arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-11-14 03:39:09 +00:00
Code Issues Releases Wiki Activity

Optimize access to s_lotus_magic_table[] in -m 8700

Browse Source
This commit is contained in:
Jens Steube 2020-02-22 20:40:47 +01:00
parent caa34924bf
commit 48fd7d039f
3 changed files with 167 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

66
OpenCL/m08700_a0-optimized.cl
View File

@ -16,7 +16,9 @@
#include "inc_simd.cl"
#endif
CONSTANT_VK u32a lotus_magic_table[256] =
// we just double this buffer so we can safe the & 0xff ;)
CONSTANT_VK u32a lotus_magic_table[512] =
{
0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
@ -50,6 +52,39 @@ CONSTANT_VK u32a lotus_magic_table[256] =
0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
0x30, 0x04, 0xb6, 0xdc, 0x7d, 0xdf, 0x32, 0x4b,
0xf7, 0xcb, 0x45, 0x9b, 0x31, 0xbb, 0x21, 0x5a,
0x41, 0x9f, 0xe1, 0xd9, 0x4a, 0x4d, 0x9e, 0xda,
0xa0, 0x68, 0x2c, 0xc3, 0x27, 0x5f, 0x80, 0x36,
0x3e, 0xee, 0xfb, 0x95, 0x1a, 0xfe, 0xce, 0xa8,
0x34, 0xa9, 0x13, 0xf0, 0xa6, 0x3f, 0xd8, 0x0c,
0x78, 0x24, 0xaf, 0x23, 0x52, 0xc1, 0x67, 0x17,
0xf5, 0x66, 0x90, 0xe7, 0xe8, 0x07, 0xb8, 0x60,
0x48, 0xe6, 0x1e, 0x53, 0xf3, 0x92, 0xa4, 0x72,
0x8c, 0x08, 0x15, 0x6e, 0x86, 0x00, 0x84, 0xfa,
0xf4, 0x7f, 0x8a, 0x42, 0x19, 0xf6, 0xdb, 0xcd,
0x14, 0x8d, 0x50, 0x12, 0xba, 0x3c, 0x06, 0x4e,
0xec, 0xb3, 0x35, 0x11, 0xa1, 0x88, 0x8e, 0x2b,
0x94, 0x99, 0xb7, 0x71, 0x74, 0xd3, 0xe4, 0xbf,
0x3a, 0xde, 0x96, 0x0e, 0xbc, 0x0a, 0xed, 0x77,
0xfc, 0x37, 0x6b, 0x03, 0x79, 0x89, 0x62, 0xc6,
0xd7, 0xc0, 0xd2, 0x7c, 0x6a, 0x8b, 0x22, 0xa3,
0x5b, 0x05, 0x5d, 0x02, 0x75, 0xd5, 0x61, 0xe3,
0x18, 0x8f, 0x55, 0x51, 0xad, 0x1f, 0x0b, 0x5e,
0x85, 0xe5, 0xc2, 0x57, 0x63, 0xca, 0x3d, 0x6c,
0xb4, 0xc5, 0xcc, 0x70, 0xb2, 0x91, 0x59, 0x0d,
0x47, 0x20, 0xc8, 0x4f, 0x58, 0xe0, 0x01, 0xe2,
0x16, 0x38, 0xc4, 0x6f, 0x3b, 0x0f, 0x65, 0x46,
0xbe, 0x7e, 0x2d, 0x7b, 0x82, 0xf9, 0x40, 0xb5,
0x1d, 0x73, 0xf8, 0xeb, 0x26, 0xc7, 0x87, 0x97,
0x25, 0x54, 0xb1, 0x28, 0xaa, 0x98, 0x9d, 0xa5,
0x64, 0x6d, 0x7a, 0xd4, 0x10, 0x81, 0x44, 0xef,
0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
};
#if VECT_SIZE == 1
@ -89,10 +124,10 @@ DECLSPEC void lotus_mix (u32x *in, LOCAL_AS u32 *s_lotus_magic_table)
u32x tmp_in = in[j];
u32x tmp_out = 0;
p = (p + s--) & 0xff; p = ((tmp_in >> 0) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 0;
p = (p + s--) & 0xff; p = ((tmp_in >> 8) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 8;
p = (p + s--) & 0xff; p = ((tmp_in >> 16) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 16;
p = (p + s--) & 0xff; p = ((tmp_in >> 24) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 24;
p = ((tmp_in >> 0) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 0;
p = ((tmp_in >> 8) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 8;
p = ((tmp_in >> 16) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 16;
p = ((tmp_in >> 24) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 24;
in[j] = tmp_out;
}
@ -110,10 +145,15 @@ DECLSPEC void lotus_transform_password (u32x *in, u32x *out, LOCAL_AS u32 *s_lot
#endif
for (int i = 0; i < 4; i++)
{
t ^= (in[i] >> 0) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 0; t = ((out[i] >> 0) & 0xff);
t ^= (in[i] >> 8) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 8; t = ((out[i] >> 8) & 0xff);
t ^= (in[i] >> 16) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 16; t = ((out[i] >> 16) & 0xff);
t ^= (in[i] >> 24) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 24; t = ((out[i] >> 24) & 0xff);
u32x tmp_in = in[i];
u32x tmp_out = 0;
t ^= (tmp_in >> 0) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 0; t = (tmp_out >> 0) & 0xff;
t ^= (tmp_in >> 8) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 8; t = (tmp_out >> 8) & 0xff;
t ^= (tmp_in >> 16) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 16; t = (tmp_out >> 16) & 0xff;
t ^= (tmp_in >> 24) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 24; t = (tmp_out >> 24) & 0xff;
out[i] = tmp_out;
}
}
@ -281,9 +321,9 @@ KERNEL_FQ void m08700_m04 (KERN_ATTR_RULES ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}
@ -468,9 +508,9 @@ KERNEL_FQ void m08700_s04 (KERN_ATTR_RULES ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}

