1
0
mirror of https://github.com/hashcat/hashcat.git synced 2024-12-23 07:08:19 +00:00

Cleanup -m 1760 -a 1 kernel to latest standard

This commit is contained in:
Jens Steube 2016-03-19 18:04:50 +01:00
parent b08e6e2371
commit 19c548c36e

View File

@ -5,6 +5,8 @@
#define _SHA512_ #define _SHA512_
#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 u64 k_sha512[80] = __constant u64 k_sha512[80] =
{ {
@ -44,33 +44,33 @@ __constant u64 k_sha512[80] =
SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f, SHA512C4c, SHA512C4d, SHA512C4e, SHA512C4f,
}; };
static void sha512_transform (const u64 w0[4], const u64 w1[4], const u64 w2[4], const u64 w3[4], u64 digest[8]) static void sha512_transform (const u64x w0[4], const u64x w1[4], const u64x w2[4], const u64x w3[4], u64x digest[8])
{ {
u64 w0_t = w0[0]; u64x w0_t = w0[0];
u64 w1_t = w0[1]; u64x w1_t = w0[1];
u64 w2_t = w0[2]; u64x w2_t = w0[2];
u64 w3_t = w0[3]; u64x w3_t = w0[3];
u64 w4_t = w1[0]; u64x w4_t = w1[0];
u64 w5_t = w1[1]; u64x w5_t = w1[1];
u64 w6_t = w1[2]; u64x w6_t = w1[2];
u64 w7_t = w1[3]; u64x w7_t = w1[3];
u64 w8_t = w2[0]; u64x w8_t = w2[0];
u64 w9_t = w2[1]; u64x w9_t = w2[1];
u64 wa_t = w2[2]; u64x wa_t = w2[2];
u64 wb_t = w2[3]; u64x wb_t = w2[3];
u64 wc_t = w3[0]; u64x wc_t = w3[0];
u64 wd_t = w3[1]; u64x wd_t = w3[1];
u64 we_t = w3[2]; u64x we_t = w3[2];
u64 wf_t = w3[3]; u64x wf_t = w3[3];
u64 a = digest[0]; u64x a = digest[0];
u64 b = digest[1]; u64x b = digest[1];
u64 c = digest[2]; u64x c = digest[2];
u64 d = digest[3]; u64x d = digest[3];
u64 e = digest[4]; u64x e = digest[4];
u64 f = digest[5]; u64x f = digest[5];
u64 g = digest[6]; u64x g = digest[6];
u64 h = digest[7]; u64x h = digest[7];
#define ROUND_EXPAND() \ #define ROUND_EXPAND() \
{ \ { \
@ -130,29 +130,29 @@ static void sha512_transform (const u64 w0[4], const u64 w1[4], const u64 w2[4],
digest[7] += h; digest[7] += h;
} }
static void hmac_sha512_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u64 ipad[8], u64 opad[8]) static void hmac_sha512_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8])
{ {
u64 w0_t[4]; u64x w0_t[4];
u64 w1_t[4]; u64x w1_t[4];
u64 w2_t[4]; u64x w2_t[4];
u64 w3_t[4]; u64x w3_t[4];
w0_t[0] = hl32_to_64 (w0[0], w0[1]) ^ 0x3636363636363636; w0_t[0] = hl32_to_64 (w0[0], w0[1]) ^ (u64x) 0x3636363636363636;
w0_t[1] = hl32_to_64 (w0[2], w0[3]) ^ 0x3636363636363636; w0_t[1] = hl32_to_64 (w0[2], w0[3]) ^ (u64x) 0x3636363636363636;
w0_t[2] = hl32_to_64 (w1[0], w1[1]) ^ 0x3636363636363636; w0_t[2] = hl32_to_64 (w1[0], w1[1]) ^ (u64x) 0x3636363636363636;
w0_t[3] = hl32_to_64 (w1[2], w1[3]) ^ 0x3636363636363636; w0_t[3] = hl32_to_64 (w1[2], w1[3]) ^ (u64x) 0x3636363636363636;
w1_t[0] = hl32_to_64 (w2[0], w2[1]) ^ 0x3636363636363636; w1_t[0] = hl32_to_64 (w2[0], w2[1]) ^ (u64x) 0x3636363636363636;
w1_t[1] = hl32_to_64 (w2[2], w2[3]) ^ 0x3636363636363636; w1_t[1] = hl32_to_64 (w2[2], w2[3]) ^ (u64x) 0x3636363636363636;
w1_t[2] = hl32_to_64 (w3[0], w3[1]) ^ 0x3636363636363636; w1_t[2] = hl32_to_64 (w3[0], w3[1]) ^ (u64x) 0x3636363636363636;
