2 * Author......: Jens Steube <jens.steube@gmail.com>
9 #include "include/constants.h"
10 #include "include/kernel_vendor.h"
29 #include "include/kernel_functions.c"
30 #include "types_amd.c"
31 #include "common_amd.c"
34 #define VECT_COMPARE_S "check_single_vect1_comp4_warp.c"
35 #define VECT_COMPARE_M "check_multi_vect1_comp4_warp.c"
39 #define VECT_COMPARE_S "check_single_vect2_comp4_warp.c"
40 #define VECT_COMPARE_M "check_multi_vect2_comp4_warp.c"
44 #define VECT_COMPARE_S "check_single_vect4_comp4_warp.c"
45 #define VECT_COMPARE_M "check_multi_vect4_comp4_warp.c"
48 #define MD5_STEP_REV(f,a,b,c,d,x,t,s) \
57 #define MD5_STEP_REV1(f,a,b,c,d,x,t,s) \
65 static void m00000m (u32 w[16], const u32 pw_len, __global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global 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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
71 const u32 gid = get_global_id (0);
72 const u32 lid = get_local_id (0);
78 const u32 F_w0c00 = 0 + MD5C00;
79 const u32 F_w1c01 = w[ 1] + MD5C01;
80 const u32 F_w2c02 = w[ 2] + MD5C02;
81 const u32 F_w3c03 = w[ 3] + MD5C03;
82 const u32 F_w4c04 = w[ 4] + MD5C04;
83 const u32 F_w5c05 = w[ 5] + MD5C05;
84 const u32 F_w6c06 = w[ 6] + MD5C06;
85 const u32 F_w7c07 = w[ 7] + MD5C07;
86 const u32 F_w8c08 = w[ 8] + MD5C08;
87 const u32 F_w9c09 = w[ 9] + MD5C09;
88 const u32 F_wac0a = w[10] + MD5C0a;
89 const u32 F_wbc0b = w[11] + MD5C0b;
90 const u32 F_wcc0c = w[12] + MD5C0c;
91 const u32 F_wdc0d = w[13] + MD5C0d;
92 const u32 F_wec0e = w[14] + MD5C0e;
93 const u32 F_wfc0f = w[15] + MD5C0f;
95 const u32 G_w1c10 = w[ 1] + MD5C10;
96 const u32 G_w6c11 = w[ 6] + MD5C11;
97 const u32 G_wbc12 = w[11] + MD5C12;
98 const u32 G_w0c13 = 0 + MD5C13;
99 const u32 G_w5c14 = w[ 5] + MD5C14;
100 const u32 G_wac15 = w[10] + MD5C15;
101 const u32 G_wfc16 = w[15] + MD5C16;
102 const u32 G_w4c17 = w[ 4] + MD5C17;
103 const u32 G_w9c18 = w[ 9] + MD5C18;
104 const u32 G_wec19 = w[14] + MD5C19;
105 const u32 G_w3c1a = w[ 3] + MD5C1a;
106 const u32 G_w8c1b = w[ 8] + MD5C1b;
107 const u32 G_wdc1c = w[13] + MD5C1c;
108 const u32 G_w2c1d = w[ 2] + MD5C1d;
109 const u32 G_w7c1e = w[ 7] + MD5C1e;
110 const u32 G_wcc1f = w[12] + MD5C1f;
112 const u32 H_w5c20 = w[ 5] + MD5C20;
113 const u32 H_w8c21 = w[ 8] + MD5C21;
114 const u32 H_wbc22 = w[11] + MD5C22;
115 const u32 H_wec23 = w[14] + MD5C23;
116 const u32 H_w1c24 = w[ 1] + MD5C24;
117 const u32 H_w4c25 = w[ 4] + MD5C25;
118 const u32 H_w7c26 = w[ 7] + MD5C26;
119 const u32 H_wac27 = w[10] + MD5C27;
120 const u32 H_wdc28 = w[13] + MD5C28;
121 const u32 H_w0c29 = 0 + MD5C29;
122 const u32 H_w3c2a = w[ 3] + MD5C2a;
