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"
33 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
34 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
38 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
39 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
43 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
44 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
47 static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
71 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
72 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
73 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
74 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
75 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
76 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
77 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
78 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
79 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
80 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
81 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
82 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
83 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
84 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
85 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
86 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
88 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
89 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
90 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
91 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
92 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
93 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
94 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
95 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
96 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
97 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
98 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
99 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
100 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
101 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
102 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
103 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
105 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
106 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
107 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
108 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
109 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
110 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
111 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
112 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
113 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
114 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
115 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
116 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
117 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
118 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
119 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
120 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
122 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
123 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
124 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
125 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
126 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
127 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
128 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
129 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
130 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
131 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
132 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
133 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
134 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
135 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
136 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
137 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
145 static void gen336 (u32x digest_pre[4], u32 salt_buf[4], u32x digest[4])
152 digest_t0[0] = digest_pre[0];
153 digest_t0[1] = digest_pre[1] & 0xff;
155 digest_t1[0] = digest_pre[0] << 8;
156 digest_t1[1] = digest_pre[0] >> 24 | digest_pre[1] << 8;
158 digest_t2[0] = digest_pre[0] << 16;
159 digest_t2[1] = digest_pre[0] >> 16 | digest_pre[1] << 16;
161 digest_t3[0] = digest_pre[0] << 24;
162 digest_t3[1] = digest_pre[0] >> 8 | digest_pre[1] << 24;
169 salt_buf_t0[0] = salt_buf[0];
170 salt_buf_t0[1] = salt_buf[1];
171 salt_buf_t0[2] = salt_buf[2];
172 salt_buf_t0[3] = salt_buf[3];
174 salt_buf_t1[0] = salt_buf[0] << 8;
175 salt_buf_t1[1] = salt_buf[0] >> 24 | salt_buf[1] << 8;
176 salt_buf_t1[2] = salt_buf[1] >> 24 | salt_buf[2] << 8;
177 salt_buf_t1[3] = salt_buf[2] >> 24 | salt_buf[3] << 8;
178 salt_buf_t1[4] = salt_buf[3] >> 24;
180 salt_buf_t2[0] = salt_buf[0] << 16;
181 salt_buf_t2[1] = salt_buf[0] >> 16 | salt_buf[1] << 16;
182 salt_buf_t2[2] = salt_buf[1] >> 16 | salt_buf[2] << 16;
183 salt_buf_t2[3] = salt_buf[2] >> 16 | salt_buf[3] << 16;
184 salt_buf_t2[4] = salt_buf[3] >> 16;
186 salt_buf_t3[0] = salt_buf[0] << 24;
187 salt_buf_t3[1] = salt_buf[0] >> 8 | salt_buf[1] << 24;
188 salt_buf_t3[2] = salt_buf[1] >> 8 | salt_buf[2] << 24;
189 salt_buf_t3[3] = salt_buf[2] >> 8 | salt_buf[3] << 24;
190 salt_buf_t3[4] = salt_buf[3] >> 8;
197 // generate the 16 * 21 buffer
217 w0_t[0] = digest_t0[0];
218 w0_t[1] = digest_t0[1];
221 w0_t[1] |= salt_buf_t1[0];
222 w0_t[2] = salt_buf_t1[1];
223 w0_t[3] = salt_buf_t1[2];
224 w1_t[0] = salt_buf_t1[3];
225 w1_t[1] = salt_buf_t1[4];
228 w1_t[1] |= digest_t1[0];
229 w1_t[2] = digest_t1[1];
232 w1_t[2] |= salt_buf_t2[0];
233 w1_t[3] = salt_buf_t2[1];
234 w2_t[0] = salt_buf_t2[2];
235 w2_t[1] = salt_buf_t2[3];
236 w2_t[2] = salt_buf_t2[4];
239 w2_t[2] |= digest_t2[0];
240 w2_t[3] = digest_t2[1];
243 w2_t[3] |= salt_buf_t3[0];
244 w3_t[0] = salt_buf_t3[1];
245 w3_t[1] = salt_buf_t3[2];
246 w3_t[2] = salt_buf_t3[3];
247 w3_t[3] = salt_buf_t3[4];
251 w3_t[3] |= digest_t3[0];
253 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
273 w0_t[0] = digest_t3[1];
276 w0_t[1] = salt_buf_t0[0];
277 w0_t[2] = salt_buf_t0[1];
278 w0_t[3] = salt_buf_t0[2];
279 w1_t[0] = salt_buf_t0[3];
282 w1_t[1] = digest_t0[0];
283 w1_t[2] = digest_t0[1];
286 w1_t[2] |= salt_buf_t1[0];
287 w1_t[3] = salt_buf_t1[1];
288 w2_t[0] = salt_buf_t1[2];
289 w2_t[1] = salt_buf_t1[3];
290 w2_t[2] = salt_buf_t1[4];
293 w2_t[2] |= digest_t1[0];
294 w2_t[3] = digest_t1[1];
297 w2_t[3] |= salt_buf_t2[0];
298 w3_t[0] = salt_buf_t2[1];
299 w3_t[1] = salt_buf_t2[2];
300 w3_t[2] = salt_buf_t2[3];
301 w3_t[3] = salt_buf_t2[4];
304 w3_t[3] |= digest_t2[0];
306 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
326 w0_t[0] = digest_t2[1];
329 w0_t[0] |= salt_buf_t3[0];
330 w0_t[1] = salt_buf_t3[1];
331 w0_t[2] = salt_buf_t3[2];
332 w0_t[3] = salt_buf_t3[3];
333 w1_t[0] = salt_buf_t3[4];
336 w1_t[0] |= digest_t3[0];
337 w1_t[1] = digest_t3[1];
340 w1_t[2] = salt_buf_t0[0];
341 w1_t[3] = salt_buf_t0[1];
342 w2_t[0] = salt_buf_t0[2];
343 w2_t[1] = salt_buf_t0[3];
346 w2_t[2] = digest_t0[0];
347 w2_t[3] = digest_t0[1];
350 w2_t[3] |= salt_buf_t1[0];
351 w3_t[0] = salt_buf_t1[1];
352 w3_t[1] = salt_buf_t1[2];
353 w3_t[2] = salt_buf_t1[3];
354 w3_t[3] = salt_buf_t1[4];
357 w3_t[3] |= digest_t1[0];
359 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
379 w0_t[0] = digest_t1[1];
382 w0_t[0] |= salt_buf_t2[0];
383 w0_t[1] = salt_buf_t2[1];
