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"
20 #define COMPARE_S "check_single_comp4.c"
21 #define COMPARE_M "check_multi_comp4.c"
23 static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
47 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
48 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
49 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
50 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
51 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
52 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
53 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
54 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
55 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
56 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
57 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
58 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
59 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
60 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
61 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
62 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
64 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
65 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
66 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
67 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
68 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
69 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
70 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
71 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
72 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
73 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
74 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
75 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
76 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
77 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
78 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
79 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
81 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
82 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
83 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
84 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
85 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
86 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
87 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
88 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
89 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
90 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
91 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
92 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
93 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
94 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
95 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
96 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
98 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
99 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
100 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
101 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
102 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
103 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
104 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
105 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
106 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
107 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
108 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
109 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
110 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
111 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
112 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
113 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
121 static void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
123 w0[0] = w0[0] ^ 0x36363636;
124 w0[1] = w0[1] ^ 0x36363636;
125 w0[2] = w0[2] ^ 0x36363636;
126 w0[3] = w0[3] ^ 0x36363636;
127 w1[0] = w1[0] ^ 0x36363636;
128 w1[1] = w1[1] ^ 0x36363636;
129 w1[2] = w1[2] ^ 0x36363636;
130 w1[3] = w1[3] ^ 0x36363636;
131 w2[0] = w2[0] ^ 0x36363636;
132 w2[1] = w2[1] ^ 0x36363636;
133 w2[2] = w2[2] ^ 0x36363636;
134 w2[3] = w2[3] ^ 0x36363636;
135 w3[0] = w3[0] ^ 0x36363636;
136 w3[1] = w3[1] ^ 0x36363636;
137 w3[2] = w3[2] ^ 0x36363636;
138 w3[3] = w3[3] ^ 0x36363636;
145 md5_transform (w0, w1, w2, w3, ipad);
147 w0[0] = w0[0] ^ 0x6a6a6a6a;
148 w0[1] = w0[1] ^ 0x6a6a6a6a;
149 w0[2] = w0[2] ^ 0x6a6a6a6a;
150 w0[3] = w0[3] ^ 0x6a6a6a6a;
151 w1[0] = w1[0] ^ 0x6a6a6a6a;
152 w1[1] = w1[1] ^ 0x6a6a6a6a;
153 w1[2] = w1[2] ^ 0x6a6a6a6a;
154 w1[3] = w1[3] ^ 0x6a6a6a6a;
155 w2[0] = w2[0] ^ 0x6a6a6a6a;
156 w2[1] = w2[1] ^ 0x6a6a6a6a;
157 w2[2] = w2[2] ^ 0x6a6a6a6a;
158 w2[3] = w2[3] ^ 0x6a6a6a6a;
159 w3[0] = w3[0] ^ 0x6a6a6a6a;
160 w3[1] = w3[1] ^ 0x6a6a6a6a;
161 w3[2] = w3[2] ^ 0x6a6a6a6a;
162 w3[3] = w3[3] ^ 0x6a6a6a6a;
169 md5_transform (w0, w1, w2, w3, opad);
172 static void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
179 md5_transform (w0, w1, w2, w3, digest);
195 w3[2] = (64 + 16) * 8;
203 md5_transform (w0, w1, w2, w3, digest);
206 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05300_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 ikepsk_t *ikepsk_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)
212 const u32 lid = get_local_id (0);
218 const u32 gid = get_global_id (0);
222 wordl0[0] = pws[gid].i[ 0];
223 wordl0[1] = pws[gid].i[ 1];
224 wordl0[2] = pws[gid].i[ 2];
225 wordl0[3] = pws[gid].i[ 3];
229 wordl1[0] = pws[gid].i[ 4];
230 wordl1[1] = pws[gid].i[ 5];
231 wordl1[2] = pws[gid].i[ 6];
232 wordl1[3] = pws[gid].i[ 7];
248 const u32 pw_l_len = pws[gid].pw_len;
250 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
252 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
259 const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
260 const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
264 salt_buf0[0] = ikepsk_bufs[salt_pos].nr_buf[ 0];
265 salt_buf0[1] = ikepsk_bufs[salt_pos].nr_buf[ 1];
266 salt_buf0[2] = ikepsk_bufs[salt_pos].nr_buf[ 2];
267 salt_buf0[3] = ikepsk_bufs[salt_pos].nr_buf[ 3];
271 salt_buf1[0] = ikepsk_bufs[salt_pos].nr_buf[ 4];
272 salt_buf1[1] = ikepsk_bufs[salt_pos].nr_buf[ 5];
273 salt_buf1[2] = ikepsk_bufs[salt_pos].nr_buf[ 6];
274 salt_buf1[3] = ikepsk_bufs[salt_pos].nr_buf[ 7];
278 salt_buf2[0] = ikepsk_bufs[salt_pos].nr_buf[ 8];
279 salt_buf2[1] = ikepsk_bufs[salt_pos].nr_buf[ 9];
280 salt_buf2[2] = ikepsk_bufs[salt_pos].nr_buf[10];
281 salt_buf2[3] = ikepsk_bufs[salt_pos].nr_buf[11];
285 salt_buf3[0] = ikepsk_bufs[salt_pos].nr_buf[12];
286 salt_buf3[1] = ikepsk_bufs[salt_pos].nr_buf[13];
290 __local u32 s_msg_buf[128];
292 const u32 lid2 = lid * 2;
294 s_msg_buf[lid2 + 0] = ikepsk_bufs[salt_pos].msg_buf[lid2 + 0];
295 s_msg_buf[lid2 + 1] = ikepsk_bufs[salt_pos].msg_buf[lid2 + 1];
297 barrier (CLK_LOCAL_MEM_FENCE);
299 if (gid >= gid_max) return;
305 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
307 const u32 pw_r_len = combs_buf[il_pos].pw_len;
309 const u32 pw_len = pw_l_len + pw_r_len;
313 wordr0[0] = combs_buf[il_pos].i[0];
314 wordr0[1] = combs_buf[il_pos].i[1];
315 wordr0[2] = combs_buf[il_pos].i[2];
316 wordr0[3] = combs_buf[il_pos].i[3];
320 wordr1[0] = combs_buf[il_pos].i[4];
321 wordr1[1] = combs_buf[il_pos].i[5];
322 wordr1[2] = combs_buf[il_pos].i[6];
323 wordr1[3] = combs_buf[il_pos].i[7];
339 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
341 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
346 w0[0] = wordl0[0] | wordr0[0];
347 w0[1] = wordl0[1] | wordr0[1];
348 w0[2] = wordl0[2] | wordr0[2];
349 w0[3] = wordl0[3] | wordr0[3];
353 w1[0] = wordl1[0] | wordr1[0];
354 w1[1] = wordl1[1] | wordr1[1];
355 w1[2] = wordl1[2] | wordr1[2];
356 w1[3] = wordl1[3] | wordr1[3];
360 w2[0] = wordl2[0] | wordr2[0];
361 w2[1] = wordl2[1] | wordr2[1];
362 w2[2] = wordl2[2] | wordr2[2];
