/**
- * Author......: Jens Steube <jens.steube@gmail.com>
+ * Authors.....: Jens Steube <jens.steube@gmail.com>
+ * Gabriele Gristina <matrix@hashcat.net>
+ * magnum <john.magnum@hushmail.com>
+ *
* License.....: MIT
*/
#define _MD5_
+//incompatible because of brances
+//#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"
-
-#define uint_to_hex_lower8(i) l_bin2asc[(i)]
-
-static u32 memcat32 (u32 block0[16], u32 block1[16], const u32 block_len, const u32 append0[4], const u32 append1[4], const u32 append2[4], const u32 append3[4], const u32 append_len)
+#include "OpenCL/simd.c"
+
+#if VECT_SIZE == 1
+#define uint_to_hex_lower8(i) (u32x) (l_bin2asc[(i)])
+#elif VECT_SIZE == 2
+#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
+
+u32 memcat32 (u32x block0[16], u32x block1[16], const u32 block_len, const u32x append0[4], const u32x append1[4], const u32x append2[4], const u32x append3[4], const u32 append_len)
{
const u32 mod = block_len & 3;
const u32 div = block_len / 4;
- #ifdef IS_AMD
+ #if defined IS_AMD || defined IS_GENERIC
const int offset_minus_4 = 4 - mod;
- u32 append0_t[4];
+ u32x append0_t[4];
append0_t[0] = amd_bytealign (append0[0], 0, offset_minus_4);
append0_t[1] = amd_bytealign (append0[1], append0[0], offset_minus_4);
append0_t[2] = amd_bytealign (append0[2], append0[1], offset_minus_4);
append0_t[3] = amd_bytealign (append0[3], append0[2], offset_minus_4);
- u32 append1_t[4];
+ u32x append1_t[4];
append1_t[0] = amd_bytealign (append1[0], append0[3], offset_minus_4);
append1_t[1] = amd_bytealign (append1[1], append1[0], offset_minus_4);
append1_t[2] = amd_bytealign (append1[2], append1[1], offset_minus_4);
append1_t[3] = amd_bytealign (append1[3], append1[2], offset_minus_4);
- u32 append2_t[4];
+ u32x append2_t[4];
append2_t[0] = amd_bytealign (append2[0], append1[3], offset_minus_4);
append2_t[1] = amd_bytealign (append2[1], append2[0], offset_minus_4);
append2_t[2] = amd_bytealign (append2[2], append2[1], offset_minus_4);
append2_t[3] = amd_bytealign (append2[3], append2[2], offset_minus_4);
- u32 append3_t[4];
+ u32x append3_t[4];
append3_t[0] = amd_bytealign (append3[0], append2[3], offset_minus_4);
append3_t[1] = amd_bytealign (append3[1], append3[0], offset_minus_4);
append3_t[2] = amd_bytealign (append3[2], append3[1], offset_minus_4);
append3_t[3] = amd_bytealign (append3[3], append3[2], offset_minus_4);
- u32 append4_t[4];
+ u32x append4_t[4];
append4_t[0] = amd_bytealign ( 0, append3[3], offset_minus_4);
append4_t[1] = 0;
const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;
- u32 append0_t[4];
+ u32x append0_t[4];
append0_t[0] = __byte_perm ( 0, append0[0], selector);
append0_t[1] = __byte_perm (append0[0], append0[1], selector);
append0_t[2] = __byte_perm (append0[1], append0[2], selector);
append0_t[3] = __byte_perm (append0[2], append0[3], selector);
- u32 append1_t[4];
+ u32x append1_t[4];
append1_t[0] = __byte_perm (append0[3], append1[0], selector);
append1_t[1] = __byte_perm (append1[0], append1[1], selector);
append1_t[2] = __byte_perm (append1[1], append1[2], selector);
append1_t[3] = __byte_perm (append1[2], append1[3], selector);
- u32 append2_t[4];
+ u32x append2_t[4];
append2_t[0] = __byte_perm (append1[3], append2[0], selector);
append2_t[1] = __byte_perm (append2[0], append2[1], selector);
append2_t[2] = __byte_perm (append2[1], append2[2], selector);
append2_t[3] = __byte_perm (append2[2], append2[3], selector);
- u32 append3_t[4];
+ u32x append3_t[4];
append3_t[0] = __byte_perm (append2[3], append3[0], selector);
append3_t[1] = __byte_perm (append3[0], append3[1], selector);
append3_t[2] = __byte_perm (append3[1], append3[2], selector);
append3_t[3] = __byte_perm (append3[2], append3[3], selector);
- u32 append4_t[4];
+ u32x append4_t[4];
append4_t[0] = __byte_perm (append3[3], 0, selector);
append4_t[1] = 0;
return new_len;
}
-static void m11400m_0_0 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 l_bin2asc[256])
+void m11400m_0_0 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 sip_t *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, __local u32 *l_bin2asc)
{
/**
* modifier
const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0);
- /**
- * 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
*/
const u32 pw_salt_len = salt_len + pw_len;
u32 salt_buf0[16];
+ u32 salt_buf1[16];
salt_buf0[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[13] = esalt_bufs[salt_pos].salt_buf[13];
salt_buf0[14] = esalt_bufs[salt_pos].salt_buf[14];
salt_buf0[15] = esalt_bufs[salt_pos].salt_buf[15];
-
- u32 salt_buf1[16];
-
salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[16];
salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[17];
salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[18];
const u32 esalt_len = esalt_bufs[salt_pos].esalt_len;
u32 esalt_buf0[16];
+ u32 esalt_buf1[16];
esalt_buf0[ 0] = esalt_bufs[salt_pos].esalt_buf[ 0];
esalt_buf0[ 1] = esalt_bufs[salt_pos].esalt_buf[ 1];
esalt_buf0[13] = esalt_bufs[salt_pos].esalt_buf[13];
esalt_buf0[14] = esalt_bufs[salt_pos].esalt_buf[14];
esalt_buf0[15] = esalt_bufs[salt_pos].esalt_buf[15];
-
- u32 esalt_buf1[16];
-
esalt_buf1[ 0] = esalt_bufs[salt_pos].esalt_buf[16];
esalt_buf1[ 1] = esalt_bufs[salt_pos].esalt_buf[17];
esalt_buf1[ 2] = esalt_bufs[salt_pos].esalt_buf[18];
u32 w0l = w0[0];
- for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 w0r = bfs_buf[il_pos].i;
+ const u32x w0r = ix_create_bft (bfs_buf, il_pos);
- w0[0] = w0l | w0r;
+ const u32x w0lr = w0l | w0r;
/*
* HA1 = md5 ($salt . $pass)
// append the pass to the salt
- u32 block0[16];
+ u32x block0[16];
+ u32x block1[16];
block0[ 0] = salt_buf0[ 0];
block0[ 1] = salt_buf0[ 1];
block0[13] = salt_buf0[13];
block0[14] = salt_buf0[14];
block0[15] = salt_buf0[15];
-
- u32 block1[16];
-
block1[ 0] = salt_buf1[ 0];
block1[ 1] = salt_buf1[ 1];
block1[ 2] = salt_buf1[ 2];
block1[14] = salt_buf1[14];
block1[15] = salt_buf1[15];
- memcat32 (block0, block1, salt_len, w0, w1, w2, w3, pw_len);
-
- u32 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[1] = w0[1];
+ w0_t[2] = w0[2];
+ w0_t[3] = w0[3];
+ w1_t[0] = w1[0];
+ w1_t[1] = w1[1];
+ w1_t[2] = w1[2];
+ w1_t[3] = w1[3];
+ w2_t[0] = w2[0];
+ w2_t[1] = w2[1];
+ w2_t[2] = w2[2];
+ w2_t[3] = w2[3];
+ w3_t[0] = w3[0];
+ w3_t[1] = w3[1];
+ w3_t[2] = w3[2];
+ w3_t[3] = w3[3];
+
+ memcat32 (block0, block1, salt_len, w0_t, w1_t, w2_t, w3_t, pw_len);
w0_t[0] = block0[ 0];
w0_t[1] = block0[ 1];
w0_t[2] = block0[ 2];
w0_t[3] = block0[ 3];
-
- u32 w1_t[4];
-
w1_t[0] = block0[ 4];
w1_t[1] = block0[ 5];
w1_t[2] = block0[ 6];
w1_t[3] = block0[ 7];
-
- u32 w2_t[4];
-
w2_t[0] = block0[ 8];
w2_t[1] = block0[ 9];
w2_t[2] = block0[10];
w2_t[3] = block0[11];
-
- u32 w3_t[4];
-
w3_t[0] = block0[12];
w3_t[1] = block0[13];
w3_t[2] = pw_salt_len * 8;
// md5
- u32 tmp2;
-
- u32 a = MD5M_A;
- u32 b = MD5M_B;
- 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_t[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
