arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-11-29 19:38:18 +00:00
Code Issues Releases Wiki Activity

Cleanup -m 27xx kernels to latest standard

Browse Source
This commit is contained in:
Jens Steube 2016-03-19 17:07:28 +01:00
parent 5d67c91a2e
commit 9f044cbdbe
3 changed files with 614 additions and 589 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

237
OpenCL/m02710_a0.cl
View File

@ -46,45 +46,6 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
const u32 lsz = get_local_size (0); const u32 lsz = get_local_size (0);
/**
* base
*/
u32 pw_buf0[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];
pw_buf1[3] = pws[gid].i[ 7];
const u32 pw_len = pws[gid].pw_len;
/**
* salt
*/
u32 s[8];
s[0] = salt_bufs[salt_pos].salt_buf[0];
s[1] = salt_bufs[salt_pos].salt_buf[1];
s[2] = salt_bufs[salt_pos].salt_buf[2];
s[3] = salt_bufs[salt_pos].salt_buf[3];
s[4] = salt_bufs[salt_pos].salt_buf[4];
s[5] = salt_bufs[salt_pos].salt_buf[5];
s[6] = salt_bufs[salt_pos].salt_buf[6];
s[7] = salt_bufs[salt_pos].salt_buf[7];
const u32 salt_len = salt_bufs[salt_pos].salt_len;
const u32 r_14 = (32 + salt_len) * 8;
/** /**
* bin2asc table * bin2asc table
*/ */
@ -104,6 +65,52 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
/**
* base
*/
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];
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_len = pws[gid].pw_len;
/**
* salt
*/
u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
u32 salt_buf3[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
salt_buf2[0] = 0;
salt_buf2[1] = 0;
salt_buf2[2] = 0;
salt_buf2[3] = 0;
salt_buf3[0] = 0;
salt_buf3[1] = 0;
salt_buf3[2] = 0;
salt_buf3[3] = 0;
const u32 salt_len = salt_bufs[salt_pos].salt_len;
/** /**
* loop * loop
*/ */
@ -120,6 +127,7 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
append_0x80_2x4_VV (w0, w1, out_len); append_0x80_2x4_VV (w0, w1, out_len);
w3[2] = out_len * 8; w3[2] = out_len * 8;
w3[3] = 0;
u32x a = MD5M_A; u32x a = MD5M_A;
u32x b = MD5M_B; u32x b = MD5M_B;
@ -216,15 +224,14 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0 w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16; | uint_to_hex_lower8 ((d >> 24) & 255) << 16;
w2[0] = s[0]; w2[0] = salt_buf0[0];
w2[1] = s[1]; w2[1] = salt_buf0[1];
w2[2] = s[2]; w2[2] = salt_buf0[2];
w2[3] = s[3]; w2[3] = salt_buf0[3];
w3[0] = salt_buf1[0];
w3[0] = s[4]; w3[1] = salt_buf1[1];
w3[1] = s[5]; w3[2] = salt_buf1[2];
w3[2] = s[6]; w3[3] = salt_buf1[3];
w3[3] = s[7];
a = MD5M_A; a = MD5M_A;
b = MD5M_B; b = MD5M_B;
@ -304,6 +311,8 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32x r_c = c + MD5M_C; const u32x r_c = c + MD5M_C;
const u32x r_d = d + MD5M_D; const u32x r_d = d + MD5M_D;
const u32x r_14 = (32 + salt_len) * 8;
a = r_a; a = r_a;
b = r_b; b = r_b;
c = r_c; c = r_c;
@ -404,57 +413,6 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
const u32 lsz = get_local_size (0); const u32 lsz = get_local_size (0);
/**
* base
*/
u32 pw_buf0[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];
pw_buf1[3] = pws[gid].i[ 7];
const u32 pw_len = pws[gid].pw_len;
/**
* salt
*/
u32 s[8];
s[0] = salt_bufs[salt_pos].salt_buf[0];
s[1] = salt_bufs[salt_pos].salt_buf[1];
s[2] = salt_bufs[salt_pos].salt_buf[2];
s[3] = salt_bufs[salt_pos].salt_buf[3];
s[4] = salt_bufs[salt_pos].salt_buf[4];
s[5] = salt_bufs[salt_pos].salt_buf[5];
s[6] = salt_bufs[salt_pos].salt_buf[6];
s[7] = salt_bufs[salt_pos].salt_buf[7];
u32 salt_len = salt_bufs[salt_pos].salt_len;
const u32 r_14 = (32 + salt_len) * 8;
/**
* 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]
};
/** /**
* bin2asc table * bin2asc table
*/ */
@ -474,6 +432,64 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
/**
