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 "OpenCL/simd.c"
23 static void m01100m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __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)
29 const u32 gid = get_global_id (0);
30 const u32 lid = get_local_id (0);
38 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
39 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
40 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
41 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
45 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
46 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
47 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
48 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
52 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
53 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
57 const u32 salt_len = salt_bufs[salt_pos].salt_len;
63 const u32 F_w0c00 = 0 + MD4C00;
64 const u32 F_w1c00 = w[ 1] + MD4C00;
65 const u32 F_w2c00 = w[ 2] + MD4C00;
66 const u32 F_w3c00 = w[ 3] + MD4C00;
67 const u32 F_w4c00 = w[ 4] + MD4C00;
68 const u32 F_w5c00 = w[ 5] + MD4C00;
69 const u32 F_w6c00 = w[ 6] + MD4C00;
70 const u32 F_w7c00 = w[ 7] + MD4C00;
71 const u32 F_w8c00 = w[ 8] + MD4C00;
72 const u32 F_w9c00 = w[ 9] + MD4C00;
73 const u32 F_wac00 = w[10] + MD4C00;
74 const u32 F_wbc00 = w[11] + MD4C00;
75 const u32 F_wcc00 = w[12] + MD4C00;
76 const u32 F_wdc00 = w[13] + MD4C00;
77 const u32 F_wec00 = w[14] + MD4C00;
78 const u32 F_wfc00 = w[15] + MD4C00;
80 const u32 G_w0c01 = 0 + MD4C01;
81 const u32 G_w4c01 = w[ 4] + MD4C01;
82 const u32 G_w8c01 = w[ 8] + MD4C01;
83 const u32 G_wcc01 = w[12] + MD4C01;
84 const u32 G_w1c01 = w[ 1] + MD4C01;
85 const u32 G_w5c01 = w[ 5] + MD4C01;
86 const u32 G_w9c01 = w[ 9] + MD4C01;
87 const u32 G_wdc01 = w[13] + MD4C01;
88 const u32 G_w2c01 = w[ 2] + MD4C01;
89 const u32 G_w6c01 = w[ 6] + MD4C01;
90 const u32 G_wac01 = w[10] + MD4C01;
91 const u32 G_wec01 = w[14] + MD4C01;
92 const u32 G_w3c01 = w[ 3] + MD4C01;
93 const u32 G_w7c01 = w[ 7] + MD4C01;
94 const u32 G_wbc01 = w[11] + MD4C01;
95 const u32 G_wfc01 = w[15] + MD4C01;
97 const u32 H_w0c02 = 0 + MD4C02;
98 const u32 H_w8c02 = w[ 8] + MD4C02;
99 const u32 H_w4c02 = w[ 4] + MD4C02;
100 const u32 H_wcc02 = w[12] + MD4C02;
101 const u32 H_w2c02 = w[ 2] + MD4C02;
102 const u32 H_wac02 = w[10] + MD4C02;
103 const u32 H_w6c02 = w[ 6] + MD4C02;
104 const u32 H_wec02 = w[14] + MD4C02;
105 const u32 H_w1c02 = w[ 1] + MD4C02;
106 const u32 H_w9c02 = w[ 9] + MD4C02;
107 const u32 H_w5c02 = w[ 5] + MD4C02;
108 const u32 H_wdc02 = w[13] + MD4C02;
109 const u32 H_w3c02 = w[ 3] + MD4C02;
110 const u32 H_wbc02 = w[11] + MD4C02;
111 const u32 H_w7c02 = w[ 7] + MD4C02;
112 const u32 H_wfc02 = w[15] + MD4C02;
120 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
122 const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
124 const u32x w0 = w0l | w0r;
131 MD4_STEP (MD4_Fo, a, b, c, d, w0, F_w0c00, MD4S00);
132 MD4_STEP0(MD4_Fo, d, a, b, c, F_w1c00, MD4S01);
133 MD4_STEP0(MD4_Fo, c, d, a, b, F_w2c00, MD4S02);
134 MD4_STEP0(MD4_Fo, b, c, d, a, F_w3c00, MD4S03);
135 MD4_STEP0(MD4_Fo, a, b, c, d, F_w4c00, MD4S00);
136 MD4_STEP0(MD4_Fo, d, a, b, c, F_w5c00, MD4S01);
137 MD4_STEP0(MD4_Fo, c, d, a, b, F_w6c00, MD4S02);
138 MD4_STEP0(MD4_Fo, b, c, d, a, F_w7c00, MD4S03);
139 MD4_STEP0(MD4_Fo, a, b, c, d, F_w8c00, MD4S00);
140 MD4_STEP0(MD4_Fo, d, a, b, c, F_w9c00, MD4S01);
141 MD4_STEP0(MD4_Fo, c, d, a, b, F_wac00, MD4S02);
142 MD4_STEP0(MD4_Fo, b, c, d, a, F_wbc00, MD4S03);
143 MD4_STEP0(MD4_Fo, a, b, c, d, F_wcc00, MD4S00);
144 MD4_STEP0(MD4_Fo, d, a, b, c, F_wdc00, MD4S01);
145 MD4_STEP0(MD4_Fo, c, d, a, b, F_wec00, MD4S02);
146 MD4_STEP0(MD4_Fo, b, c, d, a, F_wfc00, MD4S03);
148 MD4_STEP (MD4_Go, a, b, c, d, w0, G_w0c01, MD4S10);
149 MD4_STEP0(MD4_Go, d, a, b, c, G_w4c01, MD4S11);
150 MD4_STEP0(MD4_Go, c, d, a, b, G_w8c01, MD4S12);
151 MD4_STEP0(MD4_Go, b, c, d, a, G_wcc01, MD4S13);
152 MD4_STEP0(MD4_Go, a, b, c, d, G_w1c01, MD4S10);
