2015-12-04 15:47:52 +01:00
|
|
|
/**
|
2016-09-11 22:20:15 +02:00
|
|
|
* Author......: See docs/credits.txt
|
2015-12-04 15:47:52 +01:00
|
|
|
* License.....: MIT
|
|
|
|
*/
|
|
|
|
|
2017-08-29 12:01:43 +02:00
|
|
|
#define NEW_SIMD_CODE
|
|
|
|
|
2016-05-25 23:04:26 +02:00
|
|
|
#include "inc_vendor.cl"
|
2016-06-26 23:39:42 +02:00
|
|
|
#include "inc_hash_constants.h"
|
2016-05-25 23:04:26 +02:00
|
|
|
#include "inc_hash_functions.cl"
|
|
|
|
#include "inc_types.cl"
|
|
|
|
#include "inc_common.cl"
|
2017-07-03 13:00:42 +02:00
|
|
|
#include "inc_simd.cl"
|
|
|
|
#include "inc_hash_sha1.cl"
|
2017-07-09 17:39:38 +02:00
|
|
|
#include "inc_cipher_aes.cl"
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2016-05-25 23:04:26 +02:00
|
|
|
#define COMPARE_S "inc_comp_single.cl"
|
|
|
|
#define COMPARE_M "inc_comp_multi.cl"
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2019-03-08 11:14:33 +01:00
|
|
|
typedef struct agilekey_tmp
|
|
|
|
{
|
|
|
|
u32 ipad[5];
|
|
|
|
u32 opad[5];
|
|
|
|
|
|
|
|
u32 dgst[5];
|
|
|
|
u32 out[5];
|
|
|
|
|
|
|
|
} agilekey_tmp_t;
|
|
|
|
|
2018-07-22 11:47:42 +02:00
|
|
|
DECLSPEC void hmac_sha1_run_V (u32x *w0, u32x *w1, u32x *w2, u32x *w3, u32x *ipad, u32x *opad, u32x *digest)
|
2015-12-04 15:47:52 +01:00
|
|
|
{
|
|
|
|
digest[0] = ipad[0];
|
|
|
|
digest[1] = ipad[1];
|
|
|
|
digest[2] = ipad[2];
|
|
|
|
digest[3] = ipad[3];
|
|
|
|
digest[4] = ipad[4];
|
|
|
|
|
2017-07-03 13:00:42 +02:00
|
|
|
sha1_transform_vector (w0, w1, w2, w3, digest);
|
2015-12-04 15:47:52 +01:00
|
|
|
|
|
|
|
w0[0] = digest[0];
|
|
|
|
w0[1] = digest[1];
|
|
|
|
w0[2] = digest[2];
|
|
|
|
w0[3] = digest[3];
|
|
|
|
w1[0] = digest[4];
|
|
|
|
w1[1] = 0x80000000;
|
|
|
|
w1[2] = 0;
|
|
|
|
w1[3] = 0;
|
|
|
|
w2[0] = 0;
|
|
|
|
w2[1] = 0;
|
|
|
|
w2[2] = 0;
|
|
|
|
w2[3] = 0;
|
|
|
|
w3[0] = 0;
|
|
|
|
w3[1] = 0;
|
|
|
|
w3[2] = 0;
|
|
|
|
w3[3] = (64 + 20) * 8;
|
|
|
|
|
|
|
|
digest[0] = opad[0];
|
|
|
|
digest[1] = opad[1];
|
|
|
|
digest[2] = opad[2];
|
|
|
|
digest[3] = opad[3];
|
|
|
|
digest[4] = opad[4];
|
|
|
|
|
2017-07-03 13:00:42 +02:00
|
|
|
sha1_transform_vector (w0, w1, w2, w3, digest);
|
2015-12-04 15:47:52 +01:00
|
|
|
}
|
|
|
|
|
2018-11-16 10:38:22 +01:00
|
|
|
__kernel void m06600_init (KERN_ATTR_TMPS (agilekey_tmp_t))
|
2015-12-04 15:47:52 +01:00
|
|
|
{
|
|
|
|