123 const u32 H_w6c2b = w[ 6] + MD5C2b;
124 const u32 H_w9c2c = w[ 9] + MD5C2c;
125 const u32 H_wcc2d = w[12] + MD5C2d;
126 const u32 H_wfc2e = w[15] + MD5C2e;
127 const u32 H_w2c2f = w[ 2] + MD5C2f;
129 const u32 I_w0c30 = 0 + MD5C30;
130 const u32 I_w7c31 = w[ 7] + MD5C31;
131 const u32 I_wec32 = w[14] + MD5C32;
132 const u32 I_w5c33 = w[ 5] + MD5C33;
133 const u32 I_wcc34 = w[12] + MD5C34;
134 const u32 I_w3c35 = w[ 3] + MD5C35;
135 const u32 I_wac36 = w[10] + MD5C36;
136 const u32 I_w1c37 = w[ 1] + MD5C37;
137 const u32 I_w8c38 = w[ 8] + MD5C38;
138 const u32 I_wfc39 = w[15] + MD5C39;
139 const u32 I_w6c3a = w[ 6] + MD5C3a;
140 const u32 I_wdc3b = w[13] + MD5C3b;
141 const u32 I_w4c3c = w[ 4] + MD5C3c;
142 const u32 I_wbc3d = w[11] + MD5C3d;
143 const u32 I_w2c3e = w[ 2] + MD5C3e;
144 const u32 I_w9c3f = w[ 9] + MD5C3f;
150 const u32 bf_loops = ceil ((float) bfs_cnt / VECT_DIV);
154 for (u32 il_pos = 0; il_pos < bf_loops; il_pos++)
156 const u32x w0r = words_buf_r[il_pos];
158 const u32x w0 = w0l | w0r;
167 MD5_STEP (MD5_Fo, a, b, c, d, w0, F_w0c00, MD5S00);
168 MD5_STEP0(MD5_Fo, d, a, b, c, F_w1c01, MD5S01);
169 MD5_STEP0(MD5_Fo, c, d, a, b, F_w2c02, MD5S02);
170 MD5_STEP0(MD5_Fo, b, c, d, a, F_w3c03, MD5S03);
171 MD5_STEP0(MD5_Fo, a, b, c, d, F_w4c04, MD5S00);
172 MD5_STEP0(MD5_Fo, d, a, b, c, F_w5c05, MD5S01);
173 MD5_STEP0(MD5_Fo, c, d, a, b, F_w6c06, MD5S02);
174 MD5_STEP0(MD5_Fo, b, c, d, a, F_w7c07, MD5S03);
175 MD5_STEP0(MD5_Fo, a, b, c, d, F_w8c08, MD5S00);
176 MD5_STEP0(MD5_Fo, d, a, b, c, F_w9c09, MD5S01);
177 MD5_STEP0(MD5_Fo, c, d, a, b, F_wac0a, MD5S02);
178 MD5_STEP0(MD5_Fo, b, c, d, a, F_wbc0b, MD5S03);
179 MD5_STEP0(MD5_Fo, a, b, c, d, F_wcc0c, MD5S00);
180 MD5_STEP0(MD5_Fo, d, a, b, c, F_wdc0d, MD5S01);
181 MD5_STEP0(MD5_Fo, c, d, a, b, F_wec0e, MD5S02);
182 MD5_STEP0(MD5_Fo, b, c, d, a, F_wfc0f, MD5S03);
184 MD5_STEP0(MD5_Go, a, b, c, d, G_w1c10, MD5S10);
185 MD5_STEP0(MD5_Go, d, a, b, c, G_w6c11, MD5S11);
186 MD5_STEP0(MD5_Go, c, d, a, b, G_wbc12, MD5S12);
187 MD5_STEP (MD5_Go, b, c, d, a, w0, G_w0c13, MD5S13);
188 MD5_STEP0(MD5_Go, a, b, c, d, G_w5c14, MD5S10);
189 MD5_STEP0(MD5_Go, d, a, b, c, G_wac15, MD5S11);
190 MD5_STEP0(MD5_Go, c, d, a, b, G_wfc16, MD5S12);
191 MD5_STEP0(MD5_Go, b, c, d, a, G_w4c17, MD5S13);
192 MD5_STEP0(MD5_Go, a, b, c, d, G_w9c18, MD5S10);
193 MD5_STEP0(MD5_Go, d, a, b, c, G_wec19, MD5S11);
194 MD5_STEP0(MD5_Go, c, d, a, b, G_w3c1a, MD5S12);
195 MD5_STEP0(MD5_Go, b, c, d, a, G_w8c1b, MD5S13);
196 MD5_STEP0(MD5_Go, a, b, c, d, G_wdc1c, MD5S10);
197 MD5_STEP0(MD5_Go, d, a, b, c, G_w2c1d, MD5S11);
198 MD5_STEP0(MD5_Go, c, d, a, b, G_w7c1e, MD5S12);
199 MD5_STEP0(MD5_Go, b, c, d, a, G_wcc1f, MD5S13);
