#define _CRC32_
+//incompatible because of branches
+//#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"
-
-#define COMPARE_S "OpenCL/check_single_comp4.c"
-#define COMPARE_M "OpenCL/check_multi_comp4.c"
+#include "OpenCL/simd.c"
__constant u32 crc32tab[0x100] =
{
0xb40bbe37, 0xc30c8ea1, 0x5a05df1b, 0x2d02ef8d
};
-static u32 round_crc32 (u32 a, const u32 v)
+static u32x round_crc32 (u32x a, const u32x v)
{
- const u32 k = (a ^ v) & 0xff;
-
- const u32 s = a >> 8;
-
- a = crc32tab[k];
+ const u32x k = (a ^ v) & 0xff;
+
+ const u32x s = a >> 8;
+
+ #if VECT_SIZE == 1
+ a = (u32x) crc32tab[k];
+ #elif VECT_SIZE == 2
+ a = (u32x) (crc32tab[k.s0], crc32tab[k.s1]);
+ #elif VECT_SIZE == 4
+ a = (u32x) (crc32tab[k.s0], crc32tab[k.s1], crc32tab[k.s2], crc32tab[k.s3]);
+ #elif VECT_SIZE == 8
+ a = (u32x) (crc32tab[k.s0], crc32tab[k.s1], crc32tab[k.s2], crc32tab[k.s3], crc32tab[k.s4], crc32tab[k.s5], crc32tab[k.s6], crc32tab[k.s7]);
+ #elif VECT_SIZE == 16
+ a = (u32x) (crc32tab[k.s0], crc32tab[k.s1], crc32tab[k.s2], crc32tab[k.s3], crc32tab[k.s4], crc32tab[k.s5], crc32tab[k.s6], crc32tab[k.s7], crc32tab[k.s8], crc32tab[k.s9], crc32tab[k.sa], crc32tab[k.sb], crc32tab[k.sc], crc32tab[k.sd], crc32tab[k.se], crc32tab[k.sf]);
+ #endif
a ^= s;
return a;
}
-static u32 crc32 (const u32 w[16], const u32 pw_len, const u32 iv)
+static u32x crc32 (const u32x w[16], const u32 pw_len, const u32 iv)
{
- u32 a = iv ^ ~0;
+ u32x a = iv ^ ~0;
if (pw_len >= 1) a = round_crc32 (a, w[0] >> 0);
if (pw_len >= 2) a = round_crc32 (a, w[0] >> 8);
const u32 lid = get_local_id (0);
- /**
- * digest
- */
-
- const u32 iv = salt_bufs[salt_pos].salt_buf[0];
-
/**
* base
*/
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];
-
- wordl2[0] = 0;
- wordl2[1] = 0;
- wordl2[2] = 0;
- wordl2[3] = 0;
+ u32 pw_buf0[4];
+ u32 pw_buf1[4];
- 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
+ */
+
+ const u32 iv = salt_bufs[salt_pos].salt_buf[0];
/**
* 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;
-
- 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 w_t[16];
-
- w_t[ 0] = wordl0[0] | wordr0[0];
- w_t[ 1] = wordl0[1] | wordr0[1];
- w_t[ 2] = wordl0[2] | wordr0[2];
- w_t[ 3] = wordl0[3] | wordr0[3];
- w_t[ 4] = wordl1[0] | wordr1[0];
- w_t[ 5] = wordl1[1] | wordr1[1];
- w_t[ 6] = wordl1[2] | wordr1[2];
- w_t[ 7] = wordl1[3] | wordr1[3];
- w_t[ 8] = wordl2[0] | wordr2[0];
- w_t[ 9] = wordl2[1] | wordr2[1];
- w_t[10] = wordl2[2] | wordr2[2];
- w_t[11] = wordl2[3] | wordr2[3];
- w_t[12] = wordl3[0] | wordr3[0];
- w_t[13] = wordl3[1] | wordr3[1];
- w_t[14] = wordl3[2] | wordr3[2];
- w_t[15] = 0;
-
- u32 a = crc32 (w_t, pw_len, iv);
