1
0
mirror of https://github.com/hashcat/hashcat.git synced 2024-11-25 09:28:20 +00:00
hashcat/OpenCL/m08100_a1-optimized.cl

591 lines
25 KiB
Common Lisp
Raw Normal View History

2015-12-04 14:47:52 +00:00
/**
* Author......: See docs/credits.txt
2015-12-04 14:47:52 +00:00
* License.....: MIT
*/
#define NEW_SIMD_CODE
#ifdef KERNEL_STATIC
#include M2S(INCLUDE_PATH/inc_vendor.h)
#include M2S(INCLUDE_PATH/inc_types.h)
#include M2S(INCLUDE_PATH/inc_platform.cl)
#include M2S(INCLUDE_PATH/inc_common.cl)
#include M2S(INCLUDE_PATH/inc_simd.cl)
#include M2S(INCLUDE_PATH/inc_hash_sha1.cl)
#endif
2015-12-04 14:47:52 +00:00
KERNEL_FQ void m08100_m04 (KERN_ATTR_BASIC ())
2015-12-04 14:47:52 +00:00
{
/**
* modifier
*/
const u64 lid = get_local_id (0);
2015-12-04 14:47:52 +00:00
/**
* base
*/
const u64 gid = get_global_id (0);
2015-12-04 14:47:52 +00:00
if (gid >= GID_CNT) return;
2015-12-04 14:47:52 +00:00
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];
2015-12-04 14:47:52 +00:00
const u32 pw_l_len = pws[gid].pw_len & 63;
2015-12-04 14:47:52 +00:00
/**
* salt
*/
u32 salt_buf0[2];
salt_buf0[0] = salt_bufs[SALT_POS_HOST].salt_buf[0];
salt_buf0[1] = salt_bufs[SALT_POS_HOST].salt_buf[1];
2015-12-04 14:47:52 +00:00
const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len;
2015-12-04 14:47:52 +00:00
/**
* digest
*/
const u32 search[4] =
{
digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R0],
digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R1],
digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R2],
digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R3]
};
2015-12-04 14:47:52 +00:00
/**
* loop
*/
for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE)
2015-12-04 14:47:52 +00:00
{
const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos) & 63;
const u32x pw_len = (pw_l_len + pw_r_len) & 63;
/**
* concat password candidate
*/
u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
u32x wordl3[4] = { 0 };
wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
2015-12-04 14:47:52 +00:00
if (COMBS_MODE == COMBINATOR_MODE_BASE_LEFT)
2015-12-04 14:47:52 +00:00
{
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);
2015-12-04 14:47:52 +00:00
}
u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
2015-12-04 14:47:52 +00:00
w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
append_0x80_4x4_VV (w0, w1, w2, w3, pw_len + 1);
w0[0] = hc_swap32 (w0[0]);
w0[1] = hc_swap32 (w0[1]);
w0[2] = hc_swap32 (w0[2]);
w0[3] = hc_swap32 (w0[3]);
w1[0] = hc_swap32 (w1[0]);
w1[1] = hc_swap32 (w1[1]);
w1[2] = hc_swap32 (w1[2]);
w1[3] = hc_swap32 (w1[3]);
w2[0] = hc_swap32 (w2[0]);
w2[1] = hc_swap32 (w2[1]);
w2[2] = hc_swap32 (w2[2]);
w2[3] = hc_swap32 (w2[3]);
2015-12-04 14:47:52 +00:00
/**
* prepend salt
*/
const u32x pw_salt_len = pw_len + salt_len;
2015-12-04 14:47:52 +00:00
u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
2015-12-04 14:47:52 +00:00
w0_t[0] = salt_buf0[0];
w0_t[1] = salt_buf0[1];
w0_t[2] = w0[0];
w0_t[3] = w0[1];
w1_t[0] = w0[2];
w1_t[1] = w0[3];
w1_t[2] = w1[0];
w1_t[3] = w1[1];
w2_t[0] = w1[2];
w2_t[1] = w1[3];
w2_t[2] = w2[0];
w2_t[3] = w2[1];
w3_t[0] = w2[2];
w3_t[1] = w2[3];
w3_t[2] = 0;
w3_t[3] = (pw_salt_len + 1) * 8;
/**
* sha1
*/
u32x a = SHA1M_A;
u32x b = SHA1M_B;
u32x c = SHA1M_C;
u32x d = SHA1M_D;
u32x e = SHA1M_E;
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t[0]);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t[1]);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w0_t[2]);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w0_t[3]);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w1_t[0]);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w1_t[1]);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t[2]);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t[3]);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t[0]);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w2_t[1]);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w2_t[2]);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w2_t[3]);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w3_t[0]);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t[1]);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t[2]);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w3_t[3]);
