2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "include/constants.h"
11 #include "include/kernel_vendor.h"
18 #include "include/kernel_functions.c"
19 #include "OpenCL/types_ocl.c"
20 #include "OpenCL/common.c"
21 #include "OpenCL/simd.c"
23 static void md4_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
47 MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
48 MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
49 MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
50 MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
51 MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
52 MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
53 MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
54 MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
55 MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
56 MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
57 MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
58 MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
59 MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
60 MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
61 MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
62 MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
64 MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
65 MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
66 MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
67 MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
68 MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
69 MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
70 MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
71 MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
72 MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
73 MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
74 MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
75 MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
76 MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
77 MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
78 MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
79 MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
81 MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
82 MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
83 MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
84 MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
85 MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
86 MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
87 MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
88 MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
89 MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
90 MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
91 MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
92 MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
93 MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
94 MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
95 MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
96 MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
104 static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
128 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
129 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
130 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
131 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
132 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
133 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
134 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
135 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
136 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
137 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
138 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
139 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
140 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
141 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
142 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
143 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
145 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
146 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
147 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
148 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
149 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
150 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
151 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
152 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
153 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
154 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
155 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
156 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
157 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
158 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
159 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
160 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
162 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
163 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
164 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
165 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
166 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
167 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
168 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
169 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
170 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
171 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
172 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
173 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
174 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
175 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
176 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
177 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
179 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
180 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
181 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
182 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
183 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
184 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
185 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
186 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
187 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
188 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
189 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
190 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
191 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
192 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
193 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
194 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
202 static void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
204 w0[0] = w0[0] ^ 0x36363636;
205 w0[1] = w0[1] ^ 0x36363636;
206 w0[2] = w0[2] ^ 0x36363636;
207 w0[3] = w0[3] ^ 0x36363636;
208 w1[0] = w1[0] ^ 0x36363636;
209 w1[1] = w1[1] ^ 0x36363636;
210 w1[2] = w1[2] ^ 0x36363636;
211 w1[3] = w1[3] ^ 0x36363636;
212 w2[0] = w2[0] ^ 0x36363636;
213 w2[1] = w2[1] ^ 0x36363636;
214 w2[2] = w2[2] ^ 0x36363636;
215 w2[3] = w2[3] ^ 0x36363636;
216 w3[0] = w3[0] ^ 0x36363636;
217 w3[1] = w3[1] ^ 0x36363636;
218 w3[2] = w3[2] ^ 0x36363636;
219 w3[3] = w3[3] ^ 0x36363636;
226 md5_transform (w0, w1, w2, w3, ipad);
228 w0[0] = w0[0] ^ 0x6a6a6a6a;
229 w0[1] = w0[1] ^ 0x6a6a6a6a;
230 w0[2] = w0[2] ^ 0x6a6a6a6a;
231 w0[3] = w0[3] ^ 0x6a6a6a6a;
232 w1[0] = w1[0] ^ 0x6a6a6a6a;
233 w1[1] = w1[1] ^ 0x6a6a6a6a;
234 w1[2] = w1[2] ^ 0x6a6a6a6a;
235 w1[3] = w1[3] ^ 0x6a6a6a6a;
236 w2[0] = w2[0] ^ 0x6a6a6a6a;
237 w2[1] = w2[1] ^ 0x6a6a6a6a;
238 w2[2] = w2[2] ^ 0x6a6a6a6a;
239 w2[3] = w2[3] ^ 0x6a6a6a6a;
240 w3[0] = w3[0] ^ 0x6a6a6a6a;
241 w3[1] = w3[1] ^ 0x6a6a6a6a;
242 w3[2] = w3[2] ^ 0x6a6a6a6a;
243 w3[3] = w3[3] ^ 0x6a6a6a6a;
250 md5_transform (w0, w1, w2, w3, opad);
253 static void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
260 md5_transform (w0, w1, w2, w3, digest);
276 w3[2] = (64 + 16) * 8;
284 md5_transform (w0, w1, w2, w3, digest);
287 static void m05600m (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 netntlm_t *netntlm_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 s_userdomain_buf[64], __local u32 s_chall_buf[256])
293 const u32 gid = get_global_id (0);
294 const u32 lid = get_local_id (0);
300 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
301 + netntlm_bufs[salt_pos].domain_len;
303 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
304 + netntlm_bufs[salt_pos].clichall_len;
312 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
314 const u32x w0r = w0r_create_bft (bfs_buf, il_pos);
316 const u32x w0lr = w0l | w0r;
350 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
377 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
382 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
384 w0_t[0] = s_userdomain_buf[off + 0];
385 w0_t[1] = s_userdomain_buf[off + 1];
386 w0_t[2] = s_userdomain_buf[off + 2];
387 w0_t[3] = s_userdomain_buf[off + 3];
388 w1_t[0] = s_userdomain_buf[off + 4];
389 w1_t[1] = s_userdomain_buf[off + 5];
390 w1_t[2] = s_userdomain_buf[off + 6];
391 w1_t[3] = s_userdomain_buf[off + 7];
392 w2_t[0] = s_userdomain_buf[off + 8];
393 w2_t[1] = s_userdomain_buf[off + 9];
394 w2_t[2] = s_userdomain_buf[off + 10];
395 w2_t[3] = s_userdomain_buf[off + 11];
396 w3_t[0] = s_userdomain_buf[off + 12];
397 w3_t[1] = s_userdomain_buf[off + 13];
398 w3_t[2] = s_userdomain_buf[off + 14];
399 w3_t[3] = s_userdomain_buf[off + 15];
401 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
404 w0_t[0] = s_userdomain_buf[off + 0];
405 w0_t[1] = s_userdomain_buf[off + 1];
406 w0_t[2] = s_userdomain_buf[off + 2];
407 w0_t[3] = s_userdomain_buf[off + 3];
408 w1_t[0] = s_userdomain_buf[off + 4];
409 w1_t[1] = s_userdomain_buf[off + 5];
410 w1_t[2] = s_userdomain_buf[off + 6];
411 w1_t[3] = s_userdomain_buf[off + 7];
412 w2_t[0] = s_userdomain_buf[off + 8];
413 w2_t[1] = s_userdomain_buf[off + 9];
414 w2_t[2] = s_userdomain_buf[off + 10];
415 w2_t[3] = s_userdomain_buf[off + 11];
416 w3_t[0] = s_userdomain_buf[off + 12];
417 w3_t[1] = s_userdomain_buf[off + 13];
418 w3_t[2] = (64 + userdomain_len) * 8;
421 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