/**
|
|
|
|
* base
|
|
|
|
*/
|
|
|
|
|
2017-08-19 16:39:22 +02:00
|
|
|
const u64 gid = get_global_id (0);
|
2015-12-04 15:47:52 +01:00
|
|
|
|
|
|
|
if (gid >= gid_max) return;
|
|
|
|
|
2017-07-03 13:00:42 +02:00
|
|
|
sha1_hmac_ctx_t sha1_hmac_ctx;
|
|
|
|
|
2019-02-26 21:20:07 +01:00
|
|
|
sha1_hmac_init_global_swap (&sha1_hmac_ctx, pws[gid].i, pws[gid].pw_len);
|
2017-07-03 13:00:42 +02:00
|
|
|
|
|
|
|
tmps[gid].ipad[0] = sha1_hmac_ctx.ipad.h[0];
|
|
|
|
tmps[gid].ipad[1] = sha1_hmac_ctx.ipad.h[1];
|
|
|
|
tmps[gid].ipad[2] = sha1_hmac_ctx.ipad.h[2];
|
|
|
|
tmps[gid].ipad[3] = sha1_hmac_ctx.ipad.h[3];
|
|
|
|
tmps[gid].ipad[4] = sha1_hmac_ctx.ipad.h[4];
|
|
|
|
|
|
|
|
tmps[gid].opad[0] = sha1_hmac_ctx.opad.h[0];
|
|
|
|
tmps[gid].opad[1] = sha1_hmac_ctx.opad.h[1];
|
|
|
|
tmps[gid].opad[2] = sha1_hmac_ctx.opad.h[2];
|
|
|
|
tmps[gid].opad[3] = sha1_hmac_ctx.opad.h[3];
|
|
|
|
tmps[gid].opad[4] = sha1_hmac_ctx.opad.h[4];
|
|
|
|
|
|
|
|
// salt_buf[] is a bit dirty, it's used for AES IV in _comp, too
|
|
|
|
// so we have to move this data to a clean array first
|
|
|
|
|
|
|
|
u32 s[16] = { 0 };
|
|
|
|
|
|
|
|
s[0] = salt_bufs[salt_pos].salt_buf[0];
|
|
|
|
s[1] = salt_bufs[salt_pos].salt_buf[1];
|
|
|
|
|
|
|
|
sha1_hmac_update_swap (&sha1_hmac_ctx, s, salt_bufs[salt_pos].salt_len);
|
|
|
|
|
|
|
|
for (u32 i = 0, j = 1; i < 4; i += 5, j += 1)
|
2015-12-04 15:47:52 +01:00
|
|
|
{
|
2017-07-03 13:00:42 +02:00
|
|
|
sha1_hmac_ctx_t sha1_hmac_ctx2 = sha1_hmac_ctx;
|
|
|
|
|
2017-07-10 11:15:15 +02:00
|
|
|
u32 w0[4];
|
|
|
|
u32 w1[4];
|
|
|
|
u32 w2[4];
|
|
|
|
u32 w3[4];
|
|
|
|
|
2017-07-03 13:00:42 +02:00
|
|
|
w0[0] = j;
|
|
|
|
w0[1] = 0;
|
|
|
|
w0[2] = 0;
|
|
|
|
w0[3] = 0;
|
|
|
|
w1[0] = 0;
|
|
|
|
w1[1] = 0;
|
2015-12-04 15:47:52 +01:00
|
|
|
w1[2] = 0;
|
|
|
|
w1[3] = 0;
|
|
|
|
w2[0] = 0;
|
|
|
|
w2[1] = 0;
|
|
|
|
w2[2] = 0;
|
|
|
|
w2[3] = 0;
|
|
|
|
w3[0] = 0;
|
|
|
|
w3[1] = 0;
|
|
|
|
w3[2] = 0;
|
2017-07-03 13:00:42 +02:00
|
|
|
w3[3] = 0;
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2017-07-03 13:00:42 +02:00
|
|
|
sha1_hmac_update_64 (&sha1_hmac_ctx2, w0, w1, w2, w3, 4);
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2017-07-03 13:00:42 +02:00
|
|
|
sha1_hmac_final (&sha1_hmac_ctx2);
|
|
|
|
|
|
|
|
tmps[gid].dgst[i + 0] = sha1_hmac_ctx2.opad.h[0];
|
|
|
|