201 MD5_STEP0(MD5_H1, a, b, c, d, H_w5c20, MD5S20);
202 MD5_STEP0(MD5_H2, d, a, b, c, H_w8c21, MD5S21);
203 MD5_STEP0(MD5_H1, c, d, a, b, H_wbc22, MD5S22);
204 MD5_STEP0(MD5_H2, b, c, d, a, H_wec23, MD5S23);
205 MD5_STEP0(MD5_H1, a, b, c, d, H_w1c24, MD5S20);
206 MD5_STEP0(MD5_H2, d, a, b, c, H_w4c25, MD5S21);
207 MD5_STEP0(MD5_H1, c, d, a, b, H_w7c26, MD5S22);
208 MD5_STEP0(MD5_H2, b, c, d, a, H_wac27, MD5S23);
209 MD5_STEP0(MD5_H1, a, b, c, d, H_wdc28, MD5S20);
210 MD5_STEP (MD5_H2, d, a, b, c, w0, H_w0c29, MD5S21);
211 MD5_STEP0(MD5_H1, c, d, a, b, H_w3c2a, MD5S22);
212 MD5_STEP0(MD5_H2, b, c, d, a, H_w6c2b, MD5S23);
213 MD5_STEP0(MD5_H1, a, b, c, d, H_w9c2c, MD5S20);
214 MD5_STEP0(MD5_H2, d, a, b, c, H_wcc2d, MD5S21);
215 MD5_STEP0(MD5_H1, c, d, a, b, H_wfc2e, MD5S22);
216 MD5_STEP0(MD5_H2, b, c, d, a, H_w2c2f, MD5S23);
218 MD5_STEP (MD5_I , a, b, c, d, w0, I_w0c30, MD5S30);
219 MD5_STEP0(MD5_I , d, a, b, c, I_w7c31, MD5S31);
220 MD5_STEP0(MD5_I , c, d, a, b, I_wec32, MD5S32);
221 MD5_STEP0(MD5_I , b, c, d, a, I_w5c33, MD5S33);
222 MD5_STEP0(MD5_I , a, b, c, d, I_wcc34, MD5S30);
223 MD5_STEP0(MD5_I , d, a, b, c, I_w3c35, MD5S31);
224 MD5_STEP0(MD5_I , c, d, a, b, I_wac36, MD5S32);
225 MD5_STEP0(MD5_I , b, c, d, a, I_w1c37, MD5S33);
226 MD5_STEP0(MD5_I , a, b, c, d, I_w8c38, MD5S30);
227 MD5_STEP0(MD5_I , d, a, b, c, I_wfc39, MD5S31);
228 MD5_STEP0(MD5_I , c, d, a, b, I_w6c3a, MD5S32);
229 MD5_STEP0(MD5_I , b, c, d, a, I_wdc3b, MD5S33);
230 MD5_STEP0(MD5_I , a, b, c, d, I_w4c3c, MD5S30);
231 MD5_STEP0(MD5_I , d, a, b, c, I_wbc3d, MD5S31);
232 MD5_STEP0(MD5_I , c, d, a, b, I_w2c3e, MD5S32);
233 MD5_STEP0(MD5_I , b, c, d, a, I_w9c3f, MD5S33);
241 #include VECT_COMPARE_M
245 static void m00000s (u32 w[16], const u32 pw_len, __global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global 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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
251 const u32 gid = get_global_id (0);
252 const u32 lid = get_local_id (0);
258 const u32 F_w0c00 = 0 + MD5C00;
259 const u32 F_w1c01 = w[ 1] + MD5C01;
260 const u32 F_w2c02 = w[ 2] + MD5C02;
261 const u32 F_w3c03 = w[ 3] + MD5C03;
262 const u32 F_w4c04 = w[ 4] + MD5C04;
263 const u32 F_w5c05 = w[ 5] + MD5C05;
264 const u32 F_w6c06 = w[ 6] + MD5C06;
265 const u32 F_w7c07 = w[ 7] + MD5C07;
266 const u32 F_w8c08 = w[ 8] + MD5C08;
267 const u32 F_w9c09 = w[ 9] + MD5C09;
268 const u32 F_wac0a = w[10] + MD5C0a;
269 const u32 F_wbc0b = w[11] + MD5C0b;
270 const u32 F_wcc0c = w[12] + MD5C0c;
271 const u32 F_wdc0d = w[13] + MD5C0d;
272 const u32 F_wec0e = w[14] + MD5C0e;
273 const u32 F_wfc0f = w[15] + MD5C0f;
275 const u32 G_w1c10 = w[ 1] + MD5C10;