66
OpenCL/m08700_a1-optimized.cl
View File

@ -14,7 +14,9 @@
#include "inc_simd.cl"
#endif
CONSTANT_VK u32a lotus_magic_table[256] =
// we just double this buffer so we can safe the & 0xff ;)
CONSTANT_VK u32a lotus_magic_table[512] =
{
0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
@ -48,6 +50,39 @@ CONSTANT_VK u32a lotus_magic_table[256] =
0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
0x30, 0x04, 0xb6, 0xdc, 0x7d, 0xdf, 0x32, 0x4b,
0xf7, 0xcb, 0x45, 0x9b, 0x31, 0xbb, 0x21, 0x5a,
0x41, 0x9f, 0xe1, 0xd9, 0x4a, 0x4d, 0x9e, 0xda,
0xa0, 0x68, 0x2c, 0xc3, 0x27, 0x5f, 0x80, 0x36,
0x3e, 0xee, 0xfb, 0x95, 0x1a, 0xfe, 0xce, 0xa8,
0x34, 0xa9, 0x13, 0xf0, 0xa6, 0x3f, 0xd8, 0x0c,
0x78, 0x24, 0xaf, 0x23, 0x52, 0xc1, 0x67, 0x17,
0xf5, 0x66, 0x90, 0xe7, 0xe8, 0x07, 0xb8, 0x60,
0x48, 0xe6, 0x1e, 0x53, 0xf3, 0x92, 0xa4, 0x72,
0x8c, 0x08, 0x15, 0x6e, 0x86, 0x00, 0x84, 0xfa,
0xf4, 0x7f, 0x8a, 0x42, 0x19, 0xf6, 0xdb, 0xcd,
0x14, 0x8d, 0x50, 0x12, 0xba, 0x3c, 0x06, 0x4e,
0xec, 0xb3, 0x35, 0x11, 0xa1, 0x88, 0x8e, 0x2b,
0x94, 0x99, 0xb7, 0x71, 0x74, 0xd3, 0xe4, 0xbf,
0x3a, 0xde, 0x96, 0x0e, 0xbc, 0x0a, 0xed, 0x77,
0xfc, 0x37, 0x6b, 0x03, 0x79, 0x89, 0x62, 0xc6,
0xd7, 0xc0, 0xd2, 0x7c, 0x6a, 0x8b, 0x22, 0xa3,
0x5b, 0x05, 0x5d, 0x02, 0x75, 0xd5, 0x61, 0xe3,
0x18, 0x8f, 0x55, 0x51, 0xad, 0x1f, 0x0b, 0x5e,
0x85, 0xe5, 0xc2, 0x57, 0x63, 0xca, 0x3d, 0x6c,
0xb4, 0xc5, 0xcc, 0x70, 0xb2, 0x91, 0x59, 0x0d,
0x47, 0x20, 0xc8, 0x4f, 0x58, 0xe0, 0x01, 0xe2,
0x16, 0x38, 0xc4, 0x6f, 0x3b, 0x0f, 0x65, 0x46,
0xbe, 0x7e, 0x2d, 0x7b, 0x82, 0xf9, 0x40, 0xb5,
0x1d, 0x73, 0xf8, 0xeb, 0x26, 0xc7, 0x87, 0x97,
0x25, 0x54, 0xb1, 0x28, 0xaa, 0x98, 0x9d, 0xa5,
0x64, 0x6d, 0x7a, 0xd4, 0x10, 0x81, 0x44, 0xef,
0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
};
#if VECT_SIZE == 1
@ -87,10 +122,10 @@ DECLSPEC void lotus_mix (u32x *in, LOCAL_AS u32 *s_lotus_magic_table)
u32x tmp_in = in[j];
u32x tmp_out = 0;
p = (p + s--) & 0xff; p = ((tmp_in >> 0) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 0;
p = (p + s--) & 0xff; p = ((tmp_in >> 8) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 8;
p = (p + s--) & 0xff; p = ((tmp_in >> 16) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 16;
p = (p + s--) & 0xff; p = ((tmp_in >> 24) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 24;
p = ((tmp_in >> 0) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 0;
p = ((tmp_in >> 8) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 8;
p = ((tmp_in >> 16) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 16;
p = ((tmp_in >> 24) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 24;
in[j] = tmp_out;
}
@ -108,10 +143,15 @@ DECLSPEC void lotus_transform_password (u32x *in, u32x *out, LOCAL_AS u32 *s_lot
#endif
for (int i = 0; i < 4; i++)
{
t ^= (in[i] >> 0) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 0; t = ((out[i] >> 0) & 0xff);
t ^= (in[i] >> 8) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 8; t = ((out[i] >> 8) & 0xff);
t ^= (in[i] >> 16) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 16; t = ((out[i] >> 16) & 0xff);
t ^= (in[i] >> 24) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 24; t = ((out[i] >> 24) & 0xff);
u32x tmp_in = in[i];
u32x tmp_out = 0;
t ^= (tmp_in >> 0) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 0; t = (tmp_out >> 0) & 0xff;
t ^= (tmp_in >> 8) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 8; t = (tmp_out >> 8) & 0xff;
t ^= (tmp_in >> 16) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 16; t = (tmp_out >> 16) & 0xff;
t ^= (tmp_in >> 24) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 24; t = (tmp_out >> 24) & 0xff;
out[i] = tmp_out;
}
}
@ -279,9 +319,9 @@ KERNEL_FQ void m08700_m04 (KERN_ATTR_BASIC ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}
@ -526,9 +566,9 @@ KERNEL_FQ void m08700_s04 (KERN_ATTR_BASIC ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}

