arno
/
hashcat
mirror of https://github.com/hashcat/hashcat.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Refactor AES implementation to run fast on CL1.2 as well

Browse Source
pull/617/head
jsteube 8 years ago
parent ac8c4d946a
commit 56154f9f2f
1 changed files with 64 additions and 137 deletions
Show Stats Download Patch File Download Diff File

201
OpenCL/m13400.cl
Unescape View File

@ -711,53 +711,43 @@ void AES256_ExpandKey (u32 *userkey, u32 *rek, __local u32 *s_te0, __local u32 *
rek[6] = userkey[6];
rek[7] = userkey[7];
int i;
int j;
i = 0;
j = 0;
u32 run = 1;
while (run)
#ifdef _unroll
#pragma unroll
#endif
for (int i = 0, j = 0; i < 7; i += 1, j += 8)
{
u32 temp = rek[j + 7];
rek[j + 8] = rek[j + 0]
^ (s_te2[(temp >> 16) & 0xff] & 0xff000000)
^ (s_te3[(temp >> 8) & 0xff] & 0x00ff0000)
^ (s_te0[(temp >> 0) & 0xff] & 0x0000ff00)
^ (s_te1[(temp >> 24) & 0xff] & 0x000000ff)
^ rcon[i];
rek[j + 9] = rek[j + 1] ^ rek[j + 8];
rek[j + 10] = rek[j + 2] ^ rek[j + 9];
rek[j + 11] = rek[j + 3] ^ rek[j + 10];
if (++i == 7)
{
run = 0;
continue;
}
temp = rek[j + 11];
rek[j + 12] = rek[j + 4]
^ (s_te2[(temp >> 24) & 0xff] & 0xff000000)
^ (s_te3[(temp >> 16) & 0xff] & 0x00ff0000)
^ (s_te0[(temp >> 8) & 0xff] & 0x0000ff00)
^ (s_te1[(temp >> 0) & 0xff] & 0x000000ff);
rek[j + 13] = rek[j + 5] ^ rek[j + 12];
rek[j + 14] = rek[j + 6] ^ rek[j + 13];
rek[j + 15] = rek[j + 7] ^ rek[j + 14];
j += 8;
const u32 temp1 = rek[j + 7];
rek[j + 8] = rek[j + 0]
^ (s_te2[(temp1 >> 16) & 0xff] & 0xff000000)
^ (s_te3[(temp1 >> 8) & 0xff] & 0x00ff0000)
^ (s_te0[(temp1 >> 0) & 0xff] & 0x0000ff00)
^ (s_te1[(temp1 >> 24) & 0xff] & 0x000000ff)
^ rcon[i];
rek[j + 9] = rek[j + 1] ^ rek[j + 8];
rek[j + 10] = rek[j + 2] ^ rek[j + 9];
rek[j + 11] = rek[j + 3] ^ rek[j + 10];
if (i == 6) continue;
const u32 temp2 = rek[j + 11];
rek[j + 12] = rek[j + 4]
^ (s_te2[(temp2 >> 24) & 0xff] & 0xff000000)
^ (s_te3[(temp2 >> 16) & 0xff] & 0x00ff0000)
^ (s_te0[(temp2 >> 8) & 0xff] & 0x0000ff00)
^ (s_te1[(temp2 >> 0) & 0xff] & 0x000000ff);
rek[j + 13] = rek[j + 5] ^ rek[j + 12];
rek[j + 14] = rek[j + 6] ^ rek[j + 13];
rek[j + 15] = rek[j + 7] ^ rek[j + 14];
}
}
void AES256_InvertKey (u32 *rdk, __local u32 *s_td0, __local u32 *s_td1, __local u32 *s_td2, __local u32 *s_td3, __local u32 *s_td4, __local u32 *s_te0, __local u32 *s_te1, __local u32 *s_te2, __local u32 *s_te3, __local u32 *s_te4)
{
#ifdef _unroll
#pragma unroll
#endif
for (u32 i = 0, j = 56; i < j; i += 4, j -= 4)
{
u32 temp;
@ -768,6 +758,9 @@ void AES256_InvertKey (u32 *rdk, __local u32 *s_td0, __local u32 *s_td1, __local
temp = rdk[i + 3]; rdk[i + 3] = rdk[j + 3]; rdk[j + 3] = temp;
}
#ifdef _unroll
#pragma unroll
#endif
for (u32 i = 1, j = 4; i < 14; i += 1, j += 4)
{
rdk[j + 0] =
@ -812,54 +805,21 @@ void AES256_decrypt (const u32 *in, u32 *out, const u32 *rdk, __local u32 *s_td0
t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >> 8) & 0xff] ^ s_td3[s2 & 0xff] ^ rdk[ 5];
t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >> 8) & 0xff] ^ s_td3[s3 & 0xff] ^ rdk[ 6];
t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >> 8) & 0xff] ^ s_td3[s0 & 0xff] ^ rdk[ 7];
s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >> 8) & 0xff] ^ s_td3[t1 & 0xff] ^ rdk[ 8];
s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >> 8) & 0xff] ^ s_td3[t2 & 0xff] ^ rdk[ 9];
s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >> 8) & 0xff] ^ s_td3[t3 & 0xff] ^ rdk[10];
s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >> 8) & 0xff] ^ s_td3[t0 & 0xff] ^ rdk[11];
t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >> 8) & 0xff] ^ s_td3[s1 & 0xff] ^ rdk[12];
t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >> 8) & 0xff] ^ s_td3[s2 & 0xff] ^ rdk[13];
