if (gid >= gid_max) return;
u32 pw_buf0[4];
+ u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3];
-
- u32 pw_buf1[4];
-
pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6];
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
- const u32 salt_len = 16;
-
- const u32x pw_salt_len = pw_len + salt_len;
-
/**
* loop
*/
const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
- const u32x pw_salt_len = (out_len * 2) + salt_len;
-
append_0x80_2x4_VV (w0, w1, out_len);
- u32x w0_t[4];
- u32x w1_t[4];
- u32x w2_t[4];
- u32x w3_t[4];
-
- make_unicode (w0, w0_t, w1_t);
- make_unicode (w1, w2_t, w3_t);
-
- switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
- w0_t[0] = salt_buf[0];
- w0_t[1] = salt_buf[1];
- w0_t[2] = salt_buf[2];
- w0_t[3] = salt_buf[3];
- w1_t[0] = swap32 (w1_t[0]);
- w1_t[1] = swap32 (w1_t[1]);
- w1_t[2] = swap32 (w1_t[2]);
- w1_t[3] = swap32 (w1_t[3]);
- w2_t[0] = swap32 (w2_t[0]);
- w2_t[1] = swap32 (w2_t[1]);
- w2_t[2] = swap32 (w2_t[2]);
- w2_t[3] = swap32 (w2_t[3]);
- w3_t[0] = swap32 (w3_t[0]);
- w3_t[1] = swap32 (w3_t[1]);
- w3_t[2] = 0;
- w3_t[3] = pw_salt_len * 8;
-
- u32x digest[5];
+ /**
+ * sha1
+ */
+
+ make_unicode (w1, w2, w3);
+ make_unicode (w0, w0, w1);
+
+ const u32x pw_salt_len = (out_len * 2) + 16;
+
+ w3[3] = pw_salt_len * 8;
+ w3[2] = 0;
+ w3[1] = swap32 (w2[1]);
+ w3[0] = swap32 (w2[0]);
+ w2[3] = swap32 (w1[3]);
+ w2[2] = swap32 (w1[2]);
+ w2[1] = swap32 (w1[1]);
+ w2[0] = swap32 (w1[0]);
+ w1[3] = swap32 (w0[3]);
+ w1[2] = swap32 (w0[2]);
+ w1[1] = swap32 (w0[1]);
+ w1[0] = swap32 (w0[0]);
+ w0[3] = salt_buf[3];
+ w0[2] = salt_buf[2];
+ w0[1] = salt_buf[1];
+ w0[0] = salt_buf[0];
+
+ u32 digest[5];
digest[0] = SHA1M_A;
digest[1] = SHA1M_B;
digest[3] = SHA1M_D;
digest[4] = SHA1M_E;
- sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
-
- w0_t[0] = digest[0];
- w0_t[1] = digest[1];
- w0_t[2] = digest[2];
- w0_t[3] = digest[3];
- w1_t[0] = digest[4];
- w1_t[1] = 0;
- w1_t[2] = 0x80000000;
- w1_t[3] = 0;
- w2_t[0] = 0;
- w2_t[1] = 0;
- w2_t[2] = 0;
- w2_t[3] = 0;
- w3_t[0] = 0;
- w3_t[1] = 0;
- w3_t[2] = 0;
- w3_t[3] = (20 + 4) * 8;
+ sha1_transform (w0, w1, w2, w3, digest);
+
+ w0[0] = digest[0];
+ w0[1] = digest[1];
+ w0[2] = digest[2];
+ w0[3] = digest[3];
+ w1[0] = digest[4];
+ w1[1] = 0;
+ w1[2] = 0x80000000;
+ w1[3] = 0;
+ w2[0] = 0;
+ w2[1] = 0;
+ w2[2] = 0;
+ w2[3] = 0;
+ w3[0] = 0;
+ w3[1] = 0;
+ w3[2] = 0;
+ w3[3] = (20 + 4) * 8;
digest[0] = SHA1M_A;
digest[1] = SHA1M_B;
digest[3] = SHA1M_D;
digest[4] = SHA1M_E;
- sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
+ sha1_transform (w0, w1, w2, w3, digest);
- u32x a = swap32 (digest[0]);
- u32x b = swap32 (digest[1]) & 0xff;
- u32x c = 0;
- u32x d = 0;
+ digest[0] = swap32 (digest[0]);
+ digest[1] = swap32 (digest[1]) & 0xff;
+ digest[2] = 0;
+ digest[3] = 0;
- COMPARE_M_SIMD (a, b, c, d);
