arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-12-01 20:38:14 +00:00
Code Issues Releases Wiki Activity

Cleanup -m 10420 kernels to latest standard

Browse Source
This commit is contained in:
Jens Steube 2016-04-18 11:40:51 +02:00
parent 1b9e5f716b
commit 1d9a36b86e
3 changed files with 556 additions and 556 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

268
OpenCL/m10420_a0.cl
View File

@ -149,14 +149,12 @@ __kernel void m10420_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 pw_buf0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4]; pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5]; pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6]; pw_buf1[2] = pws[gid].i[ 6];
@ -201,82 +199,82 @@ __kernel void m10420_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1); const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
append_0x80_2x4_VV (w0, w1, out_len); /**
* pdf
*/
u32x w0_t[4]; u32 p0[4];
u32x w1_t[4]; u32 p1[4];
u32x w2_t[4]; u32 p2[4];
u32x w3_t[4]; u32 p3[4];
// max length supported by pdf11 is 32 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;
w0_t[0] = padding[0]; switch_buffer_by_offset_le (p0, p1, p2, p3, out_len);
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_VV (w0_t, w1_t, w2_t, w3_t, out_len);
// add password // add password
// truncate at 32 is wanted, not a bug! // truncate at 32 is wanted, not a bug!
// add o_buf // add o_buf
w0_t[0] |= w0[0]; w0[0] |= p0[0];
w0_t[1] |= w0[1]; w0[1] |= p0[1];
w0_t[2] |= w0[2]; w0[2] |= p0[2];
w0_t[3] |= w0[3]; w0[3] |= p0[3];
w1_t[0] |= w1[0]; w1[0] |= p1[0];
w1_t[1] |= w1[1]; w1[1] |= p1[1];
w1_t[2] |= w1[2]; w1[2] |= p1[2];
w1_t[3] |= w1[3]; w1[3] |= p1[3];
w2_t[0] = o_buf[0]; w2[0] = o_buf[0];
w2_t[1] = o_buf[1]; w2[1] = o_buf[1];
w2_t[2] = o_buf[2]; w2[2] = o_buf[2];
w2_t[3] = o_buf[3]; w2[3] = o_buf[3];
w3_t[0] = o_buf[4]; w3[0] = o_buf[4];
w3_t[1] = o_buf[5]; w3[1] = o_buf[5];
w3_t[2] = o_buf[6]; w3[2] = o_buf[6];
w3_t[3] = o_buf[7]; w3[3] = o_buf[7];
u32x digest[4]; u32 digest[4];
digest[0] = MD5M_A; digest[0] = MD5M_A;
digest[1] = MD5M_B; digest[1] = MD5M_B;
digest[2] = MD5M_C; digest[2] = MD5M_C;
digest[3] = MD5M_D; digest[3] = MD5M_D;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest); md5_transform (w0, w1, w2, w3, digest);
w0_t[0] = P; w0[0] = P;
w0_t[1] = id_buf[0]; w0[1] = id_buf[0];
w0_t[2] = id_buf[1]; w0[2] = id_buf[1];
w0_t[3] = id_buf[2]; w0[3] = id_buf[2];
w1_t[0] = id_buf[3]; w1[0] = id_buf[3];
w1_t[1] = 0x80; w1[1] = 0x80;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 84 * 8; w3[2] = 84 * 8;
w3_t[3] = 0; w3[3] = 0;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest); md5_transform (w0, w1, w2, w3, digest);
u32x a = digest[0]; u32x a = digest[0];
u32x b = digest[1] & 0xff; u32x b = digest[1] & 0xff;
@ -312,14 +310,12 @@ __kernel void m10420_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 pw_buf0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[ 3];
u32 pw_buf1[4];
pw_buf1[0] = pws[gid].i[ 4]; pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5]; pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6]; pw_buf1[2] = pws[gid].i[ 6];
@ -327,18 +323,6 @@ __kernel void m10420_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 pw_len = pws[gid].pw_len; const u32 pw_len = pws[gid].pw_len;
/**
* 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 * U_buf
*/ */
@ -363,6 +347,18 @@ __kernel void m10420_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
id_buf[2] = pdf_bufs[salt_pos].id_buf[2]; id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
id_buf[3] = pdf_bufs[salt_pos].id_buf[3]; id_buf[3] = pdf_bufs[salt_pos].id_buf[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 * loop
*/ */
@ -376,82 +372,82 @@ __kernel void m10420_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1); const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
append_0x80_2x4_VV (w0, w1, out_len); /**
* pdf
*/
u32x w0_t[4]; u32 p0[4];
u32x w1_t[4]; u32 p1[4];
u32x w2_t[4]; u32 p2[4];
u32x w3_t[4]; u32 p3[4];
// max length supported by pdf11 is 32 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;
w0_t[0] = padding[0]; switch_buffer_by_offset_le (p0, p1, p2, p3, out_len);
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_VV (w0_t, w1_t, w2_t, w3_t, out_len);
// add password // add password
// truncate at 32 is wanted, not a bug! // truncate at 32 is wanted, not a bug!
