2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
16 #include "include/kernel_functions.c"
17 #include "types_ocl.c"
20 #define COMPARE_S "check_single_comp4.c"
21 #define COMPARE_M "check_multi_comp4.c"
23 static void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 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 u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 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 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 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 (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 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 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05600_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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
293 const u32 lid = get_local_id (0);
299 const u32 gid = get_global_id (0);
303 wordl0[0] = pws[gid].i[ 0];
304 wordl0[1] = pws[gid].i[ 1];
305 wordl0[2] = pws[gid].i[ 2];
306 wordl0[3] = pws[gid].i[ 3];
310 wordl1[0] = pws[gid].i[ 4];
311 wordl1[1] = pws[gid].i[ 5];
312 wordl1[2] = pws[gid].i[ 6];
313 wordl1[3] = pws[gid].i[ 7];
329 const u32 pw_l_len = pws[gid].pw_len;
331 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
333 append_0x80_2 (wordl0, wordl1, pw_l_len);
335 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
342 __local u32 s_userdomain_buf[64];
343 __local u32 s_chall_buf[256];
345 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
346 + netntlm_bufs[salt_pos].domain_len;
348 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
349 + netntlm_bufs[salt_pos].clichall_len;
353 s_userdomain_buf[lid] = netntlm_bufs[salt_pos].userdomain_buf[lid];
356 const u32 lid4 = lid * 4;
358 s_chall_buf[lid4 + 0] = netntlm_bufs[salt_pos].chall_buf[lid4 + 0];
359 s_chall_buf[lid4 + 1] = netntlm_bufs[salt_pos].chall_buf[lid4 + 1];
360 s_chall_buf[lid4 + 2] = netntlm_bufs[salt_pos].chall_buf[lid4 + 2];
361 s_chall_buf[lid4 + 3] = netntlm_bufs[salt_pos].chall_buf[lid4 + 3];
363 barrier (CLK_LOCAL_MEM_FENCE);
365 if (gid >= gid_max) return;
371 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
373 const u32 pw_r_len = combs_buf[il_pos].pw_len;
375 const u32 pw_len = pw_l_len + pw_r_len;
379 wordr0[0] = combs_buf[il_pos].i[0];
380 wordr0[1] = combs_buf[il_pos].i[1];
381 wordr0[2] = combs_buf[il_pos].i[2];
382 wordr0[3] = combs_buf[il_pos].i[3];
386 wordr1[0] = combs_buf[il_pos].i[4];
387 wordr1[1] = combs_buf[il_pos].i[5];
388 wordr1[2] = combs_buf[il_pos].i[6];
389 wordr1[3] = combs_buf[il_pos].i[7];
405 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
407 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
412 w0[0] = wordl0[0] | wordr0[0];
413 w0[1] = wordl0[1] | wordr0[1];
414 w0[2] = wordl0[2] | wordr0[2];
415 w0[3] = wordl0[3] | wordr0[3];
419 w1[0] = wordl1[0] | wordr1[0];
420 w1[1] = wordl1[1] | wordr1[1];
421 w1[2] = wordl1[2] | wordr1[2];
422 w1[3] = wordl1[3] | wordr1[3];
443 make_unicode (w0, w0_t, w1_t);
444 make_unicode (w1, w2_t, w3_t);
446 w3_t[2] = pw_len * 8 * 2;
455 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
482 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
487 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
489 w0_t[0] = s_userdomain_buf[off + 0];
490 w0_t[1] = s_userdomain_buf[off + 1];
491 w0_t[2] = s_userdomain_buf[off + 2];
492 w0_t[3] = s_userdomain_buf[off + 3];
493 w1_t[0] = s_userdomain_buf[off + 4];
494 w1_t[1] = s_userdomain_buf[off + 5];
495 w1_t[2] = s_userdomain_buf[off + 6];
496 w1_t[3] = s_userdomain_buf[off + 7];
497 w2_t[0] = s_userdomain_buf[off + 8];
498 w2_t[1] = s_userdomain_buf[off + 9];
499 w2_t[2] = s_userdomain_buf[off + 10];
500 w2_t[3] = s_userdomain_buf[off + 11];
501 w3_t[0] = s_userdomain_buf[off + 12];
502 w3_t[1] = s_userdomain_buf[off + 13];
503 w3_t[2] = s_userdomain_buf[off + 14];
504 w3_t[3] = s_userdomain_buf[off + 15];
506 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
509 w0_t[0] = s_userdomain_buf[off + 0];
510 w0_t[1] = s_userdomain_buf[off + 1];
511 w0_t[2] = s_userdomain_buf[off + 2];
512 w0_t[3] = s_userdomain_buf[off + 3];
513 w1_t[0] = s_userdomain_buf[off + 4];