t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >> 8) & 0xff] ^ s_td3[s3 & 0xff] ^ rdk[14];
t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >> 8) & 0xff] ^ s_td3[s0 & 0xff] ^ rdk[15];
s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >> 8) & 0xff] ^ s_td3[t1 & 0xff] ^ rdk[16];
s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >> 8) & 0xff] ^ s_td3[t2 & 0xff] ^ rdk[17];
s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >> 8) & 0xff] ^ s_td3[t3 & 0xff] ^ rdk[18];
s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >> 8) & 0xff] ^ s_td3[t0 & 0xff] ^ rdk[19];
t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >> 8) & 0xff] ^ s_td3[s1 & 0xff] ^ rdk[20];
t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >> 8) & 0xff] ^ s_td3[s2 & 0xff] ^ rdk[21];
t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >> 8) & 0xff] ^ s_td3[s3 & 0xff] ^ rdk[22];
t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >> 8) & 0xff] ^ s_td3[s0 & 0xff] ^ rdk[23];
s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >> 8) & 0xff] ^ s_td3[t1 & 0xff] ^ rdk[24];
s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >> 8) & 0xff] ^ s_td3[t2 & 0xff] ^ rdk[25];
s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >> 8) & 0xff] ^ s_td3[t3 & 0xff] ^ rdk[26];
s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >> 8) & 0xff] ^ s_td3[t0 & 0xff] ^ rdk[27];
t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >> 8) & 0xff] ^ s_td3[s1 & 0xff] ^ rdk[28];
t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >> 8) & 0xff] ^ s_td3[s2 & 0xff] ^ rdk[29];
t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >> 8) & 0xff] ^ s_td3[s3 & 0xff] ^ rdk[30];
t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >> 8) & 0xff] ^ s_td3[s0 & 0xff] ^ rdk[31];
s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >> 8) & 0xff] ^ s_td3[t1 & 0xff] ^ rdk[32];
s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >> 8) & 0xff] ^ s_td3[t2 & 0xff] ^ rdk[33];
s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >> 8) & 0xff] ^ s_td3[t3 & 0xff] ^ rdk[34];
s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >> 8) & 0xff] ^ s_td3[t0 & 0xff] ^ rdk[35];
t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >> 8) & 0xff] ^ s_td3[s1 & 0xff] ^ rdk[36];
t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >> 8) & 0xff] ^ s_td3[s2 & 0xff] ^ rdk[37];
t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >> 8) & 0xff] ^ s_td3[s3 & 0xff] ^ rdk[38];
t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >> 8) & 0xff] ^ s_td3[s0 & 0xff] ^ rdk[39];
s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >> 8) & 0xff] ^ s_td3[t1 & 0xff] ^ rdk[40];
s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >> 8) & 0xff] ^ s_td3[t2 & 0xff] ^ rdk[41];
s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >> 8) & 0xff] ^ s_td3[t3 & 0xff] ^ rdk[42];
s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >> 8) & 0xff] ^ s_td3[t0 & 0xff] ^ rdk[43];
t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >> 8) & 0xff] ^ s_td3[s1 & 0xff] ^ rdk[44];
t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >> 8) & 0xff] ^ s_td3[s2 & 0xff] ^ rdk[45];
t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >> 8) & 0xff] ^ s_td3[s3 & 0xff] ^ rdk[46];
t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >> 8) & 0xff] ^ s_td3[s0 & 0xff] ^ rdk[47];
s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >> 8) & 0xff] ^ s_td3[t1 & 0xff] ^ rdk[48];
s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >> 8) & 0xff] ^ s_td3[t2 & 0xff] ^ rdk[49];
s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >> 8) & 0xff] ^ s_td3[t3 & 0xff] ^ rdk[50];
s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >> 8) & 0xff] ^ s_td3[t0 & 0xff] ^ rdk[51];
t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >> 8) & 0xff] ^ s_td3[s1 & 0xff] ^ rdk[52];