w1_t[3] = hl32_to_64 (w3[2], w3[3]) ^ 0x3636363636363636; w1_t[3] = hl32_to_64 (w3[2], w3[3]) ^ (u64x) 0x3636363636363636;
w2_t[0] = 0 ^ 0x3636363636363636; w2_t[0] = (u64x) 0x3636363636363636;
w2_t[1] = 0 ^ 0x3636363636363636; w2_t[1] = (u64x) 0x3636363636363636;
w2_t[2] = 0 ^ 0x3636363636363636; w2_t[2] = (u64x) 0x3636363636363636;
w2_t[3] = 0 ^ 0x3636363636363636; w2_t[3] = (u64x) 0x3636363636363636;
w3_t[0] = 0 ^ 0x3636363636363636; w3_t[0] = (u64x) 0x3636363636363636;
w3_t[1] = 0 ^ 0x3636363636363636; w3_t[1] = (u64x) 0x3636363636363636;
w3_t[2] = 0 ^ 0x3636363636363636; w3_t[2] = (u64x) 0x3636363636363636;
w3_t[3] = 0 ^ 0x3636363636363636; w3_t[3] = (u64x) 0x3636363636363636;
ipad[0] = SHA512M_A; ipad[0] = SHA512M_A;
ipad[1] = SHA512M_B; ipad[1] = SHA512M_B;
@ -165,22 +165,22 @@ static void hmac_sha512_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u64 ipa
sha512_transform (w0_t, w1_t, w2_t, w3_t, ipad); sha512_transform (w0_t, w1_t, w2_t, w3_t, ipad);
w0_t[0] = hl32_to_64 (w0[0], w0[1]) ^ 0x5c5c5c5c5c5c5c5c; w0_t[0] = hl32_to_64 (w0[0], w0[1]) ^ (u64x) 0x5c5c5c5c5c5c5c5c;
w0_t[1] = hl32_to_64 (w0[2], w0[3]) ^ 0x5c5c5c5c5c5c5c5c; w0_t[1] = hl32_to_64 (w0[2], w0[3]) ^ (u64x) 0x5c5c5c5c5c5c5c5c;
w0_t[2] = hl32_to_64 (w1[0], w1[1]) ^ 0x5c5c5c5c5c5c5c5c; w0_t[2] = hl32_to_64 (w1[0], w1[1]) ^ (u64x) 0x5c5c5c5c5c5c5c5c;
w0_t[3] = hl32_to_64 (w1[2], w1[3]) ^ 0x5c5c5c5c5c5c5c5c; w0_t[3] = hl32_to_64 (w1[2], w1[3]) ^ (u64x) 0x5c5c5c5c5c5c5c5c;
w1_t[0] = hl32_to_64 (w2[0], w2[1]) ^ 0x5c5c5c5c5c5c5c5c; w1_t[0] = hl32_to_64 (w2[0], w2[1]) ^ (u64x) 0x5c5c5c5c5c5c5c5c;
w1_t[1] = hl32_to_64 (w2[2], w2[3]) ^ 0x5c5c5c5c5c5c5c5c; w1_t[1] = hl32_to_64 (w2[2], w2[3]) ^ (u64x) 0x5c5c5c5c5c5c5c5c;
w1_t[2] = hl32_to_64 (w3[0], w3[1]) ^ 0x5c5c5c5c5c5c5c5c; w1_t[2] = hl32_to_64 (w3[0], w3[1]) ^ (u64x) 0x5c5c5c5c5c5c5c5c;
w1_t[3] = hl32_to_64 (w3[2], w3[3]) ^ 0x5c5c5c5c5c5c5c5c; w1_t[3] = hl32_to_64 (w3[2], w3[3]) ^ (u64x) 0x5c5c5c5c5c5c5c5c;
w2_t[0] = 0 ^ 0x5c5c5c5c5c5c5c5c; w2_t[0] = (u64x) 0x5c5c5c5c5c5c5c5c;
w2_t[1] = 0 ^ 0x5c5c5c5c5c5c5c5c; w2_t[1] = (u64x) 0x5c5c5c5c5c5c5c5c;
w2_t[2] = 0 ^ 0x5c5c5c5c5c5c5c5c; w2_t[2] = (u64x) 0x5c5c5c5c5c5c5c5c;
w2_t[3] = 0 ^ 0x5c5c5c5c5c5c5c5c; w2_t[3] = (u64x) 0x5c5c5c5c5c5c5c5c;
w3_t[0] = 0 ^ 0x5c5c5c5c5c5c5c5c; w3_t[0] = (u64x) 0x5c5c5c5c5c5c5c5c;
w3_t[1] = 0 ^ 0x5c5c5c5c5c5c5c5c; w3_t[1] = (u64x) 0x5c5c5c5c5c5c5c5c;
w3_t[2] = 0 ^ 0x5c5c5c5c5c5c5c5c; w3_t[2] = (u64x) 0x5c5c5c5c5c5c5c5c;
w3_t[3] = 0 ^ 0x5c5c5c5c5c5c5c5c; w3_t[3] = (u64x) 0x5c5c5c5c5c5c5c5c;
opad[0] = SHA512M_A; opad[0] = SHA512M_A;
opad[1] = SHA512M_B; opad[1] = SHA512M_B;
@ -194,12 +194,12 @@ static void hmac_sha512_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u64 ipa
sha512_transform (w0_t, w1_t, w2_t, w3_t, opad); sha512_transform (w0_t, w1_t, w2_t, w3_t, opad);
} }
static void hmac_sha512_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u64 ipad[8], u64 opad[8], u64 digest[8]) static void hmac_sha512_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u64x ipad[8], u64x opad[8], u64x digest[8])