// add o_buf // add o_buf
w0_t[0] |= w0[0]; w0[0] |= p0[0];
w0_t[1] |= w0[1]; w0[1] |= p0[1];
w0_t[2] |= w0[2]; w0[2] |= p0[2];
w0_t[3] |= w0[3]; w0[3] |= p0[3];
w1_t[0] |= w1[0]; w1[0] |= p1[0];
w1_t[1] |= w1[1]; w1[1] |= p1[1];
w1_t[2] |= w1[2]; w1[2] |= p1[2];
w1_t[3] |= w1[3]; w1[3] |= p1[3];
w2_t[0] = o_buf[0]; w2[0] = o_buf[0];
w2_t[1] = o_buf[1]; w2[1] = o_buf[1];
w2_t[2] = o_buf[2]; w2[2] = o_buf[2];
w2_t[3] = o_buf[3]; w2[3] = o_buf[3];
w3_t[0] = o_buf[4]; w3[0] = o_buf[4];
w3_t[1] = o_buf[5]; w3[1] = o_buf[5];
w3_t[2] = o_buf[6]; w3[2] = o_buf[6];
w3_t[3] = o_buf[7]; w3[3] = o_buf[7];
u32x digest[4]; u32 digest[4];
digest[0] = MD5M_A; digest[0] = MD5M_A;
digest[1] = MD5M_B; digest[1] = MD5M_B;
digest[2] = MD5M_C; digest[2] = MD5M_C;
digest[3] = MD5M_D; digest[3] = MD5M_D;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest); md5_transform (w0, w1, w2, w3, digest);
w0_t[0] = P; w0[0] = P;
w0_t[1] = id_buf[0]; w0[1] = id_buf[0];
w0_t[2] = id_buf[1]; w0[2] = id_buf[1];
w0_t[3] = id_buf[2]; w0[3] = id_buf[2];
w1_t[0] = id_buf[3]; w1[0] = id_buf[3];
w1_t[1] = 0x80; w1[1] = 0x80;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 84 * 8; w3[2] = 84 * 8;
w3_t[3] = 0; w3[3] = 0;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest); md5_transform (w0, w1, w2, w3, digest);
u32x a = digest[0]; u32x a = digest[0];
u32x b = digest[1] & 0xff; u32x b = digest[1] & 0xff;

596
OpenCL/m10420_a1.cl
View File

@ -5,6 +5,8 @@
#define _MD5_ #define _MD5_
#define NEW_SIMD_CODE
#include "include/constants.h" #include "include/constants.h"
#include "include/kernel_vendor.h" #include "include/kernel_vendor.h"
@ -16,9 +18,7 @@
#include "include/kernel_functions.c" #include "include/kernel_functions.c"
#include "OpenCL/types_ocl.c" #include "OpenCL/types_ocl.c"
#include "OpenCL/common.c" #include "OpenCL/common.c"
#include "OpenCL/simd.c"
#define COMPARE_S "OpenCL/check_single_comp4.c"
#define COMPARE_M "OpenCL/check_multi_comp4.c"
__constant u32 padding[8] = __constant u32 padding[8] =
{ {
@ -32,29 +32,29 @@ __constant u32 padding[8] =
0x7a695364 0x7a695364
}; };
static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4]) static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
{ {
u32 a = digest[0]; u32x a = digest[0];
u32 b = digest[1]; u32x b = digest[1];
u32 c = digest[2]; u32x c = digest[2];
u32 d = digest[3]; u32x d = digest[3];
u32 w0_t = w0[0]; u32x w0_t = w0[0];
u32 w1_t = w0[1]; u32x w1_t = w0[1];
u32 w2_t = w0[2]; u32x w2_t = w0[2];
u32 w3_t = w0[3]; u32x w3_t = w0[3];
u32 w4_t = w1[0]; u32x w4_t = w1[0];
u32 w5_t = w1[1]; u32x w5_t = w1[1];
u32 w6_t = w1[2]; u32x w6_t = w1[2];
u32 w7_t = w1[3]; u32x w7_t = w1[3];
u32 w8_t = w2[0]; u32x w8_t = w2[0];
u32 w9_t = w2[1]; u32x w9_t = w2[1];
u32 wa_t = w2[2]; u32x wa_t = w2[2];
u32 wb_t = w2[3]; u32x wb_t = w2[3];
u32 wc_t = w3[0]; u32x wc_t = w3[0];
u32 wd_t = w3[1]; u32x wd_t = w3[1];
u32 we_t = w3[2]; u32x we_t = w3[2];
u32 wf_t = w3[3]; u32x wf_t = w3[3];
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
@ -146,43 +146,20 @@ __kernel void m10420_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 wordl0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