445 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
447 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
449 w0_t[0] = s_chall_buf[off + 0];
450 w0_t[1] = s_chall_buf[off + 1];
451 w0_t[2] = s_chall_buf[off + 2];
452 w0_t[3] = s_chall_buf[off + 3];
453 w1_t[0] = s_chall_buf[off + 4];
454 w1_t[1] = s_chall_buf[off + 5];
455 w1_t[2] = s_chall_buf[off + 6];
456 w1_t[3] = s_chall_buf[off + 7];
457 w2_t[0] = s_chall_buf[off + 8];
458 w2_t[1] = s_chall_buf[off + 9];
459 w2_t[2] = s_chall_buf[off + 10];
460 w2_t[3] = s_chall_buf[off + 11];
461 w3_t[0] = s_chall_buf[off + 12];
462 w3_t[1] = s_chall_buf[off + 13];
463 w3_t[2] = s_chall_buf[off + 14];
464 w3_t[3] = s_chall_buf[off + 15];
466 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
469 w0_t[0] = s_chall_buf[off + 0];
470 w0_t[1] = s_chall_buf[off + 1];
471 w0_t[2] = s_chall_buf[off + 2];
472 w0_t[3] = s_chall_buf[off + 3];
473 w1_t[0] = s_chall_buf[off + 4];
474 w1_t[1] = s_chall_buf[off + 5];
475 w1_t[2] = s_chall_buf[off + 6];
476 w1_t[3] = s_chall_buf[off + 7];
477 w2_t[0] = s_chall_buf[off + 8];
478 w2_t[1] = s_chall_buf[off + 9];
479 w2_t[2] = s_chall_buf[off + 10];
480 w2_t[3] = s_chall_buf[off + 11];
481 w3_t[0] = s_chall_buf[off + 12];
482 w3_t[1] = s_chall_buf[off + 13];
483 w3_t[2] = (64 + chall_len) * 8;
486 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
488 COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
492 static void m05600s (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 netntlm_t *netntlm_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 s_userdomain_buf[64], __local u32 s_chall_buf[256])
498 const u32 gid = get_global_id (0);
499 const u32 lid = get_local_id (0);
505 const u32 search[4] =
507 digests_buf[digests_offset].digest_buf[DGST_R0],
508 digests_buf[digests_offset].digest_buf[DGST_R1],
509 digests_buf[digests_offset].digest_buf[DGST_R2],
510 digests_buf[digests_offset].digest_buf[DGST_R3]
517 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
518 + netntlm_bufs[salt_pos].domain_len;
520 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
521 + netntlm_bufs[salt_pos].clichall_len;
529 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
531 const u32x w0r = w0r_create_bft (bfs_buf, il_pos);
533 const u32x w0lr = w0l | w0r;
567 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
594 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
599 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
601 w0_t[0] = s_userdomain_buf[off + 0];
602 w0_t[1] = s_userdomain_buf[off + 1];
603 w0_t[2] = s_userdomain_buf[off + 2];
604 w0_t[3] = s_userdomain_buf[off + 3];
605 w1_t[0] = s_userdomain_buf[off + 4];
606 w1_t[1] = s_userdomain_buf[off + 5];
607 w1_t[2] = s_userdomain_buf[off + 6];
608 w1_t[3] = s_userdomain_buf[off + 7];
609 w2_t[0] = s_userdomain_buf[off + 8];
610 w2_t[1] = s_userdomain_buf[off + 9];
611 w2_t[2] = s_userdomain_buf[off + 10];
612 w2_t[3] = s_userdomain_buf[off + 11];
613 w3_t[0] = s_userdomain_buf[off + 12];
614 w3_t[1] = s_userdomain_buf[off + 13];
615 w3_t[2] = s_userdomain_buf[off + 14];
616 w3_t[3] = s_userdomain_buf[off + 15];
618 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
621 w0_t[0] = s_userdomain_buf[off + 0];
622 w0_t[1] = s_userdomain_buf[off + 1];
623 w0_t[2] = s_userdomain_buf[off + 2];
624 w0_t[3] = s_userdomain_buf[off + 3];
625 w1_t[0] = s_userdomain_buf[off + 4];
626 w1_t[1] = s_userdomain_buf[off + 5];
627 w1_t[2] = s_userdomain_buf[off + 6];
628 w1_t[3] = s_userdomain_buf[off + 7];
629 w2_t[0] = s_userdomain_buf[off + 8];
630 w2_t[1] = s_userdomain_buf[off + 9];
631 w2_t[2] = s_userdomain_buf[off + 10];
632 w2_t[3] = s_userdomain_buf[off + 11];
633 w3_t[0] = s_userdomain_buf[off + 12];
634 w3_t[1] = s_userdomain_buf[off + 13];
635 w3_t[2] = (64 + userdomain_len) * 8;
638 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
662 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
664 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
666 w0_t[0] = s_chall_buf[off + 0];
667 w0_t[1] = s_chall_buf[off + 1];
668 w0_t[2] = s_chall_buf[off + 2];
669 w0_t[3] = s_chall_buf[off + 3];
670 w1_t[0] = s_chall_buf[off + 4];
671 w1_t[1] = s_chall_buf[off + 5];
672 w1_t[2] = s_chall_buf[off + 6];
673 w1_t[3] = s_chall_buf[off + 7];
674 w2_t[0] = s_chall_buf[off + 8];
675 w2_t[1] = s_chall_buf[off + 9];
676 w2_t[2] = s_chall_buf[off + 10];
677 w2_t[3] = s_chall_buf[off + 11];
678 w3_t[0] = s_chall_buf[off + 12];
679 w3_t[1] = s_chall_buf[off + 13];
680 w3_t[2] = s_chall_buf[off + 14];
681 w3_t[3] = s_chall_buf[off + 15];
683 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
686 w0_t[0] = s_chall_buf[off + 0];
687 w0_t[1] = s_chall_buf[off + 1];
688 w0_t[2] = s_chall_buf[off + 2];
689 w0_t[3] = s_chall_buf[off + 3];
690 w1_t[0] = s_chall_buf[off + 4];
691 w1_t[1] = s_chall_buf[off + 5];
692 w1_t[2] = s_chall_buf[off + 6];
693 w1_t[3] = s_chall_buf[off + 7];
694 w2_t[0] = s_chall_buf[off + 8];
695 w2_t[1] = s_chall_buf[off + 9];
696 w2_t[2] = s_chall_buf[off + 10];
697 w2_t[3] = s_chall_buf[off + 11];
698 w3_t[0] = s_chall_buf[off + 12];
699 w3_t[1] = s_chall_buf[off + 13];
700 w3_t[2] = (64 + chall_len) * 8;
703 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
