GOSTSec
/
ccminer
mirror of https://github.com/GOSTSec/ccminer
1
0
Fork 0
Code Issues Releases Wiki Activity
GOSTcoin support for ccminer CUDA miner project, compatible with most nvidia cards
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
950 Commits
2 Branches
52 Tags
66 MiB
 
 
 
 
 
Tree: 6dc1bbdd47
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '6dc1bbdd47'
${ noResults }
ccminer/crypto/cn_jh.cuh

296 lines
18 KiB
Raw Blame History

typedef struct {
int hashbitlen;
unsigned long long databitlen;
unsigned long long datasize_in_buffer;
uint64_t x[8][2];
unsigned char buffer[64];
} jhHashState;
__constant__ unsigned char d_JH256_H0[512] = {
0xeb, 0x98, 0xa3, 0x41, 0x2c, 0x20, 0xd3, 0xeb, 0x92, 0xcd, 0xbe, 0x7b, 0x9c, 0xb2, 0x45, 0xc1,
0x1c, 0x93, 0x51, 0x91, 0x60, 0xd4, 0xc7, 0xfa, 0x26, 0x0, 0x82, 0xd6, 0x7e, 0x50, 0x8a, 0x3,
0xa4, 0x23, 0x9e, 0x26, 0x77, 0x26, 0xb9, 0x45, 0xe0, 0xfb, 0x1a, 0x48, 0xd4, 0x1a, 0x94, 0x77,
0xcd, 0xb5, 0xab, 0x26, 0x2, 0x6b, 0x17, 0x7a, 0x56, 0xf0, 0x24, 0x42, 0xf, 0xff, 0x2f, 0xa8,
0x71, 0xa3, 0x96, 0x89, 0x7f, 0x2e, 0x4d, 0x75, 0x1d, 0x14, 0x49, 0x8, 0xf7, 0x7d, 0xe2, 0x62,
0x27, 0x76, 0x95, 0xf7, 0x76, 0x24, 0x8f, 0x94, 0x87, 0xd5, 0xb6, 0x57, 0x47, 0x80, 0x29, 0x6c,
0x5c, 0x5e, 0x27, 0x2d, 0xac, 0x8e, 0xd, 0x6c, 0x51, 0x84, 0x50, 0xc6, 0x57, 0x5, 0x7a, 0xf,
0x7b, 0xe4, 0xd3, 0x67, 0x70, 0x24, 0x12, 0xea, 0x89, 0xe3, 0xab, 0x13, 0xd3, 0x1c, 0xd7, 0x69
};
__constant__ unsigned char d_E8_rc[42][32] = {
{0x72, 0xd5, 0xde, 0xa2, 0xdf, 0x15, 0xf8, 0x67, 0x7b, 0x84, 0x15, 0xa, 0xb7, 0x23, 0x15, 0x57, 0x81, 0xab, 0xd6, 0x90, 0x4d, 0x5a, 0x87, 0xf6, 0x4e, 0x9f, 0x4f, 0xc5, 0xc3, 0xd1, 0x2b, 0x40},
{0xea, 0x98, 0x3a, 0xe0, 0x5c, 0x45, 0xfa, 0x9c, 0x3, 0xc5, 0xd2, 0x99, 0x66, 0xb2, 0x99, 0x9a, 0x66, 0x2, 0x96, 0xb4, 0xf2, 0xbb, 0x53, 0x8a, 0xb5, 0x56, 0x14, 0x1a, 0x88, 0xdb, 0xa2, 0x31},
{0x3, 0xa3, 0x5a, 0x5c, 0x9a, 0x19, 0xe, 0xdb, 0x40, 0x3f, 0xb2, 0xa, 0x87, 0xc1, 0x44, 0x10, 0x1c, 0x5, 0x19, 0x80, 0x84, 0x9e, 0x95, 0x1d, 0x6f, 0x33, 0xeb, 0xad, 0x5e, 0xe7, 0xcd, 0xdc},
{0x10, 0xba, 0x13, 0x92, 0x2, 0xbf, 0x6b, 0x41, 0xdc, 0x78, 0x65, 0x15, 0xf7, 0xbb, 0x27, 0xd0, 0xa, 0x2c, 0x81, 0x39, 0x37, 0xaa, 0x78, 0x50, 0x3f, 0x1a, 0xbf, 0xd2, 0x41, 0x0, 0x91, 0xd3},