tmps[gid].dgst[i + 1] = sha1_hmac_ctx2.opad.h[1];
|
|
|
|
tmps[gid].dgst[i + 2] = sha1_hmac_ctx2.opad.h[2];
|
|
|
|
tmps[gid].dgst[i + 3] = sha1_hmac_ctx2.opad.h[3];
|
|
|
|
tmps[gid].dgst[i + 4] = sha1_hmac_ctx2.opad.h[4];
|
|
|
|
|
|
|
|
tmps[gid].out[i + 0] = tmps[gid].dgst[i + 0];
|
|
|
|
tmps[gid].out[i + 1] = tmps[gid].dgst[i + 1];
|
|
|
|
tmps[gid].out[i + 2] = tmps[gid].dgst[i + 2];
|
|
|
|
tmps[gid].out[i + 3] = tmps[gid].dgst[i + 3];
|
|
|
|
tmps[gid].out[i + 4] = tmps[gid].dgst[i + 4];
|
2015-12-04 15:47:52 +01:00
|
|
|
}
|
2017-07-03 13:00:42 +02:00
|
|
|
}
|
|
|
|
|
2018-11-16 10:38:22 +01:00
|
|
|
__kernel void m06600_loop (KERN_ATTR_TMPS (agilekey_tmp_t))
|
2017-07-03 13:00:42 +02:00
|
|
|
{
|
2017-08-19 16:39:22 +02:00
|
|
|
const u64 gid = get_global_id (0);
|
2017-07-03 13:00:42 +02:00
|
|
|
|
|
|
|
if ((gid * VECT_SIZE) >= gid_max) return;
|
|
|
|
|
|
|
|
u32x ipad[5];
|
|
|
|
u32x opad[5];
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2017-07-03 13:00:42 +02:00
|
|
|
ipad[0] = packv (tmps, ipad, gid, 0);
|
|
|
|
ipad[1] = packv (tmps, ipad, gid, 1);
|
|
|
|
ipad[2] = packv (tmps, ipad, gid, 2);
|
|
|
|
ipad[3] = packv (tmps, ipad, gid, 3);
|
|
|
|
ipad[4] = packv (tmps, ipad, gid, 4);
|
|
|
|
|
|
|
|
opad[0] = packv (tmps, opad, gid, 0);
|
|
|
|
opad[1] = packv (tmps, opad, gid, 1);
|
|
|
|
opad[2] = packv (tmps, opad, gid, 2);
|
|
|
|
opad[3] = packv (tmps, opad, gid, 3);
|
|
|
|
opad[4] = packv (tmps, opad, gid, 4);
|
|
|
|
|
|
|
|
for (u32 i = 0; i < 4; i += 5)
|
|
|
|
{
|
|
|
|
u32x dgst[5];
|
|
|
|
u32x out[5];
|
|
|
|
|
|
|
|
dgst[0] = packv (tmps, dgst, gid, i + 0);
|
|
|
|
dgst[1] = packv (tmps, dgst, gid, i + 1);
|
|
|
|
dgst[2] = packv (tmps, dgst, gid, i + 2);
|
|
|
|
dgst[3] = packv (tmps, dgst, gid, i + 3);
|
|
|
|
dgst[4] = packv (tmps, dgst, gid, i + 4);
|
|
|
|
|
|
|
|
out[0] = packv (tmps, out, gid, i + 0);
|
|
|
|
out[1] = packv (tmps, out, gid, i + 1);
|
|
|
|
out[2] = packv (tmps, out, gid, i + 2);
|
|
|
|
out[3] = packv (tmps, out, gid, i + 3);
|
|
|
|
out[4] = packv (tmps, out, gid, i + 4);
|
|
|
|
|
|
|
|
for (u32 j = 0; j < loop_cnt; j++)
|
|
|
|
{
|
|
|
|
u32x w0[4];
|
|
|
|
u32x w1[4];
|
|
|
|
u32x w2[4];
|
|
|
|
u32x w3[4];
|
|
|
|
|
|
|
|
w0[0] = dgst[0];
|
|
|
|
w0[1] = dgst[1];
|
|
|
|
w0[2] = dgst[2];
|
|
|
|
w0[3] = dgst[3];
|
|
|
|