wordl0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[0];
wordl0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[1];
wordl0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[2];
wordl0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
u32 wordl1[4]; pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
wordl1[0] = pws[gid].i[ 4]; pw_buf1[3] = pws[gid].i[7];
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;
const u32 pw_l_len = pws[gid].pw_len; const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
append_0x80_2x4 (wordl0, wordl1, pw_l_len);
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/** /**
* U_buf * U_buf
*/ */
@ -211,119 +188,123 @@ __kernel void m10420_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
* loop * 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 u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32 pw_len = pw_l_len + pw_r_len; const u32x pw_len = pw_l_len + pw_r_len;
u32 wordr0[4]; /**
* concat password candidate
*/
wordr0[0] = combs_buf[il_pos].i[0]; u32x wordl0[4] = { 0 };
wordr0[1] = combs_buf[il_pos].i[1]; u32x wordl1[4] = { 0 };
wordr0[2] = combs_buf[il_pos].i[2]; u32x wordl2[4] = { 0 };
wordr0[3] = combs_buf[il_pos].i[3]; u32x wordl3[4] = { 0 };
u32 wordr1[4]; 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];
wordr1[0] = combs_buf[il_pos].i[4]; u32x wordr0[4] = { 0 };
wordr1[1] = combs_buf[il_pos].i[5]; u32x wordr1[4] = { 0 };
wordr1[2] = combs_buf[il_pos].i[6]; u32x wordr2[4] = { 0 };
wordr1[3] = combs_buf[il_pos].i[7]; u32x wordr3[4] = { 0 };
u32 wordr2[4]; wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr2[0] = 0; wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr2[1] = 0; wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr2[2] = 0; wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr2[3] = 0; wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
u32 wordr3[4]; wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
if (combs_mode == COMBINATOR_MODE_BASE_LEFT) 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[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1]; w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2]; w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3]; w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0]; w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1]; w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2]; w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3]; w1[3] = wordl1[3] | wordr1[3];
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
u32 w2[4]; /**
* pdf
*/
w2[0] = wordl2[0] | wordr2[0]; u32 p0[4];
w2[1] = wordl2[1] | wordr2[1]; u32 p1[4];
w2[2] = wordl2[2] | wordr2[2]; u32 p2[4];
w2[3] = wordl2[3] | wordr2[3]; u32 p3[4];
u32 w3[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;
w3[0] = wordl3[0] | wordr3[0]; switch_buffer_by_offset_le (p0, p1, p2, p3, pw_len);
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);
// add password // add password
// truncate at 32 is wanted, not a bug! // truncate at 32 is wanted, not a bug!
