2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "include/constants.h"
11 #include "include/kernel_vendor.h"
18 #include "include/kernel_functions.c"
19 #include "OpenCL/types_ocl.c"
20 #include "OpenCL/common.c"
21 #include "OpenCL/simd.c"
23 static void md4_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
47 MD4_STEP (MD4_Fo, a, b, c, d, w0_t, MD4C00, MD4S00);
48 MD4_STEP (MD4_Fo, d, a, b, c, w1_t, MD4C00, MD4S01);
49 MD4_STEP (MD4_Fo, c, d, a, b, w2_t, MD4C00, MD4S02);
50 MD4_STEP (MD4_Fo, b, c, d, a, w3_t, MD4C00, MD4S03);
51 MD4_STEP (MD4_Fo, a, b, c, d, w4_t, MD4C00, MD4S00);
52 MD4_STEP (MD4_Fo, d, a, b, c, w5_t, MD4C00, MD4S01);
53 MD4_STEP (MD4_Fo, c, d, a, b, w6_t, MD4C00, MD4S02);
54 MD4_STEP (MD4_Fo, b, c, d, a, w7_t, MD4C00, MD4S03);
55 MD4_STEP (MD4_Fo, a, b, c, d, w8_t, MD4C00, MD4S00);
56 MD4_STEP (MD4_Fo, d, a, b, c, w9_t, MD4C00, MD4S01);
57 MD4_STEP (MD4_Fo, c, d, a, b, wa_t, MD4C00, MD4S02);
58 MD4_STEP (MD4_Fo, b, c, d, a, wb_t, MD4C00, MD4S03);
59 MD4_STEP (MD4_Fo, a, b, c, d, wc_t, MD4C00, MD4S00);
60 MD4_STEP (MD4_Fo, d, a, b, c, wd_t, MD4C00, MD4S01);
61 MD4_STEP (MD4_Fo, c, d, a, b, we_t, MD4C00, MD4S02);
62 MD4_STEP (MD4_Fo, b, c, d, a, wf_t, MD4C00, MD4S03);
64 MD4_STEP (MD4_Go, a, b, c, d, w0_t, MD4C01, MD4S10);
65 MD4_STEP (MD4_Go, d, a, b, c, w4_t, MD4C01, MD4S11);
66 MD4_STEP (MD4_Go, c, d, a, b, w8_t, MD4C01, MD4S12);
67 MD4_STEP (MD4_Go, b, c, d, a, wc_t, MD4C01, MD4S13);
68 MD4_STEP (MD4_Go, a, b, c, d, w1_t, MD4C01, MD4S10);
69 MD4_STEP (MD4_Go, d, a, b, c, w5_t, MD4C01, MD4S11);
70 MD4_STEP (MD4_Go, c, d, a, b, w9_t, MD4C01, MD4S12);
71 MD4_STEP (MD4_Go, b, c, d, a, wd_t, MD4C01, MD4S13);
72 MD4_STEP (MD4_Go, a, b, c, d, w2_t, MD4C01, MD4S10);
73 MD4_STEP (MD4_Go, d, a, b, c, w6_t, MD4C01, MD4S11);
74 MD4_STEP (MD4_Go, c, d, a, b, wa_t, MD4C01, MD4S12);
75 MD4_STEP (MD4_Go, b, c, d, a, we_t, MD4C01, MD4S13);
76 MD4_STEP (MD4_Go, a, b, c, d, w3_t, MD4C01, MD4S10);
77 MD4_STEP (MD4_Go, d, a, b, c, w7_t, MD4C01, MD4S11);
78 MD4_STEP (MD4_Go, c, d, a, b, wb_t, MD4C01, MD4S12);
79 MD4_STEP (MD4_Go, b, c, d, a, wf_t, MD4C01, MD4S13);
81 MD4_STEP (MD4_H , a, b, c, d, w0_t, MD4C02, MD4S20);
82 MD4_STEP (MD4_H , d, a, b, c, w8_t, MD4C02, MD4S21);
83 MD4_STEP (MD4_H , c, d, a, b, w4_t, MD4C02, MD4S22);
84 MD4_STEP (MD4_H , b, c, d, a, wc_t, MD4C02, MD4S23);
85 MD4_STEP (MD4_H , a, b, c, d, w2_t, MD4C02, MD4S20);
86 MD4_STEP (MD4_H , d, a, b, c, wa_t, MD4C02, MD4S21);
87 MD4_STEP (MD4_H , c, d, a, b, w6_t, MD4C02, MD4S22);
88 MD4_STEP (MD4_H , b, c, d, a, we_t, MD4C02, MD4S23);
89 MD4_STEP (MD4_H , a, b, c, d, w1_t, MD4C02, MD4S20);
90 MD4_STEP (MD4_H , d, a, b, c, w9_t, MD4C02, MD4S21);
91 MD4_STEP (MD4_H , c, d, a, b, w5_t, MD4C02, MD4S22);
92 MD4_STEP (MD4_H , b, c, d, a, wd_t, MD4C02, MD4S23);
93 MD4_STEP (MD4_H , a, b, c, d, w3_t, MD4C02, MD4S20);
94 MD4_STEP (MD4_H , d, a, b, c, wb_t, MD4C02, MD4S21);
95 MD4_STEP (MD4_H , c, d, a, b, w7_t, MD4C02, MD4S22);
96 MD4_STEP (MD4_H , b, c, d, a, wf_t, MD4C02, MD4S23);
104 static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
128 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
129 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
130 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
131 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
132 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
133 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
134 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
135 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
136 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
137 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
138 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
139 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
140 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
141 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
142 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
143 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