384 w0_t[2] = salt_buf_t2[2];
385 w0_t[3] = salt_buf_t2[3];
386 w1_t[0] = salt_buf_t2[4];
389 w1_t[0] |= digest_t2[0];
390 w1_t[1] = digest_t2[1];
393 w1_t[1] |= salt_buf_t3[0];
394 w1_t[2] = salt_buf_t3[1];
395 w1_t[3] = salt_buf_t3[2];
396 w2_t[0] = salt_buf_t3[3];
397 w2_t[1] = salt_buf_t3[4];
400 w2_t[1] |= digest_t3[0];
401 w2_t[2] = digest_t3[1];
404 w2_t[3] = salt_buf_t0[0];
405 w3_t[0] = salt_buf_t0[1];
406 w3_t[1] = salt_buf_t0[2];
407 w3_t[2] = salt_buf_t0[3];
410 w3_t[3] = digest_t0[0];
412 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
432 w0_t[0] = digest_t0[1];
435 w0_t[0] |= salt_buf_t1[0];
436 w0_t[1] = salt_buf_t1[1];
437 w0_t[2] = salt_buf_t1[2];
438 w0_t[3] = salt_buf_t1[3];
439 w1_t[0] = salt_buf_t1[4];
442 w1_t[0] |= digest_t1[0];
443 w1_t[1] = digest_t1[1];
446 w1_t[1] |= salt_buf_t2[0];
447 w1_t[2] = salt_buf_t2[1];
448 w1_t[3] = salt_buf_t2[2];
449 w2_t[0] = salt_buf_t2[3];
450 w2_t[1] = salt_buf_t2[4];
453 w2_t[1] |= digest_t2[0];
454 w2_t[2] = digest_t2[1];
457 w2_t[2] |= salt_buf_t3[0];
458 w2_t[3] = salt_buf_t3[1];
459 w3_t[0] = salt_buf_t3[2];
460 w3_t[1] = salt_buf_t3[3];
461 w3_t[2] = salt_buf_t3[4];
464 w3_t[2] |= digest_t3[0];
465 w3_t[3] = digest_t3[1];
467 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
469 w0_t[0] = salt_buf_t0[0];
470 w0_t[1] = salt_buf_t0[1];
471 w0_t[2] = salt_buf_t0[2];
472 w0_t[3] = salt_buf_t0[3];
483 w3_t[2] = 21 * 16 * 8;
486 md5_transform (w0_t, w1_t, w2_t, w3_t, digest);
489 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m09720_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 oldoffice01_t *oldoffice01_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)
495 const u32 lid = get_local_id (0);
501 const u32 gid = get_global_id (0);
503 if (gid >= gid_max) return;
507 wordl0[0] = pws[gid].i[ 0];
508 wordl0[1] = pws[gid].i[ 1];
509 wordl0[2] = pws[gid].i[ 2];
510 wordl0[3] = pws[gid].i[ 3];
514 wordl1[0] = pws[gid].i[ 4];
515 wordl1[1] = pws[gid].i[ 5];
516 wordl1[2] = pws[gid].i[ 6];
517 wordl1[3] = pws[gid].i[ 7];
533 const u32 pw_l_len = pws[gid].pw_len;
535 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
537 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
546 salt_buf[0] = salt_bufs[salt_pos].salt_buf[0];
547 salt_buf[1] = salt_bufs[salt_pos].salt_buf[1];
548 salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
549 salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
555 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
557 const u32 pw_r_len = combs_buf[il_pos].pw_len;
559 const u32 pw_len = pw_l_len + pw_r_len;
563 wordr0[0] = combs_buf[il_pos].i[0];
564 wordr0[1] = combs_buf[il_pos].i[1];
565 wordr0[2] = combs_buf[il_pos].i[2];
566 wordr0[3] = combs_buf[il_pos].i[3];
570 wordr1[0] = combs_buf[il_pos].i[4];
571 wordr1[1] = combs_buf[il_pos].i[5];
572 wordr1[2] = combs_buf[il_pos].i[6];
573 wordr1[3] = combs_buf[il_pos].i[7];
589 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
591 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
596 w0[0] = wordl0[0] | wordr0[0];
597 w0[1] = wordl0[1] | wordr0[1];
598 w0[2] = wordl0[2] | wordr0[2];
599 w0[3] = wordl0[3] | wordr0[3];
603 w1[0] = wordl1[0] | wordr1[0];
604 w1[1] = wordl1[1] | wordr1[1];
605 w1[2] = wordl1[2] | wordr1[2];
606 w1[3] = wordl1[3] | wordr1[3];
610 w2[0] = wordl2[0] | wordr2[0];
611 w2[1] = wordl2[1] | wordr2[1];
612 w2[2] = wordl2[2] | wordr2[2];
613 w2[3] = wordl2[3] | wordr2[3];
617 w3[0] = wordl3[0] | wordr3[0];
618 w3[1] = wordl3[1] | wordr3[1];
622 append_0x80_2 (w0, w1, pw_len);
629 make_unicode (w0, w0_t, w1_t);
630 make_unicode (w1, w2_t, w3_t);
632 w3_t[2] = pw_len * 8 * 2;
636 digest_pre[0] = MD5M_A;
637 digest_pre[1] = MD5M_B;
638 digest_pre[2] = MD5M_C;
639 digest_pre[3] = MD5M_D;
641 md5_transform (w0_t, w1_t, w2_t, w3_t, digest_pre);