* base
*/
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];
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_len = pws[gid].pw_len;
/**
* salt
*/
u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
u32 salt_buf3[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
salt_buf2[0] = 0;
salt_buf2[1] = 0;
salt_buf2[2] = 0;
salt_buf2[3] = 0;
salt_buf3[0] = 0;
salt_buf3[1] = 0;
salt_buf3[2] = 0;
salt_buf3[3] = 0;
const u32 salt_len = salt_bufs[salt_pos].salt_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]
};
/** /**
* loop * loop
*/ */
@ -490,6 +506,7 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
append_0x80_2x4_VV (w0, w1, out_len); append_0x80_2x4_VV (w0, w1, out_len);
w3[2] = out_len * 8; w3[2] = out_len * 8;
w3[3] = 0;
u32x a = MD5M_A; u32x a = MD5M_A;
u32x b = MD5M_B; u32x b = MD5M_B;
@ -586,15 +603,14 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0 w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16; | uint_to_hex_lower8 ((d >> 24) & 255) << 16;
w2[0] = s[0]; w2[0] = salt_buf0[0];
w2[1] = s[1]; w2[1] = salt_buf0[1];
w2[2] = s[2]; w2[2] = salt_buf0[2];
w2[3] = s[3]; w2[3] = salt_buf0[3];
w3[0] = salt_buf1[0];
w3[0] = s[4]; w3[1] = salt_buf1[1];
w3[1] = s[5]; w3[2] = salt_buf1[2];
w3[2] = s[6]; w3[3] = salt_buf1[3];
w3[3] = s[7];
a = MD5M_A; a = MD5M_A;
b = MD5M_B; b = MD5M_B;
@ -674,6 +690,8 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32x r_c = c + MD5M_C; const u32x r_c = c + MD5M_C;
const u32x r_d = d + MD5M_D; const u32x r_d = d + MD5M_D;
const u32x r_14 = (32 + salt_len) * 8;
a = r_a; a = r_a;
b = r_b; b = r_b;
c = r_c; c = r_c;
@ -743,6 +761,9 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
MD5_STEP0(MD5_I , c, d, a, b, MD5C3a, MD5S32); MD5_STEP0(MD5_I , c, d, a, b, MD5C3a, MD5S32);
MD5_STEP0(MD5_I , b, c, d, a, MD5C3b, MD5S33); MD5_STEP0(MD5_I , b, c, d, a, MD5C3b, MD5S33);
MD5_STEP0(MD5_I , a, b, c, d, MD5C3c, MD5S30); MD5_STEP0(MD5_I , a, b, c, d, MD5C3c, MD5S30);
if (MATCHES_NONE_VS ((a + r_a), search[0])) continue;
MD5_STEP0(MD5_I , d, a, b, c, MD5C3d, MD5S31); MD5_STEP0(MD5_I , d, a, b, c, MD5C3d, MD5S31);
MD5_STEP0(MD5_I , c, d, a, b, MD5C3e, MD5S32); MD5_STEP0(MD5_I , c, d, a, b, MD5C3e, MD5S32);
MD5_STEP0(MD5_I , b, c, d, a, MD5C3f, MD5S33); MD5_STEP0(MD5_I , b, c, d, a, MD5C3f, MD5S33);

786
OpenCL/m02710_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"
@ -16,11 +18,19 @@
#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" #if VECT_SIZE == 1
#define COMPARE_M "OpenCL/check_multi_comp4.c" #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i)])
#elif VECT_SIZE == 2
#define uint_to_hex_lower8(i) l_bin2asc[(i)] #define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1])
#elif VECT_SIZE == 4
#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3])
#elif VECT_SIZE == 8
#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7])
#elif VECT_SIZE == 16
#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
#endif
__kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max) __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
{ {
@ -32,66 +42,6 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
const u32 lsz = get_local_size (0); const u32 lsz = get_local_size (0);
/**
* base
*/
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 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;
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);
}
/**
* salt
*/
u32 s[8];
s[0] = salt_bufs[salt_pos].salt_buf[0];
s[1] = salt_bufs[salt_pos].salt_buf[1];
s[2] = salt_bufs[salt_pos].salt_buf[2];
s[3] = salt_bufs[salt_pos].salt_buf[3];
s[4] = salt_bufs[salt_pos].salt_buf[4];
s[5] = salt_bufs[salt_pos].salt_buf[5];
s[6] = salt_bufs[salt_pos].salt_buf[6];