363 w2[3] = wordl2[3] | wordr2[3];
367 w3[0] = wordl3[0] | wordr3[0];
368 w3[1] = wordl3[1] | wordr3[1];
369 w3[2] = wordl3[2] | wordr3[2];
370 w3[3] = wordl3[3] | wordr3[3];
407 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
409 w0_t[0] = salt_buf0[0];
410 w0_t[1] = salt_buf0[1];
411 w0_t[2] = salt_buf0[2];
412 w0_t[3] = salt_buf0[3];
413 w1_t[0] = salt_buf1[0];
414 w1_t[1] = salt_buf1[1];
415 w1_t[2] = salt_buf1[2];
416 w1_t[3] = salt_buf1[3];
417 w2_t[0] = salt_buf2[0];
418 w2_t[1] = salt_buf2[1];
419 w2_t[2] = salt_buf2[2];
420 w2_t[3] = salt_buf2[3];
421 w3_t[0] = salt_buf3[0];
422 w3_t[1] = salt_buf3[1];
423 w3_t[2] = (64 + nr_len) * 8;
428 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
447 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
452 for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
454 w0_t[0] = s_msg_buf[off + 0];
455 w0_t[1] = s_msg_buf[off + 1];
456 w0_t[2] = s_msg_buf[off + 2];
457 w0_t[3] = s_msg_buf[off + 3];
458 w1_t[0] = s_msg_buf[off + 4];
459 w1_t[1] = s_msg_buf[off + 5];
460 w1_t[2] = s_msg_buf[off + 6];
461 w1_t[3] = s_msg_buf[off + 7];
462 w2_t[0] = s_msg_buf[off + 8];
463 w2_t[1] = s_msg_buf[off + 9];
464 w2_t[2] = s_msg_buf[off + 10];
465 w2_t[3] = s_msg_buf[off + 11];
466 w3_t[0] = s_msg_buf[off + 12];
467 w3_t[1] = s_msg_buf[off + 13];
468 w3_t[2] = s_msg_buf[off + 14];
469 w3_t[3] = s_msg_buf[off + 15];
471 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
474 w0_t[0] = s_msg_buf[off + 0];
475 w0_t[1] = s_msg_buf[off + 1];
476 w0_t[2] = s_msg_buf[off + 2];
477 w0_t[3] = s_msg_buf[off + 3];
478 w1_t[0] = s_msg_buf[off + 4];
479 w1_t[1] = s_msg_buf[off + 5];
480 w1_t[2] = s_msg_buf[off + 6];
481 w1_t[3] = s_msg_buf[off + 7];
482 w2_t[0] = s_msg_buf[off + 8];
483 w2_t[1] = s_msg_buf[off + 9];
484 w2_t[2] = s_msg_buf[off + 10];
485 w2_t[3] = s_msg_buf[off + 11];
486 w3_t[0] = s_msg_buf[off + 12];
487 w3_t[1] = s_msg_buf[off + 13];
488 w3_t[2] = (64 + msg_len) * 8;
491 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
493 const u32 r0 = digest[0];
494 const u32 r1 = digest[3];
495 const u32 r2 = digest[2];
496 const u32 r3 = digest[1];
502 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05300_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 ikepsk_t *ikepsk_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)
506 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05300_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 ikepsk_t *ikepsk_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)
510 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05300_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 ikepsk_t *ikepsk_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)
516 const u32 lid = get_local_id (0);
522 const u32 gid = get_global_id (0);
526 wordl0[0] = pws[gid].i[ 0];
527 wordl0[1] = pws[gid].i[ 1];
528 wordl0[2] = pws[gid].i[ 2];
529 wordl0[3] = pws[gid].i[ 3];
533 wordl1[0] = pws[gid].i[ 4];
534 wordl1[1] = pws[gid].i[ 5];
535 wordl1[2] = pws[gid].i[ 6];
536 wordl1[3] = pws[gid].i[ 7];
552 const u32 pw_l_len = pws[gid].pw_len;
554 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
556 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
563 const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
564 const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
568 salt_buf0[0] = ikepsk_bufs[salt_pos].nr_buf[ 0];
569 salt_buf0[1] = ikepsk_bufs[salt_pos].nr_buf[ 1];
570 salt_buf0[2] = ikepsk_bufs[salt_pos].nr_buf[ 2];
571 salt_buf0[3] = ikepsk_bufs[salt_pos].nr_buf[ 3];
575 salt_buf1[0] = ikepsk_bufs[salt_pos].nr_buf[ 4];
576 salt_buf1[1] = ikepsk_bufs[salt_pos].nr_buf[ 5];
577 salt_buf1[2] = ikepsk_bufs[salt_pos].nr_buf[ 6];
578 salt_buf1[3] = ikepsk_bufs[salt_pos].nr_buf[ 7];
582 salt_buf2[0] = ikepsk_bufs[salt_pos].nr_buf[ 8];
583 salt_buf2[1] = ikepsk_bufs[salt_pos].nr_buf[ 9];
584 salt_buf2[2] = ikepsk_bufs[salt_pos].nr_buf[10];
585 salt_buf2[3] = ikepsk_bufs[salt_pos].nr_buf[11];
589 salt_buf3[0] = ikepsk_bufs[salt_pos].nr_buf[12];