// add o_buf // add o_buf
w0_t[0] |= w0[0]; w0[0] |= p0[0];
w0_t[1] |= w0[1]; w0[1] |= p0[1];
w0_t[2] |= w0[2]; w0[2] |= p0[2];
w0_t[3] |= w0[3]; w0[3] |= p0[3];
w1_t[0] |= w1[0]; w1[0] |= p1[0];
w1_t[1] |= w1[1]; w1[1] |= p1[1];
w1_t[2] |= w1[2]; w1[2] |= p1[2];
w1_t[3] |= w1[3]; w1[3] |= p1[3];
w2_t[0] = o_buf[0]; w2[0] = o_buf[0];
w2_t[1] = o_buf[1]; w2[1] = o_buf[1];
w2_t[2] = o_buf[2]; w2[2] = o_buf[2];
w2_t[3] = o_buf[3]; w2[3] = o_buf[3];
w3_t[0] = o_buf[4]; w3[0] = o_buf[4];
w3_t[1] = o_buf[5]; w3[1] = o_buf[5];
w3_t[2] = o_buf[6]; w3[2] = o_buf[6];
w3_t[3] = o_buf[7]; w3[3] = o_buf[7];
u32 digest[4]; u32 digest[4];
@ -332,36 +313,33 @@ __kernel void m10420_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[2] = MD5M_C; digest[2] = MD5M_C;
digest[3] = MD5M_D; digest[3] = MD5M_D;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest); md5_transform (w0, w1, w2, w3, digest);
w0_t[0] = P; w0[0] = P;
w0_t[1] = id_buf[0]; w0[1] = id_buf[0];
w0_t[2] = id_buf[1]; w0[2] = id_buf[1];
w0_t[3] = id_buf[2]; w0[3] = id_buf[2];
w1_t[0] = id_buf[3]; w1[0] = id_buf[3];
w1_t[1] = 0x80; w1[1] = 0x80;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 84 * 8; w3[2] = 84 * 8;
w3_t[3] = 0; w3[3] = 0;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest); md5_transform (w0, w1, w2, w3, digest);
u32 a = digest[0]; u32x a = digest[0];
u32 b = digest[1] & 0xff; u32x b = digest[1] & 0xff;
u32x c = 0;
u32x d = 0;
const u32 r0 = a; COMPARE_M_SIMD (a, b, c, d);
const u32 r1 = b;
const u32 r2 = 0;
const u32 r3 = 0;
#include COMPARE_M
} }
} }
@ -389,55 +367,20 @@ __kernel void m10420_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
u32 wordl0[4]; u32 pw_buf0[4];
u32 pw_buf1[4];
wordl0[0] = pws[gid].i[ 0]; pw_buf0[0] = pws[gid].i[0];
wordl0[1] = pws[gid].i[ 1]; pw_buf0[1] = pws[gid].i[1];
wordl0[2] = pws[gid].i[ 2]; pw_buf0[2] = pws[gid].i[2];
wordl0[3] = pws[gid].i[ 3]; pw_buf0[3] = pws[gid].i[3];
pw_buf1[0] = pws[gid].i[4];
u32 wordl1[4]; pw_buf1[1] = pws[gid].i[5];
pw_buf1[2] = pws[gid].i[6];
wordl1[0] = pws[gid].i[ 4]; pw_buf1[3] = pws[gid].i[7];
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;
const u32 pw_l_len = pws[gid].pw_len; const u32 pw_l_len = pws[gid].pw_len;
if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
{
append_0x80_2x4 (wordl0, wordl1, pw_l_len);
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/**
* 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 * U_buf
*/ */
@ -462,123 +405,139 @@ __kernel void m10420_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
id_buf[2] = pdf_bufs[salt_pos].id_buf[2]; id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
id_buf[3] = pdf_bufs[salt_pos].id_buf[3]; id_buf[3] = pdf_bufs[salt_pos].id_buf[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 * 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 u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
const u32 pw_len = pw_l_len + pw_r_len; const u32x pw_len = pw_l_len + pw_r_len;
u32 wordr0[4]; /**
* concat password candidate
*/
wordr0[0] = combs_buf[il_pos].i[0]; u32x wordl0[4] = { 0 };
wordr0[1] = combs_buf[il_pos].i[1]; u32x wordl1[4] = { 0 };
wordr0[2] = combs_buf[il_pos].i[2]; u32x wordl2[4] = { 0 };
wordr0[3] = combs_buf[il_pos].i[3]; u32x wordl3[4] = { 0 };
u32 wordr1[4]; 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];
wordr1[0] = combs_buf[il_pos].i[4]; u32x wordr0[4] = { 0 };
wordr1[1] = combs_buf[il_pos].i[5]; u32x wordr1[4] = { 0 };
wordr1[2] = combs_buf[il_pos].i[6]; u32x wordr2[4] = { 0 };
wordr1[3] = combs_buf[il_pos].i[7]; u32x wordr3[4] = { 0 };
u32 wordr2[4]; wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr2[0] = 0; wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr2[1] = 0; wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr2[2] = 0; wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr2[3] = 0; wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
u32 wordr3[4]; wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
wordr3[0] = 0;
wordr3[1] = 0;
wordr3[2] = 0;
wordr3[3] = 0;
if (combs_mode == COMBINATOR_MODE_BASE_LEFT) 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[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1]; w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2]; w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3]; w0[3] = wordl0[3] | wordr0[3];
u32 w1[4];
w1[0] = wordl1[0] | wordr1[0]; w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1]; w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2]; w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3]; w1[3] = wordl1[3] | wordr1[3];
w2[0] = 0;
w2[1] = 0;
w2[2] = 0;
w2[3] = 0;
w3[0] = 0;
w3[1] = 0;
w3[2] = 0;
w3[3] = 0;
u32 w2[4]; /**
* pdf
*/
w2[0] = wordl2[0] | wordr2[0]; u32 p0[4];
w2[1] = wordl2[1] | wordr2[1]; u32 p1[4];
w2[2] = wordl2[2] | wordr2[2]; u32 p2[4];
w2[3] = wordl2[3] | wordr2[3]; u32 p3[4];
u32 w3[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;
w3[0] = wordl3[0] | wordr3[0]; switch_buffer_by_offset_le (p0, p1, p2, p3, pw_len);
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);
// add password // add password
// truncate at 32 is wanted, not a bug! // truncate at 32 is wanted, not a bug!
