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 m01100_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 lid = get_local_id (0);
37 const u32 gid = get_global_id (0);
39 if (gid >= gid_max) return;
44 pw_buf0[0] = pws[gid].i[0];
45 pw_buf0[1] = pws[gid].i[1];
46 pw_buf0[2] = pws[gid].i[2];
47 pw_buf0[3] = pws[gid].i[3];
48 pw_buf1[0] = pws[gid].i[4];
49 pw_buf1[1] = pws[gid].i[5];
50 pw_buf1[2] = pws[gid].i[6];
51 pw_buf1[3] = pws[gid].i[7];
53 const u32 pw_len = pws[gid].pw_len;
59 const u32 salt_buf0 = salt_bufs[salt_pos].salt_buf[0];
60 const u32 salt_buf1 = salt_bufs[salt_pos].salt_buf[1];
61 const u32 salt_buf2 = salt_bufs[salt_pos].salt_buf[2];
62 const u32 salt_buf3 = salt_bufs[salt_pos].salt_buf[3];
63 const u32 salt_buf4 = salt_bufs[salt_pos].salt_buf[4];
64 const u32 salt_buf5 = salt_bufs[salt_pos].salt_buf[5];
65 const u32 salt_buf6 = salt_bufs[salt_pos].salt_buf[6];
66 const u32 salt_buf7 = salt_bufs[salt_pos].salt_buf[7];
67 const u32 salt_buf8 = salt_bufs[salt_pos].salt_buf[8];
68 const u32 salt_buf9 = salt_bufs[salt_pos].salt_buf[9];
70 const u32 salt_len = salt_bufs[salt_pos].salt_len;
76 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
83 const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
85 append_0x80_2x4_VV (w0, w1, out_len);
87 make_unicode (w1, w2, w3);
88 make_unicode (w0, w0, w1);
90 w3[2] = out_len * 2 * 8;
98 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
99 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
100 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
101 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
102 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
103 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
104 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
105 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
106 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
107 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
108 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
109 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
110 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
111 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
112 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
113 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
115 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
116 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
117 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
118 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
119 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
120 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
121 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
122 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
123 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
124 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
125 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
126 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
127 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
128 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
129 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
130 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
132 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
133 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
134 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
135 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
136 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
137 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
138 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
139 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
140 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
141 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
142 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
143 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
144 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
145 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
146 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
147 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
168 w3[2] = (16 + salt_len) * 8;
176 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
177 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
178 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
179 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
180 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
181 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
182 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
183 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
184 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
185 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
186 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
187 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
188 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
189 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
190 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
191 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
193 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
194 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
195 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
196 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
197 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
198 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
199 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
200 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
201 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
202 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
203 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
204 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
205 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
206 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
207 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
208 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
210 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
211 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
212 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
213 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
214 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
215 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
216 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
217 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
218 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
219 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
220 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
221 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
222 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
223 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
224 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
225 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
227 COMPARE_M_SIMD (a, d, c, b);
231 __kernel void m01100_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)
235 __kernel void m01100_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)
239 __kernel void m01100_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)
245 const u32 lid = get_local_id (0);
251 const u32 gid = get_global_id (0);