+ COMPARE_M_SIMD (digest[0], digest[1], digest[2], digest[3]);
}
}
if (gid >= gid_max) return;
u32 pw_buf0[4];
+ u32 pw_buf1[4];
pw_buf0[0] = pws[gid].i[ 0];
pw_buf0[1] = pws[gid].i[ 1];
pw_buf0[2] = pws[gid].i[ 2];
pw_buf0[3] = pws[gid].i[ 3];
-
- u32 pw_buf1[4];
-
pw_buf1[0] = pws[gid].i[ 4];
pw_buf1[1] = pws[gid].i[ 5];
pw_buf1[2] = pws[gid].i[ 6];
const u32 pw_len = pws[gid].pw_len;
- /**
- * digest
- */
-
- const u32 search[4] =
- {
- digests_buf[digests_offset].digest_buf[DGST_R0],
- digests_buf[digests_offset].digest_buf[DGST_R1],
- digests_buf[digests_offset].digest_buf[DGST_R2],
- digests_buf[digests_offset].digest_buf[DGST_R3]
- };
-
/**
* salt
*/
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
- const u32 salt_len = 16;
+ /**
+ * digest
+ */
- const u32x pw_salt_len = pw_len + salt_len;
+ 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
const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
- const u32x pw_salt_len = (out_len * 2) + salt_len;
-
append_0x80_2x4_VV (w0, w1, out_len);
- u32x w0_t[4];
- u32x w1_t[4];
- u32x w2_t[4];
- u32x w3_t[4];
-
- make_unicode (w0, w0_t, w1_t);
- make_unicode (w1, w2_t, w3_t);
-
- switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
- w0_t[0] = salt_buf[0];
- w0_t[1] = salt_buf[1];
- w0_t[2] = salt_buf[2];
- w0_t[3] = salt_buf[3];
- w1_t[0] = swap32 (w1_t[0]);
- w1_t[1] = swap32 (w1_t[1]);
- w1_t[2] = swap32 (w1_t[2]);
- w1_t[3] = swap32 (w1_t[3]);
- w2_t[0] = swap32 (w2_t[0]);
- w2_t[1] = swap32 (w2_t[1]);
- w2_t[2] = swap32 (w2_t[2]);
- w2_t[3] = swap32 (w2_t[3]);
- w3_t[0] = swap32 (w3_t[0]);
- w3_t[1] = swap32 (w3_t[1]);
- w3_t[2] = 0;
- w3_t[3] = pw_salt_len * 8;
-
- u32x digest[5];
+ /**
+ * sha1
+ */
+
+ make_unicode (w1, w2, w3);
+ make_unicode (w0, w0, w1);
+
+ const u32x pw_salt_len = (out_len * 2) + 16;
+
+ w3[3] = pw_salt_len * 8;
+ w3[2] = 0;
+ w3[1] = swap32 (w2[1]);
+ w3[0] = swap32 (w2[0]);
+ w2[3] = swap32 (w1[3]);
+ w2[2] = swap32 (w1[2]);
+ w2[1] = swap32 (w1[1]);
+ w2[0] = swap32 (w1[0]);
+ w1[3] = swap32 (w0[3]);
+ w1[2] = swap32 (w0[2]);
+ w1[1] = swap32 (w0[1]);
+ w1[0] = swap32 (w0[0]);
+ w0[3] = salt_buf[3];
+ w0[2] = salt_buf[2];
+ w0[1] = salt_buf[1];
+ w0[0] = salt_buf[0];
+
+ u32 digest[5];
digest[0] = SHA1M_A;
digest[1] = SHA1M_B;
digest[3] = SHA1M_D;
digest[4] = SHA1M_E;
- sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
-
- w0_t[0] = digest[0];
- w0_t[1] = digest[1];
- w0_t[2] = digest[2];
- w0_t[3] = digest[3];
- w1_t[0] = digest[4];
- w1_t[1] = 0;
- w1_t[2] = 0x80000000;
- w1_t[3] = 0;
- w2_t[0] = 0;
- w2_t[1] = 0;
- w2_t[2] = 0;
- w2_t[3] = 0;
- w3_t[0] = 0;
- w3_t[1] = 0;
- w3_t[2] = 0;
- w3_t[3] = (20 + 4) * 8;
+ sha1_transform (w0, w1, w2, w3, digest);
+
+ w0[0] = digest[0];
+ w0[1] = digest[1];
+ w0[2] = digest[2];
+ w0[3] = digest[3];
+ w1[0] = digest[4];
+ w1[1] = 0;
+ w1[2] = 0x80000000;
+ w1[3] = 0;
+ w2[0] = 0;