s[7] = salt_bufs[salt_pos].salt_buf[7];
const u32 salt_len = salt_bufs[salt_pos].salt_len;
const u32 r_14 = (32 + salt_len) * 8;
/** /**
* bin2asc table * bin2asc table
*/ */
@ -111,81 +61,133 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
/**
* base
*/
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];
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;
/**
* salt
*/
u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
u32 salt_buf3[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
salt_buf2[0] = 0;
salt_buf2[1] = 0;
salt_buf2[2] = 0;
salt_buf2[3] = 0;
salt_buf3[0] = 0;
salt_buf3[1] = 0;
salt_buf3[2] = 0;
salt_buf3[3] = 0;
const u32 salt_len = salt_bufs[salt_pos].salt_len;
/** /**
* 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];
u32 w2[4];
w2[0] = wordl2[0] | wordr2[0]; w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1]; w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2]; w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3]; w2[3] = wordl2[3] | wordr2[3];
u32 w3[4];
w3[0] = wordl3[0] | wordr3[0]; w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1]; w3[1] = wordl3[1] | wordr3[1];
w3[2] = pw_len * 8; w3[2] = pw_len * 8;
w3[3] = 0; w3[3] = 0;
u32 a = MD5M_A; /**
u32 b = MD5M_B; * md5
u32 c = MD5M_C; */
u32 d = MD5M_D;
u32x a = MD5M_A;
u32x b = MD5M_B;
u32x c = MD5M_C;
u32x d = MD5M_D;
MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
@ -260,109 +262,111 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
c += MD5M_C; c += MD5M_C;
d += MD5M_D; d += MD5M_D;
const u32 w0_t = uint_to_hex_lower8 ((a >> 0) & 255) << 0 w0[0] = uint_to_hex_lower8 ((a >> 0) & 255) << 0
| uint_to_hex_lower8 ((a >> 8) & 255) << 16; | uint_to_hex_lower8 ((a >> 8) & 255) << 16;
const u32 w1_t = uint_to_hex_lower8 ((a >> 16) & 255) << 0 w0[1] = uint_to_hex_lower8 ((a >> 16) & 255) << 0
| uint_to_hex_lower8 ((a >> 24) & 255) << 16; | uint_to_hex_lower8 ((a >> 24) & 255) << 16;
const u32 w2_t = uint_to_hex_lower8 ((b >> 0) & 255) << 0 w0[2] = uint_to_hex_lower8 ((b >> 0) & 255) << 0
| uint_to_hex_lower8 ((b >> 8) & 255) << 16; | uint_to_hex_lower8 ((b >> 8) & 255) << 16;
const u32 w3_t = uint_to_hex_lower8 ((b >> 16) & 255) << 0 w0[3] = uint_to_hex_lower8 ((b >> 16) & 255) << 0
| uint_to_hex_lower8 ((b >> 24) & 255) << 16; | uint_to_hex_lower8 ((b >> 24) & 255) << 16;
const u32 w4_t = uint_to_hex_lower8 ((c >> 0) & 255) << 0 w1[0] = uint_to_hex_lower8 ((c >> 0) & 255) << 0
| uint_to_hex_lower8 ((c >> 8) & 255) << 16; | uint_to_hex_lower8 ((c >> 8) & 255) << 16;
const u32 w5_t = uint_to_hex_lower8 ((c >> 16) & 255) << 0 w1[1] = uint_to_hex_lower8 ((c >> 16) & 255) << 0
| uint_to_hex_lower8 ((c >> 24) & 255) << 16; | uint_to_hex_lower8 ((c >> 24) & 255) << 16;
const u32 w6_t = uint_to_hex_lower8 ((d >> 0) & 255) << 0 w1[2] = uint_to_hex_lower8 ((d >> 0) & 255) << 0
| uint_to_hex_lower8 ((d >> 8) & 255) << 16; | uint_to_hex_lower8 ((d >> 8) & 255) << 16;
const u32 w7_t = uint_to_hex_lower8 ((d >> 16) & 255) << 0 w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16; | uint_to_hex_lower8 ((d >> 24) & 255) << 16;
const u32 w8_t = s[0]; w2[0] = salt_buf0[0];
const u32 w9_t = s[1]; w2[1] = salt_buf0[1];
const u32 wa_t = s[2]; w2[2] = salt_buf0[2];
const u32 wb_t = s[3]; w2[3] = salt_buf0[3];
const u32 wc_t = s[4]; w3[0] = salt_buf1[0];
const u32 wd_t = s[5]; w3[1] = salt_buf1[1];
const u32 we_t = s[6]; w3[2] = salt_buf1[2];
const u32 wf_t = s[7]; w3[3] = salt_buf1[3];
a = MD5M_A; a = MD5M_A;
b = MD5M_B; b = MD5M_B;
c = MD5M_C; c = MD5M_C;
d = MD5M_D; d = MD5M_D;