w0_t[0] = hc_rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t[0]);
w0_t[1] = hc_rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w0_t[1]);
w0_t[2] = hc_rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w0_t[2]);
w0_t[3] = hc_rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w0_t[3]);
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C01
w1_t[0] = hc_rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t[0]);
w1_t[1] = hc_rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[1]);
w1_t[2] = hc_rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w1_t[2]);
w1_t[3] = hc_rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[3]);
w2_t[0] = hc_rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t[0]);
w2_t[1] = hc_rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w2_t[1]);
w2_t[2] = hc_rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t[2]);
w2_t[3] = hc_rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w2_t[3]);
w3_t[0] = hc_rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[0]);
w3_t[1] = hc_rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w3_t[1]);
w3_t[2] = hc_rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t[2]);
w3_t[3] = hc_rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[3]);
w0_t[0] = hc_rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t[0]);
w0_t[1] = hc_rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w0_t[1]);
w0_t[2] = hc_rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t[2]);
w0_t[3] = hc_rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w0_t[3]);
w1_t[0] = hc_rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[0]);
w1_t[1] = hc_rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w1_t[1]);
w1_t[2] = hc_rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[2]);
w1_t[3] = hc_rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w1_t[3]);
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C02
w2_t[0] = hc_rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t[0]);
w2_t[1] = hc_rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w2_t[1]);
w2_t[2] = hc_rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w2_t[2]);
w2_t[3] = hc_rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w2_t[3]);
w3_t[0] = hc_rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w3_t[0]);
w3_t[1] = hc_rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w3_t[1]);
w3_t[2] = hc_rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t[2]);
w3_t[3] = hc_rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w3_t[3]);
w0_t[0] = hc_rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t[0]);
w0_t[1] = hc_rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w0_t[1]);
w0_t[2] = hc_rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w0_t[2]);
w0_t[3] = hc_rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w0_t[3]);
w1_t[0] = hc_rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w1_t[0]);
w1_t[1] = hc_rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w1_t[1]);
w1_t[2] = hc_rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t[2]);
w1_t[3] = hc_rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w1_t[3]);
w2_t[0] = hc_rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w2_t[0]);
w2_t[1] = hc_rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w2_t[1]);
w2_t[2] = hc_rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w2_t[2]);
w2_t[3] = hc_rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w2_t[3]);
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C03
w3_t[0] = hc_rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t[0]);