+ w2[1] = 0;
+ w2[2] = 0;
+ w2[3] = 0;
+ w3[0] = 0;
+ w3[1] = 0;
+ w3[2] = 0;
+ w3[3] = (20 + 4) * 8;
digest[0] = SHA1M_A;
digest[1] = SHA1M_B;
digest[3] = SHA1M_D;
digest[4] = SHA1M_E;
- sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
+ sha1_transform (w0, w1, w2, w3, digest);
- u32x a = swap32 (digest[0]);
- u32x b = swap32 (digest[1]) & 0xff;
- u32x c = 0;
- u32x d = 0;
+ digest[0] = swap32 (digest[0]);
+ digest[1] = swap32 (digest[1]) & 0xff;
+ digest[2] = 0;
+ digest[3] = 0;
- COMPARE_S_SIMD (a, b, c, d);
+ COMPARE_S_SIMD (digest[0], digest[1], digest[2], digest[3]);
}
}
#define _OLDOFFICE34_
+#define NEW_SIMD_CODE
+
#include "include/constants.h"
#include "include/kernel_vendor.h"
#include "include/kernel_functions.c"
#include "OpenCL/types_ocl.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"
-
-static void sha1_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[5])
+static void sha1_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[5])
{
- u32 A = digest[0];
- u32 B = digest[1];
- u32 C = digest[2];
- u32 D = digest[3];
- u32 E = digest[4];
-
- u32 w0_t = w0[0];
- u32 w1_t = w0[1];
- u32 w2_t = w0[2];
- u32 w3_t = w0[3];
- u32 w4_t = w1[0];
- u32 w5_t = w1[1];
- u32 w6_t = w1[2];
- u32 w7_t = w1[3];
- u32 w8_t = w2[0];
- u32 w9_t = w2[1];
- u32 wa_t = w2[2];
- u32 wb_t = w2[3];
- u32 wc_t = w3[0];
- u32 wd_t = w3[1];
- u32 we_t = w3[2];
- u32 wf_t = w3[3];
+ u32x A = digest[0];
+ u32x B = digest[1];
+ u32x C = digest[2];
+ u32x D = digest[3];
+ u32x E = digest[4];
+
+ u32x w0_t = w0[0];
+ u32x w1_t = w0[1];
+ u32x w2_t = w0[2];
+ u32x w3_t = w0[3];
+ u32x w4_t = w1[0];
+ u32x w5_t = w1[1];
+ u32x w6_t = w1[2];
+ u32x w7_t = w1[3];
+ u32x w8_t = w2[0];
+ u32x w9_t = w2[1];
+ u32x wa_t = w2[2];
+ u32x wb_t = w2[3];
+ u32x wc_t = w3[0];
+ u32x wd_t = w3[1];
+ u32x we_t = w3[2];
+ u32x wf_t = w3[3];
#undef K
#define K SHA1C00
if (gid >= gid_max) return;
- u32 wordl0[4];
-
- wordl0[0] = pws[gid].i[ 0];
- wordl0[1] = pws[gid].i[ 1];
- wordl0[2] = pws[gid].i[ 2];
- wordl0[3] = pws[gid].i[ 3];
-
- u32 wordl1[4];
-
- wordl1[0] = pws[gid].i[ 4];
- wordl1[1] = pws[gid].i[ 5];
- wordl1[2] = pws[gid].i[ 6];
- wordl1[3] = pws[gid].i[ 7];
+ u32 pw_buf0[4];
+ u32 pw_buf1[4];
- u32 wordl2[4];
-
- wordl2[0] = 0;
- wordl2[1] = 0;
- wordl2[2] = 0;
- wordl2[3] = 0;
-
- u32 wordl3[4];
-
- wordl3[0] = 0;
- wordl3[1] = 0;
- wordl3[2] = 0;
- wordl3[3] = 0;
+ pw_buf0[0] = pws[gid].i[0];
+ pw_buf0[1] = pws[gid].i[1];
+ pw_buf0[2] = pws[gid].i[2];
+ pw_buf0[3] = pws[gid].i[3];
+ pw_buf1[0] = pws[gid].i[4];
+ pw_buf1[1] = pws[gid].i[5];
+ pw_buf1[2] = pws[gid].i[6];
+ pw_buf1[3] = pws[gid].i[7];
const u32 pw_l_len = pws[gid].pw_len;
- if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
- {
- switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
- }
-
/**
* salt
*/
salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