-
w2_t[0] = esalt_buf0[0];
w2_t[1] = esalt_buf0[1];
w2_t[2] = esalt_buf0[2];
w2_t[3] = esalt_buf0[3];
-
w3_t[0] = esalt_buf0[4];
w3_t[1] = esalt_buf0[5];
w3_t[2] = esalt_buf0[6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += MD5M_C;
d += MD5M_D;
- u32 r_a = a;
- u32 r_b = b;
- u32 r_c = c;
- u32 r_d = d;
+ u32x r_a = a;
+ u32x r_b = b;
+ u32x r_c = c;
+ u32x r_d = d;
// 2nd transform
w0_t[1] = esalt_buf0[ 9];
w0_t[2] = esalt_buf0[10];
w0_t[3] = esalt_buf0[11];
-
w1_t[0] = esalt_buf0[12];
w1_t[1] = esalt_buf0[13];
w1_t[2] = esalt_buf0[14];
w1_t[3] = esalt_buf0[15];
-
w2_t[0] = esalt_buf1[ 0];
w2_t[1] = esalt_buf1[ 1];
w2_t[2] = esalt_buf1[ 2];
w2_t[3] = esalt_buf1[ 3];
-
w3_t[0] = esalt_buf1[ 4];
w3_t[1] = esalt_buf1[ 5];
w3_t[2] = digest_esalt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += r_c;
d += r_d;
- const u32 r0 = a;
- const u32 r1 = d;
- const u32 r2 = c;
- const u32 r3 = b;
-
- #include COMPARE_S
+ COMPARE_M_SIMD (a, d, c, b);
}
}
-static void m11400m_0_1 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 l_bin2asc[256])
+void m11400m_0_1 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 sip_t *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, __local u32 *l_bin2asc)
{
/**
* modifier
const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0);
- /**
- * 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
*/
const u32 pw_salt_len = salt_len + pw_len;
u32 salt_buf0[16];
+ u32 salt_buf1[16];
salt_buf0[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[13] = esalt_bufs[salt_pos].salt_buf[13];
salt_buf0[14] = esalt_bufs[salt_pos].salt_buf[14];
salt_buf0[15] = esalt_bufs[salt_pos].salt_buf[15];
-
- u32 salt_buf1[16];
-
salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[16];
salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[17];
salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[18];
const u32 esalt_len = esalt_bufs[salt_pos].esalt_len;
u32 esalt_buf0[16];
+ u32 esalt_buf1[16];
+ u32 esalt_buf2[16];
esalt_buf0[ 0] = esalt_bufs[salt_pos].esalt_buf[ 0];
esalt_buf0[ 1] = esalt_bufs[salt_pos].esalt_buf[ 1];
esalt_buf0[13] = esalt_bufs[salt_pos].esalt_buf[13];
esalt_buf0[14] = esalt_bufs[salt_pos].esalt_buf[14];
esalt_buf0[15] = esalt_bufs[salt_pos].esalt_buf[15];
-
- u32 esalt_buf1[16];
-
esalt_buf1[ 0] = esalt_bufs[salt_pos].esalt_buf[16];
esalt_buf1[ 1] = esalt_bufs[salt_pos].esalt_buf[17];
esalt_buf1[ 2] = esalt_bufs[salt_pos].esalt_buf[18];
esalt_buf1[13] = esalt_bufs[salt_pos].esalt_buf[29];
esalt_buf1[14] = esalt_bufs[salt_pos].esalt_buf[30];
esalt_buf1[15] = esalt_bufs[salt_pos].esalt_buf[31];
-
- u32 esalt_buf2[16];
-
esalt_buf2[ 0] = esalt_bufs[salt_pos].esalt_buf[32];
esalt_buf2[ 1] = esalt_bufs[salt_pos].esalt_buf[33];
esalt_buf2[ 2] = esalt_bufs[salt_pos].esalt_buf[34];
u32 w0l = w0[0];
- for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 w0r = bfs_buf[il_pos].i;
+ const u32x w0r = ix_create_bft (bfs_buf, il_pos);
- w0[0] = w0l | w0r;
+ const u32x w0lr = w0l | w0r;
/*
* HA1 = md5 ($salt . $pass)
// append the pass to the salt
- u32 block0[16];
+ u32x block0[16];
+ u32x block1[16];
block0[ 0] = salt_buf0[ 0];
block0[ 1] = salt_buf0[ 1];
block0[13] = salt_buf0[13];
block0[14] = salt_buf0[14];
block0[15] = salt_buf0[15];
-
- u32 block1[16];
-
block1[ 0] = salt_buf1[ 0];
block1[ 1] = salt_buf1[ 1];
block1[ 2] = salt_buf1[ 2];
block1[14] = salt_buf1[14];
block1[15] = salt_buf1[15];
- memcat32 (block0, block1, salt_len, w0, w1, w2, w3, pw_len);
-
- u32 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[1] = w0[1];
+ w0_t[2] = w0[2];
+ w0_t[3] = w0[3];
+ w1_t[0] = w1[0];
+ w1_t[1] = w1[1];
+ w1_t[2] = w1[2];
+ w1_t[3] = w1[3];
+ w2_t[0] = w2[0];
+ w2_t[1] = w2[1];
+ w2_t[2] = w2[2];
+ w2_t[3] = w2[3];
+ w3_t[0] = w3[0];
+ w3_t[1] = w3[1];
+ w3_t[2] = w3[2];
+ w3_t[3] = w3[3];
+
+ memcat32 (block0, block1, salt_len, w0_t, w1_t, w2_t, w3_t, pw_len);
w0_t[0] = block0[ 0];
w0_t[1] = block0[ 1];
w0_t[2] = block0[ 2];
w0_t[3] = block0[ 3];
-
- u32 w1_t[4];
-
w1_t[0] = block0[ 4];
w1_t[1] = block0[ 5];
w1_t[2] = block0[ 6];
w1_t[3] = block0[ 7];
-
- u32 w2_t[4];
-
w2_t[0] = block0[ 8];
w2_t[1] = block0[ 9];
w2_t[2] = block0[10];
w2_t[3] = block0[11];
-
- u32 w3_t[4];
-
w3_t[0] = block0[12];
w3_t[1] = block0[13];
w3_t[2] = pw_salt_len * 8;
// md5
- u32 tmp2;
-
- u32 a = MD5M_A;
- u32 b = MD5M_B;
- 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_t[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
-
w2_t[0] = esalt_buf0[0];
w2_t[1] = esalt_buf0[1];
w2_t[2] = esalt_buf0[2];
w2_t[3] = esalt_buf0[3];
-
w3_t[0] = esalt_buf0[4];
w3_t[1] = esalt_buf0[5];
w3_t[2] = esalt_buf0[6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += MD5M_C;
d += MD5M_D;
- u32 r_a = a;
- u32 r_b = b;
- u32 r_c = c;
- u32 r_d = d;
+ u32x r_a = a;
+ u32x r_b = b;
+ u32x r_c = c;
+ u32x r_d = d;
// 2nd transform
w0_t[1] = esalt_buf0[ 9];
w0_t[2] = esalt_buf0[10];
w0_t[3] = esalt_buf0[11];
-
w1_t[0] = esalt_buf0[12];
w1_t[1] = esalt_buf0[13];
w1_t[2] = esalt_buf0[14];
w1_t[3] = esalt_buf0[15];
-
w2_t[0] = esalt_buf1[ 0];
w2_t[1] = esalt_buf1[ 1];
w2_t[2] = esalt_buf1[ 2];
w2_t[3] = esalt_buf1[ 3];
-
w3_t[0] = esalt_buf1[ 4];
w3_t[1] = esalt_buf1[ 5];
w3_t[2] = esalt_buf1[ 6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
w0_t[1] = esalt_buf1[ 9];
w0_t[2] = esalt_buf1[10];
w0_t[3] = esalt_buf1[11];
-
w1_t[0] = esalt_buf1[12];
w1_t[1] = esalt_buf1[13];
w1_t[2] = esalt_buf1[14];
w1_t[3] = esalt_buf1[15];
-
w2_t[0] = esalt_buf2[ 0];
w2_t[1] = esalt_buf2[ 1];
w2_t[2] = esalt_buf2[ 2];
w2_t[3] = esalt_buf2[ 3];
-
w3_t[0] = esalt_buf2[ 4];
w3_t[1] = esalt_buf2[ 5];
w3_t[2] = digest_esalt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += r_c;
d += r_d;
- const u32 r0 = a;
- const u32 r1 = d;
- const u32 r2 = c;
- const u32 r3 = b;
-
- #include COMPARE_S
+ COMPARE_M_SIMD (a, d, c, b);
}
}
-static void m11400m_1_0 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 l_bin2asc[256])
+void m11400m_1_0 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 sip_t *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, __local u32 *l_bin2asc)
{
/**
* modifier
const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0);
- /**
- * 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
*/
const u32 pw_salt_len = salt_len + pw_len;
u32 salt_buf0[16];
+ u32 salt_buf1[16];
salt_buf0[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[13] = esalt_bufs[salt_pos].salt_buf[13];
salt_buf0[14] = esalt_bufs[salt_pos].salt_buf[14];
salt_buf0[15] = esalt_bufs[salt_pos].salt_buf[15];
-
- u32 salt_buf1[16];
-
salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[16];
salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[17];
salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[18];
const u32 esalt_len = esalt_bufs[salt_pos].esalt_len;
u32 esalt_buf0[16];
+ u32 esalt_buf1[16];
