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 "include/rp_kernel.h"
22 #include "OpenCL/rp.c"
23 #include "OpenCL/simd.c"
25 __kernel void m04800_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)
31 const u32 gid = get_global_id (0);
32 const u32 lid = get_local_id (0);
34 if (gid >= gid_max) return;
43 pw_buf0[0] = pws[gid].i[0];
44 pw_buf0[1] = pws[gid].i[1];
45 pw_buf0[2] = pws[gid].i[2];
46 pw_buf0[3] = pws[gid].i[3];
47 pw_buf1[0] = pws[gid].i[4];
48 pw_buf1[1] = pws[gid].i[5];
49 pw_buf1[2] = pws[gid].i[6];
50 pw_buf1[3] = pws[gid].i[7];
52 const u32 pw_len = pws[gid].pw_len;
60 salt_buf[0] = salt_bufs[salt_pos].salt_buf[0];
61 salt_buf[1] = salt_bufs[salt_pos].salt_buf[1];
62 salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
63 salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
64 salt_buf[4] = salt_bufs[salt_pos].salt_buf[4];
66 const u32 salt_len = salt_bufs[salt_pos].salt_len;
72 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
79 const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
81 const u32x salt_out_len = salt_len + out_len;
109 switch_buffer_by_offset_le_VV (s0, s1, s2, s3, out_len);
132 switch_buffer_by_offset_le (w0, w1, w2, w3, 1);
134 w0[0] |= salt_buf[4];
136 w3[2] = salt_out_len * 8;
148 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
149 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
150 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
151 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
152 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
153 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
154 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
155 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
156 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
157 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
158 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
159 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
160 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
161 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
162 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
163 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
165 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
166 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
167 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
168 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
169 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
170 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
171 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
172 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
173 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
174 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
175 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
176 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
177 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
178 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
179 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
180 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
182 MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
183 MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
184 MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
185 MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
186 MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
187 MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
188 MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
189 MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
190 MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
191 MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
192 MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
193 MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
194 MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
195 MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
196 MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
197 MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);
199 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
200 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
201 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
202 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
203 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
204 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
205 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
206 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
207 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
208 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
209 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
210 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
211 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
212 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
213 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
214 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
216 COMPARE_M_SIMD (a, d, c, b);
220 __kernel void m04800_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)
224 __kernel void m04800_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)
228 __kernel void m04800_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)
