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 m00040_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 rules_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 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];
50 pw_buf1[0] = pws[gid].i[ 4];
51 pw_buf1[1] = pws[gid].i[ 5];
52 pw_buf1[2] = pws[gid].i[ 6];
53 pw_buf1[3] = pws[gid].i[ 7];
55 const u32 pw_len = pws[gid].pw_len;
66 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
67 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
68 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
69 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
70 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
71 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
72 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
73 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
74 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
75 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
76 salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
77 salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
78 salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
79 salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
80 salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
81 salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
83 const u32 salt_len = salt_bufs[salt_pos].salt_len;
89 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos += VECT_SIZE)
96 const u32 out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
102 const u32 out_salt_len = (out_len * 2) + salt_len;
109 make_unicode (w0, w0_t, w1_t);
110 make_unicode (w1, w2_t, w3_t);
112 switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
114 w0_t[0] |= salt_buf0[0];
115 w0_t[1] |= salt_buf0[1];
116 w0_t[2] |= salt_buf0[2];
117 w0_t[3] |= salt_buf0[3];
118 w1_t[0] |= salt_buf1[0];
119 w1_t[1] |= salt_buf1[1];
120 w1_t[2] |= salt_buf1[2];
121 w1_t[3] |= salt_buf1[3];
122 w2_t[0] |= salt_buf2[0];
123 w2_t[1] |= salt_buf2[1];
124 w2_t[2] |= salt_buf2[2];
125 w2_t[3] |= salt_buf2[3];
126 w3_t[0] |= salt_buf3[0];
127 w3_t[1] |= salt_buf3[1];
128 w3_t[2] |= salt_buf3[2];
129 w3_t[3] |= salt_buf3[3];
131 append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, out_salt_len);
133 w3_t[2] = out_salt_len * 8;
145 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
146 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
147 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
148 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
149 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
150 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
151 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
152 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
153 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
154 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
155 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
156 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
157 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
158 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
159 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
160 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
162 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
163 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
164 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
165 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
166 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
167 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
168 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
169 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
170 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
171 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
172 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
173 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
174 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
175 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
176 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
177 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
179 MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
180 MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
181 MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
182 MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
183 MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
184 MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
185 MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
186 MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
187 MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
188 MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
189 MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
190 MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
191 MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
192 MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
193 MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
194 MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
196 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
197 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
198 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
199 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
200 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
201 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
202 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
203 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
204 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
205 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
206 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
207 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
208 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
209 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
210 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
211 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
213 COMPARE_M_SIMD (a, d, c, b);
217 __kernel void m00040_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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
221 __kernel void m00040_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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
225 __kernel void m00040_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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
231 const u32 lid = get_local_id (0);
237 const u32 gid = get_global_id (0);
239 if (gid >= gid_max) return;
243 pw_buf0[0] = pws[gid].i[ 0];
244 pw_buf0[1] = pws[gid].i[ 1];
245 pw_buf0[2] = pws[gid].i[ 2];
246 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;
266 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
267 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
268 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
269 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
270 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
271 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
272 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
273 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
274 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
275 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
276 salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
277 salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
278 salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
279 salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
280 salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
281 salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
283 const u32 salt_len = salt_bufs[salt_pos].salt_len;
289 const u32 search[4] =
291 digests_buf[digests_offset].digest_buf[DGST_R0],
292 digests_buf[digests_offset].digest_buf[DGST_R1],
293 digests_buf[digests_offset].digest_buf[DGST_R2],
294 digests_buf[digests_offset].digest_buf[DGST_R3]
301 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos += VECT_SIZE)
308 const u32 out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
314 const u32 out_salt_len = (out_len * 2) + salt_len;
321 make_unicode (w0, w0_t, w1_t);
322 make_unicode (w1, w2_t, w3_t);
324 switch_buffer_by_offset_le (w0_t, w1_t, w2_t, w3_t, salt_len);
326 w0_t[0] |= salt_buf0[0];
327 w0_t[1] |= salt_buf0[1];
328 w0_t[2] |= salt_buf0[2];
329 w0_t[3] |= salt_buf0[3];
330 w1_t[0] |= salt_buf1[0];
331 w1_t[1] |= salt_buf1[1];
332 w1_t[2] |= salt_buf1[2];
333 w1_t[3] |= salt_buf1[3];
334 w2_t[0] |= salt_buf2[0];
335 w2_t[1] |= salt_buf2[1];
336 w2_t[2] |= salt_buf2[2];
337 w2_t[3] |= salt_buf2[3];
338 w3_t[0] |= salt_buf3[0];
339 w3_t[1] |= salt_buf3[1];
340 w3_t[2] |= salt_buf3[2];
341 w3_t[3] |= salt_buf3[3];
343 append_0x80_4x4 (w0_t, w1_t, w2_t, w3_t, out_salt_len);
345 w3_t[2] = out_salt_len * 8;
357 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
358 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
359 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
360 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
361 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
362 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
363 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
364 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
365 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
366 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
367 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
368 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
369 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
370 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
371 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
372 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
374 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
375 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
376 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
377 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
378 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
379 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
380 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
381 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
382 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
383 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
384 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
385 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
386 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
387 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
388 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
389 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
391 MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
392 MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
393 MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
394 MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
395 MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
396 MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
397 MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
398 MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
399 MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
400 MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
401 MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
402 MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
403 MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
404 MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
405 MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
406 MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
408 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
409 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
410 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
411 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
412 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
413 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
414 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
415 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
416 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
417 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
418 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
419 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
420 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
422 if (MATCHES_NONE_VS (a, search[0])) continue;
424 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
425 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
426 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
428 COMPARE_S_SIMD (a, d, c, b);
432 __kernel void m00040_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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
436 __kernel void m00040_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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)