arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-12-27 17:08:12 +00:00
Code Issues Releases Wiki Activity

Fix -m 8500 kernel

Browse Source
This commit is contained in:
jsteube 2015-12-16 16:21:10 +01:00
parent 9d9097a12c
commit ed35c11e2a
3 changed files with 1044 additions and 939 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

151
OpenCL/m08500_a0.cl
View File

@ -84,7 +84,7 @@ __constant u8 ascii_to_ebcdic_pc[256] =
__constant u32 c_SPtrans[8][64] = __constant u32 c_SPtrans[8][64] =
{ {
/* nibble 0 */ {
0x02080800, 0x00080000, 0x02000002, 0x02080802, 0x02080800, 0x00080000, 0x02000002, 0x02080802,
0x02000000, 0x00080802, 0x00080002, 0x02000002, 0x02000000, 0x00080802, 0x00080002, 0x02000002,
0x00080802, 0x02080800, 0x02080000, 0x00000802, 0x00080802, 0x02080800, 0x02080000, 0x00000802,
@ -101,7 +101,8 @@ __constant u32 c_SPtrans[8][64] =
0x02000000, 0x00000802, 0x00080002, 0x00000000, 0x02000000, 0x00000802, 0x00080002, 0x00000000,
0x00080000, 0x02000000, 0x02000802, 0x02080800, 0x00080000, 0x02000000, 0x02000802, 0x02080800,
0x00000002, 0x02080002, 0x00000800, 0x00080802, 0x00000002, 0x02080002, 0x00000800, 0x00080802,
/* nibble 1 */ },
{
0x40108010, 0x00000000, 0x00108000, 0x40100000, 0x40108010, 0x00000000, 0x00108000, 0x40100000,
0x40000010, 0x00008010, 0x40008000, 0x00108000, 0x40000010, 0x00008010, 0x40008000, 0x00108000,
0x00008000, 0x40100010, 0x00000010, 0x40008000, 0x00008000, 0x40100010, 0x00000010, 0x40008000,
@ -118,7 +119,8 @@ __constant u32 c_SPtrans[8][64] =
0x00000010, 0x40108010, 0x00108000, 0x40100000, 0x00000010, 0x40108010, 0x00108000, 0x40100000,
0x40100010, 0x00100000, 0x00008010, 0x40008000, 0x40100010, 0x00100000, 0x00008010, 0x40008000,
0x40008010, 0x00000010, 0x40100000, 0x00108000, 0x40008010, 0x00000010, 0x40100000, 0x00108000,
/* nibble 2 */ },
{
0x04000001, 0x04040100, 0x00000100, 0x04000101, 0x04000001, 0x04040100, 0x00000100, 0x04000101,
0x00040001, 0x04000000, 0x04000101, 0x00040100, 0x00040001, 0x04000000, 0x04000101, 0x00040100,
0x04000100, 0x00040000, 0x04040000, 0x00000001, 0x04000100, 0x00040000, 0x04040000, 0x00000001,
@ -135,7 +137,8 @@ __constant u32 c_SPtrans[8][64] =
0x00040000, 0x04000100, 0x04000101, 0x00040100, 0x00040000, 0x04000100, 0x04000101, 0x00040100,
0x04000100, 0x00000000, 0x04040001, 0x00000101, 0x04000100, 0x00000000, 0x04040001, 0x00000101,
0x04000001, 0x00040101, 0x00000100, 0x04040000, 0x04000001, 0x00040101, 0x00000100, 0x04040000,
/* nibble 3 */ },
{
0x00401008, 0x10001000, 0x00000008, 0x10401008, 0x00401008, 0x10001000, 0x00000008, 0x10401008,
0x00000000, 0x10400000, 0x10001008, 0x00400008, 0x00000000, 0x10400000, 0x10001008, 0x00400008,
0x10401000, 0x10000008, 0x10000000, 0x00001008, 0x10401000, 0x10000008, 0x10000000, 0x00001008,
@ -152,7 +155,8 @@ __constant u32 c_SPtrans[8][64] =
0x10401008, 0x00000008, 0x10001000, 0x00401008, 0x10401008, 0x00000008, 0x10001000, 0x00401008,
0x00400008, 0x00401000, 0x10400000, 0x10001008, 0x00400008, 0x00401000, 0x10400000, 0x10001008,
0x00001008, 0x10000000, 0x10000008, 0x10401000, 0x00001008, 0x10000000, 0x10000008, 0x10401000,
/* nibble 4 */ },
{
0x08000000, 0x00010000, 0x00000400, 0x08010420, 0x08000000, 0x00010000, 0x00000400, 0x08010420,
0x08010020, 0x08000400, 0x00010420, 0x08010000, 0x08010020, 0x08000400, 0x00010420, 0x08010000,
0x00010000, 0x00000020, 0x08000020, 0x00010400, 0x00010000, 0x00000020, 0x08000020, 0x00010400,
@ -169,7 +173,8 @@ __constant u32 c_SPtrans[8][64] =
0x08010020, 0x08010400, 0x00000420, 0x00010000, 0x08010020, 0x08010400, 0x00000420, 0x00010000,
0x00010400, 0x08010020, 0x08000400, 0x00000420, 0x00010400, 0x08010020, 0x08000400, 0x00000420,
0x00000020, 0x00010420, 0x08010000, 0x08000020, 0x00000020, 0x00010420, 0x08010000, 0x08000020,
/* nibble 5 */ },
{
0x80000040, 0x00200040, 0x00000000, 0x80202000, 0x80000040, 0x00200040, 0x00000000, 0x80202000,
0x00200040, 0x00002000, 0x80002040, 0x00200000, 0x00200040, 0x00002000, 0x80002040, 0x00200000,
0x00002040, 0x80202040, 0x00202000, 0x80000000, 0x00002040, 0x80202040, 0x00202000, 0x80000000,
@ -186,7 +191,8 @@ __constant u32 c_SPtrans[8][64] =
0x80000040, 0x80200000, 0x00202040, 0x00000000, 0x80000040, 0x80200000, 0x00202040, 0x00000000,
0x00002000, 0x80000040, 0x80002040, 0x80202000, 0x00002000, 0x80000040, 0x80002040, 0x80202000,
0x80200000, 0x00002040, 0x00000040, 0x80200040, 0x80200000, 0x00002040, 0x00000040, 0x80200040,
/* nibble 6 */ },
{
0x00004000, 0x00000200, 0x01000200, 0x01000004, 0x00004000, 0x00000200, 0x01000200, 0x01000004,
0x01004204, 0x00004004, 0x00004200, 0x00000000, 0x01004204, 0x00004004, 0x00004200, 0x00000000,
0x01000000, 0x01000204, 0x00000204, 0x01004000, 0x01000000, 0x01000204, 0x00000204, 0x01004000,
@ -203,7 +209,8 @@ __constant u32 c_SPtrans[8][64] =
0x00000204, 0x00004000, 0x01004204, 0x01000000, 0x00000204, 0x00004000, 0x01004204, 0x01000000,
0x01004200, 0x00000004, 0x00004004, 0x01004204, 0x01004200, 0x00000004, 0x00004004, 0x01004204,
0x01000004, 0x01004200, 0x01004000, 0x00004004, 0x01000004, 0x01004200, 0x01004000, 0x00004004,
/* nibble 7 */ },
{
0x20800080, 0x20820000, 0x00020080, 0x00000000, 0x20800080, 0x20820000, 0x00020080, 0x00000000,
0x20020000, 0x00800080, 0x20800000, 0x20820080, 0x20020000, 0x00800080, 0x20800000, 0x20820080,
0x00000080, 0x20000000, 0x00820000, 0x00020080, 0x00000080, 0x20000000, 0x00820000, 0x00020080,
@ -220,11 +227,12 @@ __constant u32 c_SPtrans[8][64] =
0x00820000, 0x00020080, 0x20020080, 0x20800000, 0x00820000, 0x00020080, 0x20020080, 0x20800000,
0x00000080, 0x20820000, 0x00820080, 0x00000000, 0x00000080, 0x20820000, 0x00820080, 0x00000000,
0x20000000, 0x20800080, 0x00020000, 0x00820080, 0x20000000, 0x20800080, 0x00020000, 0x00820080,
}
}; };
__constant u32 c_skb[8][64] = __constant u32 c_skb[8][64] =
{ {