145 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
146 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
147 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
148 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
149 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
150 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
151 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
152 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
153 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
154 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
155 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
156 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
157 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
158 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
159 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
160 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
162 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
163 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
164 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
165 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
166 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
167 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
168 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
169 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
170 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
171 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
172 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
173 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
174 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
175 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
176 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
177 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
179 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
180 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
181 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
182 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
183 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
184 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
185 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
186 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
187 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
188 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
189 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
190 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
191 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
192 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
193 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
194 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
202 static void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
204 w0[0] = w0[0] ^ 0x36363636;
205 w0[1] = w0[1] ^ 0x36363636;
206 w0[2] = w0[2] ^ 0x36363636;
207 w0[3] = w0[3] ^ 0x36363636;
208 w1[0] = w1[0] ^ 0x36363636;
209 w1[1] = w1[1] ^ 0x36363636;
210 w1[2] = w1[2] ^ 0x36363636;
211 w1[3] = w1[3] ^ 0x36363636;
212 w2[0] = w2[0] ^ 0x36363636;
213 w2[1] = w2[1] ^ 0x36363636;
214 w2[2] = w2[2] ^ 0x36363636;
215 w2[3] = w2[3] ^ 0x36363636;
216 w3[0] = w3[0] ^ 0x36363636;
217 w3[1] = w3[1] ^ 0x36363636;
218 w3[2] = w3[2] ^ 0x36363636;
219 w3[3] = w3[3] ^ 0x36363636;
226 md5_transform (w0, w1, w2, w3, ipad);
228 w0[0] = w0[0] ^ 0x6a6a6a6a;
229 w0[1] = w0[1] ^ 0x6a6a6a6a;
230 w0[2] = w0[2] ^ 0x6a6a6a6a;
231 w0[3] = w0[3] ^ 0x6a6a6a6a;
232 w1[0] = w1[0] ^ 0x6a6a6a6a;
233 w1[1] = w1[1] ^ 0x6a6a6a6a;
234 w1[2] = w1[2] ^ 0x6a6a6a6a;
235 w1[3] = w1[3] ^ 0x6a6a6a6a;
236 w2[0] = w2[0] ^ 0x6a6a6a6a;
237 w2[1] = w2[1] ^ 0x6a6a6a6a;
238 w2[2] = w2[2] ^ 0x6a6a6a6a;
239 w2[3] = w2[3] ^ 0x6a6a6a6a;
240 w3[0] = w3[0] ^ 0x6a6a6a6a;
241 w3[1] = w3[1] ^ 0x6a6a6a6a;
242 w3[2] = w3[2] ^ 0x6a6a6a6a;
243 w3[3] = w3[3] ^ 0x6a6a6a6a;
250 md5_transform (w0, w1, w2, w3, opad);