esalt_buf0[ 0] = esalt_bufs[salt_pos].esalt_buf[ 0];
esalt_buf0[ 1] = esalt_bufs[salt_pos].esalt_buf[ 1];
esalt_buf0[13] = esalt_bufs[salt_pos].esalt_buf[13];
esalt_buf0[14] = esalt_bufs[salt_pos].esalt_buf[14];
esalt_buf0[15] = esalt_bufs[salt_pos].esalt_buf[15];
-
- u32 esalt_buf1[16];
-
esalt_buf1[ 0] = esalt_bufs[salt_pos].esalt_buf[16];
esalt_buf1[ 1] = esalt_bufs[salt_pos].esalt_buf[17];
esalt_buf1[ 2] = esalt_bufs[salt_pos].esalt_buf[18];
u32 w0l = w0[0];
- for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 w0r = bfs_buf[il_pos].i;
+ const u32x w0r = ix_create_bft (bfs_buf, il_pos);
- w0[0] = w0l | w0r;
+ const u32x w0lr = w0l | w0r;
/*
* HA1 = md5 ($salt . $pass)
// append the pass to the salt
- u32 block0[16];
+ u32x block0[16];
+ u32x block1[16];
block0[ 0] = salt_buf0[ 0];
block0[ 1] = salt_buf0[ 1];
block0[13] = salt_buf0[13];
block0[14] = salt_buf0[14];
block0[15] = salt_buf0[15];
-
- u32 block1[16];
-
block1[ 0] = salt_buf1[ 0];
block1[ 1] = salt_buf1[ 1];
block1[ 2] = salt_buf1[ 2];
block1[14] = salt_buf1[14];
block1[15] = salt_buf1[15];
- memcat32 (block0, block1, salt_len, w0, w1, w2, w3, pw_len);
-
- u32 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[1] = w0[1];
+ w0_t[2] = w0[2];
+ w0_t[3] = w0[3];
+ w1_t[0] = w1[0];
+ w1_t[1] = w1[1];
+ w1_t[2] = w1[2];
+ w1_t[3] = w1[3];
+ w2_t[0] = w2[0];
+ w2_t[1] = w2[1];
+ w2_t[2] = w2[2];
+ w2_t[3] = w2[3];
+ w3_t[0] = w3[0];
+ w3_t[1] = w3[1];
+ w3_t[2] = w3[2];
+ w3_t[3] = w3[3];
+
+ memcat32 (block0, block1, salt_len, w0_t, w1_t, w2_t, w3_t, pw_len);
w0_t[0] = block0[ 0];
w0_t[1] = block0[ 1];
w0_t[2] = block0[ 2];
w0_t[3] = block0[ 3];
-
- u32 w1_t[4];
-
w1_t[0] = block0[ 4];
w1_t[1] = block0[ 5];
w1_t[2] = block0[ 6];
w1_t[3] = block0[ 7];
-
- u32 w2_t[4];
-
w2_t[0] = block0[ 8];
w2_t[1] = block0[ 9];
w2_t[2] = block0[10];
w2_t[3] = block0[11];
-
- u32 w3_t[4];
-
w3_t[0] = block0[12];
w3_t[1] = block0[13];
w3_t[2] = block0[14];
// md5
- u32 tmp2;
-
- u32 a = MD5M_A;
- u32 b = MD5M_B;
- 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_t[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += MD5M_C;
d += MD5M_D;
- u32 r_a = a;
- u32 r_b = b;
- u32 r_c = c;
- u32 r_d = d;
+ u32x r_a = a;
+ u32x r_b = b;
+ u32x r_c = c;
+ u32x r_d = d;
w0_t[0] = block1[ 0];
w0_t[1] = block1[ 1];
w0_t[2] = block1[ 2];
w0_t[3] = block1[ 3];
-
w1_t[0] = block1[ 4];
w1_t[1] = block1[ 5];
w1_t[2] = block1[ 6];
w1_t[3] = block1[ 7];
-
w2_t[0] = block1[ 8];
w2_t[1] = block1[ 9];
w2_t[2] = block1[10];
w2_t[3] = block1[11];
-
w3_t[0] = block1[12];
w3_t[1] = block1[13];
w3_t[2] = pw_salt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
-
w2_t[0] = esalt_buf0[0];
w2_t[1] = esalt_buf0[1];
w2_t[2] = esalt_buf0[2];
w2_t[3] = esalt_buf0[3];
-
w3_t[0] = esalt_buf0[4];
w3_t[1] = esalt_buf0[5];
w3_t[2] = esalt_buf0[6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
w0_t[1] = esalt_buf0[ 9];
w0_t[2] = esalt_buf0[10];
w0_t[3] = esalt_buf0[11];
-
w1_t[0] = esalt_buf0[12];
w1_t[1] = esalt_buf0[13];
w1_t[2] = esalt_buf0[14];
w1_t[3] = esalt_buf0[15];
-
w2_t[0] = esalt_buf1[ 0];
w2_t[1] = esalt_buf1[ 1];
w2_t[2] = esalt_buf1[ 2];
w2_t[3] = esalt_buf1[ 3];
-
w3_t[0] = esalt_buf1[ 4];
w3_t[1] = esalt_buf1[ 5];
w3_t[2] = digest_esalt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += r_c;
d += r_d;
- const u32 r0 = a;
- const u32 r1 = d;
- const u32 r2 = c;
- const u32 r3 = b;
-
- #include COMPARE_S
+ COMPARE_M_SIMD (a, d, c, b);
}
}
-static void m11400m_1_1 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 l_bin2asc[256])
+void m11400m_1_1 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 sip_t *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, __local u32 *l_bin2asc)
{
/**
* modifier
const u32 gid = get_global_id (0);
const u32 lid = get_local_id (0);
- /**
- * 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
*/
const u32 pw_salt_len = salt_len + pw_len;
u32 salt_buf0[16];
+ u32 salt_buf1[16];
salt_buf0[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[13] = esalt_bufs[salt_pos].salt_buf[13];
salt_buf0[14] = esalt_bufs[salt_pos].salt_buf[14];
salt_buf0[15] = esalt_bufs[salt_pos].salt_buf[15];
-
- u32 salt_buf1[16];
-
salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[16];
salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[17];
salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[18];
const u32 esalt_len = esalt_bufs[salt_pos].esalt_len;
u32 esalt_buf0[16];
+ u32 esalt_buf1[16];
+ u32 esalt_buf2[16];
esalt_buf0[ 0] = esalt_bufs[salt_pos].esalt_buf[ 0];
esalt_buf0[ 1] = esalt_bufs[salt_pos].esalt_buf[ 1];
esalt_buf0[13] = esalt_bufs[salt_pos].esalt_buf[13];
esalt_buf0[14] = esalt_bufs[salt_pos].esalt_buf[14];
esalt_buf0[15] = esalt_bufs[salt_pos].esalt_buf[15];
-
- u32 esalt_buf1[16];
-
esalt_buf1[ 0] = esalt_bufs[salt_pos].esalt_buf[16];
esalt_buf1[ 1] = esalt_bufs[salt_pos].esalt_buf[17];
esalt_buf1[ 2] = esalt_bufs[salt_pos].esalt_buf[18];
esalt_buf1[13] = esalt_bufs[salt_pos].esalt_buf[29];
esalt_buf1[14] = esalt_bufs[salt_pos].esalt_buf[30];
esalt_buf1[15] = esalt_bufs[salt_pos].esalt_buf[31];
-
- u32 esalt_buf2[16];
-
esalt_buf2[ 0] = esalt_bufs[salt_pos].esalt_buf[32];
esalt_buf2[ 1] = esalt_bufs[salt_pos].esalt_buf[33];
esalt_buf2[ 2] = esalt_bufs[salt_pos].esalt_buf[34];
u32 w0l = w0[0];
- for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 w0r = bfs_buf[il_pos].i;
+ const u32x w0r = ix_create_bft (bfs_buf, il_pos);
- w0[0] = w0l | w0r;
+ const u32x w0lr = w0l | w0r;
/*
* HA1 = md5 ($salt . $pass)
// append the pass to the salt
- u32 block0[16];
+ u32x block0[16];
+ u32x block1[16];
block0[ 0] = salt_buf0[ 0];
block0[ 1] = salt_buf0[ 1];
block0[13] = salt_buf0[13];
block0[14] = salt_buf0[14];
block0[15] = salt_buf0[15];
-
- u32 block1[16];
-
block1[ 0] = salt_buf1[ 0];
block1[ 1] = salt_buf1[ 1];
block1[ 2] = salt_buf1[ 2];
block1[14] = salt_buf1[14];
block1[15] = salt_buf1[15];
- memcat32 (block0, block1, salt_len, w0, w1, w2, w3, pw_len);
-
- u32 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[1] = w0[1];
+ w0_t[2] = w0[2];
+ w0_t[3] = w0[3];
+ w1_t[0] = w1[0];
+ w1_t[1] = w1[1];
+ w1_t[2] = w1[2];
+ w1_t[3] = w1[3];
+ w2_t[0] = w2[0];
+ w2_t[1] = w2[1];
+ w2_t[2] = w2[2];
+ w2_t[3] = w2[3];
+ w3_t[0] = w3[0];
+ w3_t[1] = w3[1];
+ w3_t[2] = w3[2];
+ w3_t[3] = w3[3];
+
+ memcat32 (block0, block1, salt_len, w0_t, w1_t, w2_t, w3_t, pw_len);
w0_t[0] = block0[ 0];
w0_t[1] = block0[ 1];
w0_t[2] = block0[ 2];
w0_t[3] = block0[ 3];
-
- u32 w1_t[4];
-
w1_t[0] = block0[ 4];
w1_t[1] = block0[ 5];
w1_t[2] = block0[ 6];
w1_t[3] = block0[ 7];
-
- u32 w2_t[4];
-
w2_t[0] = block0[ 8];
w2_t[1] = block0[ 9];
w2_t[2] = block0[10];
w2_t[3] = block0[11];
-
- u32 w3_t[4];
-
w3_t[0] = block0[12];
w3_t[1] = block0[13];
w3_t[2] = block0[14];
// md5
- u32 tmp2;
-
- u32 a = MD5M_A;
- u32 b = MD5M_B;
- 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_t[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += MD5M_C;