w3_t[1] = hc_rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[1]);
w3_t[2] = hc_rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t[2]);
w3_t[3] = hc_rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[3]);
w0_t[0] = hc_rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t[0]);
w0_t[1] = hc_rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w0_t[1]);
w0_t[2] = hc_rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w0_t[2]);
w0_t[3] = hc_rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t[3]);
w1_t[0] = hc_rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[0]);
w1_t[1] = hc_rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w1_t[1]);
w1_t[2] = hc_rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t[2]);
w1_t[3] = hc_rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[3]);
w2_t[0] = hc_rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w2_t[0]);
w2_t[1] = hc_rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w2_t[1]);
w2_t[2] = hc_rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t[2]);
w2_t[3] = hc_rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w2_t[3]);
w3_t[0] = hc_rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[0]);
w3_t[1] = hc_rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t[1]);
w3_t[2] = hc_rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[2]);
w3_t[3] = hc_rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w3_t[3]);
COMPARE_M_SIMD (d, e, c, b);
2015-12-04 14:47:52 +00:00
}
}
KERNEL_FQ void m08100_m08 (KERN_ATTR_BASIC ())
2015-12-04 14:47:52 +00:00
{
}
KERNEL_FQ void m08100_m16 (KERN_ATTR_BASIC ())
2015-12-04 14:47:52 +00:00
{
}
KERNEL_FQ void m08100_s04 (KERN_ATTR_BASIC ())
2015-12-04 14:47:52 +00:00
{
/**
* modifier
*/
const u64 lid = get_local_id (0);
2015-12-04 14:47:52 +00:00
/**
* base
*/
const u64 gid = get_global_id (0);
2015-12-04 14:47:52 +00:00
if (gid >= GID_CNT) return;
2015-12-04 14:47:52 +00:00
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];
2015-12-04 14:47:52 +00:00
const u32 pw_l_len = pws[gid].pw_len & 63;
2015-12-04 14:47:52 +00:00
/**
* salt
*/
u32 salt_buf0[2];
salt_buf0[0] = salt_bufs[SALT_POS_HOST].salt_buf[0];
salt_buf0[1] = salt_bufs[SALT_POS_HOST].salt_buf[1];
2015-12-04 14:47:52 +00:00
const u32 salt_len = salt_bufs[SALT_POS_HOST].salt_len;
2015-12-04 14:47:52 +00:00
/**
* digest
*/
const u32 search[4] =
{
digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R0],
digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R1],
digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R2],
digests_buf[DIGESTS_OFFSET_HOST].digest_buf[DGST_R3]
2015-12-04 14:47:52 +00:00
};
/**
* reverse
*/
const u32 e_rev = hc_rotl32_S (search[1], 2u);
2015-12-04 14:47:52 +00:00
/**
* loop
*/
for (u32 il_pos = 0; il_pos < IL_CNT; il_pos += VECT_SIZE)
2015-12-04 14:47:52 +00:00
{
const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos) & 63;
const u32x pw_len = (pw_l_len + pw_r_len) & 63;
/**
* concat password candidate
*/
u32x wordl0[4] = { 0 };
u32x wordl1[4] = { 0 };
u32x wordl2[4] = { 0 };
u32x wordl3[4] = { 0 };
wordl0[0] = pw_buf0[0];
wordl0[1] = pw_buf0[1];
wordl0[2] = pw_buf0[2];
wordl0[3] = pw_buf0[3];
wordl1[0] = pw_buf1[0];
wordl1[1] = pw_buf1[1];
wordl1[2] = pw_buf1[2];
wordl1[3] = pw_buf1[3];
u32x wordr0[4] = { 0 };
u32x wordr1[4] = { 0 };
u32x wordr2[4] = { 0 };
u32x wordr3[4] = { 0 };
wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
2015-12-04 14:47:52 +00:00
if (COMBS_MODE == COMBINATOR_MODE_BASE_LEFT)
2015-12-04 14:47:52 +00:00
{
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);
2015-12-04 14:47:52 +00:00
}
u32x w0[4];
u32x w1[4];
u32x w2[4];
u32x w3[4];
2015-12-04 14:47:52 +00:00
w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
w2[0] = wordl2[0] | wordr2[0];
w2[1] = wordl2[1] | wordr2[1];
w2[2] = wordl2[2] | wordr2[2];
w2[3] = wordl2[3] | wordr2[3];
append_0x80_4x4_VV (w0, w1, w2, w3, pw_len + 1);
w0[0] = hc_swap32 (w0[0]);
w0[1] = hc_swap32 (w0[1]);
w0[2] = hc_swap32 (w0[2]);
