2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
28 #include "include/kernel_functions.c"
29 #include "types_amd.c"
30 #include "common_amd.c"
31 #include "include/rp_gpu.h"
35 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
36 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
40 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
41 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
45 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
46 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
49 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02410_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)
55 const u32 lid = get_local_id (0);
61 const u32 gid = get_global_id (0);
63 if (gid >= gid_max) return;
67 pw_buf0[0] = pws[gid].i[ 0];
68 pw_buf0[1] = pws[gid].i[ 1];
69 pw_buf0[2] = pws[gid].i[ 2];
70 pw_buf0[3] = pws[gid].i[ 3];
72 const u32 pw_len = pws[gid].pw_len;
80 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
85 const u32 salt_len = (salt_bufs[salt_pos].salt_len < 4) ? salt_bufs[salt_pos].salt_len : 4;
91 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
121 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
129 s0[0] = salt_buf0[0];
130 s0[1] = salt_buf0[1];
131 s0[2] = salt_buf0[2];
132 s0[3] = salt_buf0[3];
155 switch_buffer_by_offset (s0, s1, s2, s3, out_len);
162 const u32 pw_salt_len = out_len + salt_len;
164 truncate_block (w0, pw_salt_len);
176 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
177 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
178 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
179 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
180 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
181 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
182 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
183 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
184 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
185 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
186 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
187 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
188 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
189 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
190 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
191 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
193 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
194 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
195 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
196 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
197 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
198 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
199 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
200 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
201 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
202 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
203 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
204 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
205 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
206 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
207 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
208 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
210 MD5_STEP (MD5_H1, a, b, c, d, w1[1], MD5C20, MD5S20);
211 MD5_STEP (MD5_H2, d, a, b, c, w2[0], MD5C21, MD5S21);
212 MD5_STEP (MD5_H1, c, d, a, b, w2[3], MD5C22, MD5S22);
213 MD5_STEP (MD5_H2, b, c, d, a, w3[2], MD5C23, MD5S23);
214 MD5_STEP (MD5_H1, a, b, c, d, w0[1], MD5C24, MD5S20);
215 MD5_STEP (MD5_H2, d, a, b, c, w1[0], MD5C25, MD5S21);
216 MD5_STEP (MD5_H1, c, d, a, b, w1[3], MD5C26, MD5S22);
217 MD5_STEP (MD5_H2, b, c, d, a, w2[2], MD5C27, MD5S23);
218 MD5_STEP (MD5_H1, a, b, c, d, w3[1], MD5C28, MD5S20);
219 MD5_STEP (MD5_H2, d, a, b, c, w0[0], MD5C29, MD5S21);
220 MD5_STEP (MD5_H1, c, d, a, b, w0[3], MD5C2a, MD5S22);
221 MD5_STEP (MD5_H2, b, c, d, a, w1[2], MD5C2b, MD5S23);
222 MD5_STEP (MD5_H1, a, b, c, d, w2[1], MD5C2c, MD5S20);
223 MD5_STEP (MD5_H2, d, a, b, c, w3[0], MD5C2d, MD5S21);
224 MD5_STEP (MD5_H1, c, d, a, b, w3[3], MD5C2e, MD5S22);
225 MD5_STEP (MD5_H2, b, c, d, a, w0[2], MD5C2f, MD5S23);
227 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
228 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
229 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
230 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
231 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
232 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
233 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
234 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
235 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
236 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
237 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
238 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
239 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
240 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
241 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
242 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
254 #include VECT_COMPARE_M
258 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02410_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)
262 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02410_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)
266 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02410_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)
272 const u32 lid = get_local_id (0);
278 const u32 gid = get_global_id (0);
280 if (gid >= gid_max) return;
284 pw_buf0[0] = pws[gid].i[ 0];
285 pw_buf0[1] = pws[gid].i[ 1];
286 pw_buf0[2] = pws[gid].i[ 2];
287 pw_buf0[3] = pws[gid].i[ 3];
289 const u32 pw_len = pws[gid].pw_len;
297 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
302 const u32 salt_len = (salt_bufs[salt_pos].salt_len < 4) ? salt_bufs[salt_pos].salt_len : 4;
308 const u32 search[4] =
310 digests_buf[digests_offset].digest_buf[DGST_R0],
311 digests_buf[digests_offset].digest_buf[DGST_R1],
312 digests_buf[digests_offset].digest_buf[DGST_R2],
313 digests_buf[digests_offset].digest_buf[DGST_R3]
320 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
350 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
358 s0[0] = salt_buf0[0];
359 s0[1] = salt_buf0[1];
360 s0[2] = salt_buf0[2];
361 s0[3] = salt_buf0[3];
384 switch_buffer_by_offset (s0, s1, s2, s3, out_len);
391 const u32 pw_salt_len = out_len + salt_len;
393 truncate_block (w0, pw_salt_len);
405 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
406 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
407 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
408 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
409 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
410 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
411 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
412 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
413 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
414 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
415 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
416 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
417 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
418 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
419 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
420 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
422 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
423 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
424 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
425 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
426 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
427 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
428 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
429 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
430 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
431 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
432 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
433 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
434 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
435 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
436 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
437 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
439 MD5_STEP (MD5_H1, a, b, c, d, w1[1], MD5C20, MD5S20);
440 MD5_STEP (MD5_H2, d, a, b, c, w2[0], MD5C21, MD5S21);
441 MD5_STEP (MD5_H1, c, d, a, b, w2[3], MD5C22, MD5S22);
442 MD5_STEP (MD5_H2, b, c, d, a, w3[2], MD5C23, MD5S23);
443 MD5_STEP (MD5_H1, a, b, c, d, w0[1], MD5C24, MD5S20);
444 MD5_STEP (MD5_H2, d, a, b, c, w1[0], MD5C25, MD5S21);
445 MD5_STEP (MD5_H1, c, d, a, b, w1[3], MD5C26, MD5S22);
446 MD5_STEP (MD5_H2, b, c, d, a, w2[2], MD5C27, MD5S23);
447 MD5_STEP (MD5_H1, a, b, c, d, w3[1], MD5C28, MD5S20);
448 MD5_STEP (MD5_H2, d, a, b, c, w0[0], MD5C29, MD5S21);
449 MD5_STEP (MD5_H1, c, d, a, b, w0[3], MD5C2a, MD5S22);
450 MD5_STEP (MD5_H2, b, c, d, a, w1[2], MD5C2b, MD5S23);
451 MD5_STEP (MD5_H1, a, b, c, d, w2[1], MD5C2c, MD5S20);
452 MD5_STEP (MD5_H2, d, a, b, c, w3[0], MD5C2d, MD5S21);
453 MD5_STEP (MD5_H1, c, d, a, b, w3[3], MD5C2e, MD5S22);
454 MD5_STEP (MD5_H2, b, c, d, a, w0[2], MD5C2f, MD5S23);
456 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
457 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
458 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
459 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
460 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
461 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
462 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
463 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
464 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
465 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
466 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
467 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
468 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
470 bool q_cond = allx ((a & 0x00ffffff) != search[0]);
472 if (q_cond) continue;
474 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
475 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
476 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
488 #include VECT_COMPARE_S
492 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02410_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)
496 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m02410_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)