- const u32 salt_len = 16;
-
/**
* loop
*/
- for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 pw_r_len = combs_buf[il_pos].pw_len;
-
- const u32 pw_len = pw_l_len + pw_r_len;
-
- const u32 pw_salt_len = (pw_len * 2) + salt_len;
-
- u32 wordr0[4];
-
- wordr0[0] = combs_buf[il_pos].i[0];
- wordr0[1] = combs_buf[il_pos].i[1];
- wordr0[2] = combs_buf[il_pos].i[2];
- wordr0[3] = combs_buf[il_pos].i[3];
-
- u32 wordr1[4];
-
- wordr1[0] = combs_buf[il_pos].i[4];
- wordr1[1] = combs_buf[il_pos].i[5];
- wordr1[2] = combs_buf[il_pos].i[6];
- wordr1[3] = combs_buf[il_pos].i[7];
-
- u32 wordr2[4];
-
- wordr2[0] = 0;
- wordr2[1] = 0;
- wordr2[2] = 0;
- wordr2[3] = 0;
-
- u32 wordr3[4];
-
- wordr3[0] = 0;
- wordr3[1] = 0;
- wordr3[2] = 0;
- wordr3[3] = 0;
+ const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
+
+ const u32x pw_len = pw_l_len + pw_r_len;
+
+ /**
+ * concat password candidate
+ */
+
+ u32x wordl0[4] = { 0 };
+ u32x wordl1[4] = { 0 };
+ u32x wordl2[4] = { 0 };
+ u32x wordl3[4] = { 0 };
+
+ wordl0[0] = pw_buf0[0];
+ wordl0[1] = pw_buf0[1];
+ wordl0[2] = pw_buf0[2];
+ wordl0[3] = pw_buf0[3];
+ wordl1[0] = pw_buf1[0];
+ wordl1[1] = pw_buf1[1];
+ wordl1[2] = pw_buf1[2];
+ wordl1[3] = pw_buf1[3];
+
+ u32x wordr0[4] = { 0 };
+ u32x wordr1[4] = { 0 };
+ u32x wordr2[4] = { 0 };
+ u32x wordr3[4] = { 0 };
+
+ wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
+ wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
+ wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
+ wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
+ wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
+ wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
+ wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
+ wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{
- switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+ switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+ }
+ else
+ {
+ switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
}
- u32 w0[4];
+ u32x w0[4];
+ u32x w1[4];
+ u32x w2[4];
+ u32x w3[4];
w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
-
- u32 w1[4];
-
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
- u32 w2[4];
+ /**
+ * sha1
+ */
- w2[0] = wordl2[0] | wordr2[0];
- w2[1] = wordl2[1] | wordr2[1];
- w2[2] = wordl2[2] | wordr2[2];
- w2[3] = wordl2[3] | wordr2[3];
+ make_unicode (w1, w2, w3);
+ make_unicode (w0, w0, w1);
- u32 w3[4];
+ const u32x pw_salt_len = (pw_len * 2) + 16;
- w3[0] = wordl3[0] | wordr3[0];
- w3[1] = wordl3[1] | wordr3[1];
+ w3[3] = pw_salt_len * 8;
w3[2] = 0;
- w3[3] = 0;
-
- append_0x80_2x4 (w0, w1, pw_len);
-
- u32 w0_t[4];
- u32 w1_t[4];
- u32 w2_t[4];
- u32 w3_t[4];
-
- make_unicode (w0, w0_t, w1_t);
- make_unicode (w1, w2_t, w3_t);
-
- switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
- w0_t[0] = salt_buf[0];
- w0_t[1] = salt_buf[1];