590 salt_buf3[1] = ikepsk_bufs[salt_pos].nr_buf[13];
594 __local u32 s_msg_buf[128];
596 const u32 lid2 = lid * 2;
598 s_msg_buf[lid2 + 0] = ikepsk_bufs[salt_pos].msg_buf[lid2 + 0];
599 s_msg_buf[lid2 + 1] = ikepsk_bufs[salt_pos].msg_buf[lid2 + 1];
601 barrier (CLK_LOCAL_MEM_FENCE);
603 if (gid >= gid_max) return;
609 const u32 search[4] =
611 digests_buf[digests_offset].digest_buf[DGST_R0],
612 digests_buf[digests_offset].digest_buf[DGST_R1],
613 digests_buf[digests_offset].digest_buf[DGST_R2],
614 digests_buf[digests_offset].digest_buf[DGST_R3]
621 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
623 const u32 pw_r_len = combs_buf[il_pos].pw_len;
625 const u32 pw_len = pw_l_len + pw_r_len;
629 wordr0[0] = combs_buf[il_pos].i[0];
630 wordr0[1] = combs_buf[il_pos].i[1];
631 wordr0[2] = combs_buf[il_pos].i[2];
632 wordr0[3] = combs_buf[il_pos].i[3];
636 wordr1[0] = combs_buf[il_pos].i[4];
637 wordr1[1] = combs_buf[il_pos].i[5];
638 wordr1[2] = combs_buf[il_pos].i[6];
639 wordr1[3] = combs_buf[il_pos].i[7];
655 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
657 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
662 w0[0] = wordl0[0] | wordr0[0];
663 w0[1] = wordl0[1] | wordr0[1];
664 w0[2] = wordl0[2] | wordr0[2];
665 w0[3] = wordl0[3] | wordr0[3];
669 w1[0] = wordl1[0] | wordr1[0];
670 w1[1] = wordl1[1] | wordr1[1];
671 w1[2] = wordl1[2] | wordr1[2];
672 w1[3] = wordl1[3] | wordr1[3];
676 w2[0] = wordl2[0] | wordr2[0];
677 w2[1] = wordl2[1] | wordr2[1];
678 w2[2] = wordl2[2] | wordr2[2];
679 w2[3] = wordl2[3] | wordr2[3];
683 w3[0] = wordl3[0] | wordr3[0];
684 w3[1] = wordl3[1] | wordr3[1];
685 w3[2] = wordl3[2] | wordr3[2];
686 w3[3] = wordl3[3] | wordr3[3];
723 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
725 w0_t[0] = salt_buf0[0];
726 w0_t[1] = salt_buf0[1];
727 w0_t[2] = salt_buf0[2];
728 w0_t[3] = salt_buf0[3];
729 w1_t[0] = salt_buf1[0];
730 w1_t[1] = salt_buf1[1];
731 w1_t[2] = salt_buf1[2];
732 w1_t[3] = salt_buf1[3];
733 w2_t[0] = salt_buf2[0];
734 w2_t[1] = salt_buf2[1];
735 w2_t[2] = salt_buf2[2];
736 w2_t[3] = salt_buf2[3];
737 w3_t[0] = salt_buf3[0];
738 w3_t[1] = salt_buf3[1];
739 w3_t[2] = (64 + nr_len) * 8;
744 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
763 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
768 for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
770 w0_t[0] = s_msg_buf[off + 0];
771 w0_t[1] = s_msg_buf[off + 1];
772 w0_t[2] = s_msg_buf[off + 2];
773 w0_t[3] = s_msg_buf[off + 3];
774 w1_t[0] = s_msg_buf[off + 4];
775 w1_t[1] = s_msg_buf[off + 5];
776 w1_t[2] = s_msg_buf[off + 6];
777 w1_t[3] = s_msg_buf[off + 7];
778 w2_t[0] = s_msg_buf[off + 8];
779 w2_t[1] = s_msg_buf[off + 9];
780 w2_t[2] = s_msg_buf[off + 10];
781 w2_t[3] = s_msg_buf[off + 11];
782 w3_t[0] = s_msg_buf[off + 12];
783 w3_t[1] = s_msg_buf[off + 13];
784 w3_t[2] = s_msg_buf[off + 14];
785 w3_t[3] = s_msg_buf[off + 15];
787 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
790 w0_t[0] = s_msg_buf[off + 0];
791 w0_t[1] = s_msg_buf[off + 1];
792 w0_t[2] = s_msg_buf[off + 2];
793 w0_t[3] = s_msg_buf[off + 3];
794 w1_t[0] = s_msg_buf[off + 4];
795 w1_t[1] = s_msg_buf[off + 5];
796 w1_t[2] = s_msg_buf[off + 6];
797 w1_t[3] = s_msg_buf[off + 7];
798 w2_t[0] = s_msg_buf[off + 8];
799 w2_t[1] = s_msg_buf[off + 9];
800 w2_t[2] = s_msg_buf[off + 10];
801 w2_t[3] = s_msg_buf[off + 11];
802 w3_t[0] = s_msg_buf[off + 12];
803 w3_t[1] = s_msg_buf[off + 13];
804 w3_t[2] = (64 + msg_len) * 8;
807 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
809 const u32 r0 = digest[0];
810 const u32 r1 = digest[3];
811 const u32 r2 = digest[2];
812 const u32 r3 = digest[1];
818 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05300_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 ikepsk_t *ikepsk_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)
822 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05300_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 ikepsk_t *ikepsk_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)