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02); MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03); MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02); MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03); MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02); MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03); MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02); MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03); MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10); MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11); MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12); MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13); MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10); MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11); MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12); MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13); MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10); MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11); MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12); MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13); MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10); MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11); MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12); MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13); MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20); MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21); MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22); MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23); MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20); MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21); MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22); MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23); MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20); MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21); MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22); MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23); MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20); MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21); MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22); MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23); MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30); MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31); MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32); MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33); MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30); MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31); MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32); MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33); MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30); MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31); MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32); MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33); MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30); MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31); MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32); MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33); MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
const u32 r_a = a + MD5M_A; const u32x r_a = a + MD5M_A;
const u32 r_b = b + MD5M_B; const u32x r_b = b + MD5M_B;
const u32 r_c = c + MD5M_C; const u32x r_c = c + MD5M_C;
const u32 r_d = d + MD5M_D; const u32x r_d = d + MD5M_D;
const u32x r_14 = (32 + salt_len) * 8;
a = r_a; a = r_a;
b = r_b; b = r_b;
@ -442,12 +446,7 @@ __kernel void m02710_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
c += r_c; c += r_c;
d += r_d; d += r_d;
const u32 r0 = a; COMPARE_M_SIMD (a, d, c, b);
const u32 r1 = d;
const u32 r2 = c;