// add o_buf // add o_buf
w0_t[0] |= w0[0]; w0[0] |= p0[0];
w0_t[1] |= w0[1]; w0[1] |= p0[1];
w0_t[2] |= w0[2]; w0[2] |= p0[2];
w0_t[3] |= w0[3]; w0[3] |= p0[3];
w1_t[0] |= w1[0]; w1[0] |= p1[0];
w1_t[1] |= w1[1]; w1[1] |= p1[1];
w1_t[2] |= w1[2]; w1[2] |= p1[2];
w1_t[3] |= w1[3]; w1[3] |= p1[3];
w2_t[0] = o_buf[0]; w2[0] = o_buf[0];
w2_t[1] = o_buf[1]; w2[1] = o_buf[1];
w2_t[2] = o_buf[2]; w2[2] = o_buf[2];
w2_t[3] = o_buf[3]; w2[3] = o_buf[3];
w3_t[0] = o_buf[4]; w3[0] = o_buf[4];
w3_t[1] = o_buf[5]; w3[1] = o_buf[5];
w3_t[2] = o_buf[6]; w3[2] = o_buf[6];
w3_t[3] = o_buf[7]; w3[3] = o_buf[7];
u32 digest[4]; u32 digest[4];
@ -587,36 +546,33 @@ __kernel void m10420_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
digest[2] = MD5M_C; digest[2] = MD5M_C;
digest[3] = MD5M_D; digest[3] = MD5M_D;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest); md5_transform (w0, w1, w2, w3, digest);
w0_t[0] = P; w0[0] = P;
w0_t[1] = id_buf[0]; w0[1] = id_buf[0];
w0_t[2] = id_buf[1]; w0[2] = id_buf[1];
w0_t[3] = id_buf[2]; w0[3] = id_buf[2];
w1_t[0] = id_buf[3]; w1[0] = id_buf[3];
w1_t[1] = 0x80; w1[1] = 0x80;
w1_t[2] = 0; w1[2] = 0;
w1_t[3] = 0; w1[3] = 0;
w2_t[0] = 0; w2[0] = 0;
w2_t[1] = 0; w2[1] = 0;
w2_t[2] = 0; w2[2] = 0;
w2_t[3] = 0; w2[3] = 0;
w3_t[0] = 0; w3[0] = 0;
w3_t[1] = 0; w3[1] = 0;
w3_t[2] = 84 * 8; w3[2] = 84 * 8;
w3_t[3] = 0; w3[3] = 0;
md5_transform (w0_t, w1_t, w2_t, w3_t, digest); md5_transform (w0, w1, w2, w3, digest);
u32 a = digest[0]; u32x a = digest[0];
u32 b = digest[1] & 0xff; u32x b = digest[1] & 0xff;
u32x c = 0;
u32x d = 0;
const u32 r0 = a; COMPARE_S_SIMD (a, b, c, d);
const u32 r1 = b;
const u32 r2 = 0;
const u32 r3 = 0;
#include COMPARE_S
} }
} }

216
OpenCL/m10420_a3.cl
View File

@ -163,56 +163,80 @@ static void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
id_buf[2] = pdf_bufs[salt_pos].id_buf[2]; id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
id_buf[3] = pdf_bufs[salt_pos].id_buf[3]; 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 * loop
*/ */
u32 w0l = w0[0]; u32 w0l = w0[0];
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE) for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
{ {
const u32x w0r = ix_create_bft (bfs_buf, il_pos); const u32 w0r = ix_create_bft (bfs_buf, il_pos);
const u32x w0lr = w0l | w0r; w0[0] = w0l | w0r;
u32x w0_t[4]; /**
u32x w1_t[4]; * pdf
u32x w2_t[4]; */
u32x w3_t[4];
// max length supported by pdf11 is 32 u32 w0_t[4];
u32 w1_t[4];
w0_t[0] = padding[0]; u32 w2_t[4];
w0_t[1] = padding[1]; u32 w3_t[4];
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);
// add password // add password
// truncate at 32 is wanted, not a bug! // truncate at 32 is wanted, not a bug!