153 MD4_STEP0(MD4_Go, d, a, b, c, G_w5c01, MD4S11);
154 MD4_STEP0(MD4_Go, c, d, a, b, G_w9c01, MD4S12);
155 MD4_STEP0(MD4_Go, b, c, d, a, G_wdc01, MD4S13);
156 MD4_STEP0(MD4_Go, a, b, c, d, G_w2c01, MD4S10);
157 MD4_STEP0(MD4_Go, d, a, b, c, G_w6c01, MD4S11);
158 MD4_STEP0(MD4_Go, c, d, a, b, G_wac01, MD4S12);
159 MD4_STEP0(MD4_Go, b, c, d, a, G_wec01, MD4S13);
160 MD4_STEP0(MD4_Go, a, b, c, d, G_w3c01, MD4S10);
161 MD4_STEP0(MD4_Go, d, a, b, c, G_w7c01, MD4S11);
162 MD4_STEP0(MD4_Go, c, d, a, b, G_wbc01, MD4S12);
163 MD4_STEP0(MD4_Go, b, c, d, a, G_wfc01, MD4S13);
165 MD4_STEP (MD4_H , a, b, c, d, w0, H_w0c02, MD4S20);
166 MD4_STEP0(MD4_H , d, a, b, c, H_w8c02, MD4S21);
167 MD4_STEP0(MD4_H , c, d, a, b, H_w4c02, MD4S22);
168 MD4_STEP0(MD4_H , b, c, d, a, H_wcc02, MD4S23);
169 MD4_STEP0(MD4_H , a, b, c, d, H_w2c02, MD4S20);
170 MD4_STEP0(MD4_H , d, a, b, c, H_wac02, MD4S21);
171 MD4_STEP0(MD4_H , c, d, a, b, H_w6c02, MD4S22);
172 MD4_STEP0(MD4_H , b, c, d, a, H_wec02, MD4S23);
173 MD4_STEP0(MD4_H , a, b, c, d, H_w1c02, MD4S20);
174 MD4_STEP0(MD4_H , d, a, b, c, H_w9c02, MD4S21);
175 MD4_STEP0(MD4_H , c, d, a, b, H_w5c02, MD4S22);
176 MD4_STEP0(MD4_H , b, c, d, a, H_wdc02, MD4S23);
177 MD4_STEP0(MD4_H , a, b, c, d, H_w3c02, MD4S20);
178 MD4_STEP0(MD4_H , d, a, b, c, H_wbc02, MD4S21);
179 MD4_STEP0(MD4_H , c, d, a, b, H_w7c02, MD4S22);
180 MD4_STEP0(MD4_H , b, c, d, a, H_wfc02, MD4S23);
196 w1_t[0] = salt_buf0[0];
197 w1_t[1] = salt_buf0[1];
198 w1_t[2] = salt_buf0[2];
199 w1_t[3] = salt_buf0[3];
200 w2_t[0] = salt_buf1[0];
201 w2_t[1] = salt_buf1[1];
202 w2_t[2] = salt_buf1[2];
203 w2_t[3] = salt_buf1[3];
204 w3_t[0] = salt_buf2[0];
205 w3_t[1] = salt_buf2[1];
206 w3_t[2] = (16 + salt_len) * 8;
214 MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00);
215 MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01);
216 MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02);
217 MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03);
218 MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00);
219 MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01);
220 MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02);
221 MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03);
222 MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00);
223 MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01);
224 MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02);
225 MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03);
226 MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00);
227 MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01);
228 MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02);
229 MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03);
231 MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
232 MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
233 MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
234 MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
235 MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
236 MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
237 MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
238 MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
239 MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
240 MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
241 MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
242 MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
243 MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
244 MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
245 MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
246 MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
248 MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
249 MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
250 MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
251 MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
252 MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
253 MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
254 MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