/* for C bits (numbered as per FIPS 46) 1 2 3 4 5 6 */ {
0x00000000, 0x00000010, 0x20000000, 0x20000010, 0x00000000, 0x00000010, 0x20000000, 0x20000010,
0x00010000, 0x00010010, 0x20010000, 0x20010010, 0x00010000, 0x00010010, 0x20010000, 0x20010010,
0x00000800, 0x00000810, 0x20000800, 0x20000810, 0x00000800, 0x00000810, 0x20000800, 0x20000810,
@ -241,7 +249,8 @@ __constant u32 c_skb[8][64] =
0x00090020, 0x00090030, 0x20090020, 0x20090030, 0x00090020, 0x00090030, 0x20090020, 0x20090030,
0x00080820, 0x00080830, 0x20080820, 0x20080830, 0x00080820, 0x00080830, 0x20080820, 0x20080830,
0x00090820, 0x00090830, 0x20090820, 0x20090830, 0x00090820, 0x00090830, 0x20090820, 0x20090830,
/* for C bits (numbered as per FIPS 46) 7 8 10 11 12 13 */ },
{
0x00000000, 0x02000000, 0x00002000, 0x02002000, 0x00000000, 0x02000000, 0x00002000, 0x02002000,
0x00200000, 0x02200000, 0x00202000, 0x02202000, 0x00200000, 0x02200000, 0x00202000, 0x02202000,
0x00000004, 0x02000004, 0x00002004, 0x02002004, 0x00000004, 0x02000004, 0x00002004, 0x02002004,
@ -258,7 +267,8 @@ __constant u32 c_skb[8][64] =
0x10200400, 0x12200400, 0x10202400, 0x12202400, 0x10200400, 0x12200400, 0x10202400, 0x12202400,
0x10000404, 0x12000404, 0x10002404, 0x12002404, 0x10000404, 0x12000404, 0x10002404, 0x12002404,
0x10200404, 0x12200404, 0x10202404, 0x12202404, 0x10200404, 0x12200404, 0x10202404, 0x12202404,
/* for C bits (numbered as per FIPS 46) 14 15 16 17 19 20 */ },
{
0x00000000, 0x00000001, 0x00040000, 0x00040001, 0x00000000, 0x00000001, 0x00040000, 0x00040001,
0x01000000, 0x01000001, 0x01040000, 0x01040001, 0x01000000, 0x01000001, 0x01040000, 0x01040001,
0x00000002, 0x00000003, 0x00040002, 0x00040003, 0x00000002, 0x00000003, 0x00040002, 0x00040003,
@ -275,7 +285,8 @@ __constant u32 c_skb[8][64] =
0x09000200, 0x09000201, 0x09040200, 0x09040201, 0x09000200, 0x09000201, 0x09040200, 0x09040201,
0x08000202, 0x08000203, 0x08040202, 0x08040203, 0x08000202, 0x08000203, 0x08040202, 0x08040203,
0x09000202, 0x09000203, 0x09040202, 0x09040203, 0x09000202, 0x09000203, 0x09040202, 0x09040203,
/* for C bits (numbered as per FIPS 46) 21 23 24 26 27 28 */ },
{
0x00000000, 0x00100000, 0x00000100, 0x00100100, 0x00000000, 0x00100000, 0x00000100, 0x00100100,
0x00000008, 0x00100008, 0x00000108, 0x00100108, 0x00000008, 0x00100008, 0x00000108, 0x00100108,
0x00001000, 0x00101000, 0x00001100, 0x00101100, 0x00001000, 0x00101000, 0x00001100, 0x00101100,
@ -292,7 +303,8 @@ __constant u32 c_skb[8][64] =
0x04020008, 0x04120008, 0x04020108, 0x04120108, 0x04020008, 0x04120008, 0x04020108, 0x04120108,
0x04021000, 0x04121000, 0x04021100, 0x04121100, 0x04021000, 0x04121000, 0x04021100, 0x04121100,
0x04021008, 0x04121008, 0x04021108, 0x04121108, 0x04021008, 0x04121008, 0x04021108, 0x04121108,
/* for D bits (numbered as per FIPS 46) 1 2 3 4 5 6 */ },
{
0x00000000, 0x10000000, 0x00010000, 0x10010000, 0x00000000, 0x10000000, 0x00010000, 0x10010000,
0x00000004, 0x10000004, 0x00010004, 0x10010004, 0x00000004, 0x10000004, 0x00010004, 0x10010004,
0x20000000, 0x30000000, 0x20010000, 0x30010000, 0x20000000, 0x30000000, 0x20010000, 0x30010000,
@ -309,7 +321,8 @@ __constant u32 c_skb[8][64] =
0x00101004, 0x10101004, 0x00111004, 0x10111004, 0x00101004, 0x10101004, 0x00111004, 0x10111004,
0x20101000, 0x30101000, 0x20111000, 0x30111000, 0x20101000, 0x30101000, 0x20111000, 0x30111000,
0x20101004, 0x30101004, 0x20111004, 0x30111004, 0x20101004, 0x30101004, 0x20111004, 0x30111004,
/* for D bits (numbered as per FIPS 46) 8 9 11 12 13 14 */ },
{
0x00000000, 0x08000000, 0x00000008, 0x08000008, 0x00000000, 0x08000000, 0x00000008, 0x08000008,
0x00000400, 0x08000400, 0x00000408, 0x08000408, 0x00000400, 0x08000400, 0x00000408, 0x08000408,
0x00020000, 0x08020000, 0x00020008, 0x08020008, 0x00020000, 0x08020000, 0x00020008, 0x08020008,
@ -326,7 +339,8 @@ __constant u32 c_skb[8][64] =
0x02000401, 0x0A000401, 0x02000409, 0x0A000409, 0x02000401, 0x0A000401, 0x02000409, 0x0A000409,
0x02020001, 0x0A020001, 0x02020009, 0x0A020009, 0x02020001, 0x0A020001, 0x02020009, 0x0A020009,
0x02020401, 0x0A020401, 0x02020409, 0x0A020409, 0x02020401, 0x0A020401, 0x02020409, 0x0A020409,
/* for D bits (numbered as per FIPS 46) 16 17 18 19 20 21 */ },
{
0x00000000, 0x00000100, 0x00080000, 0x00080100, 0x00000000, 0x00000100, 0x00080000, 0x00080100,
0x01000000, 0x01000100, 0x01080000, 0x01080100, 0x01000000, 0x01000100, 0x01080000, 0x01080100,
0x00000010, 0x00000110, 0x00080010, 0x00080110, 0x00000010, 0x00000110, 0x00080010, 0x00080110,
@ -343,7 +357,8 @@ __constant u32 c_skb[8][64] =
0x01200200, 0x01200300, 0x01280200, 0x01280300, 0x01200200, 0x01200300, 0x01280200, 0x01280300,
0x00200210, 0x00200310, 0x00280210, 0x00280310, 0x00200210, 0x00200310, 0x00280210, 0x00280310,
0x01200210, 0x01200310, 0x01280210, 0x01280310, 0x01200210, 0x01200310, 0x01280210, 0x01280310,
/* for D bits (numbered as per FIPS 46) 22 23 24 25 27 28 */ },
{
0x00000000, 0x04000000, 0x00040000, 0x04040000, 0x00000000, 0x04000000, 0x00040000, 0x04040000,
0x00000002, 0x04000002, 0x00040002, 0x04040002, 0x00000002, 0x04000002, 0x00040002, 0x04040002,
0x00002000, 0x04002000, 0x00042000, 0x04042000, 0x00002000, 0x04002000, 0x00042000, 0x04042000,
@ -359,10 +374,11 @@ __constant u32 c_skb[8][64] =
0x00000820, 0x04000820, 0x00040820, 0x04040820, 0x00000820, 0x04000820, 0x00040820, 0x04040820,
0x00000822, 0x04000822, 0x00040822, 0x04040822, 0x00000822, 0x04000822, 0x00040822, 0x04040822,
0x00002820, 0x04002820, 0x00042820, 0x04042820, 0x00002820, 0x04002820, 0x00042820, 0x04042820,
0x00002822, 0x04002822, 0x00042822, 0x04042822, 0x00002822, 0x04002822, 0x00042822, 0x04042822
}
}; };
#define NBOX(i,n,S) (S)[(n)][(i)] #define BOX(i,n,S) (S)[(n)][(i)]
static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 s_SPtrans[8][64]) static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 s_SPtrans[8][64])