d += MD5M_D;
- u32 r_a = a;
- u32 r_b = b;
- u32 r_c = c;
- u32 r_d = d;
+ u32x r_a = a;
+ u32x r_b = b;
+ u32x r_c = c;
+ u32x r_d = d;
w0_t[0] = block1[ 0];
w0_t[1] = block1[ 1];
w0_t[2] = block1[ 2];
w0_t[3] = block1[ 3];
-
w1_t[0] = block1[ 4];
w1_t[1] = block1[ 5];
w1_t[2] = block1[ 6];
w1_t[3] = block1[ 7];
-
w2_t[0] = block1[ 8];
w2_t[1] = block1[ 9];
w2_t[2] = block1[10];
w2_t[3] = block1[11];
-
w3_t[0] = block1[12];
w3_t[1] = block1[13];
w3_t[2] = pw_salt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
-
w2_t[0] = esalt_buf0[0];
w2_t[1] = esalt_buf0[1];
w2_t[2] = esalt_buf0[2];
w2_t[3] = esalt_buf0[3];
-
w3_t[0] = esalt_buf0[4];
w3_t[1] = esalt_buf0[5];
w3_t[2] = esalt_buf0[6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
w0_t[1] = esalt_buf0[ 9];
w0_t[2] = esalt_buf0[10];
w0_t[3] = esalt_buf0[11];
-
w1_t[0] = esalt_buf0[12];
w1_t[1] = esalt_buf0[13];
w1_t[2] = esalt_buf0[14];
w1_t[3] = esalt_buf0[15];
-
w2_t[0] = esalt_buf1[ 0];
w2_t[1] = esalt_buf1[ 1];
w2_t[2] = esalt_buf1[ 2];
w2_t[3] = esalt_buf1[ 3];
-
w3_t[0] = esalt_buf1[ 4];
w3_t[1] = esalt_buf1[ 5];
w3_t[2] = esalt_buf1[ 6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
w0_t[1] = esalt_buf1[ 9];
w0_t[2] = esalt_buf1[10];
w0_t[3] = esalt_buf1[11];
-
w1_t[0] = esalt_buf1[12];
w1_t[1] = esalt_buf1[13];
w1_t[2] = esalt_buf1[14];
w1_t[3] = esalt_buf1[15];
-
w2_t[0] = esalt_buf2[ 0];
w2_t[1] = esalt_buf2[ 1];
w2_t[2] = esalt_buf2[ 2];
w2_t[3] = esalt_buf2[ 3];
-
w3_t[0] = esalt_buf2[ 4];
w3_t[1] = esalt_buf2[ 5];
w3_t[2] = digest_esalt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += r_c;
d += r_d;
- const u32 r0 = a;
- const u32 r1 = d;
- const u32 r2 = c;
- const u32 r3 = b;
-
- #include COMPARE_S
+ COMPARE_M_SIMD (a, d, c, b);
}
}
-static void m11400s_0_0 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 l_bin2asc[256])
+void m11400s_0_0 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 sip_t *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, __local u32 *l_bin2asc)
{
/**
* modifier
const u32 pw_salt_len = salt_len + pw_len;
u32 salt_buf0[16];
+ u32 salt_buf1[16];
salt_buf0[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[13] = esalt_bufs[salt_pos].salt_buf[13];
salt_buf0[14] = esalt_bufs[salt_pos].salt_buf[14];
salt_buf0[15] = esalt_bufs[salt_pos].salt_buf[15];
-
- u32 salt_buf1[16];
-
salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[16];
salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[17];
salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[18];
const u32 esalt_len = esalt_bufs[salt_pos].esalt_len;
u32 esalt_buf0[16];
+ u32 esalt_buf1[16];
esalt_buf0[ 0] = esalt_bufs[salt_pos].esalt_buf[ 0];
esalt_buf0[ 1] = esalt_bufs[salt_pos].esalt_buf[ 1];
esalt_buf0[13] = esalt_bufs[salt_pos].esalt_buf[13];
esalt_buf0[14] = esalt_bufs[salt_pos].esalt_buf[14];
esalt_buf0[15] = esalt_bufs[salt_pos].esalt_buf[15];
-
- u32 esalt_buf1[16];
-
esalt_buf1[ 0] = esalt_bufs[salt_pos].esalt_buf[16];
esalt_buf1[ 1] = esalt_bufs[salt_pos].esalt_buf[17];
esalt_buf1[ 2] = esalt_bufs[salt_pos].esalt_buf[18];
u32 w0l = w0[0];
- for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 w0r = bfs_buf[il_pos].i;
+ const u32x w0r = ix_create_bft (bfs_buf, il_pos);
- w0[0] = w0l | w0r;
+ const u32x w0lr = w0l | w0r;
/*
* HA1 = md5 ($salt . $pass)
// append the pass to the salt
- u32 block0[16];
+ u32x block0[16];
+ u32x block1[16];
block0[ 0] = salt_buf0[ 0];
block0[ 1] = salt_buf0[ 1];
block0[13] = salt_buf0[13];
block0[14] = salt_buf0[14];
block0[15] = salt_buf0[15];
-
- u32 block1[16];
-
block1[ 0] = salt_buf1[ 0];
block1[ 1] = salt_buf1[ 1];
block1[ 2] = salt_buf1[ 2];
block1[14] = salt_buf1[14];
block1[15] = salt_buf1[15];
- memcat32 (block0, block1, salt_len, w0, w1, w2, w3, pw_len);
-
- u32 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[1] = w0[1];
+ w0_t[2] = w0[2];
+ w0_t[3] = w0[3];
+ w1_t[0] = w1[0];
+ w1_t[1] = w1[1];
+ w1_t[2] = w1[2];
+ w1_t[3] = w1[3];
+ w2_t[0] = w2[0];
+ w2_t[1] = w2[1];
+ w2_t[2] = w2[2];
+ w2_t[3] = w2[3];
+ w3_t[0] = w3[0];
+ w3_t[1] = w3[1];
+ w3_t[2] = w3[2];
+ w3_t[3] = w3[3];
+
+ memcat32 (block0, block1, salt_len, w0_t, w1_t, w2_t, w3_t, pw_len);
w0_t[0] = block0[ 0];
w0_t[1] = block0[ 1];
w0_t[2] = block0[ 2];
w0_t[3] = block0[ 3];
-
- u32 w1_t[4];
-
w1_t[0] = block0[ 4];
w1_t[1] = block0[ 5];
w1_t[2] = block0[ 6];
w1_t[3] = block0[ 7];
-
- u32 w2_t[4];
-
w2_t[0] = block0[ 8];
w2_t[1] = block0[ 9];
w2_t[2] = block0[10];
w2_t[3] = block0[11];
-
- u32 w3_t[4];
-
w3_t[0] = block0[12];
w3_t[1] = block0[13];
w3_t[2] = pw_salt_len * 8;
// md5
- u32 tmp2;
-
- u32 a = MD5M_A;
- u32 b = MD5M_B;
- 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_t[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
-
w2_t[0] = esalt_buf0[0];
w2_t[1] = esalt_buf0[1];
w2_t[2] = esalt_buf0[2];
w2_t[3] = esalt_buf0[3];
-
w3_t[0] = esalt_buf0[4];
w3_t[1] = esalt_buf0[5];
w3_t[2] = esalt_buf0[6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += MD5M_C;
d += MD5M_D;
- u32 r_a = a;
- u32 r_b = b;
- u32 r_c = c;
- u32 r_d = d;
+ u32x r_a = a;
+ u32x r_b = b;
+ u32x r_c = c;
+ u32x r_d = d;
// 2nd transform
w0_t[1] = esalt_buf0[ 9];
w0_t[2] = esalt_buf0[10];
w0_t[3] = esalt_buf0[11];
-
w1_t[0] = esalt_buf0[12];
w1_t[1] = esalt_buf0[13];
w1_t[2] = esalt_buf0[14];
w1_t[3] = esalt_buf0[15];
-
w2_t[0] = esalt_buf1[ 0];
w2_t[1] = esalt_buf1[ 1];
w2_t[2] = esalt_buf1[ 2];
w2_t[3] = esalt_buf1[ 3];
-
w3_t[0] = esalt_buf1[ 4];
w3_t[1] = esalt_buf1[ 5];
w3_t[2] = digest_esalt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += r_c;
d += r_d;
- const u32 r0 = a;
- const u32 r1 = d;
- const u32 r2 = c;
- const u32 r3 = b;
-
- #include COMPARE_S
+ COMPARE_S_SIMD (a, d, c, b);
}
}
-static void m11400s_0_1 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 l_bin2asc[256])
+void m11400s_0_1 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 sip_t *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, __local u32 *l_bin2asc)
{
/**
* modifier
const u32 pw_salt_len = salt_len + pw_len;
u32 salt_buf0[16];
+ u32 salt_buf1[16];
salt_buf0[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[13] = esalt_bufs[salt_pos].salt_buf[13];
salt_buf0[14] = esalt_bufs[salt_pos].salt_buf[14];
salt_buf0[15] = esalt_bufs[salt_pos].salt_buf[15];
-
- u32 salt_buf1[16];
-
salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[16];
salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[17];
salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[18];
const u32 esalt_len = esalt_bufs[salt_pos].esalt_len;
u32 esalt_buf0[16];
+ u32 esalt_buf1[16];
+ u32 esalt_buf2[16];
esalt_buf0[ 0] = esalt_bufs[salt_pos].esalt_buf[ 0];
esalt_buf0[ 1] = esalt_bufs[salt_pos].esalt_buf[ 1];
esalt_buf0[13] = esalt_bufs[salt_pos].esalt_buf[13];
esalt_buf0[14] = esalt_bufs[salt_pos].esalt_buf[14];
esalt_buf0[15] = esalt_bufs[salt_pos].esalt_buf[15];
-
- u32 esalt_buf1[16];
-
esalt_buf1[ 0] = esalt_bufs[salt_pos].esalt_buf[16];