{0x42, 0x2d, 0x5a, 0xd, 0xf6, 0xcc, 0x7e, 0x90, 0xdd, 0x62, 0x9f, 0x9c, 0x92, 0xc0, 0x97, 0xce, 0x18, 0x5c, 0xa7, 0xb, 0xc7, 0x2b, 0x44, 0xac, 0xd1, 0xdf, 0x65, 0xd6, 0x63, 0xc6, 0xfc, 0x23},
{0x97, 0x6e, 0x6c, 0x3, 0x9e, 0xe0, 0xb8, 0x1a, 0x21, 0x5, 0x45, 0x7e, 0x44, 0x6c, 0xec, 0xa8, 0xee, 0xf1, 0x3, 0xbb, 0x5d, 0x8e, 0x61, 0xfa, 0xfd, 0x96, 0x97, 0xb2, 0x94, 0x83, 0x81, 0x97},
{0x4a, 0x8e, 0x85, 0x37, 0xdb, 0x3, 0x30, 0x2f, 0x2a, 0x67, 0x8d, 0x2d, 0xfb, 0x9f, 0x6a, 0x95, 0x8a, 0xfe, 0x73, 0x81, 0xf8, 0xb8, 0x69, 0x6c, 0x8a, 0xc7, 0x72, 0x46, 0xc0, 0x7f, 0x42, 0x14},
{0xc5, 0xf4, 0x15, 0x8f, 0xbd, 0xc7, 0x5e, 0xc4, 0x75, 0x44, 0x6f, 0xa7, 0x8f, 0x11, 0xbb, 0x80, 0x52, 0xde, 0x75, 0xb7, 0xae, 0xe4, 0x88, 0xbc, 0x82, 0xb8, 0x0, 0x1e, 0x98, 0xa6, 0xa3, 0xf4},
{0x8e, 0xf4, 0x8f, 0x33, 0xa9, 0xa3, 0x63, 0x15, 0xaa, 0x5f, 0x56, 0x24, 0xd5, 0xb7, 0xf9, 0x89, 0xb6, 0xf1, 0xed, 0x20, 0x7c, 0x5a, 0xe0, 0xfd, 0x36, 0xca, 0xe9, 0x5a, 0x6, 0x42, 0x2c, 0x36},
{0xce, 0x29, 0x35, 0x43, 0x4e, 0xfe, 0x98, 0x3d, 0x53, 0x3a, 0xf9, 0x74, 0x73, 0x9a, 0x4b, 0xa7, 0xd0, 0xf5, 0x1f, 0x59, 0x6f, 0x4e, 0x81, 0x86, 0xe, 0x9d, 0xad, 0x81, 0xaf, 0xd8, 0x5a, 0x9f},
{0xa7, 0x5, 0x6, 0x67, 0xee, 0x34, 0x62, 0x6a, 0x8b, 0xb, 0x28, 0xbe, 0x6e, 0xb9, 0x17, 0x27, 0x47, 0x74, 0x7, 0x26, 0xc6, 0x80, 0x10, 0x3f, 0xe0, 0xa0, 0x7e, 0x6f, 0xc6, 0x7e, 0x48, 0x7b},
{0xd, 0x55, 0xa, 0xa5, 0x4a, 0xf8, 0xa4, 0xc0, 0x91, 0xe3, 0xe7, 0x9f, 0x97, 0x8e, 0xf1, 0x9e, 0x86, 0x76, 0x72, 0x81, 0x50, 0x60, 0x8d, 0xd4, 0x7e, 0x9e, 0x5a, 0x41, 0xf3, 0xe5, 0xb0, 0x62},
{0xfc, 0x9f, 0x1f, 0xec, 0x40, 0x54, 0x20, 0x7a, 0xe3, 0xe4, 0x1a, 0x0, 0xce, 0xf4, 0xc9, 0x84, 0x4f, 0xd7, 0x94, 0xf5, 0x9d, 0xfa, 0x95, 0xd8, 0x55, 0x2e, 0x7e, 0x11, 0x24, 0xc3, 0x54, 0xa5},
{0x5b, 0xdf, 0x72, 0x28, 0xbd, 0xfe, 0x6e, 0x28, 0x78, 0xf5, 0x7f, 0xe2, 0xf, 0xa5, 0xc4, 0xb2, 0x5, 0x89, 0x7c, 0xef, 0xee, 0x49, 0xd3, 0x2e, 0x44, 0x7e, 0x93, 0x85, 0xeb, 0x28, 0x59, 0x7f},
{0x70, 0x5f, 0x69, 0x37, 0xb3, 0x24, 0x31, 0x4a, 0x5e, 0x86, 0x28, 0xf1, 0x1d, 0xd6, 0xe4, 0x65, 0xc7, 0x1b, 0x77, 0x4, 0x51, 0xb9, 0x20, 0xe7, 0x74, 0xfe, 0x43, 0xe8, 0x23, 0xd4, 0x87, 0x8a},