82
OpenCL/m08700_a3-optimized.cl
View File

@ -13,7 +13,9 @@
#include "inc_simd.cl"
#endif
CONSTANT_VK u32a lotus_magic_table[256] =
// we just double this buffer so we can safe the & 0xff ;)
CONSTANT_VK u32a lotus_magic_table[512] =
{
0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
@ -47,6 +49,39 @@ CONSTANT_VK u32a lotus_magic_table[256] =
0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
0x30, 0x04, 0xb6, 0xdc, 0x7d, 0xdf, 0x32, 0x4b,
0xf7, 0xcb, 0x45, 0x9b, 0x31, 0xbb, 0x21, 0x5a,
0x41, 0x9f, 0xe1, 0xd9, 0x4a, 0x4d, 0x9e, 0xda,
0xa0, 0x68, 0x2c, 0xc3, 0x27, 0x5f, 0x80, 0x36,
0x3e, 0xee, 0xfb, 0x95, 0x1a, 0xfe, 0xce, 0xa8,
0x34, 0xa9, 0x13, 0xf0, 0xa6, 0x3f, 0xd8, 0x0c,
0x78, 0x24, 0xaf, 0x23, 0x52, 0xc1, 0x67, 0x17,
0xf5, 0x66, 0x90, 0xe7, 0xe8, 0x07, 0xb8, 0x60,
0x48, 0xe6, 0x1e, 0x53, 0xf3, 0x92, 0xa4, 0x72,
0x8c, 0x08, 0x15, 0x6e, 0x86, 0x00, 0x84, 0xfa,
0xf4, 0x7f, 0x8a, 0x42, 0x19, 0xf6, 0xdb, 0xcd,
0x14, 0x8d, 0x50, 0x12, 0xba, 0x3c, 0x06, 0x4e,
0xec, 0xb3, 0x35, 0x11, 0xa1, 0x88, 0x8e, 0x2b,
0x94, 0x99, 0xb7, 0x71, 0x74, 0xd3, 0xe4, 0xbf,
0x3a, 0xde, 0x96, 0x0e, 0xbc, 0x0a, 0xed, 0x77,
0xfc, 0x37, 0x6b, 0x03, 0x79, 0x89, 0x62, 0xc6,
0xd7, 0xc0, 0xd2, 0x7c, 0x6a, 0x8b, 0x22, 0xa3,
0x5b, 0x05, 0x5d, 0x02, 0x75, 0xd5, 0x61, 0xe3,
0x18, 0x8f, 0x55, 0x51, 0xad, 0x1f, 0x0b, 0x5e,
0x85, 0xe5, 0xc2, 0x57, 0x63, 0xca, 0x3d, 0x6c,
0xb4, 0xc5, 0xcc, 0x70, 0xb2, 0x91, 0x59, 0x0d,
0x47, 0x20, 0xc8, 0x4f, 0x58, 0xe0, 0x01, 0xe2,
0x16, 0x38, 0xc4, 0x6f, 0x3b, 0x0f, 0x65, 0x46,
0xbe, 0x7e, 0x2d, 0x7b, 0x82, 0xf9, 0x40, 0xb5,
0x1d, 0x73, 0xf8, 0xeb, 0x26, 0xc7, 0x87, 0x97,
0x25, 0x54, 0xb1, 0x28, 0xaa, 0x98, 0x9d, 0xa5,
0x64, 0x6d, 0x7a, 0xd4, 0x10, 0x81, 0x44, 0xef,
0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
};
#define BOX(S,i) (S)[(i)]
@ -88,10 +123,10 @@ DECLSPEC void lotus_mix (u32x *in, LOCAL_AS u32 *s_lotus_magic_table)
u32x tmp_in = in[j];
u32x tmp_out = 0;
p = (p + s--) & 0xff; p = ((tmp_in >> 0) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 0;
p = (p + s--) & 0xff; p = ((tmp_in >> 8) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 8;
p = (p + s--) & 0xff; p = ((tmp_in >> 16) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 16;
p = (p + s--) & 0xff; p = ((tmp_in >> 24) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 24;