esalt_buf1[ 1] = esalt_bufs[salt_pos].esalt_buf[17];
esalt_buf1[ 2] = esalt_bufs[salt_pos].esalt_buf[18];
esalt_buf1[13] = esalt_bufs[salt_pos].esalt_buf[29];
esalt_buf1[14] = esalt_bufs[salt_pos].esalt_buf[30];
esalt_buf1[15] = esalt_bufs[salt_pos].esalt_buf[31];
-
- u32 esalt_buf2[16];
-
esalt_buf2[ 0] = esalt_bufs[salt_pos].esalt_buf[32];
esalt_buf2[ 1] = esalt_bufs[salt_pos].esalt_buf[33];
esalt_buf2[ 2] = esalt_bufs[salt_pos].esalt_buf[34];
u32 w0l = w0[0];
- for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 w0r = bfs_buf[il_pos].i;
+ const u32x w0r = ix_create_bft (bfs_buf, il_pos);
- w0[0] = w0l | w0r;
+ const u32x w0lr = w0l | w0r;
/*
* HA1 = md5 ($salt . $pass)
// append the pass to the salt
- u32 block0[16];
+ u32x block0[16];
+ u32x block1[16];
block0[ 0] = salt_buf0[ 0];
block0[ 1] = salt_buf0[ 1];
block0[13] = salt_buf0[13];
block0[14] = salt_buf0[14];
block0[15] = salt_buf0[15];
-
- u32 block1[16];
-
block1[ 0] = salt_buf1[ 0];
block1[ 1] = salt_buf1[ 1];
block1[ 2] = salt_buf1[ 2];
block1[14] = salt_buf1[14];
block1[15] = salt_buf1[15];
- memcat32 (block0, block1, salt_len, w0, w1, w2, w3, pw_len);
-
- u32 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[1] = w0[1];
+ w0_t[2] = w0[2];
+ w0_t[3] = w0[3];
+ w1_t[0] = w1[0];
+ w1_t[1] = w1[1];
+ w1_t[2] = w1[2];
+ w1_t[3] = w1[3];
+ w2_t[0] = w2[0];
+ w2_t[1] = w2[1];
+ w2_t[2] = w2[2];
+ w2_t[3] = w2[3];
+ w3_t[0] = w3[0];
+ w3_t[1] = w3[1];
+ w3_t[2] = w3[2];
+ w3_t[3] = w3[3];
+
+ memcat32 (block0, block1, salt_len, w0_t, w1_t, w2_t, w3_t, pw_len);
w0_t[0] = block0[ 0];
w0_t[1] = block0[ 1];
w0_t[2] = block0[ 2];
w0_t[3] = block0[ 3];
-
- u32 w1_t[4];
-
w1_t[0] = block0[ 4];
w1_t[1] = block0[ 5];
w1_t[2] = block0[ 6];
w1_t[3] = block0[ 7];
-
- u32 w2_t[4];
-
w2_t[0] = block0[ 8];
w2_t[1] = block0[ 9];
w2_t[2] = block0[10];
w2_t[3] = block0[11];
-
- u32 w3_t[4];
-
w3_t[0] = block0[12];
w3_t[1] = block0[13];
w3_t[2] = pw_salt_len * 8;
// md5
- u32 tmp2;
-
- u32 a = MD5M_A;
- u32 b = MD5M_B;
- 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_t[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
-
w2_t[0] = esalt_buf0[0];
w2_t[1] = esalt_buf0[1];
w2_t[2] = esalt_buf0[2];
w2_t[3] = esalt_buf0[3];
-
w3_t[0] = esalt_buf0[4];
w3_t[1] = esalt_buf0[5];
w3_t[2] = esalt_buf0[6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += MD5M_C;
d += MD5M_D;
- u32 r_a = a;
- u32 r_b = b;
- u32 r_c = c;
- u32 r_d = d;
+ u32x r_a = a;
+ u32x r_b = b;
+ u32x r_c = c;
+ u32x r_d = d;
// 2nd transform
w0_t[1] = esalt_buf0[ 9];
w0_t[2] = esalt_buf0[10];
w0_t[3] = esalt_buf0[11];
-
w1_t[0] = esalt_buf0[12];
w1_t[1] = esalt_buf0[13];
w1_t[2] = esalt_buf0[14];
w1_t[3] = esalt_buf0[15];
-
w2_t[0] = esalt_buf1[ 0];
w2_t[1] = esalt_buf1[ 1];
w2_t[2] = esalt_buf1[ 2];
w2_t[3] = esalt_buf1[ 3];
-
w3_t[0] = esalt_buf1[ 4];
w3_t[1] = esalt_buf1[ 5];
w3_t[2] = esalt_buf1[ 6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
w0_t[1] = esalt_buf1[ 9];
w0_t[2] = esalt_buf1[10];
w0_t[3] = esalt_buf1[11];
-
w1_t[0] = esalt_buf1[12];
w1_t[1] = esalt_buf1[13];
w1_t[2] = esalt_buf1[14];
w1_t[3] = esalt_buf1[15];
-
w2_t[0] = esalt_buf2[ 0];
w2_t[1] = esalt_buf2[ 1];
w2_t[2] = esalt_buf2[ 2];
w2_t[3] = esalt_buf2[ 3];
-
w3_t[0] = esalt_buf2[ 4];
w3_t[1] = esalt_buf2[ 5];
w3_t[2] = digest_esalt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += r_c;
d += r_d;
- const u32 r0 = a;
- const u32 r1 = d;
- const u32 r2 = c;
- const u32 r3 = b;
-
- #include COMPARE_S
+ COMPARE_S_SIMD (a, d, c, b);
}
}
-static void m11400s_1_0 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 l_bin2asc[256])
+void m11400s_1_0 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 sip_t *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, __local u32 *l_bin2asc)
{
/**
* modifier
const u32 pw_salt_len = salt_len + pw_len;
u32 salt_buf0[16];
+ u32 salt_buf1[16];
salt_buf0[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[13] = esalt_bufs[salt_pos].salt_buf[13];
salt_buf0[14] = esalt_bufs[salt_pos].salt_buf[14];
salt_buf0[15] = esalt_bufs[salt_pos].salt_buf[15];
-
- u32 salt_buf1[16];
-
salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[16];
salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[17];
salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[18];
const u32 esalt_len = esalt_bufs[salt_pos].esalt_len;
u32 esalt_buf0[16];
+ u32 esalt_buf1[16];
esalt_buf0[ 0] = esalt_bufs[salt_pos].esalt_buf[ 0];
esalt_buf0[ 1] = esalt_bufs[salt_pos].esalt_buf[ 1];
esalt_buf0[13] = esalt_bufs[salt_pos].esalt_buf[13];
esalt_buf0[14] = esalt_bufs[salt_pos].esalt_buf[14];
esalt_buf0[15] = esalt_bufs[salt_pos].esalt_buf[15];
-
- u32 esalt_buf1[16];
-
esalt_buf1[ 0] = esalt_bufs[salt_pos].esalt_buf[16];
esalt_buf1[ 1] = esalt_bufs[salt_pos].esalt_buf[17];
esalt_buf1[ 2] = esalt_bufs[salt_pos].esalt_buf[18];
u32 w0l = w0[0];
- for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 w0r = bfs_buf[il_pos].i;
+ const u32x w0r = ix_create_bft (bfs_buf, il_pos);
- w0[0] = w0l | w0r;
+ const u32x w0lr = w0l | w0r;
/*
* HA1 = md5 ($salt . $pass)
// append the pass to the salt
- u32 block0[16];
+ u32x block0[16];
+ u32x block1[16];
block0[ 0] = salt_buf0[ 0];
block0[ 1] = salt_buf0[ 1];
block0[13] = salt_buf0[13];
block0[14] = salt_buf0[14];
block0[15] = salt_buf0[15];
-
- u32 block1[16];
-
block1[ 0] = salt_buf1[ 0];
block1[ 1] = salt_buf1[ 1];
block1[ 2] = salt_buf1[ 2];
block1[14] = salt_buf1[14];
block1[15] = salt_buf1[15];
- memcat32 (block0, block1, salt_len, w0, w1, w2, w3, pw_len);
-
- u32 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[1] = w0[1];
+ w0_t[2] = w0[2];
+ w0_t[3] = w0[3];
+ w1_t[0] = w1[0];
+ w1_t[1] = w1[1];
+ w1_t[2] = w1[2];
+ w1_t[3] = w1[3];
+ w2_t[0] = w2[0];
+ w2_t[1] = w2[1];
+ w2_t[2] = w2[2];
+ w2_t[3] = w2[3];
+ w3_t[0] = w3[0];
+ w3_t[1] = w3[1];
+ w3_t[2] = w3[2];
+ w3_t[3] = w3[3];
+
+ memcat32 (block0, block1, salt_len, w0_t, w1_t, w2_t, w3_t, pw_len);
w0_t[0] = block0[ 0];
w0_t[1] = block0[ 1];
w0_t[2] = block0[ 2];
w0_t[3] = block0[ 3];
-
- u32 w1_t[4];
-
w1_t[0] = block0[ 4];
w1_t[1] = block0[ 5];
w1_t[2] = block0[ 6];
w1_t[3] = block0[ 7];
-
- u32 w2_t[4];
-
w2_t[0] = block0[ 8];
w2_t[1] = block0[ 9];
w2_t[2] = block0[10];
w2_t[3] = block0[11];
-
- u32 w3_t[4];
-
w3_t[0] = block0[12];
w3_t[1] = block0[13];
w3_t[2] = block0[14];
// md5
- u32 tmp2;
-
- u32 a = MD5M_A;
- u32 b = MD5M_B;
- 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_t[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += MD5M_C;
d += MD5M_D;
- u32 r_a = a;
- u32 r_b = b;
- u32 r_c = c;
- u32 r_d = d;
+ u32x r_a = a;
+ u32x r_b = b;
+ u32x r_c = c;
+ u32x r_d = d;
w0_t[0] = block1[ 0];
w0_t[1] = block1[ 1];
w0_t[2] = block1[ 2];
w0_t[3] = block1[ 3];
-
w1_t[0] = block1[ 4];
w1_t[1] = block1[ 5];
w1_t[2] = block1[ 6];
w1_t[3] = block1[ 7];
-
w2_t[0] = block1[ 8];
w2_t[1] = block1[ 9];
w2_t[2] = block1[10];
w2_t[3] = block1[11];
-