{ {
@ -372,23 +388,34 @@ static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16],
u32 l = data[1]; u32 l = data[1];
#pragma unroll 16 #pragma unroll 16
for (u32 i = 0; i < 16; i++) for (u32 i = 0; i < 16; i += 2)
{ {
u32 u = Kc[i] ^ r; u32 u;
u32 t = Kd[i] ^ rotl32 (r, 28u); u32 t;
l ^= NBOX (((u >> 2) & 0x3f), 0, s_SPtrans) u = Kc[i + 0] ^ r;
| NBOX (((u >> 10) & 0x3f), 2, s_SPtrans) t = Kd[i + 0] ^ rotl32 (r, 28u);
| NBOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| NBOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| NBOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| NBOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| NBOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| NBOX (((t >> 26) & 0x3f), 7, s_SPtrans);
tt = l; l ^= BOX (((u >> 2) & 0x3f), 0, s_SPtrans)
l = r; | BOX (((u >> 10) & 0x3f), 2, s_SPtrans)
r = tt; | BOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| BOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| BOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| BOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| BOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| BOX (((t >> 26) & 0x3f), 7, s_SPtrans);
u = Kc[i + 1] ^ l;
t = Kd[i + 1] ^ rotl32 (l, 28u);
r ^= BOX (((u >> 2) & 0x3f), 0, s_SPtrans)
| BOX (((u >> 10) & 0x3f), 2, s_SPtrans)
| BOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| BOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| BOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| BOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| BOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| BOX (((t >> 26) & 0x3f), 7, s_SPtrans);
} }
iv[0] = l; iv[0] = l;
@ -416,41 +443,49 @@ static void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u
#pragma unroll 16 #pragma unroll 16
for (u32 i = 0; i < 16; i++) for (u32 i = 0; i < 16; i++)
{ {
const u32 shifts3s0[16] = { 1, 1, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 1 }; if ((i < 2) || (i == 8) || (i == 15))
const u32 shifts3s1[16] = { 27, 27, 26, 26, 26, 26, 26, 26, 27, 26, 26, 26, 26, 26, 26, 27 }; {
c = ((c >> 1) | (c << 27));
c = c >> shifts3s0[i] | c << shifts3s1[i]; d = ((d >> 1) | (d << 27));
d = d >> shifts3s0[i] | d << shifts3s1[i]; }
else
{
c = ((c >> 2) | (c << 26));
d = ((d >> 2) | (d << 26));
}
c = c & 0x0fffffff; c = c & 0x0fffffff;
d = d & 0x0fffffff; d = d & 0x0fffffff;
u32 s; const u32 c00 = (c >> 0) & 0x0000003f;
u32 t; const u32 c06 = (c >> 6) & 0x00383003;
const u32 c07 = (c >> 7) & 0x0000003c;
const u32 c13 = (c >> 13) & 0x0000060f;
const u32 c20 = (c >> 20) & 0x00000001;
s = NBOX ((( c >> 0) & 0x3f), 0, s_skb) u32 s = BOX (((c00 >> 0) & 0xff), 0, s_skb)
| NBOX ((((c >> 6) & 0x03) | BOX (((c06 >> 0) & 0xff)
| ((c >> 7) & 0x3c)), 1, s_skb) |((c07 >> 0) & 0xff), 1, s_skb)
| NBOX ((((c >> 13) & 0x0f) | BOX (((c13 >> 0) & 0xff)
| ((c >> 14) & 0x30)), 2, s_skb) |((c06 >> 8) & 0xff), 2, s_skb)
| NBOX ((((c >> 20) & 0x01) | BOX (((c20 >> 0) & 0xff)
| ((c >> 21) & 0x06) |((c13 >> 8) & 0xff)
| ((c >> 22) & 0x38)), 3, s_skb); |((c06 >> 16) & 0xff), 3, s_skb);
t = NBOX ((( d >> 0) & 0x3f), 4, s_skb) const u32 d00 = (d >> 0) & 0x00003c3f;
| NBOX ((((d >> 7) & 0x03) const u32 d07 = (d >> 7) & 0x00003f03;
| ((d >> 8) & 0x3c)), 5, s_skb) const u32 d21 = (d >> 21) & 0x0000000f;
| NBOX ((((d >> 15) & 0x3f)), 6, s_skb) const u32 d22 = (d >> 22) & 0x00000030;
| NBOX ((((d >> 21) & 0x0f)
| ((d >> 22) & 0x30)), 7, s_skb); u32 t = BOX (((d00 >> 0) & 0xff), 4, s_skb)
| BOX (((d07 >> 0) & 0xff)
|((d00 >> 8) & 0xff), 5, s_skb)
| BOX (((d07 >> 8) & 0xff), 6, s_skb)
| BOX (((d21 >> 0) & 0xff)
|((d22 >> 0) & 0xff), 7, s_skb);
#if defined cl_amd_media_ops
Kc[i] = amd_bytealign (t, s << 16, 2);
Kd[i] = amd_bytealign (t >> 16, s, 2);
#else
Kc[i] = ((t << 16) | (s & 0x0000ffff)); Kc[i] = ((t << 16) | (s & 0x0000ffff));
Kd[i] = ((s >> 16) | (t & 0xffff0000)); Kd[i] = ((s >> 16) | (t & 0xffff0000));
#endif
Kc[i] = rotl32 (Kc[i], 2u); Kc[i] = rotl32 (Kc[i], 2u);
Kd[i] = rotl32 (Kd[i], 2u); Kd[i] = rotl32 (Kd[i], 2u);

151
OpenCL/m08500_a1.cl
View File

@ -82,7 +82,7 @@ __constant u8 ascii_to_ebcdic_pc[256] =
__constant u32 c_SPtrans[8][64] = __constant u32 c_SPtrans[8][64] =
{ {
/* nibble 0 */ {
0x02080800, 0x00080000, 0x02000002, 0x02080802, 0x02080800, 0x00080000, 0x02000002, 0x02080802,
0x02000000, 0x00080802, 0x00080002, 0x02000002, 0x02000000, 0x00080802, 0x00080002, 0x02000002,
0x00080802, 0x02080800, 0x02080000, 0x00000802, 0x00080802, 0x02080800, 0x02080000, 0x00000802,
@ -99,7 +99,8 @@ __constant u32 c_SPtrans[8][64] =
0x02000000, 0x00000802, 0x00080002, 0x00000000, 0x02000000, 0x00000802, 0x00080002, 0x00000000,
0x00080000, 0x02000000, 0x02000802, 0x02080800, 0x00080000, 0x02000000, 0x02000802, 0x02080800,
0x00000002, 0x02080002, 0x00000800, 0x00080802, 0x00000002, 0x02080002, 0x00000800, 0x00080802,
/* nibble 1 */ },
{
0x40108010, 0x00000000, 0x00108000, 0x40100000, 0x40108010, 0x00000000, 0x00108000, 0x40100000,
0x40000010, 0x00008010, 0x40008000, 0x00108000, 0x40000010, 0x00008010, 0x40008000, 0x00108000,
0x00008000, 0x40100010, 0x00000010, 0x40008000, 0x00008000, 0x40100010, 0x00000010, 0x40008000,
@ -116,7 +117,8 @@ __constant u32 c_SPtrans[8][64] =
0x00000010, 0x40108010, 0x00108000, 0x40100000, 0x00000010, 0x40108010, 0x00108000, 0x40100000,
0x40100010, 0x00100000, 0x00008010, 0x40008000, 0x40100010, 0x00100000, 0x00008010, 0x40008000,
0x40008010, 0x00000010, 0x40100000, 0x00108000, 0x40008010, 0x00000010, 0x40100000, 0x00108000,
/* nibble 2 */ },
{
0x04000001, 0x04040100, 0x00000100, 0x04000101, 0x04000001, 0x04040100, 0x00000100, 0x04000101,
0x00040001, 0x04000000, 0x04000101, 0x00040100, 0x00040001, 0x04000000, 0x04000101, 0x00040100,
