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

Cleanup -m 10400 kernels to latest standard

Browse Source
pull/299/head
Jens Steube 8 years ago
parent 4149ecf78f
commit 1b9e5f716b
3 changed files with 603 additions and 660 deletions
Show Stats Download Patch File Download Diff File

403
OpenCL/m10400_a0.cl
Unescape View File

@ -7,6 +7,9 @@
#define _MD5_
//too much register pressure
//#define NEW_SIMD_CODE
#include "include/constants.h"
#include "include/kernel_vendor.h"
@ -20,9 +23,7 @@
#include "OpenCL/common.c"
#include "include/rp_kernel.h"
#include "OpenCL/rp.c"
#define COMPARE_S "OpenCL/check_single_comp4.c"
#define COMPARE_M "OpenCL/check_multi_comp4.c"
#include "OpenCL/simd.c"
__constant u32 padding[8] =
{
@ -253,14 +254,12 @@ __kernel void m10400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return;
u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6];
@ -269,7 +268,7 @@ __kernel void m10400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 pw_len = pws[gid].pw_len;
/**
* key
* shared
*/
__local RC4_KEY rc4_keys[64];
@ -304,84 +303,63 @@ __kernel void m10400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* loop
*/
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
u32 w0[4];
w0[0] = pw_buf0[0];
w0[1] = pw_buf0[1];
w0[2] = pw_buf0[2];
w0[3] = pw_buf0[3];
u32 w1[4];
w1[0] = pw_buf1[0];
w1[1] = pw_buf1[1];
w1[2] = pw_buf1[2];
w1[3] = pw_buf1[3];
u32 w2[4];
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
u32 w3[4];
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
u32 w0_t[4];
u32 w1_t[4];
u32 w2_t[4];
u32 w3_t[4];
// max length supported by pdf11 is 32
w0_t[0] = padding[0];
w0_t[1] = padding[1];
w0_t[2] = padding[2];
w0_t[3] = padding[3];
w1_t[0] = padding[4];
w1_t[1] = padding[5];
w1_t[2] = padding[6];
w1_t[3] = padding[7];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, out_len);
u32x w0[4] = { 0 };
u32x w1[4] = { 0 };
u32x w2[4] = { 0 };
u32x w3[4] = { 0 };
const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
/**
* pdf
*/
u32 p0[4];
u32 p1[4];
u32 p2[4];
u32 p3[4];
p0[0] = padding[0];
p0[1] = padding[1];
p0[2] = padding[2];
p0[3] = padding[3];
p1[0] = padding[4];
p1[1] = padding[5];
p1[2] = padding[6];
p1[3] = padding[7];
p2[0] = 0;
p2[1] = 0;
p2[2] = 0;
p2[3] = 0;
p3[0] = 0;
p3[1] = 0;
p3[2] = 0;
p3[3] = 0;
switch_buffer_by_offset_le (p0, p1, p2, p3, out_len);
// add password
// truncate at 32 is wanted, not a bug!
// add o_buf
w0_t[0] |= w0[0];
w0_t[1] |= w0[1];
w0_t[2] |= w0[2];
w0_t[3] |= w0[3];
w1_t[0] |= w1[0];
w1_t[1] |= w1[1];
w1_t[2] |= w1[2];
w1_t[3] |= w1[3];
w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2];
w2_t[3] = o_buf[3];
w3_t[0] = o_buf[4];
w3_t[1] = o_buf[5];
w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7];
w0[0] |= p0[0];
w0[1] |= p0[1];
w0[2] |= p0[2];
w0[3] |= p0[3];
w1[0] |= p1[0];
w1[1] |= p1[1];
w1[2] |= p1[2];
w1[3] |= p1[3];
w2[0] = o_buf[0];
w2[1] = o_buf[1];
w2[2] = o_buf[2];
w2[3] = o_buf[3];
w3[0] = o_buf[4];
w3[1] = o_buf[5];
w3[2] = o_buf[6];
w3[3] = o_buf[7];
u32 digest[4];
@ -390,48 +368,40 @@ __kernel void m10400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[2] = MD5M_C;