const u32 r3 = b;
#include COMPARE_M
} }
} }
@ -469,78 +468,6 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
const u32 lid = get_local_id (0); const u32 lid = get_local_id (0);
const u32 lsz = get_local_size (0); const u32 lsz = get_local_size (0);
/**
* base
*/
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 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;
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);
}
/**
* salt
*/
u32 s[8];
s[0] = salt_bufs[salt_pos].salt_buf[0];
s[1] = salt_bufs[salt_pos].salt_buf[1];
s[2] = salt_bufs[salt_pos].salt_buf[2];
s[3] = salt_bufs[salt_pos].salt_buf[3];
s[4] = salt_bufs[salt_pos].salt_buf[4];
s[5] = salt_bufs[salt_pos].salt_buf[5];
s[6] = salt_bufs[salt_pos].salt_buf[6];
s[7] = salt_bufs[salt_pos].salt_buf[7];
const u32 salt_len = salt_bufs[salt_pos].salt_len;
const u32 r_14 = (32 + salt_len) * 8;
/**
* 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]
};
/** /**
* bin2asc table * bin2asc table
*/ */
@ -560,81 +487,145 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
if (gid >= gid_max) return; if (gid >= gid_max) return;
/**
* base
*/
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];
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;
/**
* salt
*/
u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
u32 salt_buf3[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
salt_buf2[0] = 0;
salt_buf2[1] = 0;
salt_buf2[2] = 0;
salt_buf2[3] = 0;
salt_buf3[0] = 0;
salt_buf3[1] = 0;
salt_buf3[2] = 0;
salt_buf3[3] = 0;
const u32 salt_len = salt_bufs[salt_pos].salt_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]
};
/** /**
* 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];
u32 w2[4];
w2[0] = wordl2[0] | wordr2[0]; w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1]; w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2]; w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3]; w2[3] = wordl2[3] | wordr2[3];
u32 w3[4];
w3[0] = wordl3[0] | wordr3[0]; w3[0] = wordl3[0] | wordr3[0];
w3[1] = wordl3[1] | wordr3[1]; w3[1] = wordl3[1] | wordr3[1];
w3[2] = pw_len * 8; w3[2] = pw_len * 8;
w3[3] = 0; w3[3] = 0;
u32 a = MD5M_A; /**
u32 b = MD5M_B; * md5
u32 c = MD5M_C; */
u32 d = MD5M_D;
u32x a = MD5M_A;
u32x b = MD5M_B;
u32x c = MD5M_C;
u32x d = MD5M_D;
MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
@ -709,109 +700,111 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
c += MD5M_C; c += MD5M_C;
d += MD5M_D; d += MD5M_D;
const u32 w0_t = uint_to_hex_lower8 ((a >> 0) & 255) << 0 w0[0] = uint_to_hex_lower8 ((a >> 0) & 255) << 0
| uint_to_hex_lower8 ((a >> 8) & 255) << 16; | uint_to_hex_lower8 ((a >> 8) & 255) << 16;
const u32 w1_t = uint_to_hex_lower8 ((a >> 16) & 255) << 0 w0[1] = uint_to_hex_lower8 ((a >> 16) & 255) << 0
| uint_to_hex_lower8 ((a >> 24) & 255) << 16; | uint_to_hex_lower8 ((a >> 24) & 255) << 16;
const u32 w2_t = uint_to_hex_lower8 ((b >> 0) & 255) << 0 w0[2] = uint_to_hex_lower8 ((b >> 0) & 255) << 0
| uint_to_hex_lower8 ((b >> 8) & 255) << 16; | uint_to_hex_lower8 ((b >> 8) & 255) << 16;
const u32 w3_t = uint_to_hex_lower8 ((b >> 16) & 255) << 0 w0[3] = uint_to_hex_lower8 ((b >> 16) & 255) << 0
| uint_to_hex_lower8 ((b >> 24) & 255) << 16; | uint_to_hex_lower8 ((b >> 24) & 255) << 16;
const u32 w4_t = uint_to_hex_lower8 ((c >> 0) & 255) << 0 w1[0] = uint_to_hex_lower8 ((c >> 0) & 255) << 0
| uint_to_hex_lower8 ((c >> 8) & 255) << 16; | uint_to_hex_lower8 ((c >> 8) & 255) << 16;
const u32 w5_t = uint_to_hex_lower8 ((c >> 16) & 255) << 0 w1[1] = uint_to_hex_lower8 ((c >> 16) & 255) << 0
| uint_to_hex_lower8 ((c >> 24) & 255) << 16; | uint_to_hex_lower8 ((c >> 24) & 255) << 16;
const u32 w6_t = uint_to_hex_lower8 ((d >> 0) & 255) << 0 w1[2] = uint_to_hex_lower8 ((d >> 0) & 255) << 0