514 w1_t[1] = s_userdomain_buf[off + 5];
515 w1_t[2] = s_userdomain_buf[off + 6];
516 w1_t[3] = s_userdomain_buf[off + 7];
517 w2_t[0] = s_userdomain_buf[off + 8];
518 w2_t[1] = s_userdomain_buf[off + 9];
519 w2_t[2] = s_userdomain_buf[off + 10];
520 w2_t[3] = s_userdomain_buf[off + 11];
521 w3_t[0] = s_userdomain_buf[off + 12];
522 w3_t[1] = s_userdomain_buf[off + 13];
523 w3_t[2] = (64 + userdomain_len) * 8;
526 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
550 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
552 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
554 w0_t[0] = s_chall_buf[off + 0];
555 w0_t[1] = s_chall_buf[off + 1];
556 w0_t[2] = s_chall_buf[off + 2];
557 w0_t[3] = s_chall_buf[off + 3];
558 w1_t[0] = s_chall_buf[off + 4];
559 w1_t[1] = s_chall_buf[off + 5];
560 w1_t[2] = s_chall_buf[off + 6];
561 w1_t[3] = s_chall_buf[off + 7];
562 w2_t[0] = s_chall_buf[off + 8];
563 w2_t[1] = s_chall_buf[off + 9];
564 w2_t[2] = s_chall_buf[off + 10];
565 w2_t[3] = s_chall_buf[off + 11];
566 w3_t[0] = s_chall_buf[off + 12];
567 w3_t[1] = s_chall_buf[off + 13];
568 w3_t[2] = s_chall_buf[off + 14];
569 w3_t[3] = s_chall_buf[off + 15];
571 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
574 w0_t[0] = s_chall_buf[off + 0];
575 w0_t[1] = s_chall_buf[off + 1];
576 w0_t[2] = s_chall_buf[off + 2];
577 w0_t[3] = s_chall_buf[off + 3];
578 w1_t[0] = s_chall_buf[off + 4];
579 w1_t[1] = s_chall_buf[off + 5];
580 w1_t[2] = s_chall_buf[off + 6];
581 w1_t[3] = s_chall_buf[off + 7];
582 w2_t[0] = s_chall_buf[off + 8];
583 w2_t[1] = s_chall_buf[off + 9];
584 w2_t[2] = s_chall_buf[off + 10];
585 w2_t[3] = s_chall_buf[off + 11];
586 w3_t[0] = s_chall_buf[off + 12];
587 w3_t[1] = s_chall_buf[off + 13];
588 w3_t[2] = (64 + chall_len) * 8;
591 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
593 const u32 r0 = digest[0];
594 const u32 r1 = digest[3];
595 const u32 r2 = digest[2];
596 const u32 r3 = digest[1];
602 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05600_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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
606 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05600_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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
610 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05600_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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
616 const u32 lid = get_local_id (0);
622 const u32 gid = get_global_id (0);
626 wordl0[0] = pws[gid].i[ 0];
627 wordl0[1] = pws[gid].i[ 1];
628 wordl0[2] = pws[gid].i[ 2];
629 wordl0[3] = pws[gid].i[ 3];
633 wordl1[0] = pws[gid].i[ 4];
634 wordl1[1] = pws[gid].i[ 5];
635 wordl1[2] = pws[gid].i[ 6];
636 wordl1[3] = pws[gid].i[ 7];
652 const u32 pw_l_len = pws[gid].pw_len;
654 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
656 append_0x80_2 (wordl0, wordl1, pw_l_len);
658 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
665 __local u32 s_userdomain_buf[64];
666 __local u32 s_chall_buf[256];
668 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
669 + netntlm_bufs[salt_pos].domain_len;
671 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
672 + netntlm_bufs[salt_pos].clichall_len;
676 s_userdomain_buf[lid] = netntlm_bufs[salt_pos].userdomain_buf[lid];
679 const u32 lid4 = lid * 4;
681 s_chall_buf[lid4 + 0] = netntlm_bufs[salt_pos].chall_buf[lid4 + 0];
682 s_chall_buf[lid4 + 1] = netntlm_bufs[salt_pos].chall_buf[lid4 + 1];
683 s_chall_buf[lid4 + 2] = netntlm_bufs[salt_pos].chall_buf[lid4 + 2];
684 s_chall_buf[lid4 + 3] = netntlm_bufs[salt_pos].chall_buf[lid4 + 3];
686 barrier (CLK_LOCAL_MEM_FENCE);
688 if (gid >= gid_max) return;
694 const u32 search[4] =
696 digests_buf[digests_offset].digest_buf[DGST_R0],
697 digests_buf[digests_offset].digest_buf[DGST_R1],
698 digests_buf[digests_offset].digest_buf[DGST_R2],
699 digests_buf[digests_offset].digest_buf[DGST_R3]
706 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
708 const u32 pw_r_len = combs_buf[il_pos].pw_len;
710 const u32 pw_len = pw_l_len + pw_r_len;
714 wordr0[0] = combs_buf[il_pos].i[0];
