2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
24 #include "include/kernel_functions.c"
26 #include "common_nv.c"
29 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
30 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
34 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
35 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
38 __device__ static void md4_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
62 MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
63 MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
64 MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
65 MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
66 MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
67 MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
68 MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
69 MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
70 MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
71 MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
72 MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
73 MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
74 MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
75 MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
76 MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
77 MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
79 MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
80 MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
81 MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
82 MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
83 MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
84 MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
85 MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
86 MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
87 MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
88 MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
89 MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
90 MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
91 MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
92 MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
93 MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
94 MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
96 MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
97 MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
98 MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
99 MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
100 MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
101 MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
102 MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
103 MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
104 MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
105 MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
106 MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
107 MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
108 MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
109 MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
110 MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
111 MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
119 __device__ static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
143 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
144 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
145 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
146 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
147 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
148 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
149 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
150 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
151 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
152 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
153 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
154 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
155 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
156 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
157 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
158 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
160 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
161 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
162 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
163 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
164 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
165 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
166 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
167 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
168 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
169 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
170 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
171 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
172 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
173 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
174 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
175 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
177 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
178 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
179 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
180 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
181 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
182 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
183 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
184 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
185 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
186 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
187 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
188 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
189 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
190 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
191 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
192 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
194 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
195 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
196 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
197 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
198 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
199 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
200 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
201 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
202 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
203 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
204 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
205 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
206 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
207 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
208 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
209 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
217 __device__ static void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
219 w0[0] = w0[0] ^ 0x36363636;
220 w0[1] = w0[1] ^ 0x36363636;
221 w0[2] = w0[2] ^ 0x36363636;
222 w0[3] = w0[3] ^ 0x36363636;
223 w1[0] = w1[0] ^ 0x36363636;
224 w1[1] = w1[1] ^ 0x36363636;
225 w1[2] = w1[2] ^ 0x36363636;
226 w1[3] = w1[3] ^ 0x36363636;
227 w2[0] = w2[0] ^ 0x36363636;
228 w2[1] = w2[1] ^ 0x36363636;
229 w2[2] = w2[2] ^ 0x36363636;
230 w2[3] = w2[3] ^ 0x36363636;
231 w3[0] = w3[0] ^ 0x36363636;
232 w3[1] = w3[1] ^ 0x36363636;
233 w3[2] = w3[2] ^ 0x36363636;
234 w3[3] = w3[3] ^ 0x36363636;
241 md5_transform (w0, w1, w2, w3, ipad);
243 w0[0] = w0[0] ^ 0x6a6a6a6a;
244 w0[1] = w0[1] ^ 0x6a6a6a6a;
245 w0[2] = w0[2] ^ 0x6a6a6a6a;
246 w0[3] = w0[3] ^ 0x6a6a6a6a;
247 w1[0] = w1[0] ^ 0x6a6a6a6a;
248 w1[1] = w1[1] ^ 0x6a6a6a6a;
249 w1[2] = w1[2] ^ 0x6a6a6a6a;
250 w1[3] = w1[3] ^ 0x6a6a6a6a;
251 w2[0] = w2[0] ^ 0x6a6a6a6a;
252 w2[1] = w2[1] ^ 0x6a6a6a6a;
253 w2[2] = w2[2] ^ 0x6a6a6a6a;
254 w2[3] = w2[3] ^ 0x6a6a6a6a;
255 w3[0] = w3[0] ^ 0x6a6a6a6a;
256 w3[1] = w3[1] ^ 0x6a6a6a6a;
257 w3[2] = w3[2] ^ 0x6a6a6a6a;
258 w3[3] = w3[3] ^ 0x6a6a6a6a;
265 md5_transform (w0, w1, w2, w3, opad);
268 __device__ 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])
275 md5_transform (w0, w1, w2, w3, digest);
291 w3[2] = (64 + 16) * 8;
299 md5_transform (w0, w1, w2, w3, digest);
302 __device__ __constant__ comb_t c_combs[1024];
304 extern "C" __global__ void __launch_bounds__ (256, 1) m05600_m04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const netntlm_t *netntlm_bufs, u32 *d_return_buf, 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)
310 const u32 lid = threadIdx.x;
316 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
320 wordl0[0] = pws[gid].i[ 0];
321 wordl0[1] = pws[gid].i[ 1];
322 wordl0[2] = pws[gid].i[ 2];
323 wordl0[3] = pws[gid].i[ 3];
327 wordl1[0] = pws[gid].i[ 4];
328 wordl1[1] = pws[gid].i[ 5];
329 wordl1[2] = pws[gid].i[ 6];
330 wordl1[3] = pws[gid].i[ 7];
346 const u32 pw_l_len = pws[gid].pw_len;
348 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
350 append_0x80_2 (wordl0, wordl1, pw_l_len);
352 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
359 __shared__ u32 s_userdomain_buf[64];
360 __shared__ u32 s_chall_buf[256];
362 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
363 + netntlm_bufs[salt_pos].domain_len;
365 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
366 + netntlm_bufs[salt_pos].clichall_len;
370 s_userdomain_buf[lid] = netntlm_bufs[salt_pos].userdomain_buf[lid];
373 s_chall_buf[lid] = netntlm_bufs[salt_pos].chall_buf[lid];
377 if (gid >= gid_max) return;
383 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
385 const u32 pw_r_len = c_combs[il_pos].pw_len;
387 const u32 pw_len = pw_l_len + pw_r_len;
391 wordr0[0] = c_combs[il_pos].i[0];
392 wordr0[1] = c_combs[il_pos].i[1];
393 wordr0[2] = c_combs[il_pos].i[2];
394 wordr0[3] = c_combs[il_pos].i[3];
398 wordr1[0] = c_combs[il_pos].i[4];