- w0_t[2] = salt_buf[2];
- w0_t[3] = salt_buf[3];
- w1_t[0] = swap32 (w1_t[0]);
- w1_t[1] = swap32 (w1_t[1]);
- w1_t[2] = swap32 (w1_t[2]);
- w1_t[3] = swap32 (w1_t[3]);
- w2_t[0] = swap32 (w2_t[0]);
- w2_t[1] = swap32 (w2_t[1]);
- w2_t[2] = swap32 (w2_t[2]);
- w2_t[3] = swap32 (w2_t[3]);
- w3_t[0] = swap32 (w3_t[0]);
- w3_t[1] = swap32 (w3_t[1]);
- w3_t[2] = 0;
- w3_t[3] = pw_salt_len * 8;
+ w3[1] = swap32 (w2[1]);
+ w3[0] = swap32 (w2[0]);
+ w2[3] = swap32 (w1[3]);
+ w2[2] = swap32 (w1[2]);
+ w2[1] = swap32 (w1[1]);
+ w2[0] = swap32 (w1[0]);
+ w1[3] = swap32 (w0[3]);
+ w1[2] = swap32 (w0[2]);
+ w1[1] = swap32 (w0[1]);
+ w1[0] = swap32 (w0[0]);
+ w0[3] = salt_buf[3];
+ w0[2] = salt_buf[2];
+ w0[1] = salt_buf[1];
+ w0[0] = salt_buf[0];
u32 digest[5];
digest[3] = SHA1M_D;
digest[4] = SHA1M_E;
- sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
-
- w0_t[0] = digest[0];
- w0_t[1] = digest[1];
- w0_t[2] = digest[2];
- w0_t[3] = digest[3];
- w1_t[0] = digest[4];
- w1_t[1] = 0;
- w1_t[2] = 0x80000000;
- w1_t[3] = 0;
- w2_t[0] = 0;
- w2_t[1] = 0;
- w2_t[2] = 0;
- w2_t[3] = 0;
- w3_t[0] = 0;
- w3_t[1] = 0;
- w3_t[2] = 0;
- w3_t[3] = (20 + 4) * 8;
+ sha1_transform (w0, w1, w2, w3, digest);
+
+ w0[0] = digest[0];
+ w0[1] = digest[1];
+ w0[2] = digest[2];
+ w0[3] = digest[3];
+ w1[0] = digest[4];
+ w1[1] = 0;
+ w1[2] = 0x80000000;
+ w1[3] = 0;
+ w2[0] = 0;
+ w2[1] = 0;
+ w2[2] = 0;
+ w2[3] = 0;
+ w3[0] = 0;
+ w3[1] = 0;
+ w3[2] = 0;
+ w3[3] = (20 + 4) * 8;
digest[0] = SHA1M_A;
digest[1] = SHA1M_B;
digest[3] = SHA1M_D;
digest[4] = SHA1M_E;
- sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
+ sha1_transform (w0, w1, w2, w3, digest);
- u32 a = swap32 (digest[0]);
- u32 b = swap32 (digest[1]) & 0xff;
+ digest[0] = swap32 (digest[0]);
+ digest[1] = swap32 (digest[1]) & 0xff;
+ digest[2] = 0;
+ digest[3] = 0;
- const u32 r0 = a;
- const u32 r1 = b;
- const u32 r2 = 0;
- const u32 r3 = 0;
-
- #include COMPARE_M
+ COMPARE_M_SIMD (digest[0], digest[1], digest[2], digest[3]);
}
}
if (gid >= gid_max) return;
- u32 wordl0[4];
-
- wordl0[0] = pws[gid].i[ 0];
- wordl0[1] = pws[gid].i[ 1];
- wordl0[2] = pws[gid].i[ 2];
- wordl0[3] = pws[gid].i[ 3];
-
- u32 wordl1[4];
-
- wordl1[0] = pws[gid].i[ 4];
- wordl1[1] = pws[gid].i[ 5];
- wordl1[2] = pws[gid].i[ 6];
- wordl1[3] = pws[gid].i[ 7];
-
- u32 wordl2[4];
+ u32 pw_buf0[4];
+ u32 pw_buf1[4];
- wordl2[0] = 0;
- wordl2[1] = 0;
- wordl2[2] = 0;
- wordl2[3] = 0;
+ pw_buf0[0] = pws[gid].i[0];
+ pw_buf0[1] = pws[gid].i[1];
+ pw_buf0[2] = pws[gid].i[2];
+ pw_buf0[3] = pws[gid].i[3];
+ pw_buf1[0] = pws[gid].i[4];
+ pw_buf1[1] = pws[gid].i[5];
+ pw_buf1[2] = pws[gid].i[6];
+ pw_buf1[3] = pws[gid].i[7];
- u32 wordl3[4];
+ const u32 pw_l_len = pws[gid].pw_len;
- wordl3[0] = 0;
- wordl3[1] = 0;
- wordl3[2] = 0;
- wordl3[3] = 0;
+ /**
+ * salt
+ */
- const u32 pw_l_len = pws[gid].pw_len;
+ u32 salt_buf[4];