w1[0] = dgst[4];
|
|
|
|
w1[1] = 0x80000000;
|
|
|
|
w1[2] = 0;
|
|
|
|
w1[3] = 0;
|
|
|
|
w2[0] = 0;
|
|
|
|
w2[1] = 0;
|
|
|
|
w2[2] = 0;
|
|
|
|
w2[3] = 0;
|
|
|
|
w3[0] = 0;
|
|
|
|
w3[1] = 0;
|
|
|
|
w3[2] = 0;
|
|
|
|
w3[3] = (64 + 20) * 8;
|
|
|
|
|
|
|
|
hmac_sha1_run_V (w0, w1, w2, w3, ipad, opad, dgst);
|
|
|
|
|
|
|
|
out[0] ^= dgst[0];
|
|
|
|
out[1] ^= dgst[1];
|
|
|
|
out[2] ^= dgst[2];
|
|
|
|
out[3] ^= dgst[3];
|
|
|
|
out[4] ^= dgst[4];
|
|
|
|
}
|
|
|
|
|
|
|
|
unpackv (tmps, dgst, gid, i + 0, dgst[0]);
|
|
|
|
unpackv (tmps, dgst, gid, i + 1, dgst[1]);
|
|
|
|
unpackv (tmps, dgst, gid, i + 2, dgst[2]);
|
|
|
|
unpackv (tmps, dgst, gid, i + 3, dgst[3]);
|
|
|
|
unpackv (tmps, dgst, gid, i + 4, dgst[4]);
|
|
|
|
|
|
|
|
unpackv (tmps, out, gid, i + 0, out[0]);
|
|
|
|
unpackv (tmps, out, gid, i + 1, out[1]);
|
|
|
|
unpackv (tmps, out, gid, i + 2, out[2]);
|
|
|
|
unpackv (tmps, out, gid, i + 3, out[3]);
|
|
|
|
unpackv (tmps, out, gid, i + 4, out[4]);
|
|
|
|
}
|
2015-12-04 15:47:52 +01:00
|
|
|
}
|
|
|
|
|
2018-11-16 10:38:22 +01:00
|
|
|
__kernel void m06600_comp (KERN_ATTR_TMPS (agilekey_tmp_t))
|
2015-12-04 15:47:52 +01:00
|
|
|
{
|
2017-08-19 16:39:22 +02:00
|
|
|
const u64 gid = get_global_id (0);
|
|
|
|
const u64 lid = get_local_id (0);
|
|
|
|
const u64 lsz = get_local_size (0);
|
2015-12-04 15:47:52 +01:00
|
|
|
|
|
|
|
/**
|
|
|
|
* aes shared
|
|
|
|
*/
|
|
|
|
|
2017-07-09 17:39:38 +02:00
|
|
|
#ifdef REAL_SHM
|
|
|
|
|
2015-12-06 11:29:29 +01:00
|
|
|
__local u32 s_td0[256];
|
|
|
|
__local u32 s_td1[256];
|
|
|
|
__local u32 s_td2[256];
|
|
|
|
__local u32 s_td3[256];
|
|
|
|
__local u32 s_td4[256];
|
|
|
|
|
|
|
|
__local u32 s_te0[256];
|
|
|
|
__local u32 s_te1[256];
|
|
|
|
__local u32 s_te2[256];
|
|
|
|
__local u32 s_te3[256];
|
|
|
|
__local u32 s_te4[256];
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2019-03-03 19:18:56 +01:00
|
|
|
for (u32 i = lid; i < 256; i += lsz)
|
2016-01-19 16:06:03 +01:00
|
|
|
{
|
|
|
|
s_td0[i] = td0[i];
|
|
|
|
s_td1[i] = td1[i];
|
|
|
|
s_td2[i] = td2[i];
|
|
|
|
s_td3[i] = td3[i];
|
|
|
|
s_td4[i] = td4[i];
|
|
|
|
|
|
|
|
s_te0[i] = te0[i];
|
|
|
|
s_te1[i] = te1[i];
|
|
|
|
s_te2[i] = te2[i];
|
|
|
|
s_te3[i] = te3[i];
|
|
|
|
s_te4[i] = te4[i];
|
|
|
|
}
|
2015-12-04 15:47:52 +01:00
|
|
|
|
|
|
|
barrier (CLK_LOCAL_MEM_FENCE);