digest[3] = MD5M_D;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = P;
w0_t[1] = id_buf[0];
w0_t[2] = id_buf[1];
w0_t[3] = id_buf[2];
w1_t[0] = id_buf[3];
w1_t[1] = 0x80;
w1_t[2] = 0;
w1_t[3] = 0;
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 84 * 8;
w3_t[3] = 0;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
md5_transform (w0, w1, w2, w3, digest);
// now the RC4 part
w0[0] = P;
w0[1] = id_buf[0];
w0[2] = id_buf[1];
w0[3] = id_buf[2];
w1[0] = id_buf[3];
w1[1] = 0x80;
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] = 84 * 8;
w3[3] = 0;
u32 key[4];
md5_transform (w0, w1, w2, w3, digest);
key[0] = digest[0];
key[1] = digest[1] & 0xff;
key[2] = 0;
key[3] = 0;
// now the RC4 part
rc4_init_16 (rc4_key, key);
digest[1] = digest[1] & 0xff;
digest[2] = 0;
digest[3] = 0;
rc4_init_16 (rc4_key, digest);
u32 out[4];
rc4_next_16 (rc4_key, 0, 0, padding, out);
const u32 r0 = out[0];
const u32 r1 = out[1];
const u32 r2 = out[2];
const u32 r3 = out[3];
#include COMPARE_M
COMPARE_M_SIMD (out[0], out[1], out[2], out[3]);
}
}
@ -460,14 +430,12 @@ __kernel void m10400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return;
u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6];
@ -476,25 +444,13 @@ __kernel void m10400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 pw_len = pws[gid].pw_len;
/**
* key
* shared
*/
__local RC4_KEY rc4_keys[64];
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/**
* digest
*/
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/**
* U_buf
*/
@ -520,87 +476,78 @@ __kernel void m10400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
/**
* loop
* digest
*/
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
const u32 search[4] =
{
u32 w0[4];
w0[0] = pw_buf0[0];
w0[1] = pw_buf0[1];
w0[2] = pw_buf0[2];
w0[3] = pw_buf0[3];
u32 w1[4];
w1[0] = pw_buf1[0];
w1[1] = pw_buf1[1];
w1[2] = pw_buf1[2];
w1[3] = pw_buf1[3];
u32 w2[4];
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
u32 w3[4];
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
/**
* loop
*/
const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
u32 w0_t[4];
u32 w1_t[4];
u32 w2_t[4];
u32 w3_t[4];
// max length supported by pdf11 is 32
w0_t[0] = padding[0];
w0_t[1] = padding[1];
w0_t[2] = padding[2];
w0_t[3] = padding[3];
w1_t[0] = padding[4];
w1_t[1] = padding[5];
w1_t[2] = padding[6];
w1_t[3] = padding[7];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, out_len);
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
u32x w0[4] = { 0 };
u32x w1[4] = { 0 };
u32x w2[4] = { 0 };
u32x w3[4] = { 0 };
const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
/**
* pdf
*/
u32 p0[4];
u32 p1[4];
u32 p2[4];
u32 p3[4];
p0[0] = padding[0];
p0[1] = padding[1];
p0[2] = padding[2];
p0[3] = padding[3];
p1[0] = padding[4];
p1[1] = padding[5];
p1[2] = padding[6];
p1[3] = padding[7];
p2[0] = 0;
p2[1] = 0;
p2[2] = 0;
p2[3] = 0;
p3[0] = 0;
p3[1] = 0;
p3[2] = 0;
p3[3] = 0;
switch_buffer_by_offset_le (p0, p1, p2, p3, out_len);
// add password
// truncate at 32 is wanted, not a bug!