276 const u32 G_w6c11 = w[ 6] + MD5C11;
277 const u32 G_wbc12 = w[11] + MD5C12;
278 const u32 G_w0c13 = 0 + MD5C13;
279 const u32 G_w5c14 = w[ 5] + MD5C14;
280 const u32 G_wac15 = w[10] + MD5C15;
281 const u32 G_wfc16 = w[15] + MD5C16;
282 const u32 G_w4c17 = w[ 4] + MD5C17;
283 const u32 G_w9c18 = w[ 9] + MD5C18;
284 const u32 G_wec19 = w[14] + MD5C19;
285 const u32 G_w3c1a = w[ 3] + MD5C1a;
286 const u32 G_w8c1b = w[ 8] + MD5C1b;
287 const u32 G_wdc1c = w[13] + MD5C1c;
288 const u32 G_w2c1d = w[ 2] + MD5C1d;
289 const u32 G_w7c1e = w[ 7] + MD5C1e;
290 const u32 G_wcc1f = w[12] + MD5C1f;
292 const u32 H_w5c20 = w[ 5] + MD5C20;
293 const u32 H_w8c21 = w[ 8] + MD5C21;
294 const u32 H_wbc22 = w[11] + MD5C22;
295 const u32 H_wec23 = w[14] + MD5C23;
296 const u32 H_w1c24 = w[ 1] + MD5C24;
297 const u32 H_w4c25 = w[ 4] + MD5C25;
298 const u32 H_w7c26 = w[ 7] + MD5C26;
299 const u32 H_wac27 = w[10] + MD5C27;
300 const u32 H_wdc28 = w[13] + MD5C28;
301 const u32 H_w0c29 = 0 + MD5C29;
302 const u32 H_w3c2a = w[ 3] + MD5C2a;
303 const u32 H_w6c2b = w[ 6] + MD5C2b;
304 const u32 H_w9c2c = w[ 9] + MD5C2c;
305 const u32 H_wcc2d = w[12] + MD5C2d;
306 const u32 H_wfc2e = w[15] + MD5C2e;
307 const u32 H_w2c2f = w[ 2] + MD5C2f;
309 const u32 I_w0c30 = 0 + MD5C30;
310 const u32 I_w7c31 = w[ 7] + MD5C31;
311 const u32 I_wec32 = w[14] + MD5C32;
312 const u32 I_w5c33 = w[ 5] + MD5C33;
313 const u32 I_wcc34 = w[12] + MD5C34;
314 const u32 I_w3c35 = w[ 3] + MD5C35;
315 const u32 I_wac36 = w[10] + MD5C36;
316 const u32 I_w1c37 = w[ 1] + MD5C37;
317 const u32 I_w8c38 = w[ 8] + MD5C38;
318 const u32 I_wfc39 = w[15] + MD5C39;
319 const u32 I_w6c3a = w[ 6] + MD5C3a;
320 const u32 I_wdc3b = w[13] + MD5C3b;
321 const u32 I_w4c3c = w[ 4] + MD5C3c;
322 const u32 I_wbc3d = w[11] + MD5C3d;
323 const u32 I_w2c3e = w[ 2] + MD5C3e;
324 const u32 I_w9c3f = w[ 9] + MD5C3f;
330 const u32 search[4] =
332 digests_buf[digests_offset].digest_buf[DGST_R0],
333 digests_buf[digests_offset].digest_buf[DGST_R1],
334 digests_buf[digests_offset].digest_buf[DGST_R2],
335 digests_buf[digests_offset].digest_buf[DGST_R3]
342 u32 a_rev = digests_buf[digests_offset].digest_buf[0];
343 u32 b_rev = digests_buf[digests_offset].digest_buf[1];
344 u32 c_rev = digests_buf[digests_offset].digest_buf[2];
345 u32 d_rev = digests_buf[digests_offset].digest_buf[3];
347 MD5_STEP_REV (MD5_I, b_rev, c_rev, d_rev, a_rev, w[ 9], MD5C3f, MD5S33);
348 MD5_STEP_REV (MD5_I, c_rev, d_rev, a_rev, b_rev, w[ 2], MD5C3e, MD5S32);
349 MD5_STEP_REV (MD5_I, d_rev, a_rev, b_rev, c_rev, w[11], MD5C3d, MD5S31);
350 MD5_STEP_REV (MD5_I, a_rev, b_rev, c_rev, d_rev, w[ 4], MD5C3c, MD5S30);