t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >> 8) & 0xff] ^ s_td3[s2 & 0xff] ^ rdk[53];
t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >> 8) & 0xff] ^ s_td3[s3 & 0xff] ^ rdk[54];
t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >> 8) & 0xff] ^ s_td3[s0 & 0xff] ^ rdk[55];
#ifdef _unroll
#pragma unroll
#endif
for (int i = 8; i < 56; i += 8)
{
s0 = s_td0[t0 >> 24] ^ s_td1[(t3 >> 16) & 0xff] ^ s_td2[(t2 >> 8) & 0xff] ^ s_td3[t1 & 0xff] ^ rdk[i + 0];
s1 = s_td0[t1 >> 24] ^ s_td1[(t0 >> 16) & 0xff] ^ s_td2[(t3 >> 8) & 0xff] ^ s_td3[t2 & 0xff] ^ rdk[i + 1];
s2 = s_td0[t2 >> 24] ^ s_td1[(t1 >> 16) & 0xff] ^ s_td2[(t0 >> 8) & 0xff] ^ s_td3[t3 & 0xff] ^ rdk[i + 2];
s3 = s_td0[t3 >> 24] ^ s_td1[(t2 >> 16) & 0xff] ^ s_td2[(t1 >> 8) & 0xff] ^ s_td3[t0 & 0xff] ^ rdk[i + 3];
t0 = s_td0[s0 >> 24] ^ s_td1[(s3 >> 16) & 0xff] ^ s_td2[(s2 >> 8) & 0xff] ^ s_td3[s1 & 0xff] ^ rdk[i + 4];
t1 = s_td0[s1 >> 24] ^ s_td1[(s0 >> 16) & 0xff] ^ s_td2[(s3 >> 8) & 0xff] ^ s_td3[s2 & 0xff] ^ rdk[i + 5];
t2 = s_td0[s2 >> 24] ^ s_td1[(s1 >> 16) & 0xff] ^ s_td2[(s0 >> 8) & 0xff] ^ s_td3[s3 & 0xff] ^ rdk[i + 6];
t3 = s_td0[s3 >> 24] ^ s_td1[(s2 >> 16) & 0xff] ^ s_td2[(s1 >> 8) & 0xff] ^ s_td3[s0 & 0xff] ^ rdk[i + 7];
}
out[0] = (s_td4[(t0 >> 24) & 0xff] & 0xff000000)
^ (s_td4[(t3 >> 16) & 0xff] & 0x00ff0000)
@ -902,54 +862,21 @@ void AES256_encrypt (const u32 *in, u32 *out, const u32 *rek, __local u32 *s_te0
t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >> 8) & 0xff] ^ s_te3[s0 & 0xff] ^ rek[ 5];
t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >> 8) & 0xff] ^ s_te3[s1 & 0xff] ^ rek[ 6];
t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >> 8) & 0xff] ^ s_te3[s2 & 0xff] ^ rek[ 7];
s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >> 8) & 0xff] ^ s_te3[t3 & 0xff] ^ rek[ 8];
s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >> 8) & 0xff] ^ s_te3[t0 & 0xff] ^ rek[ 9];
s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >> 8) & 0xff] ^ s_te3[t1 & 0xff] ^ rek[10];
s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >> 8) & 0xff] ^ s_te3[t2 & 0xff] ^ rek[11];
t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >> 8) & 0xff] ^ s_te3[s3 & 0xff] ^ rek[12];
t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >> 8) & 0xff] ^ s_te3[s0 & 0xff] ^ rek[13];
t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >> 8) & 0xff] ^ s_te3[s1 & 0xff] ^ rek[14];
t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >> 8) & 0xff] ^ s_te3[s2 & 0xff] ^ rek[15];
s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >> 8) & 0xff] ^ s_te3[t3 & 0xff] ^ rek[16];
s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >> 8) & 0xff] ^ s_te3[t0 & 0xff] ^ rek[17];
s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >> 8) & 0xff] ^ s_te3[t1 & 0xff] ^ rek[18];
s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >> 8) & 0xff] ^ s_te3[t2 & 0xff] ^ rek[19];
t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >> 8) & 0xff] ^ s_te3[s3 & 0xff] ^ rek[20];
t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >> 8) & 0xff] ^ s_te3[s0 & 0xff] ^ rek[21];
t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >> 8) & 0xff] ^ s_te3[s1 & 0xff] ^ rek[22];
t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >> 8) & 0xff] ^ s_te3[s2 & 0xff] ^ rek[23];
s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >> 8) & 0xff] ^ s_te3[t3 & 0xff] ^ rek[24];
s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >> 8) & 0xff] ^ s_te3[t0 & 0xff] ^ rek[25];
s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >> 8) & 0xff] ^ s_te3[t1 & 0xff] ^ rek[26];
s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >> 8) & 0xff] ^ s_te3[t2 & 0xff] ^ rek[27];