// add o_buf // add o_buf
w0_t[0] |= w0lr; w0_t[0] = w0[0];
w0_t[1] |= w0[1]; w0_t[1] = w0[1];
w0_t[2] |= w0[2]; w0_t[2] = w0[2];
w0_t[3] |= w0[3]; w0_t[3] = w0[3];
w1_t[0] |= w1[0]; w1_t[0] = w1[0];
w1_t[1] |= w1[1]; w1_t[1] = w1[1];
w1_t[2] |= w1[2]; w1_t[2] = w1[2];
w1_t[3] |= w1[3]; w1_t[3] = w1[3];
w2_t[0] = o_buf[0]; w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1]; w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2]; w2_t[2] = o_buf[2];
@ -222,7 +246,7 @@ static void m10420m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
w3_t[2] = o_buf[6]; w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7]; w3_t[3] = o_buf[7];
u32x digest[4]; u32 digest[4];
digest[0] = MD5M_A; digest[0] = MD5M_A;
digest[1] = MD5M_B; digest[1] = MD5M_B;
@ -268,18 +292,6 @@ static void m10420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
const u32 gid = get_global_id (0); const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
/**
* 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 * U_buf
*/ */
@ -304,56 +316,92 @@ static void m10420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
id_buf[2] = pdf_bufs[salt_pos].id_buf[2]; id_buf[2] = pdf_bufs[salt_pos].id_buf[2];
id_buf[3] = pdf_bufs[salt_pos].id_buf[3]; 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 * loop
*/ */
u32 w0l = w0[0]; u32 w0l = w0[0];
for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE) for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
{ {
const u32x w0r = ix_create_bft (bfs_buf, il_pos); const u32 w0r = ix_create_bft (bfs_buf, il_pos);
const u32x w0lr = w0l | w0r; w0[0] = w0l | w0r;
u32x w0_t[4]; /**
u32x w1_t[4]; * pdf
u32x w2_t[4]; */
u32x w3_t[4];
// max length supported by pdf11 is 32 u32 w0_t[4];
u32 w1_t[4];
w0_t[0] = padding[0]; u32 w2_t[4];
w0_t[1] = padding[1]; u32 w3_t[4];
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);
// add password // add password
// truncate at 32 is wanted, not a bug! // truncate at 32 is wanted, not a bug!
// add o_buf // add o_buf
w0_t[0] |= w0lr; w0_t[0] = w0[0];
w0_t[1] |= w0[1]; w0_t[1] = w0[1];
w0_t[2] |= w0[2]; w0_t[2] = w0[2];
w0_t[3] |= w0[3]; w0_t[3] = w0[3];
w1_t[0] |= w1[0]; w1_t[0] = w1[0];
w1_t[1] |= w1[1]; w1_t[1] = w1[1];
w1_t[2] |= w1[2]; w1_t[2] = w1[2];
w1_t[3] |= w1[3]; w1_t[3] = w1[3];
w2_t[0] = o_buf[0]; w2_t[0] = o_buf[0];
w2_t[1] = o_buf[1]; w2_t[1] = o_buf[1];
w2_t[2] = o_buf[2]; w2_t[2] = o_buf[2];
@ -363,7 +411,7 @@ static void m10420s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
w3_t[2] = o_buf[6]; w3_t[2] = o_buf[6];
w3_t[3] = o_buf[7]; w3_t[3] = o_buf[7];
u32x digest[4]; u32 digest[4];
digest[0] = MD5M_A; digest[0] = MD5M_A;
digest[1] = MD5M_B; digest[1] = MD5M_B;