// add o_buf
w0_t[0] |= w0[0];
w0_t[1] |= w0[1];
w0_t[2] |= w0[2];
w0_t[3] |= w0[3];
w1_t[0] |= w1[0];
w1_t[1] |= w1[1];
w1_t[2] |= w1[2];
w1_t[3] |= w1[3];
w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2];
w2_t[3] = o_buf[3];
w3_t[0] = o_buf[4];
w3_t[1] = o_buf[5];
w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7];
w0[0] |= p0[0];
w0[1] |= p0[1];
w0[2] |= p0[2];
w0[3] |= p0[3];
w1[0] |= p1[0];
w1[1] |= p1[1];
w1[2] |= p1[2];
w1[3] |= p1[3];
w2[0] = o_buf[0];
w2[1] = o_buf[1];
w2[2] = o_buf[2];
w2[3] = o_buf[3];
w3[0] = o_buf[4];
w3[1] = o_buf[5];
w3[2] = o_buf[6];
w3[3] = o_buf[7];
u32 digest[4];
@ -609,48 +556,40 @@ __kernel void m10400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[2] = MD5M_C;
digest[3] = MD5M_D;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = P;
w0_t[1] = id_buf[0];
w0_t[2] = id_buf[1];
w0_t[3] = id_buf[2];
w1_t[0] = id_buf[3];
w1_t[1] = 0x80;
w1_t[2] = 0;
w1_t[3] = 0;
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 84 * 8;
w3_t[3] = 0;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
md5_transform (w0, w1, w2, w3, digest);
// now the RC4 part
w0[0] = P;
w0[1] = id_buf[0];
w0[2] = id_buf[1];
w0[3] = id_buf[2];
w1[0] = id_buf[3];
w1[1] = 0x80;
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] = 84 * 8;
w3[3] = 0;
u32 key[4];
md5_transform (w0, w1, w2, w3, digest);
key[0] = digest[0];
key[1] = digest[1] & 0xff;
key[2] = 0;
key[3] = 0;
// now the RC4 part
rc4_init_16 (rc4_key, key);
digest[1] = digest[1] & 0xff;
digest[2] = 0;
digest[3] = 0;
rc4_init_16 (rc4_key, digest);
u32 out[4];
rc4_next_16 (rc4_key, 0, 0, padding, out);
const u32 r0 = out[0];
const u32 r1 = out[1];
const u32 r2 = out[2];
const u32 r3 = out[3];
#include COMPARE_S
COMPARE_S_SIMD (out[0], out[1], out[2], out[3]);
}
}

611
OpenCL/m10400_a1.cl
Unescape View File

@ -7,6 +7,9 @@
#define _MD5_
//too much register pressure
//#define NEW_SIMD_CODE
#include "include/constants.h"
#include "include/kernel_vendor.h"
@ -18,9 +21,7 @@
#include "include/kernel_functions.c"
#include "OpenCL/types_ocl.c"
#include "OpenCL/common.c"
#define COMPARE_S "OpenCL/check_single_comp4.c"
#define COMPARE_M "OpenCL/check_multi_comp4.c"
#include "OpenCL/simd.c"
__constant u32 padding[8] =
{
@ -250,43 +251,22 @@ __kernel void m10400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return;
u32 wordl0[4];
wordl0[0] = pws[gid].i[ 0];
wordl0[1] = pws[gid].i[ 1];
wordl0[2] = pws[gid].i[ 2];
wordl0[3] = pws[gid].i[ 3];
u32 wordl1[4];
wordl1[0] = pws[gid].i[ 4];
wordl1[1] = pws[gid].i[ 5];
wordl1[2] = pws[gid].i[ 6];
wordl1[3] = pws[gid].i[ 7];
u32 pw_buf0[4];
u32 pw_buf1[4];
u32 wordl2[4];
wordl2[0] = 0;
wordl2[1] = 0;
wordl2[2] = 0;
wordl2[3] = 0;
u32 wordl3[4];
wordl3[0] = 0;
wordl3[1] = 0;
wordl3[2] = 0;
wordl3[3] = 0;
pw_buf0[0] = pws[gid].i[0];
pw_buf0[1] = pws[gid].i[1];
pw_buf0[2] = pws[gid].i[2];
pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/**
* key
* shared
*/
__local RC4_KEY rc4_keys[64];
@ -321,119 +301,123 @@ __kernel void m10400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* loop
*/
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
const u32 pw_r_len = combs_buf[il_pos].pw_len;
const u32 pw_len = pw_l_len + pw_r_len;
u32 wordr0[4];
wordr0[0] = combs_buf[il_pos].i[0];
wordr0[1] = combs_buf[il_pos].i[1];
wordr0[2] = combs_buf[il_pos].i[2];
wordr0[3] = combs_buf[il_pos].i[3];
u32 wordr1[4];
wordr1[0] = combs_buf[il_pos].i[4];
wordr1[1] = combs_buf[il_pos].i[5];
wordr1[2] = combs_buf[il_pos].i[6];
wordr1[3] = combs_buf[il_pos].i[7];
u32 wordr2[4];
wordr2[0] = 0;
wordr2[1] = 0;
wordr2[2] = 0;
wordr2[3] = 0;
u32 wordr3[4];
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32x pw_len = pw_l_len + pw_r_len;
/**
* concat password candidate
*/
u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
u32x wordl3[4] = { 0 };
wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{
switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
}
u32 w0[4];
u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
u32 w2[4];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
u32 w3[4];
w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1];
w3[2] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
u32 w0_t[4];
u32 w1_t[4];
u32 w2_t[4];
u32 w3_t[4];
// max length supported by pdf11 is 32
w0_t[0] = padding[0];
w0_t[1] = padding[1];
w0_t[2] = padding[2];
w0_t[3] = padding[3];
w1_t[0] = padding[4];
w1_t[1] = padding[5];
w1_t[2] = padding[6];
w1_t[3] = padding[7];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, pw_len);
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
/**
* pdf
*/
u32 p0[4];
u32 p1[4];
u32 p2[4];
u32 p3[4];
p0[0] = padding[0];
p0[1] = padding[1];
p0[2] = padding[2];
p0[3] = padding[3];
p1[0] = padding[4];
p1[1] = padding[5];
p1[2] = padding[6];
p1[3] = padding[7];
p2[0] = 0;
p2[1] = 0;
p2[2] = 0;
p2[3] = 0;
p3[0] = 0;
p3[1] = 0;
p3[2] = 0;
p3[3] = 0;
switch_buffer_by_offset_le (p0, p1, p2, p3, pw_len);
// add password
// truncate at 32 is wanted, not a bug!