234 const u32 gid = get_global_id (0);
235 const u32 lid = get_local_id (0);
237 if (gid >= gid_max) return;
246 pw_buf0[0] = pws[gid].i[0];
247 pw_buf0[1] = pws[gid].i[1];
248 pw_buf0[2] = pws[gid].i[2];
249 pw_buf0[3] = pws[gid].i[3];
250 pw_buf1[0] = pws[gid].i[4];
251 pw_buf1[1] = pws[gid].i[5];
252 pw_buf1[2] = pws[gid].i[6];
253 pw_buf1[3] = pws[gid].i[7];
255 const u32 pw_len = pws[gid].pw_len;
263 salt_buf[0] = salt_bufs[salt_pos].salt_buf[0];
264 salt_buf[1] = salt_bufs[salt_pos].salt_buf[1];
265 salt_buf[2] = salt_bufs[salt_pos].salt_buf[2];
266 salt_buf[3] = salt_bufs[salt_pos].salt_buf[3];
267 salt_buf[4] = salt_bufs[salt_pos].salt_buf[4];
269 const u32 salt_len = salt_bufs[salt_pos].salt_len;
275 const u32 search[4] =
277 digests_buf[digests_offset].digest_buf[DGST_R0],
278 digests_buf[digests_offset].digest_buf[DGST_R1],
279 digests_buf[digests_offset].digest_buf[DGST_R2],
280 digests_buf[digests_offset].digest_buf[DGST_R3]
287 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
294 const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
296 const u32x salt_out_len = salt_len + out_len;
324 switch_buffer_by_offset_le_VV (s0, s1, s2, s3, out_len);
347 switch_buffer_by_offset_le (w0, w1, w2, w3, 1);
349 w0[0] |= salt_buf[4];
351 w3[2] = salt_out_len * 8;
363 MD5_STEP (MD5_Fo, a, b, c, d, w0[0], MD5C00, MD5S00);
364 MD5_STEP (MD5_Fo, d, a, b, c, w0[1], MD5C01, MD5S01);
365 MD5_STEP (MD5_Fo, c, d, a, b, w0[2], MD5C02, MD5S02);
366 MD5_STEP (MD5_Fo, b, c, d, a, w0[3], MD5C03, MD5S03);
367 MD5_STEP (MD5_Fo, a, b, c, d, w1[0], MD5C04, MD5S00);
368 MD5_STEP (MD5_Fo, d, a, b, c, w1[1], MD5C05, MD5S01);
369 MD5_STEP (MD5_Fo, c, d, a, b, w1[2], MD5C06, MD5S02);
370 MD5_STEP (MD5_Fo, b, c, d, a, w1[3], MD5C07, MD5S03);
371 MD5_STEP (MD5_Fo, a, b, c, d, w2[0], MD5C08, MD5S00);
372 MD5_STEP (MD5_Fo, d, a, b, c, w2[1], MD5C09, MD5S01);
373 MD5_STEP (MD5_Fo, c, d, a, b, w2[2], MD5C0a, MD5S02);
374 MD5_STEP (MD5_Fo, b, c, d, a, w2[3], MD5C0b, MD5S03);
375 MD5_STEP (MD5_Fo, a, b, c, d, w3[0], MD5C0c, MD5S00);
376 MD5_STEP (MD5_Fo, d, a, b, c, w3[1], MD5C0d, MD5S01);
377 MD5_STEP (MD5_Fo, c, d, a, b, w3[2], MD5C0e, MD5S02);
378 MD5_STEP (MD5_Fo, b, c, d, a, w3[3], MD5C0f, MD5S03);
380 MD5_STEP (MD5_Go, a, b, c, d, w0[1], MD5C10, MD5S10);
381 MD5_STEP (MD5_Go, d, a, b, c, w1[2], MD5C11, MD5S11);
382 MD5_STEP (MD5_Go, c, d, a, b, w2[3], MD5C12, MD5S12);
383 MD5_STEP (MD5_Go, b, c, d, a, w0[0], MD5C13, MD5S13);
384 MD5_STEP (MD5_Go, a, b, c, d, w1[1], MD5C14, MD5S10);
385 MD5_STEP (MD5_Go, d, a, b, c, w2[2], MD5C15, MD5S11);
386 MD5_STEP (MD5_Go, c, d, a, b, w3[3], MD5C16, MD5S12);
387 MD5_STEP (MD5_Go, b, c, d, a, w1[0], MD5C17, MD5S13);
388 MD5_STEP (MD5_Go, a, b, c, d, w2[1], MD5C18, MD5S10);
389 MD5_STEP (MD5_Go, d, a, b, c, w3[2], MD5C19, MD5S11);
390 MD5_STEP (MD5_Go, c, d, a, b, w0[3], MD5C1a, MD5S12);
391 MD5_STEP (MD5_Go, b, c, d, a, w2[0], MD5C1b, MD5S13);
392 MD5_STEP (MD5_Go, a, b, c, d, w3[1], MD5C1c, MD5S10);
393 MD5_STEP (MD5_Go, d, a, b, c, w0[2], MD5C1d, MD5S11);
394 MD5_STEP (MD5_Go, c, d, a, b, w1[3], MD5C1e, MD5S12);
395 MD5_STEP (MD5_Go, b, c, d, a, w3[0], MD5C1f, MD5S13);
397 MD5_STEP (MD5_H , a, b, c, d, w1[1], MD5C20, MD5S20);
398 MD5_STEP (MD5_H , d, a, b, c, w2[0], MD5C21, MD5S21);
399 MD5_STEP (MD5_H , c, d, a, b, w2[3], MD5C22, MD5S22);
400 MD5_STEP (MD5_H , b, c, d, a, w3[2], MD5C23, MD5S23);
401 MD5_STEP (MD5_H , a, b, c, d, w0[1], MD5C24, MD5S20);
402 MD5_STEP (MD5_H , d, a, b, c, w1[0], MD5C25, MD5S21);
403 MD5_STEP (MD5_H , c, d, a, b, w1[3], MD5C26, MD5S22);
404 MD5_STEP (MD5_H , b, c, d, a, w2[2], MD5C27, MD5S23);
405 MD5_STEP (MD5_H , a, b, c, d, w3[1], MD5C28, MD5S20);
406 MD5_STEP (MD5_H , d, a, b, c, w0[0], MD5C29, MD5S21);
407 MD5_STEP (MD5_H , c, d, a, b, w0[3], MD5C2a, MD5S22);
408 MD5_STEP (MD5_H , b, c, d, a, w1[2], MD5C2b, MD5S23);
409 MD5_STEP (MD5_H , a, b, c, d, w2[1], MD5C2c, MD5S20);
410 MD5_STEP (MD5_H , d, a, b, c, w3[0], MD5C2d, MD5S21);
411 MD5_STEP (MD5_H , c, d, a, b, w3[3], MD5C2e, MD5S22);
412 MD5_STEP (MD5_H , b, c, d, a, w0[2], MD5C2f, MD5S23);
414 MD5_STEP (MD5_I , a, b, c, d, w0[0], MD5C30, MD5S30);
415 MD5_STEP (MD5_I , d, a, b, c, w1[3], MD5C31, MD5S31);
416 MD5_STEP (MD5_I , c, d, a, b, w3[2], MD5C32, MD5S32);
417 MD5_STEP (MD5_I , b, c, d, a, w1[1], MD5C33, MD5S33);
418 MD5_STEP (MD5_I , a, b, c, d, w3[0], MD5C34, MD5S30);
419 MD5_STEP (MD5_I , d, a, b, c, w0[3], MD5C35, MD5S31);
420 MD5_STEP (MD5_I , c, d, a, b, w2[2], MD5C36, MD5S32);
421 MD5_STEP (MD5_I , b, c, d, a, w0[1], MD5C37, MD5S33);
422 MD5_STEP (MD5_I , a, b, c, d, w2[0], MD5C38, MD5S30);
423 MD5_STEP (MD5_I , d, a, b, c, w3[3], MD5C39, MD5S31);
424 MD5_STEP (MD5_I , c, d, a, b, w1[2], MD5C3a, MD5S32);
425 MD5_STEP (MD5_I , b, c, d, a, w3[1], MD5C3b, MD5S33);
426 MD5_STEP (MD5_I , a, b, c, d, w1[0], MD5C3c, MD5S30);
428 if (MATCHES_NONE_VS (a, search[0])) continue;
430 MD5_STEP (MD5_I , d, a, b, c, w2[3], MD5C3d, MD5S31);
431 MD5_STEP (MD5_I , c, d, a, b, w0[2], MD5C3e, MD5S32);
432 MD5_STEP (MD5_I , b, c, d, a, w2[1], MD5C3f, MD5S33);
434 COMPARE_S_SIMD (a, d, c, b);
438 __kernel void m04800_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)
442 __kernel void m04800_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)