255 MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
256 MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
257 MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
258 MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
259 MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
260 MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
261 MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
262 MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
263 MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
265 COMPARE_M_SIMD (a, d, c, b);
269 static void m01100s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __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)
275 const u32 gid = get_global_id (0);
276 const u32 lid = get_local_id (0);
284 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
285 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
286 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
287 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
291 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
292 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
293 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
294 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
298 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
299 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
303 const u32 salt_len = salt_bufs[salt_pos].salt_len;
309 const u32 F_w0c00 = 0 + MD4C00;
310 const u32 F_w1c00 = w[ 1] + MD4C00;
311 const u32 F_w2c00 = w[ 2] + MD4C00;
312 const u32 F_w3c00 = w[ 3] + MD4C00;
313 const u32 F_w4c00 = w[ 4] + MD4C00;
314 const u32 F_w5c00 = w[ 5] + MD4C00;
315 const u32 F_w6c00 = w[ 6] + MD4C00;
316 const u32 F_w7c00 = w[ 7] + MD4C00;
317 const u32 F_w8c00 = w[ 8] + MD4C00;
318 const u32 F_w9c00 = w[ 9] + MD4C00;
319 const u32 F_wac00 = w[10] + MD4C00;
320 const u32 F_wbc00 = w[11] + MD4C00;
321 const u32 F_wcc00 = w[12] + MD4C00;
322 const u32 F_wdc00 = w[13] + MD4C00;
323 const u32 F_wec00 = w[14] + MD4C00;
324 const u32 F_wfc00 = w[15] + MD4C00;
326 const u32 G_w0c01 = 0 + MD4C01;
327 const u32 G_w4c01 = w[ 4] + MD4C01;
328 const u32 G_w8c01 = w[ 8] + MD4C01;
329 const u32 G_wcc01 = w[12] + MD4C01;
330 const u32 G_w1c01 = w[ 1] + MD4C01;
331 const u32 G_w5c01 = w[ 5] + MD4C01;
332 const u32 G_w9c01 = w[ 9] + MD4C01;
333 const u32 G_wdc01 = w[13] + MD4C01;
334 const u32 G_w2c01 = w[ 2] + MD4C01;
335 const u32 G_w6c01 = w[ 6] + MD4C01;
336 const u32 G_wac01 = w[10] + MD4C01;
337 const u32 G_wec01 = w[14] + MD4C01;
338 const u32 G_w3c01 = w[ 3] + MD4C01;
339 const u32 G_w7c01 = w[ 7] + MD4C01;
340 const u32 G_wbc01 = w[11] + MD4C01;
341 const u32 G_wfc01 = w[15] + MD4C01;
343 const u32 H_w0c02 = 0 + MD4C02;
344 const u32 H_w8c02 = w[ 8] + MD4C02;
345 const u32 H_w4c02 = w[ 4] + MD4C02;
346 const u32 H_wcc02 = w[12] + MD4C02;
347 const u32 H_w2c02 = w[ 2] + MD4C02;
348 const u32 H_wac02 = w[10] + MD4C02;
349 const u32 H_w6c02 = w[ 6] + MD4C02;
350 const u32 H_wec02 = w[14] + MD4C02;
351 const u32 H_w1c02 = w[ 1] + MD4C02;
352 const u32 H_w9c02 = w[ 9] + MD4C02;
353 const u32 H_w5c02 = w[ 5] + MD4C02;
354 const u32 H_wdc02 = w[13] + MD4C02;
355 const u32 H_w3c02 = w[ 3] + MD4C02;
356 const u32 H_wbc02 = w[11] + MD4C02;
357 const u32 H_w7c02 = w[ 7] + MD4C02;
358 const u32 H_wfc02 = w[15] + MD4C02;
364 const u32 search[4] =
366 digests_buf[digests_offset].digest_buf[DGST_R0],
367 digests_buf[digests_offset].digest_buf[DGST_R1],
368 digests_buf[digests_offset].digest_buf[DGST_R2],
369 digests_buf[digests_offset].digest_buf[DGST_R3]
378 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
380 const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
382 const u32x w0 = w0l | w0r;
389 MD4_STEP (MD4_Fo, a, b, c, d, w0, F_w0c00, MD4S00);
390 MD4_STEP0(MD4_Fo, d, a, b, c, F_w1c00, MD4S01);
391 MD4_STEP0(MD4_Fo, c, d, a, b, F_w2c00, MD4S02);
392 MD4_STEP0(MD4_Fo, b, c, d, a, F_w3c00, MD4S03);
393 MD4_STEP0(MD4_Fo, a, b, c, d, F_w4c00, MD4S00);
394 MD4_STEP0(MD4_Fo, d, a, b, c, F_w5c00, MD4S01);