// add o_buf
w0_t[0] |= w0[0];
w0_t[1] |= w0[1];
w0_t[2] |= w0[2];
w0_t[3] |= w0[3];
w1_t[0] |= w1[0];
w1_t[1] |= w1[1];
w1_t[2] |= w1[2];
w1_t[3] |= w1[3];
w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2];
w2_t[3] = o_buf[3];
w3_t[0] = o_buf[4];
w3_t[1] = o_buf[5];
w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7];
w0[0] |= p0[0];
w0[1] |= p0[1];
w0[2] |= p0[2];
w0[3] |= p0[3];
w1[0] |= p1[0];
w1[1] |= p1[1];
w1[2] |= p1[2];
w1[3] |= p1[3];
w2[0] = o_buf[0];
w2[1] = o_buf[1];
w2[2] = o_buf[2];
w2[3] = o_buf[3];
w3[0] = o_buf[4];
w3[1] = o_buf[5];
w3[2] = o_buf[6];
w3[3] = o_buf[7];
u32 digest[4];
@ -442,48 +426,40 @@ __kernel void m10400_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[2] = MD5M_C;
digest[3] = MD5M_D;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = P;
w0_t[1] = id_buf[0];
w0_t[2] = id_buf[1];
w0_t[3] = id_buf[2];
w1_t[0] = id_buf[3];
w1_t[1] = 0x80;
w1_t[2] = 0;
w1_t[3] = 0;
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 84 * 8;
w3_t[3] = 0;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
md5_transform (w0, w1, w2, w3, digest);
w0[0] = P;
w0[1] = id_buf[0];
w0[2] = id_buf[1];
w0[3] = id_buf[2];
w1[0] = id_buf[3];
w1[1] = 0x80;
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] = 84 * 8;
w3[3] = 0;
md5_transform (w0, w1, w2, w3, digest);
// now the RC4 part
u32 key[4];
key[0] = digest[0];
key[1] = digest[1] & 0xff;
key[2] = 0;
key[3] = 0;
digest[1] = digest[1] & 0xff;
digest[2] = 0;
digest[3] = 0;
rc4_init_16 (rc4_key, key);
rc4_init_16 (rc4_key, digest);
u32 out[4];
rc4_next_16 (rc4_key, 0, 0, padding, out);
const u32 r0 = out[0];
const u32 r1 = out[1];
const u32 r2 = out[2];
const u32 r3 = out[3];
#include COMPARE_M
COMPARE_M_SIMD (out[0], out[1], out[2], out[3]);
}
}
@ -511,61 +487,28 @@ __kernel void m10400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return;
u32 wordl0[4];
wordl0[0] = pws[gid].i[ 0];
wordl0[1] = pws[gid].i[ 1];
wordl0[2] = pws[gid].i[ 2];
wordl0[3] = pws[gid].i[ 3];
u32 pw_buf0[4];
u32 pw_buf1[4];
u32 wordl1[4];
wordl1[0] = pws[gid].i[ 4];
wordl1[1] = pws[gid].i[ 5];
wordl1[2] = pws[gid].i[ 6];
wordl1[3] = pws[gid].i[ 7];
u32 wordl2[4];
wordl2[0] = 0;
wordl2[1] = 0;
wordl2[2] = 0;
wordl2[3] = 0;
u32 wordl3[4];
wordl3[0] = 0;
wordl3[1] = 0;
wordl3[2] = 0;
wordl3[3] = 0;
pw_buf0[0] = pws[gid].i[0];
pw_buf0[1] = pws[gid].i[1];
pw_buf0[2] = pws[gid].i[2];
pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
pw_buf1[3] = pws[gid].i[7];
const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/**
* key
* shared
*/
__local RC4_KEY rc4_keys[64];
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/**
* digest
*/
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/**
* U_buf
*/
@ -591,122 +534,138 @@ __kernel void m10400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
