@ -9,15 +9,18 @@ extern "C"
#include "cuda_helper.h" #include "cuda_helper.h"
static uint32_t *d_hash[MAX_GPUS]; //static uint32_t *d_hash[MAX_GPUS];
extern void x15_whirlpool_cpu_init(int thr_id, uint32_t threads, int mode); extern void x15_whirlpool_cpu_init(int thr_id, uint32_t threads, int mode);
extern void x15_whirlpool_cpu_hash_64(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order);
extern void x15_whirlpool_cpu_free(int thr_id); extern void x15_whirlpool_cpu_free(int thr_id);
extern void whirlpool512_setBlock_80(void *pdata, const void *ptarget); extern void whirlpool512_setBlock_80(void *pdata, const void *ptarget);
extern void whirlpool512_cpu_hash_80(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *d_hash, int order);
extern uint32_t whirlpool512_cpu_finalhash_64(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order); //extern void x15_whirlpool_cpu_hash_64(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order);
//extern void whirlpool512_cpu_hash_80(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *d_hash, int order);
//extern uint32_t whirlpool512_cpu_finalhash_64(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *d_nonceVector, uint32_t *d_hash, int order);
extern void whirlpool512_cpu_hash_80(int thr_id, uint32_t threads, uint32_t startNounce, uint32_t *resNonces, const uint64_t target);
//#define _DEBUG //#define _DEBUG
#define _DEBUG_PREFIX "whirl" #define _DEBUG_PREFIX "whirl"
@ -88,7 +91,7 @@ extern "C" int scanhash_whirl(int thr_id, struct work* work, uint32_t max_nonce,
} }
gpulog(LOG_INFO, thr_id, "Intensity set to %g, %u cuda threads", throughput2intensity(throughput), throughput); gpulog(LOG_INFO, thr_id, "Intensity set to %g, %u cuda threads", throughput2intensity(throughput), throughput);
CUDA_SAFE_CALL(cudaMalloc(&d_hash[thr_id], (size_t) 64 * throughput)); //CUDA_SAFE_CALL(cudaMalloc(&d_hash[thr_id], (size_t) 64 * throughput));
x15_whirlpool_cpu_init(thr_id, throughput, 1 /* old whirlpool */); x15_whirlpool_cpu_init(thr_id, throughput, 1 /* old whirlpool */);
init[thr_id] = true; init[thr_id] = true;
@ -101,18 +104,20 @@ extern "C" int scanhash_whirl(int thr_id, struct work* work, uint32_t max_nonce,
whirlpool512_setBlock_80((void*)endiandata, ptarget); whirlpool512_setBlock_80((void*)endiandata, ptarget);
do { do {
/*
int order = 0; int order = 0;
*hashes_done = pdata[19] - first_nonce + throughput;
whirlpool512_cpu_hash_80(thr_id, throughput, pdata[19], d_hash[thr_id], order++); whirlpool512_cpu_hash_80(thr_id, throughput, pdata[19], d_hash[thr_id], order++);
TRACE64(" 80 :", d_hash); TRACE64(" 80 :", d_hash);
x15_whirlpool_cpu_hash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++); x15_whirlpool_cpu_hash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++);
TRACE64(" 64 :", d_hash); TRACE64(" 64 :", d_hash);
x15_whirlpool_cpu_hash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++); x15_whirlpool_cpu_hash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++);
TRACE64(" 64 :", d_hash); TRACE64(" 64 :", d_hash);
work->nonces[0] = whirlpool512_cpu_finalhash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++); work->nonces[0] = whirlpool512_cpu_finalhash_64(thr_id, throughput, pdata[19], NULL, d_hash[thr_id], order++);
*/
*hashes_done = pdata[19] - first_nonce + throughput;
whirlpool512_cpu_hash_80(thr_id, throughput, pdata[19], work->nonces, *(uint64_t*)&ptarget[6]);
if (work->nonces[0] != UINT32_MAX && bench_algo < 0) if (work->nonces[0] != UINT32_MAX && bench_algo < 0)
{ {
const uint32_t Htarg = ptarget[7]; const uint32_t Htarg = ptarget[7];
@ -154,7 +159,7 @@ extern "C" void free_whirl(int thr_id)
cudaThreadSynchronize(); cudaThreadSynchronize();
cudaFree(d_hash[thr_id]); //cudaFree(d_hash[thr_id]);
x15_whirlpool_cpu_free(thr_id); x15_whirlpool_cpu_free(thr_id);
init[thr_id] = false; init[thr_id] = false;

Write Preview
Loading…
Cancel
Save