- if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
- {
- switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
- }
+ salt_buf[0] = salt_bufs[salt_pos].salt_buf[0];
+ salt_buf[1] = salt_bufs[salt_pos].salt_buf[1];
+ salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
+ salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
/**
* digest
digests_buf[digests_offset].digest_buf[DGST_R3]
};
- /**
- * salt
- */
-
- u32 salt_buf[4];
-
- salt_buf[0] = salt_bufs[salt_pos].salt_buf[0];
- salt_buf[1] = salt_bufs[salt_pos].salt_buf[1];
- salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
- salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
-
- const u32 salt_len = 16;
-
/**
* loop
*/
- for (u32 il_pos = 0; il_pos < il_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 pw_r_len = combs_buf[il_pos].pw_len;
-
- const u32 pw_len = pw_l_len + pw_r_len;
-
- const u32 pw_salt_len = (pw_len * 2) + salt_len;
-
- u32 wordr0[4];
-
- wordr0[0] = combs_buf[il_pos].i[0];
- wordr0[1] = combs_buf[il_pos].i[1];
- wordr0[2] = combs_buf[il_pos].i[2];
- wordr0[3] = combs_buf[il_pos].i[3];
-
- u32 wordr1[4];
-
- wordr1[0] = combs_buf[il_pos].i[4];
- wordr1[1] = combs_buf[il_pos].i[5];
- wordr1[2] = combs_buf[il_pos].i[6];
- wordr1[3] = combs_buf[il_pos].i[7];
-
- u32 wordr2[4];
-
- wordr2[0] = 0;
- wordr2[1] = 0;
- wordr2[2] = 0;
- wordr2[3] = 0;
-
- u32 wordr3[4];
-
- wordr3[0] = 0;
- wordr3[1] = 0;
- wordr3[2] = 0;
- wordr3[3] = 0;
+ const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
+
+ const u32x pw_len = pw_l_len + pw_r_len;
+
+ /**
+ * concat password candidate
+ */
+
+ u32x wordl0[4] = { 0 };
+ u32x wordl1[4] = { 0 };
+ u32x wordl2[4] = { 0 };
+ u32x wordl3[4] = { 0 };
+
+ wordl0[0] = pw_buf0[0];
+ wordl0[1] = pw_buf0[1];
+ wordl0[2] = pw_buf0[2];
+ wordl0[3] = pw_buf0[3];
+ wordl1[0] = pw_buf1[0];
+ wordl1[1] = pw_buf1[1];
+ wordl1[2] = pw_buf1[2];
+ wordl1[3] = pw_buf1[3];
+
+ u32x wordr0[4] = { 0 };
+ u32x wordr1[4] = { 0 };
+ u32x wordr2[4] = { 0 };
+ u32x wordr3[4] = { 0 };
+
+ wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
+ wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
+ wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
+ wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
+ wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
+ wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
+ wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
+ wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
{
- switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+ switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
+ }
+ else
+ {
+ switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
}
- u32 w0[4];
+ u32x w0[4];
+ u32x w1[4];
+ u32x w2[4];
+ u32x w3[4];
w0[0] = wordl0[0] | wordr0[0];
w0[1] = wordl0[1] | wordr0[1];
w0[2] = wordl0[2] | wordr0[2];
w0[3] = wordl0[3] | wordr0[3];