/**
* loop
* digest
*/
for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
const u32 search[4] =
{
const u32 pw_r_len = combs_buf[il_pos].pw_len;
const u32 pw_len = pw_l_len + pw_r_len;
u32 wordr0[4];
wordr0[0] = combs_buf[il_pos].i[0];
wordr0[1] = combs_buf[il_pos].i[1];
wordr0[2] = combs_buf[il_pos].i[2];
wordr0[3] = combs_buf[il_pos].i[3];
u32 wordr1[4];
wordr1[0] = combs_buf[il_pos].i[4];
wordr1[1] = combs_buf[il_pos].i[5];
wordr1[2] = combs_buf[il_pos].i[6];
wordr1[3] = combs_buf[il_pos].i[7];
u32 wordr2[4];
wordr2[0] = 0;
wordr2[1] = 0;
wordr2[2] = 0;
wordr2[3] = 0;
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
u32 wordr3[4];
/**
* loop
*/
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32x pw_len = pw_l_len + pw_r_len;
/**
* concat password candidate
*/
u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
u32x wordl3[4] = { 0 };
wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{
switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
}
else
{
switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
}
u32 w0[4];
u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
u32 w2[4];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
u32 w3[4];
w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1];
w3[2] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
u32 w0_t[4];
u32 w1_t[4];
u32 w2_t[4];
u32 w3_t[4];
// max length supported by pdf11 is 32
w0_t[0] = padding[0];
w0_t[1] = padding[1];
w0_t[2] = padding[2];
w0_t[3] = padding[3];
w1_t[0] = padding[4];
w1_t[1] = padding[5];
w1_t[2] = padding[6];
w1_t[3] = padding[7];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, pw_len);
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
/**
* pdf
*/
u32 p0[4];
u32 p1[4];
u32 p2[4];
u32 p3[4];
p0[0] = padding[0];
p0[1] = padding[1];
p0[2] = padding[2];
p0[3] = padding[3];
p1[0] = padding[4];
p1[1] = padding[5];
p1[2] = padding[6];
p1[3] = padding[7];
p2[0] = 0;
p2[1] = 0;
p2[2] = 0;
p2[3] = 0;
p3[0] = 0;
p3[1] = 0;
p3[2] = 0;
p3[3] = 0;
switch_buffer_by_offset_le (p0, p1, p2, p3, pw_len);
// add password
// truncate at 32 is wanted, not a bug!
// add o_buf
w0_t[0] |= w0[0];
w0_t[1] |= w0[1];
w0_t[2] |= w0[2];
w0_t[3] |= w0[3];
w1_t[0] |= w1[0];
w1_t[1] |= w1[1];
w1_t[2] |= w1[2];
w1_t[3] |= w1[3];
w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2];
w2_t[3] = o_buf[3];
w3_t[0] = o_buf[4];
w3_t[1] = o_buf[5];
w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7];
w0[0] |= p0[0];
w0[1] |= p0[1];
w0[2] |= p0[2];
w0[3] |= p0[3];
w1[0] |= p1[0];
w1[1] |= p1[1];
w1[2] |= p1[2];
w1[3] |= p1[3];
w2[0] = o_buf[0];
w2[1] = o_buf[1];
w2[2] = o_buf[2];
w2[3] = o_buf[3];
w3[0] = o_buf[4];
w3[1] = o_buf[5];
w3[2] = o_buf[6];
w3[3] = o_buf[7];
u32 digest[4];