0x04000100, 0x00040000, 0x04040000, 0x00000001, 0x04000100, 0x00040000, 0x04040000, 0x00000001,
@ -133,7 +135,8 @@ __constant u32 c_SPtrans[8][64] =
0x00040000, 0x04000100, 0x04000101, 0x00040100, 0x00040000, 0x04000100, 0x04000101, 0x00040100,
0x04000100, 0x00000000, 0x04040001, 0x00000101, 0x04000100, 0x00000000, 0x04040001, 0x00000101,
0x04000001, 0x00040101, 0x00000100, 0x04040000, 0x04000001, 0x00040101, 0x00000100, 0x04040000,
/* nibble 3 */ },
{
0x00401008, 0x10001000, 0x00000008, 0x10401008, 0x00401008, 0x10001000, 0x00000008, 0x10401008,
0x00000000, 0x10400000, 0x10001008, 0x00400008, 0x00000000, 0x10400000, 0x10001008, 0x00400008,
0x10401000, 0x10000008, 0x10000000, 0x00001008, 0x10401000, 0x10000008, 0x10000000, 0x00001008,
@ -150,7 +153,8 @@ __constant u32 c_SPtrans[8][64] =
0x10401008, 0x00000008, 0x10001000, 0x00401008, 0x10401008, 0x00000008, 0x10001000, 0x00401008,
0x00400008, 0x00401000, 0x10400000, 0x10001008, 0x00400008, 0x00401000, 0x10400000, 0x10001008,
0x00001008, 0x10000000, 0x10000008, 0x10401000, 0x00001008, 0x10000000, 0x10000008, 0x10401000,
/* nibble 4 */ },
{
0x08000000, 0x00010000, 0x00000400, 0x08010420, 0x08000000, 0x00010000, 0x00000400, 0x08010420,
0x08010020, 0x08000400, 0x00010420, 0x08010000, 0x08010020, 0x08000400, 0x00010420, 0x08010000,
0x00010000, 0x00000020, 0x08000020, 0x00010400, 0x00010000, 0x00000020, 0x08000020, 0x00010400,
@ -167,7 +171,8 @@ __constant u32 c_SPtrans[8][64] =
0x08010020, 0x08010400, 0x00000420, 0x00010000, 0x08010020, 0x08010400, 0x00000420, 0x00010000,
0x00010400, 0x08010020, 0x08000400, 0x00000420, 0x00010400, 0x08010020, 0x08000400, 0x00000420,
0x00000020, 0x00010420, 0x08010000, 0x08000020, 0x00000020, 0x00010420, 0x08010000, 0x08000020,
/* nibble 5 */ },
{
0x80000040, 0x00200040, 0x00000000, 0x80202000, 0x80000040, 0x00200040, 0x00000000, 0x80202000,
0x00200040, 0x00002000, 0x80002040, 0x00200000, 0x00200040, 0x00002000, 0x80002040, 0x00200000,
0x00002040, 0x80202040, 0x00202000, 0x80000000, 0x00002040, 0x80202040, 0x00202000, 0x80000000,
@ -184,7 +189,8 @@ __constant u32 c_SPtrans[8][64] =
0x80000040, 0x80200000, 0x00202040, 0x00000000, 0x80000040, 0x80200000, 0x00202040, 0x00000000,
0x00002000, 0x80000040, 0x80002040, 0x80202000, 0x00002000, 0x80000040, 0x80002040, 0x80202000,
0x80200000, 0x00002040, 0x00000040, 0x80200040, 0x80200000, 0x00002040, 0x00000040, 0x80200040,
/* nibble 6 */ },
{
0x00004000, 0x00000200, 0x01000200, 0x01000004, 0x00004000, 0x00000200, 0x01000200, 0x01000004,
0x01004204, 0x00004004, 0x00004200, 0x00000000, 0x01004204, 0x00004004, 0x00004200, 0x00000000,
0x01000000, 0x01000204, 0x00000204, 0x01004000, 0x01000000, 0x01000204, 0x00000204, 0x01004000,
@ -201,7 +207,8 @@ __constant u32 c_SPtrans[8][64] =
0x00000204, 0x00004000, 0x01004204, 0x01000000, 0x00000204, 0x00004000, 0x01004204, 0x01000000,
0x01004200, 0x00000004, 0x00004004, 0x01004204, 0x01004200, 0x00000004, 0x00004004, 0x01004204,
0x01000004, 0x01004200, 0x01004000, 0x00004004, 0x01000004, 0x01004200, 0x01004000, 0x00004004,
/* nibble 7 */ },
{
0x20800080, 0x20820000, 0x00020080, 0x00000000, 0x20800080, 0x20820000, 0x00020080, 0x00000000,
0x20020000, 0x00800080, 0x20800000, 0x20820080, 0x20020000, 0x00800080, 0x20800000, 0x20820080,
0x00000080, 0x20000000, 0x00820000, 0x00020080, 0x00000080, 0x20000000, 0x00820000, 0x00020080,
@ -218,11 +225,12 @@ __constant u32 c_SPtrans[8][64] =
0x00820000, 0x00020080, 0x20020080, 0x20800000, 0x00820000, 0x00020080, 0x20020080, 0x20800000,
0x00000080, 0x20820000, 0x00820080, 0x00000000, 0x00000080, 0x20820000, 0x00820080, 0x00000000,
0x20000000, 0x20800080, 0x00020000, 0x00820080, 0x20000000, 0x20800080, 0x00020000, 0x00820080,
}
}; };
__constant u32 c_skb[8][64] = __constant u32 c_skb[8][64] =
{ {
/* for C bits (numbered as per FIPS 46) 1 2 3 4 5 6 */ {
0x00000000, 0x00000010, 0x20000000, 0x20000010, 0x00000000, 0x00000010, 0x20000000, 0x20000010,
0x00010000, 0x00010010, 0x20010000, 0x20010010, 0x00010000, 0x00010010, 0x20010000, 0x20010010,
0x00000800, 0x00000810, 0x20000800, 0x20000810, 0x00000800, 0x00000810, 0x20000800, 0x20000810,
@ -239,7 +247,8 @@ __constant u32 c_skb[8][64] =
0x00090020, 0x00090030, 0x20090020, 0x20090030, 0x00090020, 0x00090030, 0x20090020, 0x20090030,
0x00080820, 0x00080830, 0x20080820, 0x20080830, 0x00080820, 0x00080830, 0x20080820, 0x20080830,
0x00090820, 0x00090830, 0x20090820, 0x20090830, 0x00090820, 0x00090830, 0x20090820, 0x20090830,
/* for C bits (numbered as per FIPS 46) 7 8 10 11 12 13 */ },
{
0x00000000, 0x02000000, 0x00002000, 0x02002000, 0x00000000, 0x02000000, 0x00002000, 0x02002000,
0x00200000, 0x02200000, 0x00202000, 0x02202000, 0x00200000, 0x02200000, 0x00202000, 0x02202000,
0x00000004, 0x02000004, 0x00002004, 0x02002004, 0x00000004, 0x02000004, 0x00002004, 0x02002004,
@ -256,7 +265,8 @@ __constant u32 c_skb[8][64] =
0x10200400, 0x12200400, 0x10202400, 0x12202400, 0x10200400, 0x12200400, 0x10202400, 0x12202400,
0x10000404, 0x12000404, 0x10002404, 0x12002404, 0x10000404, 0x12000404, 0x10002404, 0x12002404,
0x10200404, 0x12200404, 0x10202404, 0x12202404, 0x10200404, 0x12200404, 0x10202404, 0x12202404,
/* for C bits (numbered as per FIPS 46) 14 15 16 17 19 20 */ },
{
0x00000000, 0x00000001, 0x00040000, 0x00040001, 0x00000000, 0x00000001, 0x00040000, 0x00040001,
0x01000000, 0x01000001, 0x01040000, 0x01040001, 0x01000000, 0x01000001, 0x01040000, 0x01040001,
0x00000002, 0x00000003, 0x00040002, 0x00040003, 0x00000002, 0x00000003, 0x00040002, 0x00040003,
@ -273,7 +283,8 @@ __constant u32 c_skb[8][64] =
0x09000200, 0x09000201, 0x09040200, 0x09040201, 0x09000200, 0x09000201, 0x09040200, 0x09040201,
0x08000202, 0x08000203, 0x08040202, 0x08040203, 0x08000202, 0x08000203, 0x08040202, 0x08040203,
0x09000202, 0x09000203, 0x09040202, 0x09040203, 0x09000202, 0x09000203, 0x09040202, 0x09040203,