395 MD4_STEP0(MD4_Fo, c, d, a, b, F_w6c00, MD4S02);
396 MD4_STEP0(MD4_Fo, b, c, d, a, F_w7c00, MD4S03);
397 MD4_STEP0(MD4_Fo, a, b, c, d, F_w8c00, MD4S00);
398 MD4_STEP0(MD4_Fo, d, a, b, c, F_w9c00, MD4S01);
399 MD4_STEP0(MD4_Fo, c, d, a, b, F_wac00, MD4S02);
400 MD4_STEP0(MD4_Fo, b, c, d, a, F_wbc00, MD4S03);
401 MD4_STEP0(MD4_Fo, a, b, c, d, F_wcc00, MD4S00);
402 MD4_STEP0(MD4_Fo, d, a, b, c, F_wdc00, MD4S01);
403 MD4_STEP0(MD4_Fo, c, d, a, b, F_wec00, MD4S02);
404 MD4_STEP0(MD4_Fo, b, c, d, a, F_wfc00, MD4S03);
406 MD4_STEP (MD4_Go, a, b, c, d, w0, G_w0c01, MD4S10);
407 MD4_STEP0(MD4_Go, d, a, b, c, G_w4c01, MD4S11);
408 MD4_STEP0(MD4_Go, c, d, a, b, G_w8c01, MD4S12);
409 MD4_STEP0(MD4_Go, b, c, d, a, G_wcc01, MD4S13);
410 MD4_STEP0(MD4_Go, a, b, c, d, G_w1c01, MD4S10);
411 MD4_STEP0(MD4_Go, d, a, b, c, G_w5c01, MD4S11);
412 MD4_STEP0(MD4_Go, c, d, a, b, G_w9c01, MD4S12);
413 MD4_STEP0(MD4_Go, b, c, d, a, G_wdc01, MD4S13);
414 MD4_STEP0(MD4_Go, a, b, c, d, G_w2c01, MD4S10);
415 MD4_STEP0(MD4_Go, d, a, b, c, G_w6c01, MD4S11);
416 MD4_STEP0(MD4_Go, c, d, a, b, G_wac01, MD4S12);
417 MD4_STEP0(MD4_Go, b, c, d, a, G_wec01, MD4S13);
418 MD4_STEP0(MD4_Go, a, b, c, d, G_w3c01, MD4S10);
419 MD4_STEP0(MD4_Go, d, a, b, c, G_w7c01, MD4S11);
420 MD4_STEP0(MD4_Go, c, d, a, b, G_wbc01, MD4S12);
421 MD4_STEP0(MD4_Go, b, c, d, a, G_wfc01, MD4S13);
423 MD4_STEP (MD4_H , a, b, c, d, w0, H_w0c02, MD4S20);
424 MD4_STEP0(MD4_H , d, a, b, c, H_w8c02, MD4S21);
425 MD4_STEP0(MD4_H , c, d, a, b, H_w4c02, MD4S22);
426 MD4_STEP0(MD4_H , b, c, d, a, H_wcc02, MD4S23);
427 MD4_STEP0(MD4_H , a, b, c, d, H_w2c02, MD4S20);
428 MD4_STEP0(MD4_H , d, a, b, c, H_wac02, MD4S21);
429 MD4_STEP0(MD4_H , c, d, a, b, H_w6c02, MD4S22);
430 MD4_STEP0(MD4_H , b, c, d, a, H_wec02, MD4S23);
431 MD4_STEP0(MD4_H , a, b, c, d, H_w1c02, MD4S20);
432 MD4_STEP0(MD4_H , d, a, b, c, H_w9c02, MD4S21);
433 MD4_STEP0(MD4_H , c, d, a, b, H_w5c02, MD4S22);
434 MD4_STEP0(MD4_H , b, c, d, a, H_wdc02, MD4S23);
435 MD4_STEP0(MD4_H , a, b, c, d, H_w3c02, MD4S20);
436 MD4_STEP0(MD4_H , d, a, b, c, H_wbc02, MD4S21);
437 MD4_STEP0(MD4_H , c, d, a, b, H_w7c02, MD4S22);
438 MD4_STEP0(MD4_H , b, c, d, a, H_wfc02, MD4S23);
454 w1_t[0] = salt_buf0[0];
455 w1_t[1] = salt_buf0[1];
456 w1_t[2] = salt_buf0[2];
457 w1_t[3] = salt_buf0[3];
458 w2_t[0] = salt_buf1[0];
459 w2_t[1] = salt_buf1[1];
460 w2_t[2] = salt_buf1[2];
461 w2_t[3] = salt_buf1[3];
462 w3_t[0] = salt_buf2[0];
463 w3_t[1] = salt_buf2[1];
464 w3_t[2] = (16 + salt_len) * 8;
472 MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00);
473 MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01);
474 MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02);
475 MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03);
476 MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00);
477 MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01);
478 MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02);
479 MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03);
480 MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00);
481 MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01);
482 MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02);
483 MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03);
484 MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00);
485 MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01);
486 MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02);
487 MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03);
489 MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
490 MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
491 MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
492 MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
493 MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
494 MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
495 MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
496 MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
497 MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