@ -715,48 +674,40 @@ __kernel void m10400_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[2] = MD5M_C;
digest[3] = MD5M_D;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
w0_t[0] = P;
w0_t[1] = id_buf[0];
w0_t[2] = id_buf[1];
w0_t[3] = id_buf[2];
w1_t[0] = id_buf[3];
w1_t[1] = 0x80;
w1_t[2] = 0;
w1_t[3] = 0;
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 84 * 8;
w3_t[3] = 0;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
md5_transform (w0, w1, w2, w3, digest);
w0[0] = P;
w0[1] = id_buf[0];
w0[2] = id_buf[1];
w0[3] = id_buf[2];
w1[0] = id_buf[3];
w1[1] = 0x80;
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] = 84 * 8;
w3[3] = 0;
md5_transform (w0, w1, w2, w3, digest);
// now the RC4 part
u32 key[4];
digest[1] = digest[1] & 0xff;
digest[2] = 0;
digest[3] = 0;
key[0] = digest[0];
key[1] = digest[1] & 0xff;
key[2] = 0;
key[3] = 0;
rc4_init_16 (rc4_key, key);
rc4_init_16 (rc4_key, digest);
u32 out[4];
rc4_next_16 (rc4_key, 0, 0, padding, out);
const u32 r0 = out[0];
const u32 r1 = out[1];
const u32 r2 = out[2];
const u32 r3 = out[3];
#include COMPARE_S
COMPARE_S_SIMD (out[0], out[1], out[2], out[3]);
}
}

249
OpenCL/m10400_a3.cl
Unescape View File

@ -7,6 +7,9 @@
#define _MD5_
//too much register pressure
//#define NEW_SIMD_CODE
#include "include/constants.h"
#include "include/kernel_vendor.h"
@ -241,6 +244,10 @@ static void m10400m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0);
/**
* shared
*/
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/**
@ -267,6 +274,47 @@ static void m10400m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
u32 p0[4];
u32 p1[4];
u32 p2[4];
u32 p3[4];
p0[0] = padding[0];
p0[1] = padding[1];
p0[2] = padding[2];
p0[3] = padding[3];
p1[0] = padding[4];
p1[1] = padding[5];
p1[2] = padding[6];
p1[3] = padding[7];
p2[0] = 0;
p2[1] = 0;
p2[2] = 0;
p2[3] = 0;
p3[0] = 0;
p3[1] = 0;
p3[2] = 0;
p3[3] = 0;
switch_buffer_by_offset_le (p0, p1, p2, p3, pw_len);
w0[0] |= p0[0];
w0[1] |= p0[1];
w0[2] |= p0[2];
w0[3] |= p0[3];
w1[0] |= p1[0];
w1[1] |= p1[1];
w1[2] |= p1[2];
w1[3] |= p1[3];
w2[0] |= p2[0];
w2[1] |= p2[1];
w2[2] |= p2[2];
w2[3] |= p2[3];
w3[0] |= p3[0];
w3[1] |= p3[1];
w3[2] |= p3[2];
w3[3] |= p3[3];
/**
* loop
*/
@ -279,52 +327,35 @@ static void m10400m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
w0[0] = w0l | w0r;
/**
* pdf
*/
u32 w0_t[4];
u32 w1_t[4];
u32 w2_t[4];
u32 w3_t[4];
// max length supported by pdf11 is 32
w0_t[0] = padding[0];
w0_t[1] = padding[1];
w0_t[2] = padding[2];
w0_t[3] = padding[3];
w1_t[0] = padding[4];
w1_t[1] = padding[5];
w1_t[2] = padding[6];
w1_t[3] = padding[7];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
switch_buffer_by_offset_le_S (w0_t, w1_t, w2_t, w3_t, pw_len);
// add password
// truncate at 32 is wanted, not a bug!
