2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "inc_vendor.cl"
11 #include "inc_hash_constants.h"
12 #include "inc_hash_functions.cl"
13 #include "inc_types.cl"
14 #include "inc_common.cl"
17 #include "inc_simd.cl"
19 void md4_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
43 MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
44 MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
45 MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
46 MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
47 MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
48 MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
49 MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
50 MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
51 MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
52 MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
53 MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
54 MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
55 MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
56 MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
57 MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
58 MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
60 MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
61 MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
62 MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
63 MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
64 MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
65 MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
66 MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
67 MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
68 MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
69 MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
70 MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
71 MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
72 MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
73 MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
74 MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
75 MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
77 MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
78 MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
79 MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
80 MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
81 MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
82 MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
83 MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
84 MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
85 MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
86 MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
87 MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
88 MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
89 MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
90 MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
91 MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
92 MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
100 void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
124 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
125 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
126 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
127 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
128 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
129 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
130 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
131 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
132 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
133 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
134 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
135 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
136 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
137 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
138 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
139 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
141 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
142 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
143 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
144 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
145 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
146 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
147 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
148 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
149 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
150 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
151 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
152 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
153 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
154 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
155 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
156 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
158 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
159 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
160 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
161 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
162 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
163 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
164 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
165 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
166 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
167 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
168 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
169 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
170 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
171 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
172 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
173 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
175 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
176 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
177 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
178 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
179 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
180 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
181 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
182 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
183 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
184 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
185 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
186 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
187 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
188 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
189 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
190 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
198 void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
200 w0[0] = w0[0] ^ 0x36363636;
201 w0[1] = w0[1] ^ 0x36363636;
202 w0[2] = w0[2] ^ 0x36363636;
203 w0[3] = w0[3] ^ 0x36363636;
204 w1[0] = w1[0] ^ 0x36363636;
205 w1[1] = w1[1] ^ 0x36363636;
206 w1[2] = w1[2] ^ 0x36363636;
207 w1[3] = w1[3] ^ 0x36363636;
208 w2[0] = w2[0] ^ 0x36363636;
209 w2[1] = w2[1] ^ 0x36363636;
210 w2[2] = w2[2] ^ 0x36363636;
211 w2[3] = w2[3] ^ 0x36363636;
212 w3[0] = w3[0] ^ 0x36363636;
213 w3[1] = w3[1] ^ 0x36363636;
214 w3[2] = w3[2] ^ 0x36363636;
215 w3[3] = w3[3] ^ 0x36363636;
222 md5_transform (w0, w1, w2, w3, ipad);
224 w0[0] = w0[0] ^ 0x6a6a6a6a;
225 w0[1] = w0[1] ^ 0x6a6a6a6a;
226 w0[2] = w0[2] ^ 0x6a6a6a6a;
227 w0[3] = w0[3] ^ 0x6a6a6a6a;
228 w1[0] = w1[0] ^ 0x6a6a6a6a;
229 w1[1] = w1[1] ^ 0x6a6a6a6a;
230 w1[2] = w1[2] ^ 0x6a6a6a6a;
231 w1[3] = w1[3] ^ 0x6a6a6a6a;
232 w2[0] = w2[0] ^ 0x6a6a6a6a;
233 w2[1] = w2[1] ^ 0x6a6a6a6a;
234 w2[2] = w2[2] ^ 0x6a6a6a6a;
235 w2[3] = w2[3] ^ 0x6a6a6a6a;
236 w3[0] = w3[0] ^ 0x6a6a6a6a;
237 w3[1] = w3[1] ^ 0x6a6a6a6a;
238 w3[2] = w3[2] ^ 0x6a6a6a6a;
239 w3[3] = w3[3] ^ 0x6a6a6a6a;
246 md5_transform (w0, w1, w2, w3, opad);
249 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])
256 md5_transform (w0, w1, w2, w3, digest);
272 w3[2] = (64 + 16) * 8;
280 md5_transform (w0, w1, w2, w3, digest);
283 __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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
289 const u32 gid = get_global_id (0);
290 const u32 lid = get_local_id (0);
291 const u32 lsz = get_local_size (0);
297 __local u32 s_userdomain_buf[64];
299 for (u32 i = lid; i < 64; i += lsz)
301 s_userdomain_buf[i] = netntlm_bufs[salt_pos].userdomain_buf[i];
304 __local u32 s_chall_buf[256];
306 for (u32 i = lid; i < 256; i += lsz)
308 s_chall_buf[i] = netntlm_bufs[salt_pos].chall_buf[i];
311 barrier (CLK_LOCAL_MEM_FENCE);
313 if (gid >= gid_max) return;
315 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