399 wordr1[1] = c_combs[il_pos].i[5];
400 wordr1[2] = c_combs[il_pos].i[6];
401 wordr1[3] = c_combs[il_pos].i[7];
417 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
419 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
424 w0[0] = wordl0[0] | wordr0[0];
425 w0[1] = wordl0[1] | wordr0[1];
426 w0[2] = wordl0[2] | wordr0[2];
427 w0[3] = wordl0[3] | wordr0[3];
431 w1[0] = wordl1[0] | wordr1[0];
432 w1[1] = wordl1[1] | wordr1[1];
433 w1[2] = wordl1[2] | wordr1[2];
434 w1[3] = wordl1[3] | wordr1[3];
455 make_unicode (w0, w0_t, w1_t);
456 make_unicode (w1, w2_t, w3_t);
458 w3_t[2] = pw_len * 8 * 2;
467 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
494 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
499 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
501 w0_t[0] = s_userdomain_buf[off + 0];
502 w0_t[1] = s_userdomain_buf[off + 1];
503 w0_t[2] = s_userdomain_buf[off + 2];
504 w0_t[3] = s_userdomain_buf[off + 3];
505 w1_t[0] = s_userdomain_buf[off + 4];
506 w1_t[1] = s_userdomain_buf[off + 5];
507 w1_t[2] = s_userdomain_buf[off + 6];
508 w1_t[3] = s_userdomain_buf[off + 7];
509 w2_t[0] = s_userdomain_buf[off + 8];
510 w2_t[1] = s_userdomain_buf[off + 9];
511 w2_t[2] = s_userdomain_buf[off + 10];
512 w2_t[3] = s_userdomain_buf[off + 11];
513 w3_t[0] = s_userdomain_buf[off + 12];
514 w3_t[1] = s_userdomain_buf[off + 13];
515 w3_t[2] = s_userdomain_buf[off + 14];
516 w3_t[3] = s_userdomain_buf[off + 15];
518 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
521 w0_t[0] = s_userdomain_buf[off + 0];
522 w0_t[1] = s_userdomain_buf[off + 1];
523 w0_t[2] = s_userdomain_buf[off + 2];
524 w0_t[3] = s_userdomain_buf[off + 3];
525 w1_t[0] = s_userdomain_buf[off + 4];
526 w1_t[1] = s_userdomain_buf[off + 5];
527 w1_t[2] = s_userdomain_buf[off + 6];
528 w1_t[3] = s_userdomain_buf[off + 7];
529 w2_t[0] = s_userdomain_buf[off + 8];
530 w2_t[1] = s_userdomain_buf[off + 9];
531 w2_t[2] = s_userdomain_buf[off + 10];
532 w2_t[3] = s_userdomain_buf[off + 11];
533 w3_t[0] = s_userdomain_buf[off + 12];
534 w3_t[1] = s_userdomain_buf[off + 13];
535 w3_t[2] = (64 + userdomain_len) * 8;
538 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
562 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
564 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
566 w0_t[0] = s_chall_buf[off + 0];
567 w0_t[1] = s_chall_buf[off + 1];
568 w0_t[2] = s_chall_buf[off + 2];
569 w0_t[3] = s_chall_buf[off + 3];
570 w1_t[0] = s_chall_buf[off + 4];
571 w1_t[1] = s_chall_buf[off + 5];
572 w1_t[2] = s_chall_buf[off + 6];
573 w1_t[3] = s_chall_buf[off + 7];
574 w2_t[0] = s_chall_buf[off + 8];
575 w2_t[1] = s_chall_buf[off + 9];
576 w2_t[2] = s_chall_buf[off + 10];
577 w2_t[3] = s_chall_buf[off + 11];
578 w3_t[0] = s_chall_buf[off + 12];
579 w3_t[1] = s_chall_buf[off + 13];
580 w3_t[2] = s_chall_buf[off + 14];
581 w3_t[3] = s_chall_buf[off + 15];
583 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
586 w0_t[0] = s_chall_buf[off + 0];
587 w0_t[1] = s_chall_buf[off + 1];
588 w0_t[2] = s_chall_buf[off + 2];
589 w0_t[3] = s_chall_buf[off + 3];
590 w1_t[0] = s_chall_buf[off + 4];
591 w1_t[1] = s_chall_buf[off + 5];
592 w1_t[2] = s_chall_buf[off + 6];
593 w1_t[3] = s_chall_buf[off + 7];
594 w2_t[0] = s_chall_buf[off + 8];
595 w2_t[1] = s_chall_buf[off + 9];
596 w2_t[2] = s_chall_buf[off + 10];
597 w2_t[3] = s_chall_buf[off + 11];
598 w3_t[0] = s_chall_buf[off + 12];
599 w3_t[1] = s_chall_buf[off + 13];
600 w3_t[2] = (64 + chall_len) * 8;
603 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
605 const u32x r0 = digest[0];
606 const u32x r1 = digest[3];
607 const u32x r2 = digest[2];
608 const u32x r3 = digest[1];
610 #include VECT_COMPARE_M
614 extern "C" __global__ void __launch_bounds__ (256, 1) m05600_m08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const netntlm_t *netntlm_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
618 extern "C" __global__ void __launch_bounds__ (256, 1) m05600_m16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const netntlm_t *netntlm_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
622 extern "C" __global__ void __launch_bounds__ (256, 1) m05600_s04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const netntlm_t *netntlm_bufs, u32 *d_return_buf, 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)
628 const u32 lid = threadIdx.x;
634 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
638 wordl0[0] = pws[gid].i[ 0];
639 wordl0[1] = pws[gid].i[ 1];
640 wordl0[2] = pws[gid].i[ 2];
641 wordl0[3] = pws[gid].i[ 3];
645 wordl1[0] = pws[gid].i[ 4];
646 wordl1[1] = pws[gid].i[ 5];
647 wordl1[2] = pws[gid].i[ 6];
648 wordl1[3] = pws[gid].i[ 7];
664 const u32 pw_l_len = pws[gid].pw_len;
666 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
668 append_0x80_2 (wordl0, wordl1, pw_l_len);
670 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
677 __shared__ u32 s_userdomain_buf[64];
678 __shared__ u32 s_chall_buf[256];
680 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