/* for C bits (numbered as per FIPS 46) 21 23 24 26 27 28 */ },
{
0x00000000, 0x00100000, 0x00000100, 0x00100100, 0x00000000, 0x00100000, 0x00000100, 0x00100100,
0x00000008, 0x00100008, 0x00000108, 0x00100108, 0x00000008, 0x00100008, 0x00000108, 0x00100108,
0x00001000, 0x00101000, 0x00001100, 0x00101100, 0x00001000, 0x00101000, 0x00001100, 0x00101100,
@ -290,7 +301,8 @@ __constant u32 c_skb[8][64] =
0x04020008, 0x04120008, 0x04020108, 0x04120108, 0x04020008, 0x04120008, 0x04020108, 0x04120108,
0x04021000, 0x04121000, 0x04021100, 0x04121100, 0x04021000, 0x04121000, 0x04021100, 0x04121100,
0x04021008, 0x04121008, 0x04021108, 0x04121108, 0x04021008, 0x04121008, 0x04021108, 0x04121108,
/* for D bits (numbered as per FIPS 46) 1 2 3 4 5 6 */ },
{
0x00000000, 0x10000000, 0x00010000, 0x10010000, 0x00000000, 0x10000000, 0x00010000, 0x10010000,
0x00000004, 0x10000004, 0x00010004, 0x10010004, 0x00000004, 0x10000004, 0x00010004, 0x10010004,
0x20000000, 0x30000000, 0x20010000, 0x30010000, 0x20000000, 0x30000000, 0x20010000, 0x30010000,
@ -307,7 +319,8 @@ __constant u32 c_skb[8][64] =
0x00101004, 0x10101004, 0x00111004, 0x10111004, 0x00101004, 0x10101004, 0x00111004, 0x10111004,
0x20101000, 0x30101000, 0x20111000, 0x30111000, 0x20101000, 0x30101000, 0x20111000, 0x30111000,
0x20101004, 0x30101004, 0x20111004, 0x30111004, 0x20101004, 0x30101004, 0x20111004, 0x30111004,
/* for D bits (numbered as per FIPS 46) 8 9 11 12 13 14 */ },
{
0x00000000, 0x08000000, 0x00000008, 0x08000008, 0x00000000, 0x08000000, 0x00000008, 0x08000008,
0x00000400, 0x08000400, 0x00000408, 0x08000408, 0x00000400, 0x08000400, 0x00000408, 0x08000408,
0x00020000, 0x08020000, 0x00020008, 0x08020008, 0x00020000, 0x08020000, 0x00020008, 0x08020008,
@ -324,7 +337,8 @@ __constant u32 c_skb[8][64] =
0x02000401, 0x0A000401, 0x02000409, 0x0A000409, 0x02000401, 0x0A000401, 0x02000409, 0x0A000409,
0x02020001, 0x0A020001, 0x02020009, 0x0A020009, 0x02020001, 0x0A020001, 0x02020009, 0x0A020009,
0x02020401, 0x0A020401, 0x02020409, 0x0A020409, 0x02020401, 0x0A020401, 0x02020409, 0x0A020409,
/* for D bits (numbered as per FIPS 46) 16 17 18 19 20 21 */ },
{
0x00000000, 0x00000100, 0x00080000, 0x00080100, 0x00000000, 0x00000100, 0x00080000, 0x00080100,
0x01000000, 0x01000100, 0x01080000, 0x01080100, 0x01000000, 0x01000100, 0x01080000, 0x01080100,
0x00000010, 0x00000110, 0x00080010, 0x00080110, 0x00000010, 0x00000110, 0x00080010, 0x00080110,
@ -341,7 +355,8 @@ __constant u32 c_skb[8][64] =
0x01200200, 0x01200300, 0x01280200, 0x01280300, 0x01200200, 0x01200300, 0x01280200, 0x01280300,
0x00200210, 0x00200310, 0x00280210, 0x00280310, 0x00200210, 0x00200310, 0x00280210, 0x00280310,
0x01200210, 0x01200310, 0x01280210, 0x01280310, 0x01200210, 0x01200310, 0x01280210, 0x01280310,
/* for D bits (numbered as per FIPS 46) 22 23 24 25 27 28 */ },
{
0x00000000, 0x04000000, 0x00040000, 0x04040000, 0x00000000, 0x04000000, 0x00040000, 0x04040000,
0x00000002, 0x04000002, 0x00040002, 0x04040002, 0x00000002, 0x04000002, 0x00040002, 0x04040002,
0x00002000, 0x04002000, 0x00042000, 0x04042000, 0x00002000, 0x04002000, 0x00042000, 0x04042000,
@ -357,10 +372,11 @@ __constant u32 c_skb[8][64] =
0x00000820, 0x04000820, 0x00040820, 0x04040820, 0x00000820, 0x04000820, 0x00040820, 0x04040820,
0x00000822, 0x04000822, 0x00040822, 0x04040822, 0x00000822, 0x04000822, 0x00040822, 0x04040822,
0x00002820, 0x04002820, 0x00042820, 0x04042820, 0x00002820, 0x04002820, 0x00042820, 0x04042820,
0x00002822, 0x04002822, 0x00042822, 0x04042822, 0x00002822, 0x04002822, 0x00042822, 0x04042822
}
}; };
#define NBOX(i,n,S) (S)[(n)][(i)] #define BOX(i,n,S) (S)[(n)][(i)]
static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 s_SPtrans[8][64]) static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 s_SPtrans[8][64])
{ {
@ -370,23 +386,34 @@ static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16],
u32 l = data[1]; u32 l = data[1];
#pragma unroll 16 #pragma unroll 16
for (u32 i = 0; i < 16; i++) for (u32 i = 0; i < 16; i += 2)
{ {
u32 u = Kc[i] ^ r; u32 u;
u32 t = Kd[i] ^ rotl32 (r, 28u); u32 t;
l ^= NBOX (((u >> 2) & 0x3f), 0, s_SPtrans) u = Kc[i + 0] ^ r;
| NBOX (((u >> 10) & 0x3f), 2, s_SPtrans) t = Kd[i + 0] ^ rotl32 (r, 28u);
| NBOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| NBOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| NBOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| NBOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| NBOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| NBOX (((t >> 26) & 0x3f), 7, s_SPtrans);
tt = l; l ^= BOX (((u >> 2) & 0x3f), 0, s_SPtrans)
l = r; | BOX (((u >> 10) & 0x3f), 2, s_SPtrans)
r = tt; | BOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| BOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| BOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| BOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| BOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| BOX (((t >> 26) & 0x3f), 7, s_SPtrans);
u = Kc[i + 1] ^ l;
t = Kd[i + 1] ^ rotl32 (l, 28u);
r ^= BOX (((u >> 2) & 0x3f), 0, s_SPtrans)
| BOX (((u >> 10) & 0x3f), 2, s_SPtrans)
| BOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| BOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| BOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| BOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| BOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| BOX (((t >> 26) & 0x3f), 7, s_SPtrans);
} }
iv[0] = l; iv[0] = l;
@ -414,41 +441,49 @@ static void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u
#pragma unroll 16 #pragma unroll 16
for (u32 i = 0; i < 16; i++) for (u32 i = 0; i < 16; i++)
{ {
const u32 shifts3s0[16] = { 1, 1, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 1 }; if ((i < 2) || (i == 8) || (i == 15))
