2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "inc_hash_constants.h"
11 #include "inc_vendor.cl"
18 #include "inc_hash_functions.cl"
19 #include "inc_types.cl"
20 #include "inc_common.cl"
21 #include "inc_simd.cl"
23 __kernel void m01000_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 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
29 const u32 lid = get_local_id (0);
35 const u32 gid = get_global_id (0);
37 if (gid >= gid_max) return;
42 pw_buf0[0] = pws[gid].i[0];
43 pw_buf0[1] = pws[gid].i[1];
44 pw_buf0[2] = pws[gid].i[2];
45 pw_buf0[3] = pws[gid].i[3];
46 pw_buf1[0] = pws[gid].i[4];
47 pw_buf1[1] = pws[gid].i[5];
48 pw_buf1[2] = pws[gid].i[6];
49 pw_buf1[3] = pws[gid].i[7];
51 const u32 pw_l_len = pws[gid].pw_len;
57 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
59 const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
61 const u32x pw_len = pw_l_len + pw_r_len;
64 * concat password candidate
67 u32x wordl0[4] = { 0 };
68 u32x wordl1[4] = { 0 };
69 u32x wordl2[4] = { 0 };
70 u32x wordl3[4] = { 0 };
72 wordl0[0] = pw_buf0[0];
73 wordl0[1] = pw_buf0[1];
74 wordl0[2] = pw_buf0[2];
75 wordl0[3] = pw_buf0[3];
76 wordl1[0] = pw_buf1[0];
77 wordl1[1] = pw_buf1[1];
78 wordl1[2] = pw_buf1[2];
79 wordl1[3] = pw_buf1[3];
81 u32x wordr0[4] = { 0 };
82 u32x wordr1[4] = { 0 };
83 u32x wordr2[4] = { 0 };
84 u32x wordr3[4] = { 0 };
86 wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
87 wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
88 wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
89 wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
90 wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
91 wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
92 wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
93 wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
95 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
97 switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
101 switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
109 w0[0] = wordl0[0] | wordr0[0];
110 w0[1] = wordl0[1] | wordr0[1];
111 w0[2] = wordl0[2] | wordr0[2];
112 w0[3] = wordl0[3] | wordr0[3];
113 w1[0] = wordl1[0] | wordr1[0];
114 w1[1] = wordl1[1] | wordr1[1];
115 w1[2] = wordl1[2] | wordr1[2];
116 w1[3] = wordl1[3] | wordr1[3];
117 w2[0] = wordl2[0] | wordr2[0];
118 w2[1] = wordl2[1] | wordr2[1];
119 w2[2] = wordl2[2] | wordr2[2];
120 w2[3] = wordl2[3] | wordr2[3];
121 w3[0] = wordl3[0] | wordr3[0];
122 w3[1] = wordl3[1] | wordr3[1];
123 w3[2] = wordl3[2] | wordr3[2];
124 w3[3] = wordl3[3] | wordr3[3];
126 make_unicode (w1, w2, w3);
127 make_unicode (w0, w0, w1);
129 w3[2] = pw_len * 8 * 2;
141 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
142 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
143 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
144 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
145 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
146 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
147 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
148 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
149 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
150 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
151 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
152 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
153 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
154 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
155 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
156 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
158 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
159 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
160 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
161 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
162 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
163 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
164 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
165 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
166 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
167 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
168 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
169 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
170 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
171 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
172 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
173 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
175 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
176 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
177 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
178 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
179 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
180 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
181 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
182 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
183 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
184 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
185 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
186 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
187 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
188 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
189 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
190 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
192 COMPARE_M_SIMD (a, d, c, b);
196 __kernel void m01000_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 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
200 __kernel void m01000_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 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