|
|
|
|
|
2017-07-09 17:39:38 +02:00
|
|
|
#else
|
|
|
|
|
|
|
|
__constant u32a *s_td0 = td0;
|
|
|
|
__constant u32a *s_td1 = td1;
|
|
|
|
__constant u32a *s_td2 = td2;
|
|
|
|
__constant u32a *s_td3 = td3;
|
|
|
|
__constant u32a *s_td4 = td4;
|
|
|
|
|
|
|
|
__constant u32a *s_te0 = te0;
|
|
|
|
__constant u32a *s_te1 = te1;
|
|
|
|
__constant u32a *s_te2 = te2;
|
|
|
|
__constant u32a *s_te3 = te3;
|
|
|
|
__constant u32a *s_te4 = te4;
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
2015-12-04 15:47:52 +01:00
|
|
|
if (gid >= gid_max) return;
|
|
|
|
|
2016-01-19 16:06:03 +01:00
|
|
|
/**
|
|
|
|
* salt
|
|
|
|
*/
|
|
|
|
|
|
|
|
const u32 iv[4] =
|
|
|
|
{
|
|
|
|
salt_bufs[salt_pos].salt_buf[ 4],
|
|
|
|
salt_bufs[salt_pos].salt_buf[ 5],
|
|
|
|
salt_bufs[salt_pos].salt_buf[ 6],
|
|
|
|
salt_bufs[salt_pos].salt_buf[ 7]
|
|
|
|
};
|
|
|
|
|
|
|
|
const u32 data[4] =
|
|
|
|
{
|
|
|
|
salt_bufs[salt_pos].salt_buf[ 8],
|
|
|
|
salt_bufs[salt_pos].salt_buf[ 9],
|
|
|
|
salt_bufs[salt_pos].salt_buf[10],
|
|
|
|
salt_bufs[salt_pos].salt_buf[11]
|
|
|
|
};
|
|
|
|
|
2015-12-04 15:47:52 +01:00
|
|
|
/**
|
2017-07-09 17:39:38 +02:00
|
|
|
* aes
|
2015-12-04 15:47:52 +01:00
|
|
|
*/
|
|
|
|
|
2017-07-09 17:39:38 +02:00
|
|
|
u32 ukey[4];
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2017-07-09 17:39:38 +02:00
|
|
|
ukey[0] = tmps[gid].out[0];
|
|
|
|
ukey[1] = tmps[gid].out[1];
|
|
|
|
ukey[2] = tmps[gid].out[2];
|
|
|
|
ukey[3] = tmps[gid].out[3];
|
2015-12-04 15:47:52 +01:00
|
|
|
|
|
|
|
#define KEYLEN 44
|
|
|
|
|
2017-07-09 17:39:38 +02:00
|
|
|
u32 ks[KEYLEN];
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2017-07-09 17:39:38 +02:00
|
|
|
AES128_set_decrypt_key (ks, ukey, s_te0, s_te1, s_te2, s_te3, s_te4, s_td0, s_td1, s_td2, s_td3, s_td4);
|
2015-12-04 15:47:52 +01:00
|
|
|
|
|
|
|
u32 out[4];
|
|
|
|
|
2017-07-09 17:39:38 +02:00
|
|
|
AES128_decrypt (ks, data, out, s_td0, s_td1, s_td2, s_td3, s_td4);
|
2015-12-04 15:47:52 +01:00
|
|
|
|
2017-07-09 17:39:38 +02:00
|
|
|
const u32 r0 = out[0] ^ iv[0];
|
|
|
|
const u32 r1 = out[1] ^ iv[1];
|
|
|
|
const u32 r2 = out[2] ^ iv[2];
|
|
|
|
const u32 r3 = out[3] ^ iv[3];
|
2015-12-04 15:47:52 +01:00
|
|
|
|
|
|
|
#define il_pos 0
|
|
|
|
|
2015-12-15 12:04:22 +01:00
|
|
|
#include COMPARE_M
|
2015-12-04 15:47:52 +01:00
|
|
|
}
|