253 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])
260 md5_transform (w0, w1, w2, w3, digest);
276 w3[2] = (64 + 16) * 8;
284 md5_transform (w0, w1, w2, w3, digest);
287 __kernel void m05600_m04 (__global pw_t *pws, __global kernel_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 netntlm_t *netntlm_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)
293 const u32 gid = get_global_id (0);
294 const u32 lid = get_local_id (0);
295 const u32 lsz = get_local_size (0);
301 __local u32 s_userdomain_buf[64];
303 for (u32 i = lid; i < 64; i += lsz)
305 s_userdomain_buf[i] = netntlm_bufs[salt_pos].userdomain_buf[i];
308 __local u32 s_chall_buf[256];
310 for (u32 i = lid; i < 256; i += lsz)
312 s_chall_buf[i] = netntlm_bufs[salt_pos].chall_buf[i];
315 barrier (CLK_LOCAL_MEM_FENCE);
317 if (gid >= gid_max) return;
319 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
320 + netntlm_bufs[salt_pos].domain_len;
322 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
323 + netntlm_bufs[salt_pos].clichall_len;
332 pws0[0] = pws[gid].i[0];
333 pws0[1] = pws[gid].i[1];
334 pws0[2] = pws[gid].i[2];
335 pws0[3] = pws[gid].i[3];
336 pws1[0] = pws[gid].i[4];
337 pws1[1] = pws[gid].i[5];
338 pws1[2] = pws[gid].i[6];
339 pws1[3] = pws[gid].i[7];
341 const u32 pw_l_len = pws[gid].pw_len;
347 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos += VECT_SIZE)
349 const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
351 const u32x pw_len = pw_l_len + pw_r_len;
353 u32x wordr0[4] = { 0 };
354 u32x wordr1[4] = { 0 };
355 u32x wordr2[4] = { 0 };
356 u32x wordr3[4] = { 0 };
358 wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
359 wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
360 wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
361 wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
362 wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
363 wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
364 wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
365 wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
367 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
369 switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
374 w0[0] = wordl0[0] | wordr0[0];
375 w0[1] = wordl0[1] | wordr0[1];
376 w0[2] = wordl0[2] | wordr0[2];
377 w0[3] = wordl0[3] | wordr0[3];
381 w1[0] = wordl1[0] | wordr1[0];
382 w1[1] = wordl1[1] | wordr1[1];
383 w1[2] = wordl1[2] | wordr1[2];
384 w1[3] = wordl1[3] | wordr1[3];
405 make_unicode (w0, w0_t, w1_t);
406 make_unicode (w1, w2_t, w3_t);
408 w3_t[2] = pw_len * 8 * 2;
417 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
444 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
449 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
451 w0_t[0] = s_userdomain_buf[off + 0];
452 w0_t[1] = s_userdomain_buf[off + 1];
453 w0_t[2] = s_userdomain_buf[off + 2];
454 w0_t[3] = s_userdomain_buf[off + 3];
455 w1_t[0] = s_userdomain_buf[off + 4];
456 w1_t[1] = s_userdomain_buf[off + 5];
457 w1_t[2] = s_userdomain_buf[off + 6];
458 w1_t[3] = s_userdomain_buf[off + 7];
459 w2_t[0] = s_userdomain_buf[off + 8];
460 w2_t[1] = s_userdomain_buf[off + 9];
461 w2_t[2] = s_userdomain_buf[off + 10];
462 w2_t[3] = s_userdomain_buf[off + 11];
463 w3_t[0] = s_userdomain_buf[off + 12];
464 w3_t[1] = s_userdomain_buf[off + 13];
465 w3_t[2] = s_userdomain_buf[off + 14];
466 w3_t[3] = s_userdomain_buf[off + 15];
468 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
471 w0_t[0] = s_userdomain_buf[off + 0];
472 w0_t[1] = s_userdomain_buf[off + 1];
473 w0_t[2] = s_userdomain_buf[off + 2];
474 w0_t[3] = s_userdomain_buf[off + 3];
475 w1_t[0] = s_userdomain_buf[off + 4];
476 w1_t[1] = s_userdomain_buf[off + 5];
477 w1_t[2] = s_userdomain_buf[off + 6];
478 w1_t[3] = s_userdomain_buf[off + 7];
479 w2_t[0] = s_userdomain_buf[off + 8];
480 w2_t[1] = s_userdomain_buf[off + 9];
481 w2_t[2] = s_userdomain_buf[off + 10];
482 w2_t[3] = s_userdomain_buf[off + 11];
483 w3_t[0] = s_userdomain_buf[off + 12];