204 __kernel void m01000_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 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
210 const u32 lid = get_local_id (0);
216 const u32 gid = get_global_id (0);
218 if (gid >= gid_max) return;
223 pw_buf0[0] = pws[gid].i[0];
224 pw_buf0[1] = pws[gid].i[1];
225 pw_buf0[2] = pws[gid].i[2];
226 pw_buf0[3] = pws[gid].i[3];
227 pw_buf1[0] = pws[gid].i[4];
228 pw_buf1[1] = pws[gid].i[5];
229 pw_buf1[2] = pws[gid].i[6];
230 pw_buf1[3] = pws[gid].i[7];
232 const u32 pw_l_len = pws[gid].pw_len;
238 const u32 search[4] =
240 digests_buf[digests_offset].digest_buf[DGST_R0],
241 digests_buf[digests_offset].digest_buf[DGST_R1],
242 digests_buf[digests_offset].digest_buf[DGST_R2],
243 digests_buf[digests_offset].digest_buf[DGST_R3]
250 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
252 const u32x pw_r_len = pwlenx_create_combt (combs_buf, il_pos);
254 const u32x pw_len = pw_l_len + pw_r_len;
257 * concat password candidate
260 u32x wordl0[4] = { 0 };
261 u32x wordl1[4] = { 0 };
262 u32x wordl2[4] = { 0 };
263 u32x wordl3[4] = { 0 };
265 wordl0[0] = pw_buf0[0];
266 wordl0[1] = pw_buf0[1];
267 wordl0[2] = pw_buf0[2];
268 wordl0[3] = pw_buf0[3];
269 wordl1[0] = pw_buf1[0];
270 wordl1[1] = pw_buf1[1];
271 wordl1[2] = pw_buf1[2];
272 wordl1[3] = pw_buf1[3];
274 u32x wordr0[4] = { 0 };
275 u32x wordr1[4] = { 0 };
276 u32x wordr2[4] = { 0 };
277 u32x wordr3[4] = { 0 };
279 wordr0[0] = ix_create_combt (combs_buf, il_pos, 0);
280 wordr0[1] = ix_create_combt (combs_buf, il_pos, 1);
281 wordr0[2] = ix_create_combt (combs_buf, il_pos, 2);
282 wordr0[3] = ix_create_combt (combs_buf, il_pos, 3);
283 wordr1[0] = ix_create_combt (combs_buf, il_pos, 4);
284 wordr1[1] = ix_create_combt (combs_buf, il_pos, 5);
285 wordr1[2] = ix_create_combt (combs_buf, il_pos, 6);
286 wordr1[3] = ix_create_combt (combs_buf, il_pos, 7);
288 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
290 switch_buffer_by_offset_le_VV (wordr0, wordr1, wordr2, wordr3, pw_l_len);
294 switch_buffer_by_offset_le_VV (wordl0, wordl1, wordl2, wordl3, pw_r_len);
302 w0[0] = wordl0[0] | wordr0[0];
303 w0[1] = wordl0[1] | wordr0[1];
304 w0[2] = wordl0[2] | wordr0[2];
305 w0[3] = wordl0[3] | wordr0[3];
306 w1[0] = wordl1[0] | wordr1[0];
307 w1[1] = wordl1[1] | wordr1[1];
308 w1[2] = wordl1[2] | wordr1[2];
309 w1[3] = wordl1[3] | wordr1[3];
310 w2[0] = wordl2[0] | wordr2[0];
311 w2[1] = wordl2[1] | wordr2[1];
312 w2[2] = wordl2[2] | wordr2[2];
313 w2[3] = wordl2[3] | wordr2[3];
314 w3[0] = wordl3[0] | wordr3[0];
315 w3[1] = wordl3[1] | wordr3[1];
316 w3[2] = wordl3[2] | wordr3[2];
317 w3[3] = wordl3[3] | wordr3[3];
319 make_unicode (w1, w2, w3);
320 make_unicode (w0, w0, w1);
322 w3[2] = pw_len * 8 * 2;
334 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
335 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
336 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
337 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
338 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
339 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
340 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
341 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
342 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
343 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
344 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
345 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
346 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
347 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
348 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
349 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
351 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
352 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
353 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
354 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
355 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
356 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
357 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
358 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
359 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
360 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
361 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
362 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
363 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
364 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
365 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
366 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
368 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
369 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
370 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
371 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
372 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
373 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
374 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
375 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
376 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
377 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
378 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
379 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
380 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
382 if (MATCHES_NONE_VS (a, search[0])) continue;
384 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
385 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
386 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
388 COMPARE_S_SIMD (a, d, c, b);
392 __kernel void m01000_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 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
396 __kernel void m01000_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 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)