-
- u32 w1[4];
-
w1[0] = wordl1[0] | wordr1[0];
w1[1] = wordl1[1] | wordr1[1];
w1[2] = wordl1[2] | wordr1[2];
w1[3] = wordl1[3] | wordr1[3];
- u32 w2[4];
+ /**
+ * sha1
+ */
- w2[0] = wordl2[0] | wordr2[0];
- w2[1] = wordl2[1] | wordr2[1];
- w2[2] = wordl2[2] | wordr2[2];
- w2[3] = wordl2[3] | wordr2[3];
+ make_unicode (w1, w2, w3);
+ make_unicode (w0, w0, w1);
- u32 w3[4];
+ const u32x pw_salt_len = (pw_len * 2) + 16;
- w3[0] = wordl3[0] | wordr3[0];
- w3[1] = wordl3[1] | wordr3[1];
+ w3[3] = pw_salt_len * 8;
w3[2] = 0;
- w3[3] = 0;
-
- append_0x80_2x4 (w0, w1, pw_len);
-
- u32 w0_t[4];
- u32 w1_t[4];
- u32 w2_t[4];
- u32 w3_t[4];
-
- make_unicode (w0, w0_t, w1_t);
- make_unicode (w1, w2_t, w3_t);
-
- switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
-
- w0_t[0] = salt_buf[0];
- w0_t[1] = salt_buf[1];
- w0_t[2] = salt_buf[2];
- w0_t[3] = salt_buf[3];
- w1_t[0] = swap32 (w1_t[0]);
- w1_t[1] = swap32 (w1_t[1]);
- w1_t[2] = swap32 (w1_t[2]);
- w1_t[3] = swap32 (w1_t[3]);
- w2_t[0] = swap32 (w2_t[0]);
- w2_t[1] = swap32 (w2_t[1]);
- w2_t[2] = swap32 (w2_t[2]);
- w2_t[3] = swap32 (w2_t[3]);
- w3_t[0] = swap32 (w3_t[0]);
- w3_t[1] = swap32 (w3_t[1]);
- w3_t[2] = 0;
- w3_t[3] = pw_salt_len * 8;
+ w3[1] = swap32 (w2[1]);
+ w3[0] = swap32 (w2[0]);
+ w2[3] = swap32 (w1[3]);
+ w2[2] = swap32 (w1[2]);
+ w2[1] = swap32 (w1[1]);
+ w2[0] = swap32 (w1[0]);
+ w1[3] = swap32 (w0[3]);
+ w1[2] = swap32 (w0[2]);
+ w1[1] = swap32 (w0[1]);
+ w1[0] = swap32 (w0[0]);
+ w0[3] = salt_buf[3];
+ w0[2] = salt_buf[2];
+ w0[1] = salt_buf[1];
+ w0[0] = salt_buf[0];
u32 digest[5];
digest[3] = SHA1M_D;
digest[4] = SHA1M_E;
- sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
-
- w0_t[0] = digest[0];
- w0_t[1] = digest[1];
- w0_t[2] = digest[2];
- w0_t[3] = digest[3];
- w1_t[0] = digest[4];
- w1_t[1] = 0;
- w1_t[2] = 0x80000000;
- w1_t[3] = 0;
- w2_t[0] = 0;
- w2_t[1] = 0;
- w2_t[2] = 0;
- w2_t[3] = 0;
- w3_t[0] = 0;
- w3_t[1] = 0;
- w3_t[2] = 0;
- w3_t[3] = (20 + 4) * 8;
+ sha1_transform (w0, w1, w2, w3, digest);
+
+ w0[0] = digest[0];
+ w0[1] = digest[1];
+ w0[2] = digest[2];
+ w0[3] = digest[3];
+ w1[0] = digest[4];
+ w1[1] = 0;
+ w1[2] = 0x80000000;
+ w1[3] = 0;
+ w2[0] = 0;
+ w2[1] = 0;
+ w2[2] = 0;
+ w2[3] = 0;
+ w3[0] = 0;
+ w3[1] = 0;
+ w3[2] = 0;
+ w3[3] = (20 + 4) * 8;
digest[0] = SHA1M_A;
digest[1] = SHA1M_B;
digest[3] = SHA1M_D;
digest[4] = SHA1M_E;
- sha1_transform (w0_t, w1_t, w2_t, w3_t, digest);
-
- u32 a = swap32 (digest[0]);
- u32 b = swap32 (digest[1]) & 0xff;
+ sha1_transform (w0, w1, w2, w3, digest);
- const u32 r0 = a;
- const u32 r1 = b;
- const u32 r2 = 0;
- const u32 r3 = 0;
+ digest[0] = swap32 (digest[0]);
+ digest[1] = swap32 (digest[1]) & 0xff;
+ digest[2] = 0;
+ digest[3] = 0;
- #include COMPARE_S
+ COMPARE_S_SIMD (digest[0], digest[1], digest[2], digest[3]);
}
}
projects / hashcat.git / commitdiff