const u32 shifts3s1[16] = { 27, 27, 26, 26, 26, 26, 26, 26, 27, 26, 26, 26, 26, 26, 26, 27 }; {
c = ((c >> 1) | (c << 27));
c = c >> shifts3s0[i] | c << shifts3s1[i]; d = ((d >> 1) | (d << 27));
d = d >> shifts3s0[i] | d << shifts3s1[i]; }
else
{
c = ((c >> 2) | (c << 26));
d = ((d >> 2) | (d << 26));
}
c = c & 0x0fffffff; c = c & 0x0fffffff;
d = d & 0x0fffffff; d = d & 0x0fffffff;
u32 s; const u32 c00 = (c >> 0) & 0x0000003f;
u32 t; const u32 c06 = (c >> 6) & 0x00383003;
const u32 c07 = (c >> 7) & 0x0000003c;
const u32 c13 = (c >> 13) & 0x0000060f;
const u32 c20 = (c >> 20) & 0x00000001;
s = NBOX ((( c >> 0) & 0x3f), 0, s_skb) u32 s = BOX (((c00 >> 0) & 0xff), 0, s_skb)
| NBOX ((((c >> 6) & 0x03) | BOX (((c06 >> 0) & 0xff)
| ((c >> 7) & 0x3c)), 1, s_skb) |((c07 >> 0) & 0xff), 1, s_skb)
| NBOX ((((c >> 13) & 0x0f) | BOX (((c13 >> 0) & 0xff)
| ((c >> 14) & 0x30)), 2, s_skb) |((c06 >> 8) & 0xff), 2, s_skb)
| NBOX ((((c >> 20) & 0x01) | BOX (((c20 >> 0) & 0xff)
| ((c >> 21) & 0x06) |((c13 >> 8) & 0xff)
| ((c >> 22) & 0x38)), 3, s_skb); |((c06 >> 16) & 0xff), 3, s_skb);
t = NBOX ((( d >> 0) & 0x3f), 4, s_skb) const u32 d00 = (d >> 0) & 0x00003c3f;
| NBOX ((((d >> 7) & 0x03) const u32 d07 = (d >> 7) & 0x00003f03;
| ((d >> 8) & 0x3c)), 5, s_skb) const u32 d21 = (d >> 21) & 0x0000000f;
| NBOX ((((d >> 15) & 0x3f)), 6, s_skb) const u32 d22 = (d >> 22) & 0x00000030;
| NBOX ((((d >> 21) & 0x0f)
| ((d >> 22) & 0x30)), 7, s_skb); u32 t = BOX (((d00 >> 0) & 0xff), 4, s_skb)
| BOX (((d07 >> 0) & 0xff)
|((d00 >> 8) & 0xff), 5, s_skb)
| BOX (((d07 >> 8) & 0xff), 6, s_skb)
| BOX (((d21 >> 0) & 0xff)
|((d22 >> 0) & 0xff), 7, s_skb);
#if defined cl_amd_media_ops
Kc[i] = amd_bytealign (t, s << 16, 2);
Kd[i] = amd_bytealign (t >> 16, s, 2);
#else
Kc[i] = ((t << 16) | (s & 0x0000ffff)); Kc[i] = ((t << 16) | (s & 0x0000ffff));
Kd[i] = ((s >> 16) | (t & 0xffff0000)); Kd[i] = ((s >> 16) | (t & 0xffff0000));
#endif
Kc[i] = rotl32 (Kc[i], 2u); Kc[i] = rotl32 (Kc[i], 2u);
Kd[i] = rotl32 (Kd[i], 2u); Kd[i] = rotl32 (Kd[i], 2u);

151
OpenCL/m08500_a3.cl
View File

@ -82,7 +82,7 @@ __constant u8 ascii_to_ebcdic_pc[256] =
__constant u32 c_SPtrans[8][64] = __constant u32 c_SPtrans[8][64] =
{ {
/* nibble 0 */ {
0x02080800, 0x00080000, 0x02000002, 0x02080802, 0x02080800, 0x00080000, 0x02000002, 0x02080802,
0x02000000, 0x00080802, 0x00080002, 0x02000002, 0x02000000, 0x00080802, 0x00080002, 0x02000002,
0x00080802, 0x02080800, 0x02080000, 0x00000802, 0x00080802, 0x02080800, 0x02080000, 0x00000802,
@ -99,7 +99,8 @@ __constant u32 c_SPtrans[8][64] =
0x02000000, 0x00000802, 0x00080002, 0x00000000, 0x02000000, 0x00000802, 0x00080002, 0x00000000,
0x00080000, 0x02000000, 0x02000802, 0x02080800, 0x00080000, 0x02000000, 0x02000802, 0x02080800,
0x00000002, 0x02080002, 0x00000800, 0x00080802, 0x00000002, 0x02080002, 0x00000800, 0x00080802,
/* nibble 1 */ },
{
0x40108010, 0x00000000, 0x00108000, 0x40100000, 0x40108010, 0x00000000, 0x00108000, 0x40100000,
0x40000010, 0x00008010, 0x40008000, 0x00108000, 0x40000010, 0x00008010, 0x40008000, 0x00108000,
0x00008000, 0x40100010, 0x00000010, 0x40008000, 0x00008000, 0x40100010, 0x00000010, 0x40008000,
@ -116,7 +117,8 @@ __constant u32 c_SPtrans[8][64] =
0x00000010, 0x40108010, 0x00108000, 0x40100000, 0x00000010, 0x40108010, 0x00108000, 0x40100000,
0x40100010, 0x00100000, 0x00008010, 0x40008000, 0x40100010, 0x00100000, 0x00008010, 0x40008000,
0x40008010, 0x00000010, 0x40100000, 0x00108000, 0x40008010, 0x00000010, 0x40100000, 0x00108000,
/* nibble 2 */ },
{
0x04000001, 0x04040100, 0x00000100, 0x04000101, 0x04000001, 0x04040100, 0x00000100, 0x04000101,
0x00040001, 0x04000000, 0x04000101, 0x00040100, 0x00040001, 0x04000000, 0x04000101, 0x00040100,
0x04000100, 0x00040000, 0x04040000, 0x00000001, 0x04000100, 0x00040000, 0x04040000, 0x00000001,
@ -133,7 +135,8 @@ __constant u32 c_SPtrans[8][64] =
0x00040000, 0x04000100, 0x04000101, 0x00040100, 0x00040000, 0x04000100, 0x04000101, 0x00040100,
0x04000100, 0x00000000, 0x04040001, 0x00000101, 0x04000100, 0x00000000, 0x04040001, 0x00000101,
0x04000001, 0x00040101, 0x00000100, 0x04040000, 0x04000001, 0x00040101, 0x00000100, 0x04040000,
/* nibble 3 */ },
{
0x00401008, 0x10001000, 0x00000008, 0x10401008, 0x00401008, 0x10001000, 0x00000008, 0x10401008,
0x00000000, 0x10400000, 0x10001008, 0x00400008, 0x00000000, 0x10400000, 0x10001008, 0x00400008,
0x10401000, 0x10000008, 0x10000000, 0x00001008, 0x10401000, 0x10000008, 0x10000000, 0x00001008,
@ -150,7 +153,8 @@ __constant u32 c_SPtrans[8][64] =
0x10401008, 0x00000008, 0x10001000, 0x00401008, 0x10401008, 0x00000008, 0x10001000, 0x00401008,
0x00400008, 0x00401000, 0x10400000, 0x10001008, 0x00400008, 0x00401000, 0x10400000, 0x10001008,
0x00001008, 0x10000000, 0x10000008, 0x10401000, 0x00001008, 0x10000000, 0x10000008, 0x10401000,
/* nibble 4 */ },
{
0x08000000, 0x00010000, 0x00000400, 0x08010420, 0x08000000, 0x00010000, 0x00000400, 0x08010420,
0x08010020, 0x08000400, 0x00010420, 0x08010000, 0x08010020, 0x08000400, 0x00010420, 0x08010000,
0x00010000, 0x00000020, 0x08000020, 0x00010400, 0x00010000, 0x00000020, 0x08000020, 0x00010400,
@ -167,7 +171,8 @@ __constant u32 c_SPtrans[8][64] =
0x08010020, 0x08010400, 0x00000420, 0x00010000, 0x08010020, 0x08010400, 0x00000420, 0x00010000,
0x00010400, 0x08010020, 0x08000400, 0x00000420, 0x00010400, 0x08010020, 0x08000400, 0x00000420,
0x00000020, 0x00010420, 0x08010000, 0x08000020, 0x00000020, 0x00010420, 0x08010000, 0x08000020,
/* nibble 5 */ },
{
0x80000040, 0x00200040, 0x00000000, 0x80202000, 0x80000040, 0x00200040, 0x00000000, 0x80202000,
0x00200040, 0x00002000, 0x80002040, 0x00200000, 0x00200040, 0x00002000, 0x80002040, 0x00200000,
0x00002040, 0x80202040, 0x00202000, 0x80000000, 0x00002040, 0x80202040, 0x00202000, 0x80000000,