{0x7d, 0x29, 0xe8, 0xa3, 0x92, 0x76, 0x94, 0xf2, 0xdd, 0xcb, 0x7a, 0x9, 0x9b, 0x30, 0xd9, 0xc1, 0x1d, 0x1b, 0x30, 0xfb, 0x5b, 0xdc, 0x1b, 0xe0, 0xda, 0x24, 0x49, 0x4f, 0xf2, 0x9c, 0x82, 0xbf},
{0xa4, 0xe7, 0xba, 0x31, 0xb4, 0x70, 0xbf, 0xff, 0xd, 0x32, 0x44, 0x5, 0xde, 0xf8, 0xbc, 0x48, 0x3b, 0xae, 0xfc, 0x32, 0x53, 0xbb, 0xd3, 0x39, 0x45, 0x9f, 0xc3, 0xc1, 0xe0, 0x29, 0x8b, 0xa0},
{0xe5, 0xc9, 0x5, 0xfd, 0xf7, 0xae, 0x9, 0xf, 0x94, 0x70, 0x34, 0x12, 0x42, 0x90, 0xf1, 0x34, 0xa2, 0x71, 0xb7, 0x1, 0xe3, 0x44, 0xed, 0x95, 0xe9, 0x3b, 0x8e, 0x36, 0x4f, 0x2f, 0x98, 0x4a},
{0x88, 0x40, 0x1d, 0x63, 0xa0, 0x6c, 0xf6, 0x15, 0x47, 0xc1, 0x44, 0x4b, 0x87, 0x52, 0xaf, 0xff, 0x7e, 0xbb, 0x4a, 0xf1, 0xe2, 0xa, 0xc6, 0x30, 0x46, 0x70, 0xb6, 0xc5, 0xcc, 0x6e, 0x8c, 0xe6},
{0xa4, 0xd5, 0xa4, 0x56, 0xbd, 0x4f, 0xca, 0x0, 0xda, 0x9d, 0x84, 0x4b, 0xc8, 0x3e, 0x18, 0xae, 0x73, 0x57, 0xce, 0x45, 0x30, 0x64, 0xd1, 0xad, 0xe8, 0xa6, 0xce, 0x68, 0x14, 0x5c, 0x25, 0x67},
{0xa3, 0xda, 0x8c, 0xf2, 0xcb, 0xe, 0xe1, 0x16, 0x33, 0xe9, 0x6, 0x58, 0x9a, 0x94, 0x99, 0x9a, 0x1f, 0x60, 0xb2, 0x20, 0xc2, 0x6f, 0x84, 0x7b, 0xd1, 0xce, 0xac, 0x7f, 0xa0, 0xd1, 0x85, 0x18},
{0x32, 0x59, 0x5b, 0xa1, 0x8d, 0xdd, 0x19, 0xd3, 0x50, 0x9a, 0x1c, 0xc0, 0xaa, 0xa5, 0xb4, 0x46, 0x9f, 0x3d, 0x63, 0x67, 0xe4, 0x4, 0x6b, 0xba, 0xf6, 0xca, 0x19, 0xab, 0xb, 0x56, 0xee, 0x7e},
{0x1f, 0xb1, 0x79, 0xea, 0xa9, 0x28, 0x21, 0x74, 0xe9, 0xbd, 0xf7, 0x35, 0x3b, 0x36, 0x51, 0xee, 0x1d, 0x57, 0xac, 0x5a, 0x75, 0x50, 0xd3, 0x76, 0x3a, 0x46, 0xc2, 0xfe, 0xa3, 0x7d, 0x70, 0x1},
{0xf7, 0x35, 0xc1, 0xaf, 0x98, 0xa4, 0xd8, 0x42, 0x78, 0xed, 0xec, 0x20, 0x9e, 0x6b, 0x67, 0x79, 0x41, 0x83, 0x63, 0x15, 0xea, 0x3a, 0xdb, 0xa8, 0xfa, 0xc3, 0x3b, 0x4d, 0x32, 0x83, 0x2c, 0x83},
{0xa7, 0x40, 0x3b, 0x1f, 0x1c, 0x27, 0x47, 0xf3, 0x59, 0x40, 0xf0, 0x34, 0xb7, 0x2d, 0x76, 0x9a, 0xe7, 0x3e, 0x4e, 0x6c, 0xd2, 0x21, 0x4f, 0xfd, 0xb8, 0xfd, 0x8d, 0x39, 0xdc, 0x57, 0x59, 0xef},
{0x8d, 0x9b, 0xc, 0x49, 0x2b, 0x49, 0xeb, 0xda, 0x5b, 0xa2, 0xd7, 0x49, 0x68, 0xf3, 0x70, 0xd, 0x7d, 0x3b, 0xae, 0xd0, 0x7a, 0x8d, 0x55, 0x84, 0xf5, 0xa5, 0xe9, 0xf0, 0xe4, 0xf8, 0x8e, 0x65},
{0xa0, 0xb8, 0xa2, 0xf4, 0x36, 0x10, 0x3b, 0x53, 0xc, 0xa8, 0x7, 0x9e, 0x75, 0x3e, 0xec, 0x5a, 0x91, 0x68, 0x94, 0x92, 0x56, 0xe8, 0x88, 0x4f, 0x5b, 0xb0, 0x5c, 0x55, 0xf8, 0xba, 0xbc, 0x4c},