p = ((tmp_in >> 0) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 0;
p = ((tmp_in >> 8) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 8;
p = ((tmp_in >> 16) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 16;
p = ((tmp_in >> 24) & 0xff) ^ BOX1 (s_lotus_magic_table, (p + s--)); tmp_out |= p << 24;
in[j] = tmp_out;
}
@ -109,10 +144,15 @@ DECLSPEC void lotus_transform_password (u32x *in, u32x *out, LOCAL_AS u32 *s_lot
#endif
for (int i = 0; i < 4; i++)
{
t ^= (in[i] >> 0) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 0; t = ((out[i] >> 0) & 0xff);
t ^= (in[i] >> 8) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 8; t = ((out[i] >> 8) & 0xff);
t ^= (in[i] >> 16) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 16; t = ((out[i] >> 16) & 0xff);
t ^= (in[i] >> 24) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 24; t = ((out[i] >> 24) & 0xff);
u32x tmp_in = in[i];
u32x tmp_out = 0;
t ^= (tmp_in >> 0) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 0; t = (tmp_out >> 0) & 0xff;
t ^= (tmp_in >> 8) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 8; t = (tmp_out >> 8) & 0xff;
t ^= (tmp_in >> 16) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 16; t = (tmp_out >> 16) & 0xff;
t ^= (tmp_in >> 24) & 0xff; c = BOX1 (s_lotus_magic_table, t); tmp_out ^= c << 24; t = (tmp_out >> 24) & 0xff;
out[i] = tmp_out;
}
}
@ -558,9 +598,9 @@ KERNEL_FQ void m08700_m04 (KERN_ATTR_VECTOR ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}
@ -626,9 +666,9 @@ KERNEL_FQ void m08700_m08 (KERN_ATTR_VECTOR ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}
@ -694,9 +734,9 @@ KERNEL_FQ void m08700_m16 (KERN_ATTR_VECTOR ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}
@ -762,9 +802,9 @@ KERNEL_FQ void m08700_s04 (KERN_ATTR_VECTOR ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}
@ -830,9 +870,9 @@ KERNEL_FQ void m08700_s08 (KERN_ATTR_VECTOR ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}
@ -898,9 +938,9 @@ KERNEL_FQ void m08700_s16 (KERN_ATTR_VECTOR ())
* sbox
*/
LOCAL_VK u32 s_lotus_magic_table[256];
LOCAL_VK u32 s_lotus_magic_table[512];
for (u32 i = lid; i < 256; i += lsz)
for (u32 i = lid; i < 512; i += lsz)
{
s_lotus_magic_table[i] = lotus_magic_table[i];
}