- u32 b = 0;
-
- const u32 r0 = a;
- const u32 r1 = b;
- const u32 r2 = 0;
- const u32 r3 = 0;
-
- #include COMPARE_M
+ 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];
+ 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];
+ w3[0] = wordl3[0] | wordr3[0];
+ w3[1] = wordl3[1] | wordr3[1];
+ w3[2] = wordl3[2] | wordr3[2];
+ w3[3] = wordl3[3] | wordr3[3];
+
+ /**
+ * crc32
+ */
+
+ u32x w[16];
+
+ w[ 0] = w0[0];
+ w[ 1] = w0[1];
+ w[ 2] = w0[2];
+ w[ 3] = w0[3];
+ w[ 4] = w1[0];
+ w[ 5] = w1[1];
+ w[ 6] = w1[2];
+ w[ 7] = w1[3];
+ w[ 8] = w2[0];
+ w[ 9] = w2[1];
+ w[10] = w2[2];
+ w[11] = w2[3];
+ w[12] = w3[0];
+ w[13] = w3[1];
+ w[14] = w3[2];
+ w[15] = w3[3];
+
+ u32x a = crc32 (w, pw_len, iv);
+ u32x b = 0;
+ u32x c = 0;
+ u32x d = 0;
+
+ COMPARE_M_SIMD (a, b, c, d);
}
}
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;
-
- 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);
- }
-
/**
- * digest
+ * salt
*/
const u32 iv = salt_bufs[salt_pos].salt_buf[0];
+ /**
+ * digest
+ */
+
const u32 search[4] =
{
digests_buf[digests_offset].digest_buf[DGST_R0],
* 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;
-
- 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 w_t[16];
-
- w_t[ 0] = wordl0[0] | wordr0[0];
- w_t[ 1] = wordl0[1] | wordr0[1];
- w_t[ 2] = wordl0[2] | wordr0[2];
- w_t[ 3] = wordl0[3] | wordr0[3];
- w_t[ 4] = wordl1[0] | wordr1[0];
- w_t[ 5] = wordl1[1] | wordr1[1];
- w_t[ 6] = wordl1[2] | wordr1[2];
- w_t[ 7] = wordl1[3] | wordr1[3];
- w_t[ 8] = wordl2[0] | wordr2[0];
- w_t[ 9] = wordl2[1] | wordr2[1];
- w_t[10] = wordl2[2] | wordr2[2];
- w_t[11] = wordl2[3] | wordr2[3];
- w_t[12] = wordl3[0] | wordr3[0];
- w_t[13] = wordl3[1] | wordr3[1];
- w_t[14] = wordl3[2] | wordr3[2];
- w_t[15] = 0;
-
- u32 a = crc32 (w_t, pw_len, iv);
- u32 b = 0;
-
- const u32 r0 = a;
- const u32 r1 = b;
- const u32 r2 = 0;
- const u32 r3 = 0;
-
- #include COMPARE_S
+ 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];
+ 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];
+ w3[0] = wordl3[0] | wordr3[0];
+ w3[1] = wordl3[1] | wordr3[1];
+ w3[2] = wordl3[2] | wordr3[2];
+ w3[3] = wordl3[3] | wordr3[3];
+
+ /**
+ * crc32
+ */
+
+ u32x w[16];
+
+ w[ 0] = w0[0];
+ w[ 1] = w0[1];
+ w[ 2] = w0[2];
+ w[ 3] = w0[3];
+ w[ 4] = w1[0];
+ w[ 5] = w1[1];
+ w[ 6] = w1[2];
+ w[ 7] = w1[3];
+ w[ 8] = w2[0];
+ w[ 9] = w2[1];
+ w[10] = w2[2];
+ w[11] = w2[3];
+ w[12] = w3[0];
+ w[13] = w3[1];
+ w[14] = w3[2];
+ w[15] = w3[3];
+
+ u32x a = crc32 (w, pw_len, iv);
+ u32x b = 0;
+ u32x c = 0;
+ u32x d = 0;
+
+ COMPARE_S_SIMD (a, b, c, d);
}
}
projects / hashcat.git / commitdiff