{0xe3, 0xbb, 0x3b, 0x99, 0xf3, 0x87, 0x94, 0x7b, 0x75, 0xda, 0xf4, 0xd6, 0x72, 0x6b, 0x1c, 0x5d, 0x64, 0xae, 0xac, 0x28, 0xdc, 0x34, 0xb3, 0x6d, 0x6c, 0x34, 0xa5, 0x50, 0xb8, 0x28, 0xdb, 0x71},
{0xf8, 0x61, 0xe2, 0xf2, 0x10, 0x8d, 0x51, 0x2a, 0xe3, 0xdb, 0x64, 0x33, 0x59, 0xdd, 0x75, 0xfc, 0x1c, 0xac, 0xbc, 0xf1, 0x43, 0xce, 0x3f, 0xa2, 0x67, 0xbb, 0xd1, 0x3c, 0x2, 0xe8, 0x43, 0xb0},
{0x33, 0xa, 0x5b, 0xca, 0x88, 0x29, 0xa1, 0x75, 0x7f, 0x34, 0x19, 0x4d, 0xb4, 0x16, 0x53, 0x5c, 0x92, 0x3b, 0x94, 0xc3, 0xe, 0x79, 0x4d, 0x1e, 0x79, 0x74, 0x75, 0xd7, 0xb6, 0xee, 0xaf, 0x3f},
{0xea, 0xa8, 0xd4, 0xf7, 0xbe, 0x1a, 0x39, 0x21, 0x5c, 0xf4, 0x7e, 0x9, 0x4c, 0x23, 0x27, 0x51, 0x26, 0xa3, 0x24, 0x53, 0xba, 0x32, 0x3c, 0xd2, 0x44, 0xa3, 0x17, 0x4a, 0x6d, 0xa6, 0xd5, 0xad},
{0xb5, 0x1d, 0x3e, 0xa6, 0xaf, 0xf2, 0xc9, 0x8, 0x83, 0x59, 0x3d, 0x98, 0x91, 0x6b, 0x3c, 0x56, 0x4c, 0xf8, 0x7c, 0xa1, 0x72, 0x86, 0x60, 0x4d, 0x46, 0xe2, 0x3e, 0xcc, 0x8, 0x6e, 0xc7, 0xf6},
{0x2f, 0x98, 0x33, 0xb3, 0xb1, 0xbc, 0x76, 0x5e, 0x2b, 0xd6, 0x66, 0xa5, 0xef, 0xc4, 0xe6, 0x2a, 0x6, 0xf4, 0xb6, 0xe8, 0xbe, 0xc1, 0xd4, 0x36, 0x74, 0xee, 0x82, 0x15, 0xbc, 0xef, 0x21, 0x63},
{0xfd, 0xc1, 0x4e, 0xd, 0xf4, 0x53, 0xc9, 0x69, 0xa7, 0x7d, 0x5a, 0xc4, 0x6, 0x58, 0x58, 0x26, 0x7e, 0xc1, 0x14, 0x16, 0x6, 0xe0, 0xfa, 0x16, 0x7e, 0x90, 0xaf, 0x3d, 0x28, 0x63, 0x9d, 0x3f},
{0xd2, 0xc9, 0xf2, 0xe3, 0x0, 0x9b, 0xd2, 0xc, 0x5f, 0xaa, 0xce, 0x30, 0xb7, 0xd4, 0xc, 0x30, 0x74, 0x2a, 0x51, 0x16, 0xf2, 0xe0, 0x32, 0x98, 0xd, 0xeb, 0x30, 0xd8, 0xe3, 0xce, 0xf8, 0x9a},
{0x4b, 0xc5, 0x9e, 0x7b, 0xb5, 0xf1, 0x79, 0x92, 0xff, 0x51, 0xe6, 0x6e, 0x4, 0x86, 0x68, 0xd3, 0x9b, 0x23, 0x4d, 0x57, 0xe6, 0x96, 0x67, 0x31, 0xcc, 0xe6, 0xa6, 0xf3, 0x17, 0xa, 0x75, 0x5},
{0xb1, 0x76, 0x81, 0xd9, 0x13, 0x32, 0x6c, 0xce, 0x3c, 0x17, 0x52, 0x84, 0xf8, 0x5, 0xa2, 0x62, 0xf4, 0x2b, 0xcb, 0xb3, 0x78, 0x47, 0x15, 0x47, 0xff, 0x46, 0x54, 0x82, 0x23, 0x93, 0x6a, 0x48},
{0x38, 0xdf, 0x58, 0x7, 0x4e, 0x5e, 0x65, 0x65, 0xf2, 0xfc, 0x7c, 0x89, 0xfc, 0x86, 0x50, 0x8e, 0x31, 0x70, 0x2e, 0x44, 0xd0, 0xb, 0xca, 0x86, 0xf0, 0x40, 0x9, 0xa2, 0x30, 0x78, 0x47, 0x4e},