| uint_to_hex_lower8 ((d >> 8) & 255) << 16; | uint_to_hex_lower8 ((d >> 8) & 255) << 16;
const u32 w7_t = uint_to_hex_lower8 ((d >> 16) & 255) << 0 w1[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16; | uint_to_hex_lower8 ((d >> 24) & 255) << 16;
const u32 w8_t = s[0]; w2[0] = salt_buf0[0];
const u32 w9_t = s[1]; w2[1] = salt_buf0[1];
const u32 wa_t = s[2]; w2[2] = salt_buf0[2];
const u32 wb_t = s[3]; w2[3] = salt_buf0[3];
const u32 wc_t = s[4]; w3[0] = salt_buf1[0];
const u32 wd_t = s[5]; w3[1] = salt_buf1[1];
const u32 we_t = s[6]; w3[2] = salt_buf1[2];
const u32 wf_t = s[7]; w3[3] = salt_buf1[3];
a = MD5M_A; a = MD5M_A;
b = MD5M_B; b = MD5M_B;
c = MD5M_C; c = MD5M_C;
d = MD5M_D; d = MD5M_D;
MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02); MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03); MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02); MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03); MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02); MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03); MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00); MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01); MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02); MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03); MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10); MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11); MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12); MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13); MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10); MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11); MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12); MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13); MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10); MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11); MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12); MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13); MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10); MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11); MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12); MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13); MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20); MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21); MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22); MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23); MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20); MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21); MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22); MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23); MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20); MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21); MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22); MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23); MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20); MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21); MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22); MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23); MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30); MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31); MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32); MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33); MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30); MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31); MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32); MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33); MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30); MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31); MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32); MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33); MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30); MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31); MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32); MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33); MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