w3_t[0] = block1[12];
w3_t[1] = block1[13];
w3_t[2] = pw_salt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
-
w2_t[0] = esalt_buf0[0];
w2_t[1] = esalt_buf0[1];
w2_t[2] = esalt_buf0[2];
w2_t[3] = esalt_buf0[3];
-
w3_t[0] = esalt_buf0[4];
w3_t[1] = esalt_buf0[5];
w3_t[2] = esalt_buf0[6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
w0_t[1] = esalt_buf0[ 9];
w0_t[2] = esalt_buf0[10];
w0_t[3] = esalt_buf0[11];
-
w1_t[0] = esalt_buf0[12];
w1_t[1] = esalt_buf0[13];
w1_t[2] = esalt_buf0[14];
w1_t[3] = esalt_buf0[15];
-
w2_t[0] = esalt_buf1[ 0];
w2_t[1] = esalt_buf1[ 1];
w2_t[2] = esalt_buf1[ 2];
w2_t[3] = esalt_buf1[ 3];
-
w3_t[0] = esalt_buf1[ 4];
w3_t[1] = esalt_buf1[ 5];
w3_t[2] = digest_esalt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += r_c;
d += r_d;
- const u32 r0 = a;
- const u32 r1 = d;
- const u32 r2 = c;
- const u32 r3 = b;
-
- #include COMPARE_S
+ COMPARE_S_SIMD (a, d, c, b);
}
}
-static void m11400s_1_1 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 l_bin2asc[256])
+void m11400s_1_1 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 sip_t *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, __local u32 *l_bin2asc)
{
/**
* modifier
const u32 pw_salt_len = salt_len + pw_len;
u32 salt_buf0[16];
+ u32 salt_buf1[16];
salt_buf0[ 0] = esalt_bufs[salt_pos].salt_buf[ 0];
salt_buf0[ 1] = esalt_bufs[salt_pos].salt_buf[ 1];
salt_buf0[13] = esalt_bufs[salt_pos].salt_buf[13];
salt_buf0[14] = esalt_bufs[salt_pos].salt_buf[14];
salt_buf0[15] = esalt_bufs[salt_pos].salt_buf[15];
-
- u32 salt_buf1[16];
-
salt_buf1[ 0] = esalt_bufs[salt_pos].salt_buf[16];
salt_buf1[ 1] = esalt_bufs[salt_pos].salt_buf[17];
salt_buf1[ 2] = esalt_bufs[salt_pos].salt_buf[18];
const u32 esalt_len = esalt_bufs[salt_pos].esalt_len;
u32 esalt_buf0[16];
+ u32 esalt_buf1[16];
+ u32 esalt_buf2[16];
esalt_buf0[ 0] = esalt_bufs[salt_pos].esalt_buf[ 0];
esalt_buf0[ 1] = esalt_bufs[salt_pos].esalt_buf[ 1];
esalt_buf0[13] = esalt_bufs[salt_pos].esalt_buf[13];
esalt_buf0[14] = esalt_bufs[salt_pos].esalt_buf[14];
esalt_buf0[15] = esalt_bufs[salt_pos].esalt_buf[15];
-
- u32 esalt_buf1[16];
-
esalt_buf1[ 0] = esalt_bufs[salt_pos].esalt_buf[16];
esalt_buf1[ 1] = esalt_bufs[salt_pos].esalt_buf[17];
esalt_buf1[ 2] = esalt_bufs[salt_pos].esalt_buf[18];
esalt_buf1[13] = esalt_bufs[salt_pos].esalt_buf[29];
esalt_buf1[14] = esalt_bufs[salt_pos].esalt_buf[30];
esalt_buf1[15] = esalt_bufs[salt_pos].esalt_buf[31];
-
- u32 esalt_buf2[16];
-
esalt_buf2[ 0] = esalt_bufs[salt_pos].esalt_buf[32];
esalt_buf2[ 1] = esalt_bufs[salt_pos].esalt_buf[33];
esalt_buf2[ 2] = esalt_bufs[salt_pos].esalt_buf[34];
u32 w0l = w0[0];
- for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
+ for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
{
- const u32 w0r = bfs_buf[il_pos].i;
+ const u32x w0r = ix_create_bft (bfs_buf, il_pos);
- w0[0] = w0l | w0r;
+ const u32x w0lr = w0l | w0r;
/*
* HA1 = md5 ($salt . $pass)
// append the pass to the salt
- u32 block0[16];
+ u32x block0[16];
+ u32x block1[16];
block0[ 0] = salt_buf0[ 0];
block0[ 1] = salt_buf0[ 1];
block0[13] = salt_buf0[13];
block0[14] = salt_buf0[14];
block0[15] = salt_buf0[15];
-
- u32 block1[16];
-
block1[ 0] = salt_buf1[ 0];
block1[ 1] = salt_buf1[ 1];
block1[ 2] = salt_buf1[ 2];
block1[14] = salt_buf1[14];
block1[15] = salt_buf1[15];
- memcat32 (block0, block1, salt_len, w0, w1, w2, w3, pw_len);
-
- u32 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[1] = w0[1];
+ w0_t[2] = w0[2];
+ w0_t[3] = w0[3];
+ w1_t[0] = w1[0];
+ w1_t[1] = w1[1];
+ w1_t[2] = w1[2];
+ w1_t[3] = w1[3];
+ w2_t[0] = w2[0];
+ w2_t[1] = w2[1];
+ w2_t[2] = w2[2];
+ w2_t[3] = w2[3];
+ w3_t[0] = w3[0];
+ w3_t[1] = w3[1];
+ w3_t[2] = w3[2];
+ w3_t[3] = w3[3];
+
+ memcat32 (block0, block1, salt_len, w0_t, w1_t, w2_t, w3_t, pw_len);
w0_t[0] = block0[ 0];
w0_t[1] = block0[ 1];
w0_t[2] = block0[ 2];
w0_t[3] = block0[ 3];
-
- u32 w1_t[4];
-
w1_t[0] = block0[ 4];
w1_t[1] = block0[ 5];
w1_t[2] = block0[ 6];
w1_t[3] = block0[ 7];
-
- u32 w2_t[4];
-
w2_t[0] = block0[ 8];
w2_t[1] = block0[ 9];
w2_t[2] = block0[10];
w2_t[3] = block0[11];
-
- u32 w3_t[4];
-
w3_t[0] = block0[12];
w3_t[1] = block0[13];
w3_t[2] = block0[14];
// md5
- u32 tmp2;
-
- u32 a = MD5M_A;
- u32 b = MD5M_B;
- 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_t[0], MD5C00, MD5S00);
MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += MD5M_C;
d += MD5M_D;
- u32 r_a = a;
- u32 r_b = b;
- u32 r_c = c;
- u32 r_d = d;
+ u32x r_a = a;
+ u32x r_b = b;
+ u32x r_c = c;
+ u32x r_d = d;
w0_t[0] = block1[ 0];
w0_t[1] = block1[ 1];
w0_t[2] = block1[ 2];
w0_t[3] = block1[ 3];
-
w1_t[0] = block1[ 4];
w1_t[1] = block1[ 5];
w1_t[2] = block1[ 6];
w1_t[3] = block1[ 7];
-
w2_t[0] = block1[ 8];
w2_t[1] = block1[ 9];
w2_t[2] = block1[10];
w2_t[3] = block1[11];
-
w3_t[0] = block1[12];
w3_t[1] = block1[13];
w3_t[2] = pw_salt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
| uint_to_hex_lower8 ((d >> 8) & 255) << 16;
w1_t[3] = uint_to_hex_lower8 ((d >> 16) & 255) << 0
| uint_to_hex_lower8 ((d >> 24) & 255) << 16;
-
w2_t[0] = esalt_buf0[0];
w2_t[1] = esalt_buf0[1];
w2_t[2] = esalt_buf0[2];
w2_t[3] = esalt_buf0[3];
-
w3_t[0] = esalt_buf0[4];
w3_t[1] = esalt_buf0[5];
w3_t[2] = esalt_buf0[6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
w0_t[1] = esalt_buf0[ 9];
w0_t[2] = esalt_buf0[10];
w0_t[3] = esalt_buf0[11];
-
w1_t[0] = esalt_buf0[12];
w1_t[1] = esalt_buf0[13];
w1_t[2] = esalt_buf0[14];
w1_t[3] = esalt_buf0[15];
-
w2_t[0] = esalt_buf1[ 0];
w2_t[1] = esalt_buf1[ 1];
w2_t[2] = esalt_buf1[ 2];
w2_t[3] = esalt_buf1[ 3];
-
w3_t[0] = esalt_buf1[ 4];
w3_t[1] = esalt_buf1[ 5];
w3_t[2] = esalt_buf1[ 6];
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
w0_t[1] = esalt_buf1[ 9];
w0_t[2] = esalt_buf1[10];
w0_t[3] = esalt_buf1[11];
-
w1_t[0] = esalt_buf1[12];
w1_t[1] = esalt_buf1[13];
w1_t[2] = esalt_buf1[14];
w1_t[3] = esalt_buf1[15];
-
w2_t[0] = esalt_buf2[ 0];
w2_t[1] = esalt_buf2[ 1];
w2_t[2] = esalt_buf2[ 2];
w2_t[3] = esalt_buf2[ 3];
-
w3_t[0] = esalt_buf2[ 4];
w3_t[1] = esalt_buf2[ 5];
w3_t[2] = digest_esalt_len * 8;
MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
- MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
- MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
- MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
- MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
- MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
+ MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
+ MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
+ MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
+ MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
c += r_c;
d += r_d;
- const u32 r0 = a;
- const u32 r1 = d;
- const u32 r2 = c;
- const u32 r3 = b;
-
- #include COMPARE_S
+ COMPARE_S_SIMD (a, d, c, b);
}