{0x65, 0xa0, 0xee, 0x39, 0xd1, 0xf7, 0x38, 0x83, 0xf7, 0x5e, 0xe9, 0x37, 0xe4, 0x2c, 0x3a, 0xbd, 0x21, 0x97, 0xb2, 0x26, 0x1, 0x13, 0xf8, 0x6f, 0xa3, 0x44, 0xed, 0xd1, 0xef, 0x9f, 0xde, 0xe7},
{0x8b, 0xa0, 0xdf, 0x15, 0x76, 0x25, 0x92, 0xd9, 0x3c, 0x85, 0xf7, 0xf6, 0x12, 0xdc, 0x42, 0xbe, 0xd8, 0xa7, 0xec, 0x7c, 0xab, 0x27, 0xb0, 0x7e, 0x53, 0x8d, 0x7d, 0xda, 0xaa, 0x3e, 0xa8, 0xde},
{0xaa, 0x25, 0xce, 0x93, 0xbd, 0x2, 0x69, 0xd8, 0x5a, 0xf6, 0x43, 0xfd, 0x1a, 0x73, 0x8, 0xf9, 0xc0, 0x5f, 0xef, 0xda, 0x17, 0x4a, 0x19, 0xa5, 0x97, 0x4d, 0x66, 0x33, 0x4c, 0xfd, 0x21, 0x6a},
{0x35, 0xb4, 0x98, 0x31, 0xdb, 0x41, 0x15, 0x70, 0xea, 0x1e, 0xf, 0xbb, 0xed, 0xcd, 0x54, 0x9b, 0x9a, 0xd0, 0x63, 0xa1, 0x51, 0x97, 0x40, 0x72, 0xf6, 0x75, 0x9d, 0xbf, 0x91, 0x47, 0x6f, 0xe2}
};
#define JH_SWAP1(x) (x) = ((((x) & 0x5555555555555555ULL) << 1) | (((x) & 0xaaaaaaaaaaaaaaaaULL) >> 1));
#define JH_SWAP2(x) (x) = ((((x) & 0x3333333333333333ULL) << 2) | (((x) & 0xccccccccccccccccULL) >> 2));
#define JH_SWAP4(x) (x) = ((((x) & 0x0f0f0f0f0f0f0f0fULL) << 4) | (((x) & 0xf0f0f0f0f0f0f0f0ULL) >> 4));
#define JH_SWAP8(x) (x) = ((((x) & 0x00ff00ff00ff00ffULL) << 8) | (((x) & 0xff00ff00ff00ff00ULL) >> 8));
#define JH_SWAP16(x) (x) = ((((x) & 0x0000ffff0000ffffULL) << 16) | (((x) & 0xffff0000ffff0000ULL) >> 16));
#define JH_SWAP32(x) (x) = (((x) << 32) | ((x) >> 32));
#define JH_L(m0,m1,m2,m3,m4,m5,m6,m7) \
(m4) ^= (m1); \
(m5) ^= (m2); \
(m6) ^= (m0) ^ (m3); \
(m7) ^= (m0); \
(m0) ^= (m5); \
(m1) ^= (m6); \
(m2) ^= (m4) ^ (m7); \
(m3) ^= (m4);
#define JH_SS(m0,m1,m2,m3,m4,m5,m6,m7,cc0,cc1) \
m3 = ~(m3); \
m7 = ~(m7); \
m0 ^= ((~(m2)) & (cc0)); \
m4 ^= ((~(m6)) & (cc1)); \
temp0 = (cc0) ^ ((m0) & (m1));\
temp1 = (cc1) ^ ((m4) & (m5));\
m0 ^= ((m2) & (m3)); \
m4 ^= ((m6) & (m7)); \
m3 ^= ((~(m1)) & (m2)); \
m7 ^= ((~(m5)) & (m6)); \
m1 ^= ((m0) & (m2)); \
m5 ^= ((m4) & (m6)); \
m2 ^= ((m0) & (~(m3))); \
m6 ^= ((m4) & (~(m7))); \
m0 ^= ((m1) | (m3)); \
m4 ^= ((m5) | (m7)); \
m3 ^= ((m1) & (m2)); \
m7 ^= ((m5) & (m6)); \
m1 ^= (temp0 & (m0)); \
m5 ^= (temp1 & (m4)); \
m2 ^= temp0; \
m6 ^= temp1;
__device__
void cn_jh_E8(jhHashState *state)
{
uint64_t i,roundnumber,temp0,temp1;
for (roundnumber = 0; roundnumber < 42; roundnumber = roundnumber+7)
{
for (i = 0; i < 2; i++) {
