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"
19 #include "include/rp_gpu.h"
22 #define COMPARE_S "check_single_comp4.c"
23 #define COMPARE_M "check_multi_comp4.c"
25 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04800_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 void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
31 const u32 lid = get_local_id (0);
37 const u32 gid = get_global_id (0);
39 if (gid >= gid_max) return;
43 pw_buf0[0] = pws[gid].i[ 0];
44 pw_buf0[1] = pws[gid].i[ 1];
45 pw_buf0[2] = pws[gid].i[ 2];
46 pw_buf0[3] = pws[gid].i[ 3];
50 pw_buf1[0] = pws[gid].i[ 4];
51 pw_buf1[1] = pws[gid].i[ 5];
52 pw_buf1[2] = pws[gid].i[ 6];
53 pw_buf1[3] = pws[gid].i[ 7];
55 const u32 pw_len = pws[gid].pw_len;
63 salt_buf[0] = salt_bufs[salt_pos].salt_buf[0];
64 salt_buf[1] = salt_bufs[salt_pos].salt_buf[1];
65 salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
66 salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
67 salt_buf[4] = salt_bufs[salt_pos].salt_buf[4];
69 const u32 salt_len = salt_bufs[salt_pos].salt_len;
75 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
105 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
107 u32 pw_salt_len = out_len + salt_len;
141 switch_buffer_by_offset (s0, s1, s2, s3, out_len);
167 switch_buffer_by_offset (w0, w1, w2, w3, 1);
169 w0[0] |= salt_buf[4];
171 w3[2] = pw_salt_len * 8;
182 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
183 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
184 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
185 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
186 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
187 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
188 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
189 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
190 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
191 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
192 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
193 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
194 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
195 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
196 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
197 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
199 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
200 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
201 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
202 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
203 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
204 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
205 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
206 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
207 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
208 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
209 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
210 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
211 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
212 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
213 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
214 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
216 MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
217 MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
218 MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
219 MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
220 MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
221 MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
222 MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
223 MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
224 MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
225 MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
226 MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
227 MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
228 MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
229 MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
230 MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
231 MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);
233 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
234 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
235 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
236 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
237 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
238 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
239 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
240 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
241 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
242 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
243 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
244 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
245 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
246 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
247 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
248 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
259 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04800_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 void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
263 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04800_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 void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
267 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04800_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 void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
273 const u32 lid = get_local_id (0);
279 const u32 gid = get_global_id (0);
281 if (gid >= gid_max) return;
285 pw_buf0[0] = pws[gid].i[ 0];
286 pw_buf0[1] = pws[gid].i[ 1];
287 pw_buf0[2] = pws[gid].i[ 2];
288 pw_buf0[3] = pws[gid].i[ 3];
292 pw_buf1[0] = pws[gid].i[ 4];
293 pw_buf1[1] = pws[gid].i[ 5];
294 pw_buf1[2] = pws[gid].i[ 6];
295 pw_buf1[3] = pws[gid].i[ 7];
297 const u32 pw_len = pws[gid].pw_len;
305 salt_buf[0] = salt_bufs[salt_pos].salt_buf[0];
306 salt_buf[1] = salt_bufs[salt_pos].salt_buf[1];
307 salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
308 salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
309 salt_buf[4] = salt_bufs[salt_pos].salt_buf[4];
311 const u32 salt_len = salt_bufs[salt_pos].salt_len;
317 const u32 search[4] =
319 digests_buf[digests_offset].digest_buf[DGST_R0],
320 digests_buf[digests_offset].digest_buf[DGST_R1],
321 digests_buf[digests_offset].digest_buf[DGST_R2],
322 digests_buf[digests_offset].digest_buf[DGST_R3]
329 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
359 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
361 u32 pw_salt_len = out_len + salt_len;
395 switch_buffer_by_offset (s0, s1, s2, s3, out_len);
421 switch_buffer_by_offset (w0, w1, w2, w3, 1);
423 w0[0] |= salt_buf[4];
425 w3[2] = pw_salt_len * 8;
436 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
437 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
438 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
439 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
440 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
441 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
442 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
443 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
444 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
445 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
446 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
447 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
448 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
449 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
450 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
451 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
453 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
454 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
455 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
456 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
457 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
458 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
459 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
460 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
461 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
462 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
463 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
464 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
465 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
466 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
467 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
468 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
470 MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
471 MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
472 MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
473 MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
474 MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
475 MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
476 MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
477 MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
478 MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
479 MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
480 MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
481 MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
482 MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
483 MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
484 MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
485 MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);
487 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
488 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
489 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
490 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
491 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
492 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
493 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
494 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
495 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
496 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
497 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
498 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
499 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
501 bool q_cond = allx (search[0] != a);
503 if (q_cond) continue;
505 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
506 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
507 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
518 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04800_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 void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
522 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m04800_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 void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)