643 digest_pre[0] &= 0xffffffff;
644 digest_pre[1] &= 0x000000ff;
645 digest_pre[2] &= 0x00000000;
646 digest_pre[3] &= 0x00000000;
655 gen336 (digest_pre, salt_buf, digest);
658 u32x b = digest[1] & 0xff;
665 #include VECT_COMPARE_M
669 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m09720_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 oldoffice01_t *oldoffice01_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)
673 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m09720_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 oldoffice01_t *oldoffice01_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)
677 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m09720_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 oldoffice01_t *oldoffice01_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)
683 const u32 lid = get_local_id (0);
689 const u32 gid = get_global_id (0);
691 if (gid >= gid_max) return;
695 wordl0[0] = pws[gid].i[ 0];
696 wordl0[1] = pws[gid].i[ 1];
697 wordl0[2] = pws[gid].i[ 2];
698 wordl0[3] = pws[gid].i[ 3];
702 wordl1[0] = pws[gid].i[ 4];
703 wordl1[1] = pws[gid].i[ 5];
704 wordl1[2] = pws[gid].i[ 6];
705 wordl1[3] = pws[gid].i[ 7];
721 const u32 pw_l_len = pws[gid].pw_len;
723 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
725 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
732 const u32 search[4] =
734 digests_buf[digests_offset].digest_buf[DGST_R0],
735 digests_buf[digests_offset].digest_buf[DGST_R1],
736 digests_buf[digests_offset].digest_buf[DGST_R2],
737 digests_buf[digests_offset].digest_buf[DGST_R3]
746 salt_buf[0] = salt_bufs[salt_pos].salt_buf[0];
747 salt_buf[1] = salt_bufs[salt_pos].salt_buf[1];
748 salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
749 salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
755 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
757 const u32 pw_r_len = combs_buf[il_pos].pw_len;
759 const u32 pw_len = pw_l_len + pw_r_len;
763 wordr0[0] = combs_buf[il_pos].i[0];
764 wordr0[1] = combs_buf[il_pos].i[1];
765 wordr0[2] = combs_buf[il_pos].i[2];
766 wordr0[3] = combs_buf[il_pos].i[3];
770 wordr1[0] = combs_buf[il_pos].i[4];
771 wordr1[1] = combs_buf[il_pos].i[5];
772 wordr1[2] = combs_buf[il_pos].i[6];
773 wordr1[3] = combs_buf[il_pos].i[7];
789 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
791 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
796 w0[0] = wordl0[0] | wordr0[0];
797 w0[1] = wordl0[1] | wordr0[1];
798 w0[2] = wordl0[2] | wordr0[2];
799 w0[3] = wordl0[3] | wordr0[3];
803 w1[0] = wordl1[0] | wordr1[0];
804 w1[1] = wordl1[1] | wordr1[1];
805 w1[2] = wordl1[2] | wordr1[2];
806 w1[3] = wordl1[3] | wordr1[3];
810 w2[0] = wordl2[0] | wordr2[0];
811 w2[1] = wordl2[1] | wordr2[1];
812 w2[2] = wordl2[2] | wordr2[2];
813 w2[3] = wordl2[3] | wordr2[3];
817 w3[0] = wordl3[0] | wordr3[0];
818 w3[1] = wordl3[1] | wordr3[1];
822 append_0x80_2 (w0, w1, pw_len);
829 make_unicode (w0, w0_t, w1_t);
830 make_unicode (w1, w2_t, w3_t);
832 w3_t[2] = pw_len * 8 * 2;
836 digest_pre[0] = MD5M_A;
837 digest_pre[1] = MD5M_B;
838 digest_pre[2] = MD5M_C;
839 digest_pre[3] = MD5M_D;
841 md5_transform (w0_t, w1_t, w2_t, w3_t, digest_pre);
843 digest_pre[0] &= 0xffffffff;
844 digest_pre[1] &= 0x000000ff;
845 digest_pre[2] &= 0x00000000;
846 digest_pre[3] &= 0x00000000;
855 gen336 (digest_pre, salt_buf, digest);
858 u32x b = digest[1] & 0xff;
865 #include VECT_COMPARE_S
869 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m09720_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 oldoffice01_t *oldoffice01_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)
873 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m09720_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 oldoffice01_t *oldoffice01_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)