484 w3_t[1] = s_userdomain_buf[off + 13];
485 w3_t[2] = (64 + userdomain_len) * 8;
488 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
512 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
514 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
516 w0_t[0] = s_chall_buf[off + 0];
517 w0_t[1] = s_chall_buf[off + 1];
518 w0_t[2] = s_chall_buf[off + 2];
519 w0_t[3] = s_chall_buf[off + 3];
520 w1_t[0] = s_chall_buf[off + 4];
521 w1_t[1] = s_chall_buf[off + 5];
522 w1_t[2] = s_chall_buf[off + 6];
523 w1_t[3] = s_chall_buf[off + 7];
524 w2_t[0] = s_chall_buf[off + 8];
525 w2_t[1] = s_chall_buf[off + 9];
526 w2_t[2] = s_chall_buf[off + 10];
527 w2_t[3] = s_chall_buf[off + 11];
528 w3_t[0] = s_chall_buf[off + 12];
529 w3_t[1] = s_chall_buf[off + 13];
530 w3_t[2] = s_chall_buf[off + 14];
531 w3_t[3] = s_chall_buf[off + 15];
533 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
536 w0_t[0] = s_chall_buf[off + 0];
537 w0_t[1] = s_chall_buf[off + 1];
538 w0_t[2] = s_chall_buf[off + 2];
539 w0_t[3] = s_chall_buf[off + 3];
540 w1_t[0] = s_chall_buf[off + 4];
541 w1_t[1] = s_chall_buf[off + 5];
542 w1_t[2] = s_chall_buf[off + 6];
543 w1_t[3] = s_chall_buf[off + 7];
544 w2_t[0] = s_chall_buf[off + 8];
545 w2_t[1] = s_chall_buf[off + 9];
546 w2_t[2] = s_chall_buf[off + 10];
547 w2_t[3] = s_chall_buf[off + 11];
548 w3_t[0] = s_chall_buf[off + 12];
549 w3_t[1] = s_chall_buf[off + 13];
550 w3_t[2] = (64 + chall_len) * 8;
553 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
555 COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
559 __kernel void m05600_m08 (__global pw_t *pws, __global kernel_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 netntlm_t *netntlm_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)
563 __kernel void m05600_m16 (__global pw_t *pws, __global kernel_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 netntlm_t *netntlm_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)
567 __kernel void m05600_s04 (__global pw_t *pws, __global kernel_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 netntlm_t *netntlm_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)
573 const u32 gid = get_global_id (0);
574 const u32 lid = get_local_id (0);
575 const u32 lsz = get_local_size (0);
581 __local u32 s_userdomain_buf[64];
583 for (u32 i = lid; i < 64; i += lsz)
585 s_userdomain_buf[i] = netntlm_bufs[salt_pos].userdomain_buf[i];
588 __local u32 s_chall_buf[256];
590 for (u32 i = lid; i < 256; i += lsz)
592 s_chall_buf[i] = netntlm_bufs[salt_pos].chall_buf[i];
595 barrier (CLK_LOCAL_MEM_FENCE);
597 if (gid >= gid_max) return;
599 const u32 userdomain_len = netntlm_bufs[salt_pos].user_len
600 + netntlm_bufs[salt_pos].domain_len;
602 const u32 chall_len = netntlm_bufs[salt_pos].srvchall_len
603 + netntlm_bufs[salt_pos].clichall_len;
612 pws0[0] = pws[gid].i[0];
613 pws0[1] = pws[gid].i[1];
614 pws0[2] = pws[gid].i[2];
615 pws0[3] = pws[gid].i[3];
616 pws1[0] = pws[gid].i[4];
617 pws1[1] = pws[gid].i[5];
618 pws1[2] = pws[gid].i[6];
619 pws1[3] = pws[gid].i[7];
621 const u32 pw_l_len = pws[gid].pw_len;
627 const u32 search[4] =
629 digests_buf[digests_offset].digest_buf[DGST_R0],
630 digests_buf[digests_offset].digest_buf[DGST_R1],
631 digests_buf[digests_offset].digest_buf[DGST_R2],
632 digests_buf[digests_offset].digest_buf[DGST_R3]
639 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos += VECT_SIZE)
641 const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
643 const u32x pw_len = pw_l_len + pw_r_len;
645 u32x wordr0[4] = { 0 };
646 u32x wordr1[4] = { 0 };
647 u32x wordr2[4] = { 0 };
648 u32x wordr3[4] = { 0 };
650 wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
651 wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
652 wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
653 wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
654 wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
655 wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