t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >> 8) & 0xff] ^ s_te3[s3 & 0xff] ^ rek[28];
t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >> 8) & 0xff] ^ s_te3[s0 & 0xff] ^ rek[29];
t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >> 8) & 0xff] ^ s_te3[s1 & 0xff] ^ rek[30];
t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >> 8) & 0xff] ^ s_te3[s2 & 0xff] ^ rek[31];
s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >> 8) & 0xff] ^ s_te3[t3 & 0xff] ^ rek[32];
s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >> 8) & 0xff] ^ s_te3[t0 & 0xff] ^ rek[33];
s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >> 8) & 0xff] ^ s_te3[t1 & 0xff] ^ rek[34];
s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >> 8) & 0xff] ^ s_te3[t2 & 0xff] ^ rek[35];
t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >> 8) & 0xff] ^ s_te3[s3 & 0xff] ^ rek[36];
t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >> 8) & 0xff] ^ s_te3[s0 & 0xff] ^ rek[37];
t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >> 8) & 0xff] ^ s_te3[s1 & 0xff] ^ rek[38];
t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >> 8) & 0xff] ^ s_te3[s2 & 0xff] ^ rek[39];
s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >> 8) & 0xff] ^ s_te3[t3 & 0xff] ^ rek[40];
s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >> 8) & 0xff] ^ s_te3[t0 & 0xff] ^ rek[41];
s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >> 8) & 0xff] ^ s_te3[t1 & 0xff] ^ rek[42];
s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >> 8) & 0xff] ^ s_te3[t2 & 0xff] ^ rek[43];
t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >> 8) & 0xff] ^ s_te3[s3 & 0xff] ^ rek[44];
t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >> 8) & 0xff] ^ s_te3[s0 & 0xff] ^ rek[45];
t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >> 8) & 0xff] ^ s_te3[s1 & 0xff] ^ rek[46];
t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >> 8) & 0xff] ^ s_te3[s2 & 0xff] ^ rek[47];
s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >> 8) & 0xff] ^ s_te3[t3 & 0xff] ^ rek[48];
s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >> 8) & 0xff] ^ s_te3[t0 & 0xff] ^ rek[49];
s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >> 8) & 0xff] ^ s_te3[t1 & 0xff] ^ rek[50];
s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >> 8) & 0xff] ^ s_te3[t2 & 0xff] ^ rek[51];
t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >> 8) & 0xff] ^ s_te3[s3 & 0xff] ^ rek[52];
t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >> 8) & 0xff] ^ s_te3[s0 & 0xff] ^ rek[53];
t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >> 8) & 0xff] ^ s_te3[s1 & 0xff] ^ rek[54];
t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >> 8) & 0xff] ^ s_te3[s2 & 0xff] ^ rek[55];
#ifdef _unroll
#pragma unroll
#endif
for (int i = 8; i < 56; i += 8)
{
s0 = s_te0[t0 >> 24] ^ s_te1[(t1 >> 16) & 0xff] ^ s_te2[(t2 >> 8) & 0xff] ^ s_te3[t3 & 0xff] ^ rek[i + 0];
s1 = s_te0[t1 >> 24] ^ s_te1[(t2 >> 16) & 0xff] ^ s_te2[(t3 >> 8) & 0xff] ^ s_te3[t0 & 0xff] ^ rek[i + 1];
s2 = s_te0[t2 >> 24] ^ s_te1[(t3 >> 16) & 0xff] ^ s_te2[(t0 >> 8) & 0xff] ^ s_te3[t1 & 0xff] ^ rek[i + 2];
s3 = s_te0[t3 >> 24] ^ s_te1[(t0 >> 16) & 0xff] ^ s_te2[(t1 >> 8) & 0xff] ^ s_te3[t2 & 0xff] ^ rek[i + 3];
t0 = s_te0[s0 >> 24] ^ s_te1[(s1 >> 16) & 0xff] ^ s_te2[(s2 >> 8) & 0xff] ^ s_te3[s3 & 0xff] ^ rek[i + 4];
t1 = s_te0[s1 >> 24] ^ s_te1[(s2 >> 16) & 0xff] ^ s_te2[(s3 >> 8) & 0xff] ^ s_te3[s0 & 0xff] ^ rek[i + 5];
t2 = s_te0[s2 >> 24] ^ s_te1[(s3 >> 16) & 0xff] ^ s_te2[(s0 >> 8) & 0xff] ^ s_te3[s1 & 0xff] ^ rek[i + 6];
t3 = s_te0[s3 >> 24] ^ s_te1[(s0 >> 16) & 0xff] ^ s_te2[(s1 >> 8) & 0xff] ^ s_te3[s2 & 0xff] ^ rek[i + 7];
}
out[0] = (s_te4[(t0 >> 24) & 0xff] & 0xff000000)
^ (s_te4[(t1 >> 16) & 0xff] & 0x00ff0000)

Write Preview
Loading…
Cancel
Save