253 if (gid >= gid_max) return;
258 pw_buf0[0] = pws[gid].i[0];
259 pw_buf0[1] = pws[gid].i[1];
260 pw_buf0[2] = pws[gid].i[2];
261 pw_buf0[3] = pws[gid].i[3];
262 pw_buf1[0] = pws[gid].i[4];
263 pw_buf1[1] = pws[gid].i[5];
264 pw_buf1[2] = pws[gid].i[6];
265 pw_buf1[3] = pws[gid].i[7];
267 const u32 pw_len = pws[gid].pw_len;
273 const u32 search[4] =
275 digests_buf[digests_offset].digest_buf[DGST_R0],
276 digests_buf[digests_offset].digest_buf[DGST_R1],
277 digests_buf[digests_offset].digest_buf[DGST_R2],
278 digests_buf[digests_offset].digest_buf[DGST_R3]
285 const u32 salt_buf0 = salt_bufs[salt_pos].salt_buf[0];
286 const u32 salt_buf1 = salt_bufs[salt_pos].salt_buf[1];
287 const u32 salt_buf2 = salt_bufs[salt_pos].salt_buf[2];
288 const u32 salt_buf3 = salt_bufs[salt_pos].salt_buf[3];
289 const u32 salt_buf4 = salt_bufs[salt_pos].salt_buf[4];
290 const u32 salt_buf5 = salt_bufs[salt_pos].salt_buf[5];
291 const u32 salt_buf6 = salt_bufs[salt_pos].salt_buf[6];
292 const u32 salt_buf7 = salt_bufs[salt_pos].salt_buf[7];
293 const u32 salt_buf8 = salt_bufs[salt_pos].salt_buf[8];
294 const u32 salt_buf9 = salt_bufs[salt_pos].salt_buf[9];
296 const u32 salt_len = salt_bufs[salt_pos].salt_len;
302 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
309 const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);
311 append_0x80_2x4_VV (w0, w1, out_len);
313 make_unicode (w1, w2, w3);
314 make_unicode (w0, w0, w1);
316 w3[2] = out_len * 2 * 8;
324 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
325 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
326 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
327 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
328 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
329 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
330 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
331 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
332 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
333 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
334 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
335 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
336 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
337 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
338 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
339 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
341 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
342 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
343 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
344 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
345 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
346 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
347 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
348 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
349 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
350 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
351 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
352 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
353 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
354 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
355 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
356 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
358 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
359 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
360 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
361 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
362 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
363 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
364 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
365 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
366 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
367 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
368 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
369 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
370 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
371 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
372 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
373 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
394 w3[2] = (16 + salt_len) * 8;
402 MD4_STEP (MD4_Fo, a, b, c, d, w0[0], MD4C00, MD4S00);
403 MD4_STEP (MD4_Fo, d, a, b, c, w0[1], MD4C00, MD4S01);
404 MD4_STEP (MD4_Fo, c, d, a, b, w0[2], MD4C00, MD4S02);
405 MD4_STEP (MD4_Fo, b, c, d, a, w0[3], MD4C00, MD4S03);
406 MD4_STEP (MD4_Fo, a, b, c, d, w1[0], MD4C00, MD4S00);
407 MD4_STEP (MD4_Fo, d, a, b, c, w1[1], MD4C00, MD4S01);
408 MD4_STEP (MD4_Fo, c, d, a, b, w1[2], MD4C00, MD4S02);
409 MD4_STEP (MD4_Fo, b, c, d, a, w1[3], MD4C00, MD4S03);
410 MD4_STEP (MD4_Fo, a, b, c, d, w2[0], MD4C00, MD4S00);
411 MD4_STEP (MD4_Fo, d, a, b, c, w2[1], MD4C00, MD4S01);
412 MD4_STEP (MD4_Fo, c, d, a, b, w2[2], MD4C00, MD4S02);
413 MD4_STEP (MD4_Fo, b, c, d, a, w2[3], MD4C00, MD4S03);
414 MD4_STEP (MD4_Fo, a, b, c, d, w3[0], MD4C00, MD4S00);
415 MD4_STEP (MD4_Fo, d, a, b, c, w3[1], MD4C00, MD4S01);
416 MD4_STEP (MD4_Fo, c, d, a, b, w3[2], MD4C00, MD4S02);
417 MD4_STEP (MD4_Fo, b, c, d, a, w3[3], MD4C00, MD4S03);
419 MD4_STEP (MD4_Go, a, b, c, d, w0[0], MD4C01, MD4S10);
420 MD4_STEP (MD4_Go, d, a, b, c, w1[0], MD4C01, MD4S11);
421 MD4_STEP (MD4_Go, c, d, a, b, w2[0], MD4C01, MD4S12);
422 MD4_STEP (MD4_Go, b, c, d, a, w3[0], MD4C01, MD4S13);
423 MD4_STEP (MD4_Go, a, b, c, d, w0[1], MD4C01, MD4S10);
424 MD4_STEP (MD4_Go, d, a, b, c, w1[1], MD4C01, MD4S11);
425 MD4_STEP (MD4_Go, c, d, a, b, w2[1], MD4C01, MD4S12);
426 MD4_STEP (MD4_Go, b, c, d, a, w3[1], MD4C01, MD4S13);
427 MD4_STEP (MD4_Go, a, b, c, d, w0[2], MD4C01, MD4S10);
428 MD4_STEP (MD4_Go, d, a, b, c, w1[2], MD4C01, MD4S11);
429 MD4_STEP (MD4_Go, c, d, a, b, w2[2], MD4C01, MD4S12);
430 MD4_STEP (MD4_Go, b, c, d, a, w3[2], MD4C01, MD4S13);
431 MD4_STEP (MD4_Go, a, b, c, d, w0[3], MD4C01, MD4S10);
432 MD4_STEP (MD4_Go, d, a, b, c, w1[3], MD4C01, MD4S11);
433 MD4_STEP (MD4_Go, c, d, a, b, w2[3], MD4C01, MD4S12);
434 MD4_STEP (MD4_Go, b, c, d, a, w3[3], MD4C01, MD4S13);
436 MD4_STEP (MD4_H , a, b, c, d, w0[0], MD4C02, MD4S20);
437 MD4_STEP (MD4_H , d, a, b, c, w2[0], MD4C02, MD4S21);
438 MD4_STEP (MD4_H , c, d, a, b, w1[0], MD4C02, MD4S22);
439 MD4_STEP (MD4_H , b, c, d, a, w3[0], MD4C02, MD4S23);
440 MD4_STEP (MD4_H , a, b, c, d, w0[2], MD4C02, MD4S20);
441 MD4_STEP (MD4_H , d, a, b, c, w2[2], MD4C02, MD4S21);
442 MD4_STEP (MD4_H , c, d, a, b, w1[2], MD4C02, MD4S22);
443 MD4_STEP (MD4_H , b, c, d, a, w3[2], MD4C02, MD4S23);
444 MD4_STEP (MD4_H , a, b, c, d, w0[1], MD4C02, MD4S20);
445 MD4_STEP (MD4_H , d, a, b, c, w2[1], MD4C02, MD4S21);
446 MD4_STEP (MD4_H , c, d, a, b, w1[1], MD4C02, MD4S22);
447 MD4_STEP (MD4_H , b, c, d, a, w3[1], MD4C02, MD4S23);
448 MD4_STEP (MD4_H , a, b, c, d, w0[3], MD4C02, MD4S20);
450 if (MATCHES_NONE_VS (a, search[0])) continue;
452 MD4_STEP (MD4_H , d, a, b, c, w2[3], MD4C02, MD4S21);
453 MD4_STEP (MD4_H , c, d, a, b, w1[3], MD4C02, MD4S22);
454 MD4_STEP (MD4_H , b, c, d, a, w3[3], MD4C02, MD4S23);
456 COMPARE_S_SIMD (a, d, c, b);
460 __kernel void m01100_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)
464 __kernel void m01100_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)