JH_SS(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i],
((uint64_t *)d_E8_rc[roundnumber+0])[i],((uint64_t *)d_E8_rc[roundnumber+0])[i+2] );
JH_L(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i]);
JH_SWAP1(state->x[1][i]); JH_SWAP1(state->x[3][i]); JH_SWAP1(state->x[5][i]); JH_SWAP1(state->x[7][i]);
}
for (i = 0; i < 2; i++) {
JH_SS(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i],
((uint64_t *)d_E8_rc[roundnumber+1])[i],((uint64_t *)d_E8_rc[roundnumber+1])[i+2] );
JH_L(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i]);
JH_SWAP2(state->x[1][i]); JH_SWAP2(state->x[3][i]); JH_SWAP2(state->x[5][i]); JH_SWAP2(state->x[7][i]);
}
for (i = 0; i < 2; i++) {
JH_SS(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i],
((uint64_t *)d_E8_rc[roundnumber+2])[i],((uint64_t *)d_E8_rc[roundnumber+2])[i+2] );
JH_L(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i]);
JH_SWAP4(state->x[1][i]); JH_SWAP4(state->x[3][i]); JH_SWAP4(state->x[5][i]); JH_SWAP4(state->x[7][i]);
}
for (i = 0; i < 2; i++) {
JH_SS(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i],
((uint64_t *)d_E8_rc[roundnumber+3])[i],((uint64_t *)d_E8_rc[roundnumber+3])[i+2] );
JH_L(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i]);
JH_SWAP8(state->x[1][i]); JH_SWAP8(state->x[3][i]); JH_SWAP8(state->x[5][i]); JH_SWAP8(state->x[7][i]);
}
for (i = 0; i < 2; i++) {
JH_SS(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i],
((uint64_t *)d_E8_rc[roundnumber+4])[i],((uint64_t *)d_E8_rc[roundnumber+4])[i+2] );
JH_L(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i]);
JH_SWAP16(state->x[1][i]); JH_SWAP16(state->x[3][i]); JH_SWAP16(state->x[5][i]); JH_SWAP16(state->x[7][i]);
}
for (i = 0; i < 2; i++) {
JH_SS(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i],
((uint64_t *)d_E8_rc[roundnumber+5])[i],((uint64_t *)d_E8_rc[roundnumber+5])[i+2] );
JH_L(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i]);
JH_SWAP32(state->x[1][i]); JH_SWAP32(state->x[3][i]); JH_SWAP32(state->x[5][i]); JH_SWAP32(state->x[7][i]);
}
for (i = 0; i < 2; i++) {
JH_SS(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i],
((uint64_t *)d_E8_rc[roundnumber+6])[i],((uint64_t *)d_E8_rc[roundnumber+6])[i+2] );