const u32 r_a = a + MD5M_A; const u32x r_a = a + MD5M_A;
const u32 r_b = b + MD5M_B; const u32x r_b = b + MD5M_B;
const u32 r_c = c + MD5M_C; const u32x r_c = c + MD5M_C;
const u32 r_d = d + MD5M_D; const u32x r_d = d + MD5M_D;
const u32x r_14 = (32 + salt_len) * 8;
a = r_a; a = r_a;
b = r_b; b = r_b;
@ -883,7 +876,7 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
MD5_STEP0(MD5_I , b, c, d, a, MD5C3b, MD5S33); MD5_STEP0(MD5_I , b, c, d, a, MD5C3b, MD5S33);
MD5_STEP0(MD5_I , a, b, c, d, MD5C3c, MD5S30); MD5_STEP0(MD5_I , a, b, c, d, MD5C3c, MD5S30);
if (allx ((a + r_a) != search[0])) continue; if (MATCHES_NONE_VS ((a + r_a), search[0])) continue;
MD5_STEP0(MD5_I , d, a, b, c, MD5C3d, MD5S31); MD5_STEP0(MD5_I , d, a, b, c, MD5C3d, MD5S31);
MD5_STEP0(MD5_I , c, d, a, b, MD5C3e, MD5S32); MD5_STEP0(MD5_I , c, d, a, b, MD5C3e, MD5S32);
@ -894,12 +887,7 @@ __kernel void m02710_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf,
c += r_c; c += r_c;
d += r_d; d += r_d;
const u32 r0 = a; COMPARE_S_SIMD (a, d, c, b);
const u32 r1 = d;
const u32 r2 = c;
const u32 r3 = b;
#include COMPARE_S
} }
} }

148
OpenCL/m02710_a3.cl
View File

@ -48,21 +48,30 @@ static void m02710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
* salt * salt
*/ */
u32 s[8]; u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
u32 salt_buf3[4];
s[0] = salt_bufs[salt_pos].salt_buf[0]; salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
s[1] = salt_bufs[salt_pos].salt_buf[1]; salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
s[2] = salt_bufs[salt_pos].salt_buf[2]; salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
s[3] = salt_bufs[salt_pos].salt_buf[3]; salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
s[4] = salt_bufs[salt_pos].salt_buf[4]; salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
s[5] = salt_bufs[salt_pos].salt_buf[5]; salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
s[6] = salt_bufs[salt_pos].salt_buf[6]; salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
s[7] = salt_bufs[salt_pos].salt_buf[7]; salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
salt_buf2[0] = 0;
salt_buf2[1] = 0;
salt_buf2[2] = 0;
salt_buf2[3] = 0;
salt_buf3[0] = 0;
salt_buf3[1] = 0;
salt_buf3[2] = 0;
salt_buf3[3] = 0;
const u32 salt_len = salt_bufs[salt_pos].salt_len; const u32 salt_len = salt_bufs[salt_pos].salt_len;
const u32 r_14 = (32 + salt_len) * 8;
/** /**
* loop * loop
*/ */
@ -76,33 +85,31 @@ static void m02710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
const u32x w0lr = w0l | w0r; const u32x w0lr = w0l | w0r;
u32x w0_t[4]; u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
w0_t[0] = w0lr; w0_t[0] = w0lr;
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];
u32x w1_t[4];
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];
u32x w2_t[4];
w2_t[0] = w2[0]; w2_t[0] = w2[0];
w2_t[1] = w2[1]; w2_t[1] = w2[1];
w2_t[2] = w2[2]; w2_t[2] = w2[2];
w2_t[3] = w2[3]; w2_t[3] = w2[3];
u32x w3_t[4];
w3_t[0] = w3[0]; w3_t[0] = w3[0];
w3_t[1] = w3[1]; w3_t[1] = w3[1];
w3_t[2] = w3[2]; w3_t[2] = w3[2];
w3_t[3] = w3[3]; w3_t[3] = w3[3];
/**
* md5
*/
u32x a = MD5M_A; u32x a = MD5M_A;
u32x b = MD5M_B; u32x b = MD5M_B;