// add o_buf
w0_t[0] |= w0[0];
w0_t[1] |= w0[1];
w0_t[2] |= w0[2];
w0_t[3] |= w0[3];
w1_t[0] |= w1[0];
w1_t[1] |= w1[1];
w1_t[2] |= w1[2];
w1_t[3] |= w1[3];
w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2];
w2_t[3] = o_buf[3];
w3_t[0] = o_buf[4];
w3_t[1] = o_buf[5];
w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7];
w0_t[0] = w0[0];
w0_t[1] = w0[1];
w0_t[2] = w0[2];
w0_t[3] = w0[3];
w1_t[0] = w1[0];
w1_t[1] = w1[1];
w1_t[2] = w1[2];
w1_t[3] = w1[3];
w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2];
w2_t[3] = o_buf[3];
w3_t[0] = o_buf[4];
w3_t[1] = o_buf[5];
w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7];
u32 digest[4];
@ -356,14 +387,11 @@ static void m10400m (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
// now the RC4 part
u32 key[4];
digest[1] = digest[1] & 0xff;
digest[2] = 0;
digest[3] = 0;
key[0] = digest[0];
key[1] = digest[1] & 0xff;
key[2] = 0;
key[3] = 0;
rc4_init_16 (rc4_key, key);
rc4_init_16 (rc4_key, digest);
u32 out[4];
@ -382,19 +410,11 @@ static void m10400s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0);
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/**
* digest
* shared
*/
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
__local RC4_KEY *rc4_key = &rc4_keys[lid];
/**
* U_buf
@ -420,6 +440,59 @@ static void m10400s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
id_buf[3] = pdf_bufs[salt_pos].id_buf[3];
u32 p0[4];
u32 p1[4];
u32 p2[4];
u32 p3[4];
p0[0] = padding[0];
p0[1] = padding[1];
p0[2] = padding[2];
p0[3] = padding[3];
p1[0] = padding[4];
p1[1] = padding[5];
p1[2] = padding[6];
p1[3] = padding[7];
p2[0] = 0;
p2[1] = 0;
p2[2] = 0;
p2[3] = 0;
p3[0] = 0;
p3[1] = 0;
p3[2] = 0;
p3[3] = 0;
switch_buffer_by_offset_le (p0, p1, p2, p3, pw_len);
w0[0] |= p0[0];
w0[1] |= p0[1];
w0[2] |= p0[2];
w0[3] |= p0[3];
w1[0] |= p1[0];
w1[1] |= p1[1];
w1[2] |= p1[2];
w1[3] |= p1[3];
w2[0] |= p2[0];
w2[1] |= p2[1];
w2[2] |= p2[2];
w2[3] |= p2[3];
w3[0] |= p3[0];
w3[1] |= p3[1];
w3[2] |= p3[2];
w3[3] |= p3[3];
/**
* digest
*/
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
digests_buf[digests_offset].digest_buf[DGST_R1],
digests_buf[digests_offset].digest_buf[DGST_R2],
digests_buf[digests_offset].digest_buf[DGST_R3]
};
/**
* loop
*/
@ -432,52 +505,35 @@ static void m10400s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
w0[0] = w0l | w0r;
/**
* pdf
*/
u32 w0_t[4];
u32 w1_t[4];
u32 w2_t[4];
u32 w3_t[4];
// max length supported by pdf11 is 32
w0_t[0] = padding[0];
w0_t[1] = padding[1];
w0_t[2] = padding[2];
w0_t[3] = padding[3];
w1_t[0] = padding[4];
w1_t[1] = padding[5];
w1_t[2] = padding[6];
w1_t[3] = padding[7];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
switch_buffer_by_offset_le_S (w0_t, w1_t, w2_t, w3_t, pw_len);
// add password
// truncate at 32 is wanted, not a bug!
// add o_buf
w0_t[0] |= w0[0];
w0_t[1] |= w0[1];
w0_t[2] |= w0[2];
w0_t[3] |= w0[3];
w1_t[0] |= w1[0];
w1_t[1] |= w1[1];
w1_t[2] |= w1[2];
w1_t[3] |= w1[3];
w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2];
w2_t[3] = o_buf[3];
w3_t[0] = o_buf[4];
w3_t[1] = o_buf[5];
w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7];
w0_t[0] = w0[0];
w0_t[1] = w0[1];
w0_t[2] = w0[2];
w0_t[3] = w0[3];
w1_t[0] = w1[0];
w1_t[1] = w1[1];
w1_t[2] = w1[2];
w1_t[3] = w1[3];
w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2];
w2_t[3] = o_buf[3];
w3_t[0] = o_buf[4];
w3_t[1] = o_buf[5];
w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7];
u32 digest[4];
@ -509,14 +565,11 @@ static void m10400s (__local RC4_KEY *rc4_keys, u32 w0[4], u32 w1[4], u32 w2[4],
// now the RC4 part
u32 key[4];
key[0] = digest[0];
key[1] = digest[1] & 0xff;
key[2] = 0;
key[3] = 0;
digest[1] = digest[1] & 0xff;
digest[2] = 0;
digest[3] = 0;
rc4_init_16 (rc4_key, key);
rc4_init_16 (rc4_key, digest);
u32 out[4];

Write Preview
Loading…
Cancel
Save