{ {
u64 w0_t[4]; u64x w0_t[4];
u64 w1_t[4]; u64x w1_t[4];
u64 w2_t[4]; u64x w2_t[4];
u64 w3_t[4]; u64x w3_t[4];
w0_t[0] = hl32_to_64 (w0[0], w0[1]); w0_t[0] = hl32_to_64 (w0[0], w0[1]);
w0_t[1] = hl32_to_64 (w0[2], w0[3]); w0_t[1] = hl32_to_64 (w0[2], w0[3]);
@ -274,93 +274,76 @@ __kernel void m01760_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)
{
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/** /**
* salt * salt
*/ */
u32 salt_buf0[4]; u32 salt_buf0[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];
u32 salt_buf1[4]; u32 salt_buf1[4];
u32 salt_buf2[4];
u32 salt_buf3[4];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4]; salt_buf0[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 0]);
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5]; salt_buf0[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 1]);
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6]; salt_buf0[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 2]);
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7]; salt_buf0[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 3]);
salt_buf1[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 4]);
salt_buf1[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 5]);
salt_buf1[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 6]);
salt_buf1[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 7]);
salt_buf2[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 8]);
salt_buf2[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 9]);
salt_buf2[2] = swap32_S (salt_bufs[salt_pos].salt_buf[10]);
salt_buf2[3] = swap32_S (salt_bufs[salt_pos].salt_buf[11]);
salt_buf3[0] = swap32_S (salt_bufs[salt_pos].salt_buf[12]);
salt_buf3[1] = swap32_S (salt_bufs[salt_pos].salt_buf[13]);
salt_buf3[2] = swap32_S (salt_bufs[salt_pos].salt_buf[14]);
salt_buf3[3] = swap32_S (salt_bufs[salt_pos].salt_buf[15]);
const u32 salt_len = salt_bufs[salt_pos].salt_len;
/** /**
* pads * pads
*/ */
u32 w0_t[4]; u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
w0_t[0] = swap32 (salt_buf0[0]); w0_t[0] = salt_buf0[0];
w0_t[1] = swap32 (salt_buf0[1]); w0_t[1] = salt_buf0[1];
w0_t[2] = swap32 (salt_buf0[2]); w0_t[2] = salt_buf0[2];
w0_t[3] = swap32 (salt_buf0[3]); w0_t[3] = salt_buf0[3];
w1_t[0] = salt_buf1[0];
w1_t[1] = salt_buf1[1];
w1_t[2] = salt_buf1[2];
w1_t[3] = salt_buf1[3];
w2_t[0] = salt_buf2[0];
w2_t[1] = salt_buf2[1];
w2_t[2] = salt_buf2[2];
w2_t[3] = salt_buf2[3];
w3_t[0] = salt_buf3[0];
w3_t[1] = salt_buf3[1];
w3_t[2] = salt_buf3[2];
w3_t[3] = salt_buf3[3];
u32 w1_t[4]; u64x ipad[8];
u64x opad[8];
w1_t[0] = swap32 (salt_buf1[0]);
w1_t[1] = swap32 (salt_buf1[1]);
w1_t[2] = swap32 (salt_buf1[2]);
w1_t[3] = swap32 (salt_buf1[3]);
u32 w2_t[4];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
u32 w3_t[4];
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
u64 ipad[8];
u64 opad[8];
hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad); hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