u32x c = MD5M_C; u32x c = MD5M_C;
@ -198,15 +205,14 @@ static void m02710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0 w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16; | uint_to_hex_lower8 ((d >> 24) & 255) << 16;
w2_t[0] = s[0]; w2_t[0] = salt_buf0[0];
w2_t[1] = s[1]; w2_t[1] = salt_buf0[1];
w2_t[2] = s[2]; w2_t[2] = salt_buf0[2];
w2_t[3] = s[3]; w2_t[3] = salt_buf0[3];
w3_t[0] = salt_buf1[0];
w3_t[0] = s[4]; w3_t[1] = salt_buf1[1];
w3_t[1] = s[5]; w3_t[2] = salt_buf1[2];
w3_t[2] = s[6]; w3_t[3] = salt_buf1[3];
w3_t[3] = s[7];
a = MD5M_A; a = MD5M_A;
b = MD5M_B; b = MD5M_B;
@ -286,6 +292,8 @@ static void m02710m (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
const u32x r_c = c + MD5M_C; const u32x r_c = c + MD5M_C;
const u32x r_d = d + MD5M_D; const u32x r_d = d + MD5M_D;
const u32x r_14 = (32 + salt_len) * 8;
a = r_a; a = r_a;
b = r_b; b = r_b;
c = r_c; c = r_c;
@ -377,6 +385,34 @@ static void m02710s (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);
/**
* salt
*/
u32 salt_buf0[4];
u32 salt_buf1[4];
u32 salt_buf2[4];
u32 salt_buf3[4];
salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
salt_buf2[0] = 0;
salt_buf2[1] = 0;
salt_buf2[2] = 0;
salt_buf2[3] = 0;
salt_buf3[0] = 0;
salt_buf3[1] = 0;
salt_buf3[2] = 0;
salt_buf3[3] = 0;
const u32 salt_len = salt_bufs[salt_pos].salt_len;
/** /**
* digest * digest
*/ */
@ -389,25 +425,6 @@ static void m02710s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
digests_buf[digests_offset].digest_buf[DGST_R3] digests_buf[digests_offset].digest_buf[DGST_R3]
}; };
/**
* salt
*/
u32 s[8];
s[0] = salt_bufs[salt_pos].salt_buf[0];
s[1] = salt_bufs[salt_pos].salt_buf[1];
s[2] = salt_bufs[salt_pos].salt_buf[2];
s[3] = salt_bufs[salt_pos].salt_buf[3];
s[4] = salt_bufs[salt_pos].salt_buf[4];
s[5] = salt_bufs[salt_pos].salt_buf[5];
s[6] = salt_bufs[salt_pos].salt_buf[6];
s[7] = salt_bufs[salt_pos].salt_buf[7];
const u32 salt_len = salt_bufs[salt_pos].salt_len;
const u32 r_14 = (32 + salt_len) * 8;
/** /**
* loop * loop
*/ */
@ -421,33 +438,31 @@ static void m02710s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
const u32x w0lr = w0l | w0r; const u32x w0lr = w0l | w0r;
u32x w0_t[4]; u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
w0_t[0] = w0lr; w0_t[0] = w0lr;
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];
u32x w1_t[4];
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];
u32x w2_t[4];
w2_t[0] = w2[0]; w2_t[0] = w2[0];
w2_t[1] = w2[1]; w2_t[1] = w2[1];
w2_t[2] = w2[2]; w2_t[2] = w2[2];
w2_t[3] = w2[3]; w2_t[3] = w2[3];
u32x w3_t[4];
w3_t[0] = w3[0]; w3_t[0] = w3[0];
w3_t[1] = w3[1]; w3_t[1] = w3[1];
w3_t[2] = w3[2]; w3_t[2] = w3[2];
w3_t[3] = w3[3]; w3_t[3] = w3[3];
/**
* md5
*/
u32x a = MD5M_A; u32x a = MD5M_A;
u32x b = MD5M_B; u32x b = MD5M_B;
u32x c = MD5M_C; u32x c = MD5M_C;
@ -543,15 +558,14 @@ static void m02710s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0 w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16; | uint_to_hex_lower8 ((d >> 24) & 255) << 16;
w2_t[0] = s[0]; w2_t[0] = salt_buf0[0];
w2_t[1] = s[1]; w2_t[1] = salt_buf0[1];
w2_t[2] = s[2]; w2_t[2] = salt_buf0[2];
w2_t[3] = s[3]; w2_t[3] = salt_buf0[3];
w3_t[0] = salt_buf1[0];
w3_t[0] = s[4]; w3_t[1] = salt_buf1[1];
w3_t[1] = s[5]; w3_t[2] = salt_buf1[2];
w3_t[2] = s[6]; w3_t[3] = salt_buf1[3];
w3_t[3] = s[7];
a = MD5M_A; a = MD5M_A;
b = MD5M_B; b = MD5M_B;
@ -631,6 +645,8 @@ static void m02710s (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_le
const u32x r_c = c + MD5M_C; const u32x r_c = c + MD5M_C;
const u32x r_d = d + MD5M_D; const u32x r_d = d + MD5M_D;
const u32x r_14 = (32 + salt_len) * 8;
a = r_a; a = r_a;
b = r_b; b = r_b;
c = r_c; c = r_c;