@ -184,7 +189,8 @@ __constant u32 c_SPtrans[8][64] =
0x80000040, 0x80200000, 0x00202040, 0x00000000, 0x80000040, 0x80200000, 0x00202040, 0x00000000,
0x00002000, 0x80000040, 0x80002040, 0x80202000, 0x00002000, 0x80000040, 0x80002040, 0x80202000,
0x80200000, 0x00002040, 0x00000040, 0x80200040, 0x80200000, 0x00002040, 0x00000040, 0x80200040,
/* nibble 6 */ },
{
0x00004000, 0x00000200, 0x01000200, 0x01000004, 0x00004000, 0x00000200, 0x01000200, 0x01000004,
0x01004204, 0x00004004, 0x00004200, 0x00000000, 0x01004204, 0x00004004, 0x00004200, 0x00000000,
0x01000000, 0x01000204, 0x00000204, 0x01004000, 0x01000000, 0x01000204, 0x00000204, 0x01004000,
@ -201,7 +207,8 @@ __constant u32 c_SPtrans[8][64] =
0x00000204, 0x00004000, 0x01004204, 0x01000000, 0x00000204, 0x00004000, 0x01004204, 0x01000000,
0x01004200, 0x00000004, 0x00004004, 0x01004204, 0x01004200, 0x00000004, 0x00004004, 0x01004204,
0x01000004, 0x01004200, 0x01004000, 0x00004004, 0x01000004, 0x01004200, 0x01004000, 0x00004004,
/* nibble 7 */ },
{
0x20800080, 0x20820000, 0x00020080, 0x00000000, 0x20800080, 0x20820000, 0x00020080, 0x00000000,
0x20020000, 0x00800080, 0x20800000, 0x20820080, 0x20020000, 0x00800080, 0x20800000, 0x20820080,
0x00000080, 0x20000000, 0x00820000, 0x00020080, 0x00000080, 0x20000000, 0x00820000, 0x00020080,
@ -218,11 +225,12 @@ __constant u32 c_SPtrans[8][64] =
0x00820000, 0x00020080, 0x20020080, 0x20800000, 0x00820000, 0x00020080, 0x20020080, 0x20800000,
0x00000080, 0x20820000, 0x00820080, 0x00000000, 0x00000080, 0x20820000, 0x00820080, 0x00000000,
0x20000000, 0x20800080, 0x00020000, 0x00820080, 0x20000000, 0x20800080, 0x00020000, 0x00820080,
}
}; };
__constant u32 c_skb[8][64] = __constant u32 c_skb[8][64] =
{ {
/* for C bits (numbered as per FIPS 46) 1 2 3 4 5 6 */ {
0x00000000, 0x00000010, 0x20000000, 0x20000010, 0x00000000, 0x00000010, 0x20000000, 0x20000010,
0x00010000, 0x00010010, 0x20010000, 0x20010010, 0x00010000, 0x00010010, 0x20010000, 0x20010010,
0x00000800, 0x00000810, 0x20000800, 0x20000810, 0x00000800, 0x00000810, 0x20000800, 0x20000810,
@ -239,7 +247,8 @@ __constant u32 c_skb[8][64] =
0x00090020, 0x00090030, 0x20090020, 0x20090030, 0x00090020, 0x00090030, 0x20090020, 0x20090030,
0x00080820, 0x00080830, 0x20080820, 0x20080830, 0x00080820, 0x00080830, 0x20080820, 0x20080830,
0x00090820, 0x00090830, 0x20090820, 0x20090830, 0x00090820, 0x00090830, 0x20090820, 0x20090830,
/* for C bits (numbered as per FIPS 46) 7 8 10 11 12 13 */ },
{
0x00000000, 0x02000000, 0x00002000, 0x02002000, 0x00000000, 0x02000000, 0x00002000, 0x02002000,
0x00200000, 0x02200000, 0x00202000, 0x02202000, 0x00200000, 0x02200000, 0x00202000, 0x02202000,
0x00000004, 0x02000004, 0x00002004, 0x02002004, 0x00000004, 0x02000004, 0x00002004, 0x02002004,
@ -256,7 +265,8 @@ __constant u32 c_skb[8][64] =
0x10200400, 0x12200400, 0x10202400, 0x12202400, 0x10200400, 0x12200400, 0x10202400, 0x12202400,
0x10000404, 0x12000404, 0x10002404, 0x12002404, 0x10000404, 0x12000404, 0x10002404, 0x12002404,
0x10200404, 0x12200404, 0x10202404, 0x12202404, 0x10200404, 0x12200404, 0x10202404, 0x12202404,
/* for C bits (numbered as per FIPS 46) 14 15 16 17 19 20 */ },
{
0x00000000, 0x00000001, 0x00040000, 0x00040001, 0x00000000, 0x00000001, 0x00040000, 0x00040001,
0x01000000, 0x01000001, 0x01040000, 0x01040001, 0x01000000, 0x01000001, 0x01040000, 0x01040001,
0x00000002, 0x00000003, 0x00040002, 0x00040003, 0x00000002, 0x00000003, 0x00040002, 0x00040003,
@ -273,7 +283,8 @@ __constant u32 c_skb[8][64] =
0x09000200, 0x09000201, 0x09040200, 0x09040201, 0x09000200, 0x09000201, 0x09040200, 0x09040201,
0x08000202, 0x08000203, 0x08040202, 0x08040203, 0x08000202, 0x08000203, 0x08040202, 0x08040203,
0x09000202, 0x09000203, 0x09040202, 0x09040203, 0x09000202, 0x09000203, 0x09040202, 0x09040203,
/* for C bits (numbered as per FIPS 46) 21 23 24 26 27 28 */ },
{
0x00000000, 0x00100000, 0x00000100, 0x00100100, 0x00000000, 0x00100000, 0x00000100, 0x00100100,
0x00000008, 0x00100008, 0x00000108, 0x00100108, 0x00000008, 0x00100008, 0x00000108, 0x00100108,
0x00001000, 0x00101000, 0x00001100, 0x00101100, 0x00001000, 0x00101000, 0x00001100, 0x00101100,
@ -290,7 +301,8 @@ __constant u32 c_skb[8][64] =
0x04020008, 0x04120008, 0x04020108, 0x04120108, 0x04020008, 0x04120008, 0x04020108, 0x04120108,
0x04021000, 0x04121000, 0x04021100, 0x04121100, 0x04021000, 0x04121000, 0x04021100, 0x04121100,
0x04021008, 0x04121008, 0x04021108, 0x04121108, 0x04021008, 0x04121008, 0x04021108, 0x04121108,
/* for D bits (numbered as per FIPS 46) 1 2 3 4 5 6 */ },
{
0x00000000, 0x10000000, 0x00010000, 0x10010000, 0x00000000, 0x10000000, 0x00010000, 0x10010000,
0x00000004, 0x10000004, 0x00010004, 0x10010004, 0x00000004, 0x10000004, 0x00010004, 0x10010004,
0x20000000, 0x30000000, 0x20010000, 0x30010000, 0x20000000, 0x30000000, 0x20010000, 0x30010000,
@ -307,7 +319,8 @@ __constant u32 c_skb[8][64] =
0x00101004, 0x10101004, 0x00111004, 0x10111004, 0x00101004, 0x10101004, 0x00111004, 0x10111004,
0x20101000, 0x30101000, 0x20111000, 0x30111000, 0x20101000, 0x30101000, 0x20111000, 0x30111000,
0x20101004, 0x30101004, 0x20111004, 0x30111004, 0x20101004, 0x30101004, 0x20111004, 0x30111004,
/* for D bits (numbered as per FIPS 46) 8 9 11 12 13 14 */ },
{
0x00000000, 0x08000000, 0x00000008, 0x08000008, 0x00000000, 0x08000000, 0x00000008, 0x08000008,
0x00000400, 0x08000400, 0x00000408, 0x08000408, 0x00000400, 0x08000400, 0x00000408, 0x08000408,
0x00020000, 0x08020000, 0x00020008, 0x08020008, 0x00020000, 0x08020000, 0x00020008, 0x08020008,
@ -324,7 +337,8 @@ __constant u32 c_skb[8][64] =
0x02000401, 0x0A000401, 0x02000409, 0x0A000409, 0x02000401, 0x0A000401, 0x02000409, 0x0A000409,
0x02020001, 0x0A020001, 0x02020009, 0x0A020009, 0x02020001, 0x0A020001, 0x02020009, 0x0A020009,