681 + netntlm_bufs[salt_pos].domain_len;
683 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
684 + netntlm_bufs[salt_pos].clichall_len;
688 s_userdomain_buf[lid] = netntlm_bufs[salt_pos].userdomain_buf[lid];
691 s_chall_buf[lid] = netntlm_bufs[salt_pos].chall_buf[lid];
695 if (gid >= gid_max) return;
701 const u32 search[4] =
703 digests_buf[digests_offset].digest_buf[DGST_R0],
704 digests_buf[digests_offset].digest_buf[DGST_R1],
705 digests_buf[digests_offset].digest_buf[DGST_R2],
706 digests_buf[digests_offset].digest_buf[DGST_R3]
713 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
715 const u32 pw_r_len = c_combs[il_pos].pw_len;
717 const u32 pw_len = pw_l_len + pw_r_len;
721 wordr0[0] = c_combs[il_pos].i[0];
722 wordr0[1] = c_combs[il_pos].i[1];
723 wordr0[2] = c_combs[il_pos].i[2];
724 wordr0[3] = c_combs[il_pos].i[3];
728 wordr1[0] = c_combs[il_pos].i[4];
729 wordr1[1] = c_combs[il_pos].i[5];
730 wordr1[2] = c_combs[il_pos].i[6];
731 wordr1[3] = c_combs[il_pos].i[7];
747 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
749 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
754 w0[0] = wordl0[0] | wordr0[0];
755 w0[1] = wordl0[1] | wordr0[1];
756 w0[2] = wordl0[2] | wordr0[2];
757 w0[3] = wordl0[3] | wordr0[3];
761 w1[0] = wordl1[0] | wordr1[0];
762 w1[1] = wordl1[1] | wordr1[1];
763 w1[2] = wordl1[2] | wordr1[2];
764 w1[3] = wordl1[3] | wordr1[3];
785 make_unicode (w0, w0_t, w1_t);
786 make_unicode (w1, w2_t, w3_t);
788 w3_t[2] = pw_len * 8 * 2;
797 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
824 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
829 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
831 w0_t[0] = s_userdomain_buf[off + 0];
832 w0_t[1] = s_userdomain_buf[off + 1];
833 w0_t[2] = s_userdomain_buf[off + 2];
834 w0_t[3] = s_userdomain_buf[off + 3];
835 w1_t[0] = s_userdomain_buf[off + 4];
836 w1_t[1] = s_userdomain_buf[off + 5];
837 w1_t[2] = s_userdomain_buf[off + 6];
838 w1_t[3] = s_userdomain_buf[off + 7];
839 w2_t[0] = s_userdomain_buf[off + 8];
840 w2_t[1] = s_userdomain_buf[off + 9];
841 w2_t[2] = s_userdomain_buf[off + 10];
842 w2_t[3] = s_userdomain_buf[off + 11];
843 w3_t[0] = s_userdomain_buf[off + 12];
844 w3_t[1] = s_userdomain_buf[off + 13];
845 w3_t[2] = s_userdomain_buf[off + 14];
846 w3_t[3] = s_userdomain_buf[off + 15];
848 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
851 w0_t[0] = s_userdomain_buf[off + 0];
852 w0_t[1] = s_userdomain_buf[off + 1];
853 w0_t[2] = s_userdomain_buf[off + 2];
854 w0_t[3] = s_userdomain_buf[off + 3];
855 w1_t[0] = s_userdomain_buf[off + 4];
856 w1_t[1] = s_userdomain_buf[off + 5];
857 w1_t[2] = s_userdomain_buf[off + 6];
858 w1_t[3] = s_userdomain_buf[off + 7];
859 w2_t[0] = s_userdomain_buf[off + 8];
860 w2_t[1] = s_userdomain_buf[off + 9];
861 w2_t[2] = s_userdomain_buf[off + 10];
862 w2_t[3] = s_userdomain_buf[off + 11];
863 w3_t[0] = s_userdomain_buf[off + 12];
864 w3_t[1] = s_userdomain_buf[off + 13];
865 w3_t[2] = (64 + userdomain_len) * 8;
868 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
892 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
894 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
896 w0_t[0] = s_chall_buf[off + 0];
897 w0_t[1] = s_chall_buf[off + 1];
898 w0_t[2] = s_chall_buf[off + 2];
899 w0_t[3] = s_chall_buf[off + 3];
900 w1_t[0] = s_chall_buf[off + 4];
901 w1_t[1] = s_chall_buf[off + 5];
902 w1_t[2] = s_chall_buf[off + 6];
903 w1_t[3] = s_chall_buf[off + 7];
904 w2_t[0] = s_chall_buf[off + 8];
905 w2_t[1] = s_chall_buf[off + 9];
906 w2_t[2] = s_chall_buf[off + 10];
907 w2_t[3] = s_chall_buf[off + 11];
908 w3_t[0] = s_chall_buf[off + 12];
909 w3_t[1] = s_chall_buf[off + 13];
910 w3_t[2] = s_chall_buf[off + 14];
911 w3_t[3] = s_chall_buf[off + 15];
913 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
916 w0_t[0] = s_chall_buf[off + 0];
917 w0_t[1] = s_chall_buf[off + 1];
918 w0_t[2] = s_chall_buf[off + 2];
919 w0_t[3] = s_chall_buf[off + 3];
920 w1_t[0] = s_chall_buf[off + 4];
921 w1_t[1] = s_chall_buf[off + 5];
922 w1_t[2] = s_chall_buf[off + 6];
923 w1_t[3] = s_chall_buf[off + 7];
924 w2_t[0] = s_chall_buf[off + 8];
925 w2_t[1] = s_chall_buf[off + 9];
926 w2_t[2] = s_chall_buf[off + 10];
927 w2_t[3] = s_chall_buf[off + 11];
928 w3_t[0] = s_chall_buf[off + 12];
929 w3_t[1] = s_chall_buf[off + 13];
930 w3_t[2] = (64 + chall_len) * 8;
933 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
935 const u32x r0 = digest[0];
936 const u32x r1 = digest[3];
937 const u32x r2 = digest[2];
938 const u32x r3 = digest[1];
940 #include VECT_COMPARE_S
944 extern "C" __global__ void __launch_bounds__ (256, 1) m05600_s08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const netntlm_t *netntlm_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
948 extern "C" __global__ void __launch_bounds__ (256, 1) m05600_s16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const netntlm_t *netntlm_bufs, u32 *d_return_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)