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m11400_m04 (__global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m11400_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 sip_t *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)
{
/**
- * base
+ * modifier
*/
const u32 gid = get_global_id (0);
+ const u32 lid = get_local_id (0);
+ const u32 lsz = get_local_size (0);
/**
- * modifier
+ * bin2asc table
*/
- const u32 lid = get_local_id (0);
+ __local u32 l_bin2asc[256];
+
+ for (u32 i = lid; i < 256; i += lsz)
+ {
+ const u32 i0 = (i >> 0) & 15;
+ const u32 i1 = (i >> 4) & 15;
+
+ l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 8
+ | ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 0;
+ }
+ barrier (CLK_LOCAL_MEM_FENCE);
+
+ if (gid >= gid_max) return;
+
+ /**
+ * modifier
+ */
u32 w0[4];
const u32 pw_len = pws[gid].pw_len;
- /**
- * bin2asc table
- */
-
- __local u32 l_bin2asc[256];
-
- const u32 lid4 = lid * 4;
-
- const u32 lid40 = lid4 + 0;
- const u32 lid41 = lid4 + 1;
- const u32 lid42 = lid4 + 2;
- const u32 lid43 = lid4 + 3;
-
- const u32 v400 = (lid40 >> 0) & 15;
- const u32 v401 = (lid40 >> 4) & 15;
- const u32 v410 = (lid41 >> 0) & 15;
- const u32 v411 = (lid41 >> 4) & 15;
- const u32 v420 = (lid42 >> 0) & 15;
- const u32 v421 = (lid42 >> 4) & 15;
- const u32 v430 = (lid43 >> 0) & 15;
- const u32 v431 = (lid43 >> 4) & 15;
-
- l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
- | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
- l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
- | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
- l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
- | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
- l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
- | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
-
- barrier (CLK_LOCAL_MEM_FENCE);
-
- if (gid >= gid_max) return;
-
/**
* main
*/
switch (sw_1 | sw_2)
{
case 0:
- m11400m_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 1:
- m11400m_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 2:
- m11400m_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 3:
- m11400m_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
}
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m11400_m08 (__global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m11400_m08 (__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 sip_t *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)
{
/**
- * base
+ * modifier
*/
const u32 gid = get_global_id (0);
+ const u32 lid = get_local_id (0);
+ const u32 lsz = get_local_size (0);
/**
- * modifier
+ * bin2asc table
*/
- const u32 lid = get_local_id (0);
+ __local u32 l_bin2asc[256];
+
+ for (u32 i = lid; i < 256; i += lsz)
+ {
+ const u32 i0 = (i >> 0) & 15;
+ const u32 i1 = (i >> 4) & 15;
+
+ l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 8
+ | ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 0;
+ }
+
+ barrier (CLK_LOCAL_MEM_FENCE);
+
+ if (gid >= gid_max) return;
+
+ /**
+ * modifier
+ */
u32 w0[4];
const u32 pw_len = pws[gid].pw_len;
- /**
- * bin2asc table
- */
-
- __local u32 l_bin2asc[256];
-
- const u32 lid4 = lid * 4;
-
- const u32 lid40 = lid4 + 0;
- const u32 lid41 = lid4 + 1;
- const u32 lid42 = lid4 + 2;
- const u32 lid43 = lid4 + 3;
-
- const u32 v400 = (lid40 >> 0) & 15;
- const u32 v401 = (lid40 >> 4) & 15;
- const u32 v410 = (lid41 >> 0) & 15;
- const u32 v411 = (lid41 >> 4) & 15;
- const u32 v420 = (lid42 >> 0) & 15;
- const u32 v421 = (lid42 >> 4) & 15;
- const u32 v430 = (lid43 >> 0) & 15;
- const u32 v431 = (lid43 >> 4) & 15;
-
- l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
- | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
- l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
- | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
- l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
- | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
- l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
- | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
-
- barrier (CLK_LOCAL_MEM_FENCE);
-
- if (gid >= gid_max) return;
-
/**
* main
*/
switch (sw_1 | sw_2)
{
case 0:
- m11400m_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 1:
- m11400m_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 2:
- m11400m_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 3:
- m11400m_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
}
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m11400_m16 (__global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m11400_m16 (__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 sip_t *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)
{
/**
- * base
+ * modifier
*/
const u32 gid = get_global_id (0);
+ const u32 lid = get_local_id (0);
+ const u32 lsz = get_local_size (0);
/**
- * modifier
+ * bin2asc table
*/
- const u32 lid = get_local_id (0);
+ __local u32 l_bin2asc[256];
+
+ for (u32 i = lid; i < 256; i += lsz)
+ {
+ const u32 i0 = (i >> 0) & 15;
+ const u32 i1 = (i >> 4) & 15;
+
+ l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 8
+ | ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 0;
+ }
+
+ barrier (CLK_LOCAL_MEM_FENCE);
+
+ if (gid >= gid_max) return;
+
+ /**
+ * modifier
+ */
u32 w0[4];
const u32 pw_len = pws[gid].pw_len;
- /**
- * bin2asc table
- */
-
- __local u32 l_bin2asc[256];
-
- const u32 lid4 = lid * 4;
-
- const u32 lid40 = lid4 + 0;
- const u32 lid41 = lid4 + 1;
- const u32 lid42 = lid4 + 2;
- const u32 lid43 = lid4 + 3;
-
- const u32 v400 = (lid40 >> 0) & 15;
- const u32 v401 = (lid40 >> 4) & 15;
- const u32 v410 = (lid41 >> 0) & 15;
- const u32 v411 = (lid41 >> 4) & 15;
- const u32 v420 = (lid42 >> 0) & 15;
- const u32 v421 = (lid42 >> 4) & 15;
- const u32 v430 = (lid43 >> 0) & 15;
- const u32 v431 = (lid43 >> 4) & 15;
-
- l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
- | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
- l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
- | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
- l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
- | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
- l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
- | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
-
- barrier (CLK_LOCAL_MEM_FENCE);
-
- if (gid >= gid_max) return;
-
/**
* main
*/
switch (sw_1 | sw_2)
{
case 0:
- m11400m_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 1:
- m11400m_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 2:
- m11400m_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 3:
- m11400m_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400m_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
}
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m11400_s04 (__global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m11400_s04 (__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 sip_t *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)
{
/**
- * base