w0[3] = hc_swap32 (w0[3]);
w1[0] = hc_swap32 (w1[0]);
w1[1] = hc_swap32 (w1[1]);
w1[2] = hc_swap32 (w1[2]);
w1[3] = hc_swap32 (w1[3]);
w2[0] = hc_swap32 (w2[0]);
w2[1] = hc_swap32 (w2[1]);
w2[2] = hc_swap32 (w2[2]);
w2[3] = hc_swap32 (w2[3]);
2015-12-04 14:47:52 +00:00
/**
* prepend salt
*/
const u32x pw_salt_len = pw_len + salt_len;
2015-12-04 14:47:52 +00:00
u32x w0_t[4];
u32x w1_t[4];
u32x w2_t[4];
u32x w3_t[4];
2015-12-04 14:47:52 +00:00
w0_t[0] = salt_buf0[0];
w0_t[1] = salt_buf0[1];
w0_t[2] = w0[0];
w0_t[3] = w0[1];
w1_t[0] = w0[2];
w1_t[1] = w0[3];
w1_t[2] = w1[0];
w1_t[3] = w1[1];
w2_t[0] = w1[2];
w2_t[1] = w1[3];
w2_t[2] = w2[0];
w2_t[3] = w2[1];
w3_t[0] = w2[2];
w3_t[1] = w2[3];
w3_t[2] = 0;
w3_t[3] = (pw_salt_len + 1) * 8;
/**
* sha1
*/
u32x a = SHA1M_A;
u32x b = SHA1M_B;
u32x c = SHA1M_C;
u32x d = SHA1M_D;
u32x e = SHA1M_E;
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C00
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t[0]);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t[1]);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w0_t[2]);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w0_t[3]);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w1_t[0]);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w1_t[1]);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t[2]);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t[3]);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t[0]);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w2_t[1]);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w2_t[2]);
SHA1_STEP (SHA1_F0o, e, a, b, c, d, w2_t[3]);
SHA1_STEP (SHA1_F0o, d, e, a, b, c, w3_t[0]);
SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t[1]);
SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t[2]);
SHA1_STEP (SHA1_F0o, a, b, c, d, e, w3_t[3]);
w0_t[0] = hc_rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t[0]);
w0_t[1] = hc_rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w0_t[1]);
w0_t[2] = hc_rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w0_t[2]);
w0_t[3] = hc_rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w0_t[3]);
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C01
w1_t[0] = hc_rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t[0]);
w1_t[1] = hc_rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[1]);
w1_t[2] = hc_rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w1_t[2]);
w1_t[3] = hc_rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[3]);
w2_t[0] = hc_rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t[0]);
w2_t[1] = hc_rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w2_t[1]);
w2_t[2] = hc_rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t[2]);
w2_t[3] = hc_rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w2_t[3]);
w3_t[0] = hc_rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[0]);
w3_t[1] = hc_rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w3_t[1]);
w3_t[2] = hc_rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t[2]);
w3_t[3] = hc_rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[3]);
w0_t[0] = hc_rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t[0]);
w0_t[1] = hc_rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w0_t[1]);
w0_t[2] = hc_rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t[2]);
w0_t[3] = hc_rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w0_t[3]);
w1_t[0] = hc_rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[0]);
w1_t[1] = hc_rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w1_t[1]);
w1_t[2] = hc_rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[2]);
w1_t[3] = hc_rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w1_t[3]);
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C02
w2_t[0] = hc_rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t[0]);
w2_t[1] = hc_rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w2_t[1]);