705 COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
709 __kernel void m05600_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 netntlm_t *netntlm_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)
715 const u32 gid = get_global_id (0);
716 const u32 lid = get_local_id (0);
717 const u32 lsz = get_local_size (0);
723 __local u32 s_userdomain_buf[64];
725 for (u32 i = lid; i < 64; i += lsz)
727 s_userdomain_buf[i] = netntlm_bufs[salt_pos].userdomain_buf[i];
730 __local u32 s_chall_buf[256];
732 for (u32 i = lid; i < 256; i += lsz)
734 s_chall_buf[i] = netntlm_bufs[salt_pos].chall_buf[i];
737 barrier (CLK_LOCAL_MEM_FENCE);
739 if (gid >= gid_max) return;
741 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
742 + netntlm_bufs[salt_pos].domain_len;
744 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
745 + netntlm_bufs[salt_pos].clichall_len;
753 w0[0] = pws[gid].i[ 0];
754 w0[1] = pws[gid].i[ 1];
755 w0[2] = pws[gid].i[ 2];
756 w0[3] = pws[gid].i[ 3];
776 w3[2] = pws[gid].i[14];
779 const u32 pw_len = pws[gid].pw_len;
785 m05600m (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, netntlm_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, s_userdomain_buf, s_chall_buf);
788 __kernel void m05600_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 netntlm_t *netntlm_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)
794 const u32 gid = get_global_id (0);
795 const u32 lid = get_local_id (0);
796 const u32 lsz = get_local_size (0);
802 __local u32 s_userdomain_buf[64];
804 for (u32 i = lid; i < 64; i += lsz)
806 s_userdomain_buf[i] = netntlm_bufs[salt_pos].userdomain_buf[i];
809 __local u32 s_chall_buf[256];
811 for (u32 i = lid; i < 256; i += lsz)
813 s_chall_buf[i] = netntlm_bufs[salt_pos].chall_buf[i];
816 barrier (CLK_LOCAL_MEM_FENCE);
818 if (gid >= gid_max) return;
820 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
821 + netntlm_bufs[salt_pos].domain_len;
823 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
824 + netntlm_bufs[salt_pos].clichall_len;
832 w0[0] = pws[gid].i[ 0];
833 w0[1] = pws[gid].i[ 1];
834 w0[2] = pws[gid].i[ 2];
835 w0[3] = pws[gid].i[ 3];
839 w1[0] = pws[gid].i[ 4];
840 w1[1] = pws[gid].i[ 5];
841 w1[2] = pws[gid].i[ 6];
842 w1[3] = pws[gid].i[ 7];
855 w3[2] = pws[gid].i[14];
858 const u32 pw_len = pws[gid].pw_len;
864 m05600m (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, netntlm_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, s_userdomain_buf, s_chall_buf);
867 __kernel void m05600_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 netntlm_t *netntlm_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)
871 __kernel void m05600_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 netntlm_t *netntlm_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)
877 const u32 gid = get_global_id (0);
878 const u32 lid = get_local_id (0);
879 const u32 lsz = get_local_size (0);
885 __local u32 s_userdomain_buf[64];
887 for (u32 i = lid; i < 64; i += lsz)
889 s_userdomain_buf[i] = netntlm_bufs[salt_pos].userdomain_buf[i];
892 __local u32 s_chall_buf[256];
894 for (u32 i = lid; i < 256; i += lsz)
896 s_chall_buf[i] = netntlm_bufs[salt_pos].chall_buf[i];
899 barrier (CLK_LOCAL_MEM_FENCE);
901 if (gid >= gid_max) return;
903 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
904 + netntlm_bufs[salt_pos].domain_len;
906 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
907 + netntlm_bufs[salt_pos].clichall_len;
915 w0[0] = pws[gid].i[ 0];
916 w0[1] = pws[gid].i[ 1];
917 w0[2] = pws[gid].i[ 2];
918 w0[3] = pws[gid].i[ 3];
938 w3[2] = pws[gid].i[14];
941 const u32 pw_len = pws[gid].pw_len;
947 m05600s (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, netntlm_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, s_userdomain_buf, s_chall_buf);
950 __kernel void m05600_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 netntlm_t *netntlm_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)
956 const u32 gid = get_global_id (0);
957 const u32 lid = get_local_id (0);
958 const u32 lsz = get_local_size (0);
964 __local u32 s_userdomain_buf[64];
966 for (u32 i = lid; i < 64; i += lsz)
968 s_userdomain_buf[i] = netntlm_bufs[salt_pos].userdomain_buf[i];
971 __local u32 s_chall_buf[256];
973 for (u32 i = lid; i < 256; i += lsz)
975 s_chall_buf[i] = netntlm_bufs[salt_pos].chall_buf[i];
978 barrier (CLK_LOCAL_MEM_FENCE);
980 if (gid >= gid_max) return;
982 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
983 + netntlm_bufs[salt_pos].domain_len;
985 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
986 + netntlm_bufs[salt_pos].clichall_len;
994 w0[0] = pws[gid].i[ 0];
995 w0[1] = pws[gid].i[ 1];
996 w0[2] = pws[gid].i[ 2];
997 w0[3] = pws[gid].i[ 3];
1001 w1[0] = pws[gid].i[ 4];
1002 w1[1] = pws[gid].i[ 5];
1003 w1[2] = pws[gid].i[ 6];
1004 w1[3] = pws[gid].i[ 7];
1017 w3[2] = pws[gid].i[14];
1020 const u32 pw_len = pws[gid].pw_len;
1026 m05600s (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, netntlm_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, s_userdomain_buf, s_chall_buf);
1029 __kernel void m05600_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 netntlm_t *netntlm_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)