+ * modifier
*/
const u32 gid = get_global_id (0);
+ const u32 lid = get_local_id (0);
+ const u32 lsz = get_local_size (0);
/**
- * modifier
+ * bin2asc table
*/
- const u32 lid = get_local_id (0);
+ __local u32 l_bin2asc[256];
+
+ for (u32 i = lid; i < 256; i += lsz)
+ {
+ const u32 i0 = (i >> 0) & 15;
+ const u32 i1 = (i >> 4) & 15;
+
+ l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 8
+ | ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 0;
+ }
+
+ barrier (CLK_LOCAL_MEM_FENCE);
+
+ if (gid >= gid_max) return;
+
+ /**
+ * modifier
+ */
u32 w0[4];
const u32 pw_len = pws[gid].pw_len;
- /**
- * bin2asc table
- */
-
- __local u32 l_bin2asc[256];
-
- const u32 lid4 = lid * 4;
-
- const u32 lid40 = lid4 + 0;
- const u32 lid41 = lid4 + 1;
- const u32 lid42 = lid4 + 2;
- const u32 lid43 = lid4 + 3;
-
- const u32 v400 = (lid40 >> 0) & 15;
- const u32 v401 = (lid40 >> 4) & 15;
- const u32 v410 = (lid41 >> 0) & 15;
- const u32 v411 = (lid41 >> 4) & 15;
- const u32 v420 = (lid42 >> 0) & 15;
- const u32 v421 = (lid42 >> 4) & 15;
- const u32 v430 = (lid43 >> 0) & 15;
- const u32 v431 = (lid43 >> 4) & 15;
-
- l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
- | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
- l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
- | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
- l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
- | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
- l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
- | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
-
- barrier (CLK_LOCAL_MEM_FENCE);
-
- if (gid >= gid_max) return;
-
/**
* main
*/
switch (sw_1 | sw_2)
{
case 0:
- m11400s_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 1:
- m11400s_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 2:
- m11400s_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 3:
- m11400s_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
}
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m11400_s08 (__global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m11400_s08 (__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 sip_t *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)
{
/**
- * base
+ * modifier
*/
const u32 gid = get_global_id (0);
+ const u32 lid = get_local_id (0);
+ const u32 lsz = get_local_size (0);
/**
- * modifier
+ * bin2asc table
*/
- const u32 lid = get_local_id (0);
+ __local u32 l_bin2asc[256];
+
+ for (u32 i = lid; i < 256; i += lsz)
+ {
+ const u32 i0 = (i >> 0) & 15;
+ const u32 i1 = (i >> 4) & 15;
+
+ l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 8
+ | ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 0;
+ }
+
+ barrier (CLK_LOCAL_MEM_FENCE);
+
+ if (gid >= gid_max) return;
+
+ /**
+ * modifier
+ */
u32 w0[4];
const u32 pw_len = pws[gid].pw_len;
- /**
- * bin2asc table
- */
-
- __local u32 l_bin2asc[256];
-
- const u32 lid4 = lid * 4;
-
- const u32 lid40 = lid4 + 0;
- const u32 lid41 = lid4 + 1;
- const u32 lid42 = lid4 + 2;
- const u32 lid43 = lid4 + 3;
-
- const u32 v400 = (lid40 >> 0) & 15;
- const u32 v401 = (lid40 >> 4) & 15;
- const u32 v410 = (lid41 >> 0) & 15;
- const u32 v411 = (lid41 >> 4) & 15;
- const u32 v420 = (lid42 >> 0) & 15;
- const u32 v421 = (lid42 >> 4) & 15;
- const u32 v430 = (lid43 >> 0) & 15;
- const u32 v431 = (lid43 >> 4) & 15;
-
- l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
- | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
- l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
- | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
- l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
- | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
- l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
- | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
-
- barrier (CLK_LOCAL_MEM_FENCE);
-
- if (gid >= gid_max) return;
-
/**
* main
*/
switch (sw_1 | sw_2)
{
case 0:
- m11400s_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 1:
- m11400s_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 2:
- m11400s_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 3:
- m11400s_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
}
}
-__kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m11400_s16 (__global pw_t *pws, __global gpu_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 sip_t *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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
+__kernel void m11400_s16 (__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 sip_t *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)
{
/**
- * base
+ * modifier
*/
const u32 gid = get_global_id (0);
+ const u32 lid = get_local_id (0);
+ const u32 lsz = get_local_size (0);
/**
- * modifier
+ * bin2asc table
*/
- const u32 lid = get_local_id (0);
+ __local u32 l_bin2asc[256];
+
+ for (u32 i = lid; i < 256; i += lsz)
+ {
+ const u32 i0 = (i >> 0) & 15;
+ const u32 i1 = (i >> 4) & 15;
+
+ l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'a' - 10 + i0) << 8
+ | ((i1 < 10) ? '0' + i1 : 'a' - 10 + i1) << 0;
+ }
+
+ barrier (CLK_LOCAL_MEM_FENCE);
+
+ if (gid >= gid_max) return;
+
+ /**
+ * modifier
+ */
u32 w0[4];
const u32 pw_len = pws[gid].pw_len;
- /**
- * bin2asc table
- */
-
- __local u32 l_bin2asc[256];
-
- const u32 lid4 = lid * 4;
-
- const u32 lid40 = lid4 + 0;
- const u32 lid41 = lid4 + 1;
- const u32 lid42 = lid4 + 2;
- const u32 lid43 = lid4 + 3;
-
- const u32 v400 = (lid40 >> 0) & 15;
- const u32 v401 = (lid40 >> 4) & 15;
- const u32 v410 = (lid41 >> 0) & 15;
- const u32 v411 = (lid41 >> 4) & 15;
- const u32 v420 = (lid42 >> 0) & 15;
- const u32 v421 = (lid42 >> 4) & 15;
- const u32 v430 = (lid43 >> 0) & 15;
- const u32 v431 = (lid43 >> 4) & 15;
-
- l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
- | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
- l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
- | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
- l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
- | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
- l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
- | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
-
- barrier (CLK_LOCAL_MEM_FENCE);
-
- if (gid >= gid_max) return;
-
/**
* main
*/
switch (sw_1 | sw_2)
{
case 0:
- m11400s_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_0_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 1:
- m11400s_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_0_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 2:
- m11400s_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_1_0 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
case 3:
- m11400s_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, l_bin2asc);
+ m11400s_1_1 (w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset, l_bin2asc);
break;
}
}
projects / hashcat.git / blobdiff