@ -368,103 +351,104 @@ __kernel void m01760_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];
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] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
append_0x80_4x4_VV (w0, w1, w2, w3, pw_len);
w0[0] = swap32 (w0[0]);
w0[1] = swap32 (w0[1]);
w0[2] = swap32 (w0[2]);
w0[3] = swap32 (w0[3]);
w1[0] = swap32 (w1[0]);
w1[1] = swap32 (w1[1]);
w1[2] = swap32 (w1[2]);
w1[3] = swap32 (w1[3]);
w2[0] = swap32 (w2[0]);
w2[1] = swap32 (w2[1]);
w2[2] = swap32 (w2[2]);
w2[3] = swap32 (w2[3]);
w3[0] = swap32 (w3[0]);
w3[1] = swap32 (w3[1]);
w3[2] = 0; w3[2] = 0;
w3[3] = 0; w3[3] = (128 + pw_len) * 8;
append_0x80_4x4 (w0, w1, w2, w3, pw_len); u64x digest[8];
w0_t[0] = swap32 (w0[0]); hmac_sha512_run (w0, w1, w2, w3, ipad, opad, digest);
w0_t[1] = swap32 (w0[1]);
w0_t[2] = swap32 (w0[2]);
w0_t[3] = swap32 (w0[3]);
w1_t[0] = swap32 (w1[0]);
w1_t[1] = swap32 (w1[1]);
w1_t[2] = swap32 (w1[2]);
w1_t[3] = swap32 (w1[3]);
w2_t[0] = swap32 (w2[0]);
w2_t[1] = swap32 (w2[1]);
w2_t[2] = swap32 (w2[2]);
w2_t[3] = swap32 (w2[3]);
w3_t[0] = swap32 (w3[0]);
w3_t[1] = swap32 (w3[1]);
w3_t[2] = 0;
w3_t[3] = (128 + pw_len) * 8;
u64 digest[8]; const u32x r0 = l32_from_64 (digest[7]);
const u32x r1 = h32_from_64 (digest[7]);
const u32x r2 = l32_from_64 (digest[3]);
const u32x r3 = h32_from_64 (digest[3]);
hmac_sha512_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest); COMPARE_M_SIMD (r0, r1, r2, r3);
const u32 r0 = l32_from_64 (digest[7]);
const u32 r1 = h32_from_64 (digest[7]);
const u32 r2 = l32_from_64 (digest[3]);
const u32 r3 = h32_from_64 (digest[3]);
#include COMPARE_M
} }
} }
@ -492,93 +476,76 @@ __kernel void m01760_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)
{
switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
}
/** /**
* salt * salt
*/ */
u32 salt_buf0[4]; u32 salt_buf0[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];
u32 salt_buf1[4]; u32 salt_buf1[4];
u32 salt_buf2[4];
u32 salt_buf3[4];
salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4]; salt_buf0[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 0]);
salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5]; salt_buf0[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 1]);
salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6]; salt_buf0[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 2]);
salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7]; salt_buf0[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 3]);
salt_buf1[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 4]);
salt_buf1[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 5]);
salt_buf1[2] = swap32_S (salt_bufs[salt_pos].salt_buf[ 6]);
salt_buf1[3] = swap32_S (salt_bufs[salt_pos].salt_buf[ 7]);
salt_buf2[0] = swap32_S (salt_bufs[salt_pos].salt_buf[ 8]);
salt_buf2[1] = swap32_S (salt_bufs[salt_pos].salt_buf[ 9]);
salt_buf2[2] = swap32_S (salt_bufs[salt_pos].salt_buf[10]);
salt_buf2[3] = swap32_S (salt_bufs[salt_pos].salt_buf[11]);