JH_L(state->x[0][i],state->x[2][i],state->x[4][i],state->x[6][i],state->x[1][i],state->x[3][i],state->x[5][i],state->x[7][i]);
}
for (i = 1; i < 8; i = i+2) {
temp0 = state->x[i][0]; state->x[i][0] = state->x[i][1]; state->x[i][1] = temp0;
}
}
}
__device__
void cn_jh_F8(jhHashState *state)
{
uint64_t i;
for (i = 0; i < 8; i++) {
state->x[i >> 1][i & 1] ^= ((uint64_t *)state->buffer)[i];
}
cn_jh_E8(state);
for (i = 0; i < 8; i++) {
state->x[(8+i) >> 1][(8+i) & 1] ^= ((uint64_t *)state->buffer)[i];
}
}
__device__
void cn_jh_update(jhHashState * __restrict__ state, const uint8_t * __restrict__ data, DataLength databitlen)
{
DataLength index;
state->databitlen += databitlen;
index = 0;
if ( (state->datasize_in_buffer > 0 ) && (( state->datasize_in_buffer + databitlen) < 512) )
{
if ( (databitlen & 7) == 0 ) {
memcpy(state->buffer + (state->datasize_in_buffer >> 3), data, 64-(state->datasize_in_buffer >> 3));
}
else memcpy(state->buffer + (state->datasize_in_buffer >> 3), data, 64-(state->datasize_in_buffer >> 3)+1);
state->datasize_in_buffer += databitlen;
databitlen = 0;
}
if ( (state->datasize_in_buffer > 0 ) && (( state->datasize_in_buffer + databitlen) >= 512) )
{
memcpy( state->buffer + (state->datasize_in_buffer >> 3), data, 64-(state->datasize_in_buffer >> 3) );
index = 64-(state->datasize_in_buffer >> 3);
databitlen = databitlen - (512 - state->datasize_in_buffer);
cn_jh_F8(state);
state->datasize_in_buffer = 0;
}
for ( ; databitlen >= 512; index = index+64, databitlen = databitlen - 512) {
memcpy(state->buffer, data+index, 64);
cn_jh_F8(state);
}
if ( databitlen > 0) {
if ((databitlen & 7) == 0)
memcpy(state->buffer, data+index, (databitlen & 0x1ff) >> 3);
else
memcpy(state->buffer, data+index, ((databitlen & 0x1ff) >> 3)+1);
state->datasize_in_buffer = databitlen;
}
}
/* pad the message, process the padded block(s), truncate the hash value H to obtain the message digest */
__device__
void cn_jh_final(jhHashState * __restrict__ state, uint8_t * __restrict__ hashval)
{
unsigned int i;
//uint32_t *bufptr = (uint32_t *)state->buffer;
if ( (state->databitlen & 0x1ff) == 0 ) {
/* pad the message when databitlen is multiple of 512 bits, then process the padded block */
