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 "OpenCL/types_ocl.c"
18 #include "OpenCL/common.c"
20 #define COMPARE_S "OpenCL/check_single_comp4.c"
21 #define COMPARE_M "OpenCL/check_multi_comp4.c"
23 __kernel void m00900_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 combs_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;
41 wordl0[0] = pws[gid].i[ 0];
42 wordl0[1] = pws[gid].i[ 1];
43 wordl0[2] = pws[gid].i[ 2];
44 wordl0[3] = pws[gid].i[ 3];
48 wordl1[0] = pws[gid].i[ 4];
49 wordl1[1] = pws[gid].i[ 5];
50 wordl1[2] = pws[gid].i[ 6];
51 wordl1[3] = pws[gid].i[ 7];
67 const u32 pw_l_len = pws[gid].pw_len;
69 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
71 append_0x80_2x4 (wordl0, wordl1, pw_l_len);
73 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
80 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
82 const u32 pw_r_len = combs_buf[il_pos].pw_len;
84 const u32 pw_len = pw_l_len + pw_r_len;
88 wordr0[0] = combs_buf[il_pos].i[0];
89 wordr0[1] = combs_buf[il_pos].i[1];
90 wordr0[2] = combs_buf[il_pos].i[2];
91 wordr0[3] = combs_buf[il_pos].i[3];
95 wordr1[0] = combs_buf[il_pos].i[4];
96 wordr1[1] = combs_buf[il_pos].i[5];
97 wordr1[2] = combs_buf[il_pos].i[6];
98 wordr1[3] = combs_buf[il_pos].i[7];
114 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
116 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
121 w0[0] = wordl0[0] | wordr0[0];
122 w0[1] = wordl0[1] | wordr0[1];
123 w0[2] = wordl0[2] | wordr0[2];
124 w0[3] = wordl0[3] | wordr0[3];
128 w1[0] = wordl1[0] | wordr1[0];
129 w1[1] = wordl1[1] | wordr1[1];
130 w1[2] = wordl1[2] | wordr1[2];
131 w1[3] = wordl1[3] | wordr1[3];
135 w2[0] = wordl2[0] | wordr2[0];
136 w2[1] = wordl2[1] | wordr2[1];
137 w2[2] = wordl2[2] | wordr2[2];
138 w2[3] = wordl2[3] | wordr2[3];
142 w3[0] = wordl3[0] | wordr3[0];
143 w3[1] = wordl3[1] | wordr3[1];
152 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
153 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
154 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
155 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
156 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
157 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
158 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
159 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
160 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
161 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
162 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
163 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
164 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
165 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
166 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
167 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
169 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
170 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
171 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
172 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
173 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
174 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
175 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
176 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
177 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
178 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
179 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
180 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
181 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
182 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
183 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
184 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
186 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
187 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
188 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
189 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
190 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
191 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
192 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
193 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
194 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
195 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
196 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
197 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
198 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
199 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
200 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
201 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
212 __kernel void m00900_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
216 __kernel void m00900_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
220 __kernel void m00900_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
226 const u32 lid = get_local_id (0);
232 const u32 gid = get_global_id (0);
234 if (gid >= gid_max) return;
238 wordl0[0] = pws[gid].i[ 0];
239 wordl0[1] = pws[gid].i[ 1];
240 wordl0[2] = pws[gid].i[ 2];
241 wordl0[3] = pws[gid].i[ 3];
245 wordl1[0] = pws[gid].i[ 4];
246 wordl1[1] = pws[gid].i[ 5];
247 wordl1[2] = pws[gid].i[ 6];
248 wordl1[3] = pws[gid].i[ 7];
264 const u32 pw_l_len = pws[gid].pw_len;
266 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
268 append_0x80_2x4 (wordl0, wordl1, pw_l_len);
270 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
277 const u32 search[4] =
279 digests_buf[digests_offset].digest_buf[DGST_R0],
280 digests_buf[digests_offset].digest_buf[DGST_R1],
281 digests_buf[digests_offset].digest_buf[DGST_R2],
282 digests_buf[digests_offset].digest_buf[DGST_R3]
289 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
291 const u32 pw_r_len = combs_buf[il_pos].pw_len;
293 const u32 pw_len = pw_l_len + pw_r_len;
297 wordr0[0] = combs_buf[il_pos].i[0];
298 wordr0[1] = combs_buf[il_pos].i[1];
299 wordr0[2] = combs_buf[il_pos].i[2];
300 wordr0[3] = combs_buf[il_pos].i[3];
304 wordr1[0] = combs_buf[il_pos].i[4];
305 wordr1[1] = combs_buf[il_pos].i[5];
306 wordr1[2] = combs_buf[il_pos].i[6];
307 wordr1[3] = combs_buf[il_pos].i[7];
323 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
325 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
330 w0[0] = wordl0[0] | wordr0[0];
331 w0[1] = wordl0[1] | wordr0[1];
332 w0[2] = wordl0[2] | wordr0[2];
333 w0[3] = wordl0[3] | wordr0[3];
337 w1[0] = wordl1[0] | wordr1[0];
338 w1[1] = wordl1[1] | wordr1[1];
339 w1[2] = wordl1[2] | wordr1[2];
340 w1[3] = wordl1[3] | wordr1[3];
344 w2[0] = wordl2[0] | wordr2[0];
345 w2[1] = wordl2[1] | wordr2[1];
346 w2[2] = wordl2[2] | wordr2[2];
347 w2[3] = wordl2[3] | wordr2[3];
351 w3[0] = wordl3[0] | wordr3[0];
352 w3[1] = wordl3[1] | wordr3[1];
361 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
362 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
363 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
364 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
365 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
366 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
367 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
368 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
369 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
370 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
371 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
372 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
373 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
374 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
375 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
376 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
378 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
379 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
380 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
381 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
382 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
383 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
384 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
385 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
386 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
387 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
388 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
389 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
390 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
391 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
392 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
393 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
395 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
396 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
397 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
398 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
399 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
400 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
401 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
402 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
403 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
404 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
405 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
406 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
407 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
408 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
409 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
410 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
421 __kernel void m00900_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
425 __kernel void m00900_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)