656 wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
657 wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
659 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
661 switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
666 w0[0] = wordl0[0] | wordr0[0];
667 w0[1] = wordl0[1] | wordr0[1];
668 w0[2] = wordl0[2] | wordr0[2];
669 w0[3] = wordl0[3] | wordr0[3];
673 w1[0] = wordl1[0] | wordr1[0];
674 w1[1] = wordl1[1] | wordr1[1];
675 w1[2] = wordl1[2] | wordr1[2];
676 w1[3] = wordl1[3] | wordr1[3];
697 make_unicode (w0, w0_t, w1_t);
698 make_unicode (w1, w2_t, w3_t);
700 w3_t[2] = pw_len * 8 * 2;
709 md4_transform (w0_t, w1_t, w2_t, w3_t, digest);
736 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
741 for (left = userdomain_len, off = 0; left >= 56; left -= 64, off += 16)
743 w0_t[0] = s_userdomain_buf[off + 0];
744 w0_t[1] = s_userdomain_buf[off + 1];
745 w0_t[2] = s_userdomain_buf[off + 2];
746 w0_t[3] = s_userdomain_buf[off + 3];
747 w1_t[0] = s_userdomain_buf[off + 4];
748 w1_t[1] = s_userdomain_buf[off + 5];
749 w1_t[2] = s_userdomain_buf[off + 6];
750 w1_t[3] = s_userdomain_buf[off + 7];
751 w2_t[0] = s_userdomain_buf[off + 8];
752 w2_t[1] = s_userdomain_buf[off + 9];
753 w2_t[2] = s_userdomain_buf[off + 10];
754 w2_t[3] = s_userdomain_buf[off + 11];
755 w3_t[0] = s_userdomain_buf[off + 12];
756 w3_t[1] = s_userdomain_buf[off + 13];
757 w3_t[2] = s_userdomain_buf[off + 14];
758 w3_t[3] = s_userdomain_buf[off + 15];
760 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
763 w0_t[0] = s_userdomain_buf[off + 0];
764 w0_t[1] = s_userdomain_buf[off + 1];
765 w0_t[2] = s_userdomain_buf[off + 2];
766 w0_t[3] = s_userdomain_buf[off + 3];
767 w1_t[0] = s_userdomain_buf[off + 4];
768 w1_t[1] = s_userdomain_buf[off + 5];
769 w1_t[2] = s_userdomain_buf[off + 6];
770 w1_t[3] = s_userdomain_buf[off + 7];
771 w2_t[0] = s_userdomain_buf[off + 8];
772 w2_t[1] = s_userdomain_buf[off + 9];
773 w2_t[2] = s_userdomain_buf[off + 10];
774 w2_t[3] = s_userdomain_buf[off + 11];
775 w3_t[0] = s_userdomain_buf[off + 12];
776 w3_t[1] = s_userdomain_buf[off + 13];
777 w3_t[2] = (64 + userdomain_len) * 8;
780 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
804 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
806 for (left = chall_len, off = 0; left >= 56; left -= 64, off += 16)
808 w0_t[0] = s_chall_buf[off + 0];
809 w0_t[1] = s_chall_buf[off + 1];
810 w0_t[2] = s_chall_buf[off + 2];
811 w0_t[3] = s_chall_buf[off + 3];
812 w1_t[0] = s_chall_buf[off + 4];
813 w1_t[1] = s_chall_buf[off + 5];
814 w1_t[2] = s_chall_buf[off + 6];
815 w1_t[3] = s_chall_buf[off + 7];
816 w2_t[0] = s_chall_buf[off + 8];
817 w2_t[1] = s_chall_buf[off + 9];
818 w2_t[2] = s_chall_buf[off + 10];
819 w2_t[3] = s_chall_buf[off + 11];
820 w3_t[0] = s_chall_buf[off + 12];
821 w3_t[1] = s_chall_buf[off + 13];
822 w3_t[2] = s_chall_buf[off + 14];
823 w3_t[3] = s_chall_buf[off + 15];
825 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
828 w0_t[0] = s_chall_buf[off + 0];
829 w0_t[1] = s_chall_buf[off + 1];
830 w0_t[2] = s_chall_buf[off + 2];
831 w0_t[3] = s_chall_buf[off + 3];
832 w1_t[0] = s_chall_buf[off + 4];
833 w1_t[1] = s_chall_buf[off + 5];
834 w1_t[2] = s_chall_buf[off + 6];
835 w1_t[3] = s_chall_buf[off + 7];
836 w2_t[0] = s_chall_buf[off + 8];
837 w2_t[1] = s_chall_buf[off + 9];
838 w2_t[2] = s_chall_buf[off + 10];
839 w2_t[3] = s_chall_buf[off + 11];
840 w3_t[0] = s_chall_buf[off + 12];
841 w3_t[1] = s_chall_buf[off + 13];
842 w3_t[2] = (64 + chall_len) * 8;
845 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
847 COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
851 __kernel void m05600_s08 (__global pw_t *pws, __global kernel_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 netntlm_t *netntlm_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)
855 __kernel void m05600_s16 (__global pw_t *pws, __global kernel_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 netntlm_t *netntlm_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)