498 MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
499 MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
500 MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
501 MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
502 MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
503 MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
504 MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
506 MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
507 MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
508 MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
509 MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
510 MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
511 MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
512 MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
513 MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
514 MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
515 MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
516 MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
517 MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
518 MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
520 if (MATCHES_NONE_VS (a, search[0])) continue;
522 MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
523 MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
524 MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
526 COMPARE_S_SIMD (a, d, c, b);
530 __kernel void m01100_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __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)
536 const u32 gid = get_global_id (0);
538 if (gid >= gid_max) return;
542 w[ 0] = pws[gid].i[ 0];
543 w[ 1] = pws[gid].i[ 1];
544 w[ 2] = pws[gid].i[ 2];
545 w[ 3] = pws[gid].i[ 3];
556 w[14] = pws[gid].i[14];
559 const u32 pw_len = pws[gid].pw_len;
565 m01100m (w, pw_len, pws, rules_buf, combs_buf, words_buf_r, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
568 __kernel void m01100_m08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __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)
574 const u32 gid = get_global_id (0);
576 if (gid >= gid_max) return;
580 w[ 0] = pws[gid].i[ 0];
581 w[ 1] = pws[gid].i[ 1];
582 w[ 2] = pws[gid].i[ 2];
583 w[ 3] = pws[gid].i[ 3];
584 w[ 4] = pws[gid].i[ 4];
585 w[ 5] = pws[gid].i[ 5];
586 w[ 6] = pws[gid].i[ 6];
587 w[ 7] = pws[gid].i[ 7];
594 w[14] = pws[gid].i[14];
597 const u32 pw_len = pws[gid].pw_len;
603 m01100m (w, pw_len, pws, rules_buf, combs_buf, words_buf_r, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
606 __kernel void m01100_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __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)
610 __kernel void m01100_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __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)
616 const u32 gid = get_global_id (0);
618 if (gid >= gid_max) return;
622 w[ 0] = pws[gid].i[ 0];
623 w[ 1] = pws[gid].i[ 1];
624 w[ 2] = pws[gid].i[ 2];
625 w[ 3] = pws[gid].i[ 3];
636 w[14] = pws[gid].i[14];
639 const u32 pw_len = pws[gid].pw_len;
645 m01100s (w, pw_len, pws, rules_buf, combs_buf, words_buf_r, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
648 __kernel void m01100_s08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __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)
654 const u32 gid = get_global_id (0);
656 if (gid >= gid_max) return;
660 w[ 0] = pws[gid].i[ 0];
661 w[ 1] = pws[gid].i[ 1];
662 w[ 2] = pws[gid].i[ 2];
663 w[ 3] = pws[gid].i[ 3];
664 w[ 4] = pws[gid].i[ 4];
665 w[ 5] = pws[gid].i[ 5];
666 w[ 6] = pws[gid].i[ 6];
667 w[ 7] = pws[gid].i[ 7];
674 w[14] = pws[gid].i[14];
677 const u32 pw_len = pws[gid].pw_len;
683 m01100s (w, pw_len, pws, rules_buf, combs_buf, words_buf_r, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, esalt_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, il_cnt, digests_cnt, digests_offset);
686 __kernel void m01100_s16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32x * words_buf_r, __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)