memset(state->buffer, 0, 64);
//for( i = 0; i < 16; i++ ) *(bufptr+i) = 0x00000000;
state->buffer[0] = 0x80;
state->buffer[63] = state->databitlen & 0xff;
state->buffer[62] = (state->databitlen >> 8) & 0xff;
state->buffer[61] = (state->databitlen >> 16) & 0xff;
state->buffer[60] = (state->databitlen >> 24) & 0xff;
state->buffer[59] = (state->databitlen >> 32) & 0xff;
state->buffer[58] = (state->databitlen >> 40) & 0xff;
state->buffer[57] = (state->databitlen >> 48) & 0xff;
state->buffer[56] = (state->databitlen >> 56) & 0xff;
cn_jh_F8(state);
} else {
/* set the rest of the bytes in the buffer to 0 */
if ( (state->datasize_in_buffer & 7) == 0) {
for (i = (state->databitlen & 0x1ff) >> 3; i < 64; i++) state->buffer[i] = 0;
} else {
for (i = ((state->databitlen & 0x1ff) >> 3)+1; i < 64; i++) state->buffer[i] = 0;
}
/*pad and process the partial block when databitlen is not multiple of 512 bits, then hash the padded blocks*/
state->buffer[((state->databitlen & 0x1ff) >> 3)] |= 1 << (7- (state->databitlen & 7));
cn_jh_F8(state);
memset(state->buffer, 0, 64);
//for( i = 0; i < 16; i++ ) *(bufptr+i) = 0x00000000;
state->buffer[63] = state->databitlen & 0xff;
state->buffer[62] = (state->databitlen >> 8) & 0xff;
state->buffer[61] = (state->databitlen >> 16) & 0xff;
state->buffer[60] = (state->databitlen >> 24) & 0xff;
state->buffer[59] = (state->databitlen >> 32) & 0xff;
state->buffer[58] = (state->databitlen >> 40) & 0xff;
state->buffer[57] = (state->databitlen >> 48) & 0xff;
state->buffer[56] = (state->databitlen >> 56) & 0xff;
cn_jh_F8(state);
}
memcpy(hashval, ((unsigned char*)state->x) + 64 + 32, 32);
//MEMCPY4(hashval, ((unsigned char*)state->x) + 64 + 32, 8);
}
__device__
void cn_jh_init(jhHashState *state, int hashbitlen)
{
state->databitlen = 0;
state->datasize_in_buffer = 0;
state->hashbitlen = hashbitlen;
memcpy(state->x, d_JH256_H0, 128);
//MEMCPY8(state->x, d_JH256_H0, 128 / 8);
}
__device__
void cn_jh(const uint8_t * __restrict__ data, DataLength len, uint32_t * hashval)
{
const int hashbitlen = 256;
DataLength databitlen = len << 3;
jhHashState state;
cn_jh_init(&state, hashbitlen);
cn_jh_update(&state, data, databitlen);
cn_jh_final(&state, (uint8_t*) hashval);
}