0x02020401, 0x0A020401, 0x02020409, 0x0A020409, 0x02020401, 0x0A020401, 0x02020409, 0x0A020409,
/* for D bits (numbered as per FIPS 46) 16 17 18 19 20 21 */ },
{
0x00000000, 0x00000100, 0x00080000, 0x00080100, 0x00000000, 0x00000100, 0x00080000, 0x00080100,
0x01000000, 0x01000100, 0x01080000, 0x01080100, 0x01000000, 0x01000100, 0x01080000, 0x01080100,
0x00000010, 0x00000110, 0x00080010, 0x00080110, 0x00000010, 0x00000110, 0x00080010, 0x00080110,
@ -341,7 +355,8 @@ __constant u32 c_skb[8][64] =
0x01200200, 0x01200300, 0x01280200, 0x01280300, 0x01200200, 0x01200300, 0x01280200, 0x01280300,
0x00200210, 0x00200310, 0x00280210, 0x00280310, 0x00200210, 0x00200310, 0x00280210, 0x00280310,
0x01200210, 0x01200310, 0x01280210, 0x01280310, 0x01200210, 0x01200310, 0x01280210, 0x01280310,
/* for D bits (numbered as per FIPS 46) 22 23 24 25 27 28 */ },
{
0x00000000, 0x04000000, 0x00040000, 0x04040000, 0x00000000, 0x04000000, 0x00040000, 0x04040000,
0x00000002, 0x04000002, 0x00040002, 0x04040002, 0x00000002, 0x04000002, 0x00040002, 0x04040002,
0x00002000, 0x04002000, 0x00042000, 0x04042000, 0x00002000, 0x04002000, 0x00042000, 0x04042000,
@ -357,10 +372,11 @@ __constant u32 c_skb[8][64] =
0x00000820, 0x04000820, 0x00040820, 0x04040820, 0x00000820, 0x04000820, 0x00040820, 0x04040820,
0x00000822, 0x04000822, 0x00040822, 0x04040822, 0x00000822, 0x04000822, 0x00040822, 0x04040822,
0x00002820, 0x04002820, 0x00042820, 0x04042820, 0x00002820, 0x04002820, 0x00042820, 0x04042820,
0x00002822, 0x04002822, 0x00042822, 0x04042822, 0x00002822, 0x04002822, 0x00042822, 0x04042822
}
}; };
#define NBOX(i,n,S) (S)[(n)][(i)] #define BOX(i,n,S) (S)[(n)][(i)]
static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 s_SPtrans[8][64]) static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16], __local u32 s_SPtrans[8][64])
{ {
@ -370,23 +386,34 @@ static void _des_crypt_encrypt (u32 iv[2], u32 data[2], u32 Kc[16], u32 Kd[16],
u32 l = data[1]; u32 l = data[1];
#pragma unroll 16 #pragma unroll 16
for (u32 i = 0; i < 16; i++) for (u32 i = 0; i < 16; i += 2)
{ {
u32 u = Kc[i] ^ r; u32 u;
u32 t = Kd[i] ^ rotl32 (r, 28u); u32 t;
l ^= NBOX (((u >> 2) & 0x3f), 0, s_SPtrans) u = Kc[i + 0] ^ r;
| NBOX (((u >> 10) & 0x3f), 2, s_SPtrans) t = Kd[i + 0] ^ rotl32 (r, 28u);
| NBOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| NBOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| NBOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| NBOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| NBOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| NBOX (((t >> 26) & 0x3f), 7, s_SPtrans);
tt = l; l ^= BOX (((u >> 2) & 0x3f), 0, s_SPtrans)
l = r; | BOX (((u >> 10) & 0x3f), 2, s_SPtrans)
r = tt; | BOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| BOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| BOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| BOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| BOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| BOX (((t >> 26) & 0x3f), 7, s_SPtrans);
u = Kc[i + 1] ^ l;
t = Kd[i + 1] ^ rotl32 (l, 28u);
r ^= BOX (((u >> 2) & 0x3f), 0, s_SPtrans)
| BOX (((u >> 10) & 0x3f), 2, s_SPtrans)
| BOX (((u >> 18) & 0x3f), 4, s_SPtrans)
| BOX (((u >> 26) & 0x3f), 6, s_SPtrans)
| BOX (((t >> 2) & 0x3f), 1, s_SPtrans)
| BOX (((t >> 10) & 0x3f), 3, s_SPtrans)
| BOX (((t >> 18) & 0x3f), 5, s_SPtrans)
| BOX (((t >> 26) & 0x3f), 7, s_SPtrans);
} }
iv[0] = l; iv[0] = l;
@ -414,41 +441,49 @@ static void _des_crypt_keysetup (u32 c, u32 d, u32 Kc[16], u32 Kd[16], __local u
#pragma unroll 16 #pragma unroll 16
for (u32 i = 0; i < 16; i++) for (u32 i = 0; i < 16; i++)
{ {
const u32 shifts3s0[16] = { 1, 1, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 1 }; if ((i < 2) || (i == 8) || (i == 15))
const u32 shifts3s1[16] = { 27, 27, 26, 26, 26, 26, 26, 26, 27, 26, 26, 26, 26, 26, 26, 27 }; {
c = ((c >> 1) | (c << 27));
c = c >> shifts3s0[i] | c << shifts3s1[i]; d = ((d >> 1) | (d << 27));
d = d >> shifts3s0[i] | d << shifts3s1[i]; }
else
{
c = ((c >> 2) | (c << 26));
d = ((d >> 2) | (d << 26));
}
c = c & 0x0fffffff; c = c & 0x0fffffff;
d = d & 0x0fffffff; d = d & 0x0fffffff;
u32 s; const u32 c00 = (c >> 0) & 0x0000003f;
u32 t; const u32 c06 = (c >> 6) & 0x00383003;
const u32 c07 = (c >> 7) & 0x0000003c;
const u32 c13 = (c >> 13) & 0x0000060f;
const u32 c20 = (c >> 20) & 0x00000001;
s = NBOX ((( c >> 0) & 0x3f), 0, s_skb) u32 s = BOX (((c00 >> 0) & 0xff), 0, s_skb)
| NBOX ((((c >> 6) & 0x03) | BOX (((c06 >> 0) & 0xff)
| ((c >> 7) & 0x3c)), 1, s_skb) |((c07 >> 0) & 0xff), 1, s_skb)
| NBOX ((((c >> 13) & 0x0f) | BOX (((c13 >> 0) & 0xff)
| ((c >> 14) & 0x30)), 2, s_skb) |((c06 >> 8) & 0xff), 2, s_skb)
| NBOX ((((c >> 20) & 0x01) | BOX (((c20 >> 0) & 0xff)
| ((c >> 21) & 0x06) |((c13 >> 8) & 0xff)
| ((c >> 22) & 0x38)), 3, s_skb); |((c06 >> 16) & 0xff), 3, s_skb);
t = NBOX ((( d >> 0) & 0x3f), 4, s_skb) const u32 d00 = (d >> 0) & 0x00003c3f;
| NBOX ((((d >> 7) & 0x03) const u32 d07 = (d >> 7) & 0x00003f03;
| ((d >> 8) & 0x3c)), 5, s_skb) const u32 d21 = (d >> 21) & 0x0000000f;
| NBOX ((((d >> 15) & 0x3f)), 6, s_skb) const u32 d22 = (d >> 22) & 0x00000030;
| NBOX ((((d >> 21) & 0x0f)
| ((d >> 22) & 0x30)), 7, s_skb); u32 t = BOX (((d00 >> 0) & 0xff), 4, s_skb)
| BOX (((d07 >> 0) & 0xff)
|((d00 >> 8) & 0xff), 5, s_skb)
| BOX (((d07 >> 8) & 0xff), 6, s_skb)
| BOX (((d21 >> 0) & 0xff)
|((d22 >> 0) & 0xff), 7, s_skb);
#if defined cl_amd_media_ops
Kc[i] = amd_bytealign (t, s << 16, 2);
Kd[i] = amd_bytealign (t >> 16, s, 2);
#else
Kc[i] = ((t << 16) | (s & 0x0000ffff)); Kc[i] = ((t << 16) | (s & 0x0000ffff));
Kd[i] = ((s >> 16) | (t & 0xffff0000)); Kd[i] = ((s >> 16) | (t & 0xffff0000));
#endif
Kc[i] = rotl32 (Kc[i], 2u); Kc[i] = rotl32 (Kc[i], 2u);
Kd[i] = rotl32 (Kd[i], 2u); Kd[i] = rotl32 (Kd[i], 2u);