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"
31 #include "include/rp_gpu.h"
35 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
36 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
40 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
41 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
45 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
46 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
49 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_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 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
55 const u32 lid = get_local_id (0);
61 const u32 gid = get_global_id (0);
63 if (gid >= gid_max) return;
67 pw_buf0[0] = pws[gid].i[ 0];
68 pw_buf0[1] = pws[gid].i[ 1];
69 pw_buf0[2] = pws[gid].i[ 2];
70 pw_buf0[3] = pws[gid].i[ 3];
74 pw_buf1[0] = pws[gid].i[ 4];
75 pw_buf1[1] = pws[gid].i[ 5];
76 pw_buf1[2] = pws[gid].i[ 6];
77 pw_buf1[3] = pws[gid].i[ 7];
79 const u32 pw_len = pws[gid].pw_len;
85 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
115 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
117 append_0x80_2 (w0, w1, out_len);
128 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
129 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
130 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
131 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
132 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
133 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
134 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
135 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
136 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
137 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
138 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
139 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
140 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
141 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
142 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
143 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
145 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
146 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
147 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
148 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
149 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
150 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
151 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
152 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
153 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
154 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
155 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
156 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
157 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
158 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
159 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
160 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
162 MD5_STEP (MD5_H1, a, b, c, d, w1[1], MD5C20, MD5S20);
163 MD5_STEP (MD5_H2, d, a, b, c, w2[0], MD5C21, MD5S21);
164 MD5_STEP (MD5_H1, c, d, a, b, w2[3], MD5C22, MD5S22);
165 MD5_STEP (MD5_H2, b, c, d, a, w3[2], MD5C23, MD5S23);
166 MD5_STEP (MD5_H1, a, b, c, d, w0[1], MD5C24, MD5S20);
167 MD5_STEP (MD5_H2, d, a, b, c, w1[0], MD5C25, MD5S21);
168 MD5_STEP (MD5_H1, c, d, a, b, w1[3], MD5C26, MD5S22);
169 MD5_STEP (MD5_H2, b, c, d, a, w2[2], MD5C27, MD5S23);
170 MD5_STEP (MD5_H1, a, b, c, d, w3[1], MD5C28, MD5S20);
171 MD5_STEP (MD5_H2, d, a, b, c, w0[0], MD5C29, MD5S21);
172 MD5_STEP (MD5_H1, c, d, a, b, w0[3], MD5C2a, MD5S22);
173 MD5_STEP (MD5_H2, b, c, d, a, w1[2], MD5C2b, MD5S23);
174 MD5_STEP (MD5_H1, a, b, c, d, w2[1], MD5C2c, MD5S20);
175 MD5_STEP (MD5_H2, d, a, b, c, w3[0], MD5C2d, MD5S21);
176 MD5_STEP (MD5_H1, c, d, a, b, w3[3], MD5C2e, MD5S22);
177 MD5_STEP (MD5_H2, b, c, d, a, w0[2], MD5C2f, MD5S23);
179 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
180 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
181 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
182 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
183 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
184 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
185 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
186 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
187 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
188 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
189 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
190 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
191 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
192 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
193 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
194 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
202 #include VECT_COMPARE_M
206 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_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 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)
210 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_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 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)
214 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_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 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
220 const u32 lid = get_local_id (0);
226 const u32 gid = get_global_id (0);
228 if (gid >= gid_max) return;
232 pw_buf0[0] = pws[gid].i[ 0];
233 pw_buf0[1] = pws[gid].i[ 1];
234 pw_buf0[2] = pws[gid].i[ 2];
235 pw_buf0[3] = pws[gid].i[ 3];
239 pw_buf1[0] = pws[gid].i[ 4];
240 pw_buf1[1] = pws[gid].i[ 5];
241 pw_buf1[2] = pws[gid].i[ 6];
242 pw_buf1[3] = pws[gid].i[ 7];
244 const u32 pw_len = pws[gid].pw_len;
250 const u32 search[4] =
252 digests_buf[digests_offset].digest_buf[DGST_R0],
253 digests_buf[digests_offset].digest_buf[DGST_R1],
254 digests_buf[digests_offset].digest_buf[DGST_R2],
255 digests_buf[digests_offset].digest_buf[DGST_R3]
262 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
292 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
294 append_0x80_2 (w0, w1, out_len);
305 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
306 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
307 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
308 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
309 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
310 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
311 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
312 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
313 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
314 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
315 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
316 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
317 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
318 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
319 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
320 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
322 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
323 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
324 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
325 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
326 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
327 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
328 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
329 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
330 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
331 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
332 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
333 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
334 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
335 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
336 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
337 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
339 MD5_STEP (MD5_H1, a, b, c, d, w1[1], MD5C20, MD5S20);
340 MD5_STEP (MD5_H2, d, a, b, c, w2[0], MD5C21, MD5S21);
341 MD5_STEP (MD5_H1, c, d, a, b, w2[3], MD5C22, MD5S22);
342 MD5_STEP (MD5_H2, b, c, d, a, w3[2], MD5C23, MD5S23);
343 MD5_STEP (MD5_H1, a, b, c, d, w0[1], MD5C24, MD5S20);
344 MD5_STEP (MD5_H2, d, a, b, c, w1[0], MD5C25, MD5S21);
345 MD5_STEP (MD5_H1, c, d, a, b, w1[3], MD5C26, MD5S22);
346 MD5_STEP (MD5_H2, b, c, d, a, w2[2], MD5C27, MD5S23);
347 MD5_STEP (MD5_H1, a, b, c, d, w3[1], MD5C28, MD5S20);
348 MD5_STEP (MD5_H2, d, a, b, c, w0[0], MD5C29, MD5S21);
349 MD5_STEP (MD5_H1, c, d, a, b, w0[3], MD5C2a, MD5S22);
350 MD5_STEP (MD5_H2, b, c, d, a, w1[2], MD5C2b, MD5S23);
351 MD5_STEP (MD5_H1, a, b, c, d, w2[1], MD5C2c, MD5S20);
352 MD5_STEP (MD5_H2, d, a, b, c, w3[0], MD5C2d, MD5S21);
353 MD5_STEP (MD5_H1, c, d, a, b, w3[3], MD5C2e, MD5S22);
354 MD5_STEP (MD5_H2, b, c, d, a, w0[2], MD5C2f, MD5S23);
356 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
357 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
358 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
359 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
360 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
361 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
362 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
363 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
364 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
365 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
366 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
367 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
368 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
370 bool q_cond = allx (search[0] != a);
372 if (q_cond) continue;
374 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
375 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
376 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
384 #include VECT_COMPARE_S
388 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_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 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)
392 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_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 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)