w2_t[2] = hc_rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w2_t[2]);
w2_t[3] = hc_rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w2_t[3]);
w3_t[0] = hc_rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w3_t[0]);
w3_t[1] = hc_rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w3_t[1]);
w3_t[2] = hc_rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t[2]);
w3_t[3] = hc_rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w3_t[3]);
w0_t[0] = hc_rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t[0]);
w0_t[1] = hc_rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w0_t[1]);
w0_t[2] = hc_rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w0_t[2]);
w0_t[3] = hc_rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w0_t[3]);
w1_t[0] = hc_rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w1_t[0]);
w1_t[1] = hc_rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w1_t[1]);
w1_t[2] = hc_rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t[2]);
w1_t[3] = hc_rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w1_t[3]);
w2_t[0] = hc_rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w2_t[0]);
w2_t[1] = hc_rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w2_t[1]);
w2_t[2] = hc_rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w2_t[2]);
w2_t[3] = hc_rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w2_t[3]);
2015-12-04 14:47:52 +00:00
#undef K
#define K SHA1C03
w3_t[0] = hc_rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t[0]);
w3_t[1] = hc_rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[1]);
w3_t[2] = hc_rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t[2]);
w3_t[3] = hc_rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[3]);
w0_t[0] = hc_rotl32 ((w3_t[1] ^ w2_t[0] ^ w0_t[2] ^ w0_t[0]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t[0]);
w0_t[1] = hc_rotl32 ((w3_t[2] ^ w2_t[1] ^ w0_t[3] ^ w0_t[1]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w0_t[1]);
w0_t[2] = hc_rotl32 ((w3_t[3] ^ w2_t[2] ^ w1_t[0] ^ w0_t[2]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w0_t[2]);
w0_t[3] = hc_rotl32 ((w0_t[0] ^ w2_t[3] ^ w1_t[1] ^ w0_t[3]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t[3]);
w1_t[0] = hc_rotl32 ((w0_t[1] ^ w3_t[0] ^ w1_t[2] ^ w1_t[0]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t[0]);
w1_t[1] = hc_rotl32 ((w0_t[2] ^ w3_t[1] ^ w1_t[3] ^ w1_t[1]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w1_t[1]);
w1_t[2] = hc_rotl32 ((w0_t[3] ^ w3_t[2] ^ w2_t[0] ^ w1_t[2]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t[2]);
w1_t[3] = hc_rotl32 ((w1_t[0] ^ w3_t[3] ^ w2_t[1] ^ w1_t[3]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w1_t[3]);
w2_t[0] = hc_rotl32 ((w1_t[1] ^ w0_t[0] ^ w2_t[2] ^ w2_t[0]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w2_t[0]);
w2_t[1] = hc_rotl32 ((w1_t[2] ^ w0_t[1] ^ w2_t[3] ^ w2_t[1]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w2_t[1]);
w2_t[2] = hc_rotl32 ((w1_t[3] ^ w0_t[2] ^ w3_t[0] ^ w2_t[2]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t[2]);
w2_t[3] = hc_rotl32 ((w2_t[0] ^ w0_t[3] ^ w3_t[1] ^ w2_t[3]), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w2_t[3]);
2015-12-04 14:47:52 +00:00
if (MATCHES_NONE_VS (e, e_rev)) continue;
2015-12-04 14:47:52 +00:00
w3_t[0] = hc_rotl32 ((w2_t[1] ^ w1_t[0] ^ w3_t[2] ^ w3_t[0]), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w3_t[0]);
w3_t[1] = hc_rotl32 ((w2_t[2] ^ w1_t[1] ^ w3_t[3] ^ w3_t[1]), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t[1]);
w3_t[2] = hc_rotl32 ((w2_t[3] ^ w1_t[2] ^ w0_t[0] ^ w3_t[2]), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w3_t[2]);
w3_t[3] = hc_rotl32 ((w3_t[0] ^ w1_t[3] ^ w0_t[1] ^ w3_t[3]), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w3_t[3]);
COMPARE_S_SIMD (d, e, c, b);
2015-12-04 14:47:52 +00:00
}
}
KERNEL_FQ void m08100_s08 (KERN_ATTR_BASIC ())
2015-12-04 14:47:52 +00:00
{
}
KERNEL_FQ void m08100_s16 (KERN_ATTR_BASIC ())
2015-12-04 14:47:52 +00:00
{
}