316 + netntlm_bufs[salt_pos].domain_len;
318 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
319 + netntlm_bufs[salt_pos].clichall_len;
328 pw_buf0[0] = pws[gid].i[0];
329 pw_buf0[1] = pws[gid].i[1];
330 pw_buf0[2] = pws[gid].i[2];
331 pw_buf0[3] = pws[gid].i[3];
332 pw_buf1[0] = pws[gid].i[4];
333 pw_buf1[1] = pws[gid].i[5];
334 pw_buf1[2] = pws[gid].i[6];
335 pw_buf1[3] = pws[gid].i[7];
337 const u32 pw_len = pws[gid].pw_len;
343 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
350 const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
352 append_0x80_2x4_VV (w0, w1, out_len);
359 make_unicode (w0, w0_t, w1_t);
360 make_unicode (w1, w2_t, w3_t);
362 w3_t[2] = out_len * 8 * 2;
372 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
399 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
404 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
406 w0_t[0] = s_userdomain_buf[off + 0];
407 w0_t[1] = s_userdomain_buf[off + 1];
408 w0_t[2] = s_userdomain_buf[off + 2];
409 w0_t[3] = s_userdomain_buf[off + 3];
410 w1_t[0] = s_userdomain_buf[off + 4];
411 w1_t[1] = s_userdomain_buf[off + 5];
412 w1_t[2] = s_userdomain_buf[off + 6];
413 w1_t[3] = s_userdomain_buf[off + 7];
414 w2_t[0] = s_userdomain_buf[off + 8];
415 w2_t[1] = s_userdomain_buf[off + 9];
416 w2_t[2] = s_userdomain_buf[off + 10];
417 w2_t[3] = s_userdomain_buf[off + 11];
418 w3_t[0] = s_userdomain_buf[off + 12];
419 w3_t[1] = s_userdomain_buf[off + 13];
420 w3_t[2] = s_userdomain_buf[off + 14];
421 w3_t[3] = s_userdomain_buf[off + 15];
423 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
426 w0_t[0] = s_userdomain_buf[off + 0];
427 w0_t[1] = s_userdomain_buf[off + 1];
428 w0_t[2] = s_userdomain_buf[off + 2];
429 w0_t[3] = s_userdomain_buf[off + 3];
430 w1_t[0] = s_userdomain_buf[off + 4];
431 w1_t[1] = s_userdomain_buf[off + 5];
432 w1_t[2] = s_userdomain_buf[off + 6];
433 w1_t[3] = s_userdomain_buf[off + 7];
434 w2_t[0] = s_userdomain_buf[off + 8];
435 w2_t[1] = s_userdomain_buf[off + 9];
436 w2_t[2] = s_userdomain_buf[off + 10];
437 w2_t[3] = s_userdomain_buf[off + 11];
438 w3_t[0] = s_userdomain_buf[off + 12];
439 w3_t[1] = s_userdomain_buf[off + 13];
440 w3_t[2] = (64 + userdomain_len) * 8;
443 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
467 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
469 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
471 w0_t[0] = s_chall_buf[off + 0];
472 w0_t[1] = s_chall_buf[off + 1];
473 w0_t[2] = s_chall_buf[off + 2];
474 w0_t[3] = s_chall_buf[off + 3];
475 w1_t[0] = s_chall_buf[off + 4];
476 w1_t[1] = s_chall_buf[off + 5];
477 w1_t[2] = s_chall_buf[off + 6];
478 w1_t[3] = s_chall_buf[off + 7];
479 w2_t[0] = s_chall_buf[off + 8];
480 w2_t[1] = s_chall_buf[off + 9];
481 w2_t[2] = s_chall_buf[off + 10];
482 w2_t[3] = s_chall_buf[off + 11];
483 w3_t[0] = s_chall_buf[off + 12];
484 w3_t[1] = s_chall_buf[off + 13];
485 w3_t[2] = s_chall_buf[off + 14];
486 w3_t[3] = s_chall_buf[off + 15];
488 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
491 w0_t[0] = s_chall_buf[off + 0];
492 w0_t[1] = s_chall_buf[off + 1];
493 w0_t[2] = s_chall_buf[off + 2];
494 w0_t[3] = s_chall_buf[off + 3];
495 w1_t[0] = s_chall_buf[off + 4];
496 w1_t[1] = s_chall_buf[off + 5];
497 w1_t[2] = s_chall_buf[off + 6];
498 w1_t[3] = s_chall_buf[off + 7];
499 w2_t[0] = s_chall_buf[off + 8];
500 w2_t[1] = s_chall_buf[off + 9];
501 w2_t[2] = s_chall_buf[off + 10];
502 w2_t[3] = s_chall_buf[off + 11];
503 w3_t[0] = s_chall_buf[off + 12];
504 w3_t[1] = s_chall_buf[off + 13];
505 w3_t[2] = (64 + chall_len) * 8;
508 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
510 COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
514 __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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
518 __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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
522 __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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
528 const u32 gid = get_global_id (0);
529 const u32 lid = get_local_id (0);
530 const u32 lsz = get_local_size (0);
536 __local u32 s_userdomain_buf[64];
538 for (u32 i = lid; i < 64; i += lsz)
540 s_userdomain_buf[i] = netntlm_bufs[salt_pos].userdomain_buf[i];
543 __local u32 s_chall_buf[256];
545 for (u32 i = lid; i < 256; i += lsz)
547 s_chall_buf[i] = netntlm_bufs[salt_pos].chall_buf[i];