715 wordr0[1] = combs_buf[il_pos].i[1];
716 wordr0[2] = combs_buf[il_pos].i[2];
717 wordr0[3] = combs_buf[il_pos].i[3];
721 wordr1[0] = combs_buf[il_pos].i[4];
722 wordr1[1] = combs_buf[il_pos].i[5];
723 wordr1[2] = combs_buf[il_pos].i[6];
724 wordr1[3] = combs_buf[il_pos].i[7];
740 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
742 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
747 w0[0] = wordl0[0] | wordr0[0];
748 w0[1] = wordl0[1] | wordr0[1];
749 w0[2] = wordl0[2] | wordr0[2];
750 w0[3] = wordl0[3] | wordr0[3];
754 w1[0] = wordl1[0] | wordr1[0];
755 w1[1] = wordl1[1] | wordr1[1];
756 w1[2] = wordl1[2] | wordr1[2];
757 w1[3] = wordl1[3] | wordr1[3];
778 make_unicode (w0, w0_t, w1_t);
779 make_unicode (w1, w2_t, w3_t);
781 w3_t[2] = pw_len * 8 * 2;
790 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
817 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
822 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
824 w0_t[0] = s_userdomain_buf[off + 0];
825 w0_t[1] = s_userdomain_buf[off + 1];
826 w0_t[2] = s_userdomain_buf[off + 2];
827 w0_t[3] = s_userdomain_buf[off + 3];
828 w1_t[0] = s_userdomain_buf[off + 4];
829 w1_t[1] = s_userdomain_buf[off + 5];
830 w1_t[2] = s_userdomain_buf[off + 6];
831 w1_t[3] = s_userdomain_buf[off + 7];
832 w2_t[0] = s_userdomain_buf[off + 8];
833 w2_t[1] = s_userdomain_buf[off + 9];
834 w2_t[2] = s_userdomain_buf[off + 10];
835 w2_t[3] = s_userdomain_buf[off + 11];
836 w3_t[0] = s_userdomain_buf[off + 12];
837 w3_t[1] = s_userdomain_buf[off + 13];
838 w3_t[2] = s_userdomain_buf[off + 14];
839 w3_t[3] = s_userdomain_buf[off + 15];
841 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
844 w0_t[0] = s_userdomain_buf[off + 0];
845 w0_t[1] = s_userdomain_buf[off + 1];
846 w0_t[2] = s_userdomain_buf[off + 2];
847 w0_t[3] = s_userdomain_buf[off + 3];
848 w1_t[0] = s_userdomain_buf[off + 4];
849 w1_t[1] = s_userdomain_buf[off + 5];
850 w1_t[2] = s_userdomain_buf[off + 6];
851 w1_t[3] = s_userdomain_buf[off + 7];
852 w2_t[0] = s_userdomain_buf[off + 8];
853 w2_t[1] = s_userdomain_buf[off + 9];
854 w2_t[2] = s_userdomain_buf[off + 10];
855 w2_t[3] = s_userdomain_buf[off + 11];
856 w3_t[0] = s_userdomain_buf[off + 12];
857 w3_t[1] = s_userdomain_buf[off + 13];
858 w3_t[2] = (64 + userdomain_len) * 8;
861 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
885 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
887 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
889 w0_t[0] = s_chall_buf[off + 0];
890 w0_t[1] = s_chall_buf[off + 1];
891 w0_t[2] = s_chall_buf[off + 2];
892 w0_t[3] = s_chall_buf[off + 3];
893 w1_t[0] = s_chall_buf[off + 4];
894 w1_t[1] = s_chall_buf[off + 5];
895 w1_t[2] = s_chall_buf[off + 6];
896 w1_t[3] = s_chall_buf[off + 7];
897 w2_t[0] = s_chall_buf[off + 8];
898 w2_t[1] = s_chall_buf[off + 9];
899 w2_t[2] = s_chall_buf[off + 10];
900 w2_t[3] = s_chall_buf[off + 11];
901 w3_t[0] = s_chall_buf[off + 12];
902 w3_t[1] = s_chall_buf[off + 13];
903 w3_t[2] = s_chall_buf[off + 14];
904 w3_t[3] = s_chall_buf[off + 15];
906 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
909 w0_t[0] = s_chall_buf[off + 0];
910 w0_t[1] = s_chall_buf[off + 1];
911 w0_t[2] = s_chall_buf[off + 2];
912 w0_t[3] = s_chall_buf[off + 3];
913 w1_t[0] = s_chall_buf[off + 4];
914 w1_t[1] = s_chall_buf[off + 5];
915 w1_t[2] = s_chall_buf[off + 6];
916 w1_t[3] = s_chall_buf[off + 7];
917 w2_t[0] = s_chall_buf[off + 8];
918 w2_t[1] = s_chall_buf[off + 9];
919 w2_t[2] = s_chall_buf[off + 10];
920 w2_t[3] = s_chall_buf[off + 11];
921 w3_t[0] = s_chall_buf[off + 12];
922 w3_t[1] = s_chall_buf[off + 13];
923 w3_t[2] = (64 + chall_len) * 8;
926 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
928 const u32 r0 = digest[0];
929 const u32 r1 = digest[3];
930 const u32 r2 = digest[2];
931 const u32 r3 = digest[1];
937 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05600_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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
941 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05600_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 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)