salt_buf3[0] = swap32_S (salt_bufs[salt_pos].salt_buf[12]);
salt_buf3[1] = swap32_S (salt_bufs[salt_pos].salt_buf[13]);
salt_buf3[2] = swap32_S (salt_bufs[salt_pos].salt_buf[14]);
salt_buf3[3] = swap32_S (salt_bufs[salt_pos].salt_buf[15]);
const u32 salt_len = salt_bufs[salt_pos].salt_len;
/** /**
* pads * pads
*/ */
u32 w0_t[4]; u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
w0_t[0] = swap32 (salt_buf0[0]); w0_t[0] = salt_buf0[0];
w0_t[1] = swap32 (salt_buf0[1]); w0_t[1] = salt_buf0[1];
w0_t[2] = swap32 (salt_buf0[2]); w0_t[2] = salt_buf0[2];
w0_t[3] = swap32 (salt_buf0[3]); w0_t[3] = salt_buf0[3];
w1_t[0] = salt_buf1[0];
w1_t[1] = salt_buf1[1];
w1_t[2] = salt_buf1[2];
w1_t[3] = salt_buf1[3];
w2_t[0] = salt_buf2[0];
w2_t[1] = salt_buf2[1];
w2_t[2] = salt_buf2[2];
w2_t[3] = salt_buf2[3];
w3_t[0] = salt_buf3[0];
w3_t[1] = salt_buf3[1];
w3_t[2] = salt_buf3[2];
w3_t[3] = salt_buf3[3];
u32 w1_t[4]; u64x ipad[8];
u64x opad[8];
w1_t[0] = swap32 (salt_buf1[0]);
w1_t[1] = swap32 (salt_buf1[1]);
w1_t[2] = swap32 (salt_buf1[2]);
w1_t[3] = swap32 (salt_buf1[3]);
u32 w2_t[4];
w2_t[0] = 0;
w2_t[1] = 0;
w2_t[2] = 0;
w2_t[3] = 0;
u32 w3_t[4];
w3_t[0] = 0;
w3_t[1] = 0;
w3_t[2] = 0;
w3_t[3] = 0;
u64 ipad[8];
u64 opad[8];
hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad); hmac_sha512_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
@ -598,103 +565,104 @@ __kernel void m01760_s04 (__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];
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] = wordl3[2] | wordr3[2];
w3[3] = wordl3[3] | wordr3[3];
append_0x80_4x4_VV (w0, w1, w2, w3, pw_len);
w0[0] = swap32 (w0[0]);
w0[1] = swap32 (w0[1]);
w0[2] = swap32 (w0[2]);
w0[3] = swap32 (w0[3]);
w1[0] = swap32 (w1[0]);
w1[1] = swap32 (w1[1]);
w1[2] = swap32 (w1[2]);
w1[3] = swap32 (w1[3]);
w2[0] = swap32 (w2[0]);
w2[1] = swap32 (w2[1]);
w2[2] = swap32 (w2[2]);
w2[3] = swap32 (w2[3]);
w3[0] = swap32 (w3[0]);
w3[1] = swap32 (w3[1]);
w3[2] = 0; w3[2] = 0;
w3[3] = 0; w3[3] = (128 + pw_len) * 8;
append_0x80_4x4 (w0, w1, w2, w3, pw_len); u64x digest[8];
w0_t[0] = swap32 (w0[0]); hmac_sha512_run (w0, w1, w2, w3, ipad, opad, digest);
w0_t[1] = swap32 (w0[1]);
w0_t[2] = swap32 (w0[2]);
w0_t[3] = swap32 (w0[3]);
w1_t[0] = swap32 (w1[0]);
w1_t[1] = swap32 (w1[1]);
w1_t[2] = swap32 (w1[2]);
w1_t[3] = swap32 (w1[3]);
w2_t[0] = swap32 (w2[0]);
w2_t[1] = swap32 (w2[1]);
w2_t[2] = swap32 (w2[2]);
w2_t[3] = swap32 (w2[3]);
w3_t[0] = swap32 (w3[0]);
w3_t[1] = swap32 (w3[1]);
w3_t[2] = 0;
w3_t[3] = (128 + pw_len) * 8;
u64 digest[8]; const u32x r0 = l32_from_64 (digest[7]);
const u32x r1 = h32_from_64 (digest[7]);
const u32x r2 = l32_from_64 (digest[3]);
const u32x r3 = h32_from_64 (digest[3]);
hmac_sha512_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest); COMPARE_S_SIMD (r0, r1, r2, r3);
const u32 r0 = l32_from_64 (digest[7]);
const u32 r1 = h32_from_64 (digest[7]);
const u32 r2 = l32_from_64 (digest[3]);
const u32 r3 = h32_from_64 (digest[3]);
#include COMPARE_S
} }
} }