550 barrier (CLK_LOCAL_MEM_FENCE);
552 if (gid >= gid_max) return;
554 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
555 + netntlm_bufs[salt_pos].domain_len;
557 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
558 + netntlm_bufs[salt_pos].clichall_len;
567 pw_buf0[0] = pws[gid].i[0];
568 pw_buf0[1] = pws[gid].i[1];
569 pw_buf0[2] = pws[gid].i[2];
570 pw_buf0[3] = pws[gid].i[3];
571 pw_buf1[0] = pws[gid].i[4];
572 pw_buf1[1] = pws[gid].i[5];
573 pw_buf1[2] = pws[gid].i[6];
574 pw_buf1[3] = pws[gid].i[7];
576 const u32 pw_len = pws[gid].pw_len;
582 const u32 search[4] =
584 digests_buf[digests_offset].digest_buf[DGST_R0],
585 digests_buf[digests_offset].digest_buf[DGST_R1],
586 digests_buf[digests_offset].digest_buf[DGST_R2],
587 digests_buf[digests_offset].digest_buf[DGST_R3]
594 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
601 const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
603 append_0x80_2x4_VV (w0, w1, out_len);
610 make_unicode (w0, w0_t, w1_t);
611 make_unicode (w1, w2_t, w3_t);
613 w3_t[2] = out_len * 8 * 2;
623 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
650 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
655 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
657 w0_t[0] = s_userdomain_buf[off + 0];
658 w0_t[1] = s_userdomain_buf[off + 1];
659 w0_t[2] = s_userdomain_buf[off + 2];
660 w0_t[3] = s_userdomain_buf[off + 3];
661 w1_t[0] = s_userdomain_buf[off + 4];
662 w1_t[1] = s_userdomain_buf[off + 5];
663 w1_t[2] = s_userdomain_buf[off + 6];
664 w1_t[3] = s_userdomain_buf[off + 7];
665 w2_t[0] = s_userdomain_buf[off + 8];
666 w2_t[1] = s_userdomain_buf[off + 9];
667 w2_t[2] = s_userdomain_buf[off + 10];
668 w2_t[3] = s_userdomain_buf[off + 11];
669 w3_t[0] = s_userdomain_buf[off + 12];
670 w3_t[1] = s_userdomain_buf[off + 13];
671 w3_t[2] = s_userdomain_buf[off + 14];
672 w3_t[3] = s_userdomain_buf[off + 15];
674 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
677 w0_t[0] = s_userdomain_buf[off + 0];
678 w0_t[1] = s_userdomain_buf[off + 1];
679 w0_t[2] = s_userdomain_buf[off + 2];
680 w0_t[3] = s_userdomain_buf[off + 3];
681 w1_t[0] = s_userdomain_buf[off + 4];
682 w1_t[1] = s_userdomain_buf[off + 5];
683 w1_t[2] = s_userdomain_buf[off + 6];
684 w1_t[3] = s_userdomain_buf[off + 7];
685 w2_t[0] = s_userdomain_buf[off + 8];
686 w2_t[1] = s_userdomain_buf[off + 9];
687 w2_t[2] = s_userdomain_buf[off + 10];
688 w2_t[3] = s_userdomain_buf[off + 11];
689 w3_t[0] = s_userdomain_buf[off + 12];
690 w3_t[1] = s_userdomain_buf[off + 13];
691 w3_t[2] = (64 + userdomain_len) * 8;
694 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
718 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
720 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
722 w0_t[0] = s_chall_buf[off + 0];
723 w0_t[1] = s_chall_buf[off + 1];
724 w0_t[2] = s_chall_buf[off + 2];
725 w0_t[3] = s_chall_buf[off + 3];
726 w1_t[0] = s_chall_buf[off + 4];
727 w1_t[1] = s_chall_buf[off + 5];
728 w1_t[2] = s_chall_buf[off + 6];
729 w1_t[3] = s_chall_buf[off + 7];
730 w2_t[0] = s_chall_buf[off + 8];
731 w2_t[1] = s_chall_buf[off + 9];
732 w2_t[2] = s_chall_buf[off + 10];
733 w2_t[3] = s_chall_buf[off + 11];
734 w3_t[0] = s_chall_buf[off + 12];
735 w3_t[1] = s_chall_buf[off + 13];
736 w3_t[2] = s_chall_buf[off + 14];
737 w3_t[3] = s_chall_buf[off + 15];
739 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
742 w0_t[0] = s_chall_buf[off + 0];
743 w0_t[1] = s_chall_buf[off + 1];
744 w0_t[2] = s_chall_buf[off + 2];
745 w0_t[3] = s_chall_buf[off + 3];
746 w1_t[0] = s_chall_buf[off + 4];
747 w1_t[1] = s_chall_buf[off + 5];
748 w1_t[2] = s_chall_buf[off + 6];
749 w1_t[3] = s_chall_buf[off + 7];
750 w2_t[0] = s_chall_buf[off + 8];
751 w2_t[1] = s_chall_buf[off + 9];
752 w2_t[2] = s_chall_buf[off + 10];
753 w2_t[3] = s_chall_buf[off + 11];
754 w3_t[0] = s_chall_buf[off + 12];
755 w3_t[1] = s_chall_buf[off + 13];
756 w3_t[2] = (64 + chall_len) * 8;
759 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
761 COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
765 __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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)
769 __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_scryptV0_buf, __global u32 *d_scryptV1_buf, __global u32 *d_scryptV2_buf, __global u32 *d_scryptV3_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)