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 "types_ocl.c"
20 #define COMPARE_S "check_single_comp4.c"
21 #define COMPARE_M "check_multi_comp4.c"
25 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00040_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
31 const u32 lid = get_local_id (0);
37 const u32 gid = get_global_id (0);
39 if (gid >= gid_max) return;
43 wordl0[0] = pws[gid].i[ 0];
44 wordl0[1] = pws[gid].i[ 1];
45 wordl0[2] = pws[gid].i[ 2];
46 wordl0[3] = pws[gid].i[ 3];
50 wordl1[0] = pws[gid].i[ 4];
51 wordl1[1] = pws[gid].i[ 5];
52 wordl1[2] = pws[gid].i[ 6];
53 wordl1[3] = pws[gid].i[ 7];
69 const u32 pw_l_len = pws[gid].pw_len;
71 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
73 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
82 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
83 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
84 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
85 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
89 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
90 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
91 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
92 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
94 const u32 salt_len = salt_bufs[salt_pos].salt_len;
100 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
102 const u32 pw_r_len = combs_buf[il_pos].pw_len;
104 const u32 pw_len = pw_l_len + pw_r_len;
108 wordr0[0] = combs_buf[il_pos].i[0];
109 wordr0[1] = combs_buf[il_pos].i[1];
110 wordr0[2] = combs_buf[il_pos].i[2];
111 wordr0[3] = combs_buf[il_pos].i[3];
115 wordr1[0] = combs_buf[il_pos].i[4];
116 wordr1[1] = combs_buf[il_pos].i[5];
117 wordr1[2] = combs_buf[il_pos].i[6];
118 wordr1[3] = combs_buf[il_pos].i[7];
134 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
136 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
143 const u32 pw_salt_len = (pw_len * 2) + salt_len;
150 w0[0] = wordl0[0] | wordr0[0];
151 w0[1] = wordl0[1] | wordr0[1];
152 w0[2] = wordl0[2] | wordr0[2];
153 w0[3] = wordl0[3] | wordr0[3];
154 w1[0] = wordl1[0] | wordr1[0];
155 w1[1] = wordl1[1] | wordr1[1];
156 w1[2] = wordl1[2] | wordr1[2];
157 w1[3] = wordl1[3] | wordr1[3];
172 make_unicode (w0, w0_t, w1_t);
173 make_unicode (w1, w2_t, w3_t);
175 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
177 w0_t[0] |= salt_buf0[0];
178 w0_t[1] |= salt_buf0[1];
179 w0_t[2] |= salt_buf0[2];
180 w0_t[3] |= salt_buf0[3];
181 w1_t[0] |= salt_buf1[0];
182 w1_t[1] |= salt_buf1[1];
183 w1_t[2] |= salt_buf1[2];
184 w1_t[3] |= salt_buf1[3];
186 append_0x80_4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
188 w3_t[2] = pw_salt_len * 8;
199 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
200 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
201 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
202 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
203 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
204 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
205 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
206 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
207 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
208 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
209 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
210 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
211 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
212 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
213 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
214 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
216 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
217 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
218 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
219 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
220 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
221 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
222 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
223 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
224 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
225 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
226 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
227 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
228 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
229 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
230 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
231 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
233 MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
234 MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
235 MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
236 MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
237 MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
238 MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
239 MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
240 MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
241 MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
242 MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
243 MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
244 MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
245 MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
246 MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
247 MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
248 MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
250 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
251 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
252 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
253 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
254 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
255 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
256 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
257 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
258 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
259 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
260 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
261 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
262 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
263 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
264 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
265 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
277 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00040_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)
281 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00040_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)
285 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00040_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
291 const u32 lid = get_local_id (0);
297 const u32 gid = get_global_id (0);
299 if (gid >= gid_max) return;
303 wordl0[0] = pws[gid].i[ 0];
304 wordl0[1] = pws[gid].i[ 1];
305 wordl0[2] = pws[gid].i[ 2];
306 wordl0[3] = pws[gid].i[ 3];
310 wordl1[0] = pws[gid].i[ 4];
311 wordl1[1] = pws[gid].i[ 5];
312 wordl1[2] = pws[gid].i[ 6];
313 wordl1[3] = pws[gid].i[ 7];
329 const u32 pw_l_len = pws[gid].pw_len;
331 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
333 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
342 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
343 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
344 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
345 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
349 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
350 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
351 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
352 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
354 const u32 salt_len = salt_bufs[salt_pos].salt_len;
360 const u32 search[4] =
362 digests_buf[digests_offset].digest_buf[DGST_R0],
363 digests_buf[digests_offset].digest_buf[DGST_R1],
364 digests_buf[digests_offset].digest_buf[DGST_R2],
365 digests_buf[digests_offset].digest_buf[DGST_R3]
372 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
374 const u32 pw_r_len = combs_buf[il_pos].pw_len;
376 const u32 pw_len = pw_l_len + pw_r_len;
380 wordr0[0] = combs_buf[il_pos].i[0];
381 wordr0[1] = combs_buf[il_pos].i[1];
382 wordr0[2] = combs_buf[il_pos].i[2];
383 wordr0[3] = combs_buf[il_pos].i[3];
387 wordr1[0] = combs_buf[il_pos].i[4];
388 wordr1[1] = combs_buf[il_pos].i[5];
389 wordr1[2] = combs_buf[il_pos].i[6];
390 wordr1[3] = combs_buf[il_pos].i[7];
406 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
408 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
415 const u32 pw_salt_len = (pw_len * 2) + salt_len;
422 w0[0] = wordl0[0] | wordr0[0];
423 w0[1] = wordl0[1] | wordr0[1];
424 w0[2] = wordl0[2] | wordr0[2];
425 w0[3] = wordl0[3] | wordr0[3];
426 w1[0] = wordl1[0] | wordr1[0];
427 w1[1] = wordl1[1] | wordr1[1];
428 w1[2] = wordl1[2] | wordr1[2];
429 w1[3] = wordl1[3] | wordr1[3];
444 make_unicode (w0, w0_t, w1_t);
445 make_unicode (w1, w2_t, w3_t);
447 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
449 w0_t[0] |= salt_buf0[0];
450 w0_t[1] |= salt_buf0[1];
451 w0_t[2] |= salt_buf0[2];
452 w0_t[3] |= salt_buf0[3];
453 w1_t[0] |= salt_buf1[0];
454 w1_t[1] |= salt_buf1[1];
455 w1_t[2] |= salt_buf1[2];
456 w1_t[3] |= salt_buf1[3];
458 append_0x80_4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
460 w3_t[2] = pw_salt_len * 8;
471 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
472 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
473 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
474 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
475 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
476 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
477 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
478 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
479 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
480 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
481 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
482 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
483 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
484 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
485 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
486 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
488 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
489 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
490 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
491 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
492 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
493 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
494 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
495 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
496 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
497 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
498 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
499 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
500 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
501 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
502 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
503 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
505 MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
506 MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
507 MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
508 MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
509 MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
510 MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
511 MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
512 MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
513 MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
514 MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
515 MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
516 MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
517 MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
518 MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
519 MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
520 MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
522 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
523 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
524 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
525 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
526 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
527 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
528 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
529 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
530 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
531 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
532 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
533 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
534 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
536 bool q_cond = allx (search[0] != a);
538 if (q_cond) continue;
540 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
541 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
542 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
554 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00040_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)
558 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00040_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)