351 MD5_STEP_REV (MD5_I, b_rev, c_rev, d_rev, a_rev, w[13], MD5C3b, MD5S33);
352 MD5_STEP_REV (MD5_I, c_rev, d_rev, a_rev, b_rev, w[ 6], MD5C3a, MD5S32);
353 MD5_STEP_REV (MD5_I, d_rev, a_rev, b_rev, c_rev, w[15], MD5C39, MD5S31);
354 MD5_STEP_REV (MD5_I, a_rev, b_rev, c_rev, d_rev, w[ 8], MD5C38, MD5S30);
355 MD5_STEP_REV (MD5_I, b_rev, c_rev, d_rev, a_rev, w[ 1], MD5C37, MD5S33);
356 MD5_STEP_REV (MD5_I, c_rev, d_rev, a_rev, b_rev, w[10], MD5C36, MD5S32);
357 MD5_STEP_REV (MD5_I, d_rev, a_rev, b_rev, c_rev, w[ 3], MD5C35, MD5S31);
358 MD5_STEP_REV (MD5_I, a_rev, b_rev, c_rev, d_rev, w[12], MD5C34, MD5S30);
359 MD5_STEP_REV (MD5_I, b_rev, c_rev, d_rev, a_rev, w[ 5], MD5C33, MD5S33);
360 MD5_STEP_REV (MD5_I, c_rev, d_rev, a_rev, b_rev, w[14], MD5C32, MD5S32);
361 MD5_STEP_REV (MD5_I, d_rev, a_rev, b_rev, c_rev, w[ 7], MD5C31, MD5S31);
362 MD5_STEP_REV (MD5_I, a_rev, b_rev, c_rev, d_rev, 0, MD5C30, MD5S30);
364 const u32x pre_cd = c_rev ^ d_rev;
366 MD5_STEP_REV1(MD5_H, b_rev, c_rev, d_rev, a_rev, w[ 2], MD5C2f, MD5S23);
367 MD5_STEP_REV1(MD5_H, c_rev, d_rev, a_rev, b_rev, w[15], MD5C2e, MD5S22);
373 const u32 bf_loops = ceil ((float) bfs_cnt / VECT_DIV);
377 for (u32 il_pos = 0; il_pos < bf_loops; il_pos++)
379 const u32x w0r = words_buf_r[il_pos];
381 const u32x w0 = w0l | w0r;
383 const u32x pre_d = d_rev;
384 const u32x pre_a = a_rev - w0;
385 const u32x pre_b = b_rev - (pre_a ^ pre_cd);
386 const u32x pre_c = c_rev - (pre_a ^ pre_b ^ pre_d);
395 MD5_STEP (MD5_Fo, a, b, c, d, w0, F_w0c00, MD5S00);
396 MD5_STEP0(MD5_Fo, d, a, b, c, F_w1c01, MD5S01);
397 MD5_STEP0(MD5_Fo, c, d, a, b, F_w2c02, MD5S02);
398 MD5_STEP0(MD5_Fo, b, c, d, a, F_w3c03, MD5S03);
399 MD5_STEP0(MD5_Fo, a, b, c, d, F_w4c04, MD5S00);
400 MD5_STEP0(MD5_Fo, d, a, b, c, F_w5c05, MD5S01);
401 MD5_STEP0(MD5_Fo, c, d, a, b, F_w6c06, MD5S02);
402 MD5_STEP0(MD5_Fo, b, c, d, a, F_w7c07, MD5S03);
403 MD5_STEP0(MD5_Fo, a, b, c, d, F_w8c08, MD5S00);
404 MD5_STEP0(MD5_Fo, d, a, b, c, F_w9c09, MD5S01);
405 MD5_STEP0(MD5_Fo, c, d, a, b, F_wac0a, MD5S02);
406 MD5_STEP0(MD5_Fo, b, c, d, a, F_wbc0b, MD5S03);
407 MD5_STEP0(MD5_Fo, a, b, c, d, F_wcc0c, MD5S00);
408 MD5_STEP0(MD5_Fo, d, a, b, c, F_wdc0d, MD5S01);
409 MD5_STEP0(MD5_Fo, c, d, a, b, F_wec0e, MD5S02);
410 MD5_STEP0(MD5_Fo, b, c, d, a, F_wfc0f, MD5S03);
412 MD5_STEP0(MD5_Go, a, b, c, d, G_w1c10, MD5S10);
413 MD5_STEP0(MD5_Go, d, a, b, c, G_w6c11, MD5S11);
414 MD5_STEP0(MD5_Go, c, d, a, b, G_wbc12, MD5S12);
415 MD5_STEP (MD5_Go, b, c, d, a, w0, G_w0c13, MD5S13);
416 MD5_STEP0(MD5_Go, a, b, c, d, G_w5c14, MD5S10);
417 MD5_STEP0(MD5_Go, d, a, b, c, G_wac15, MD5S11);
418 MD5_STEP0(MD5_Go, c, d, a, b, G_wfc16, MD5S12);
419 MD5_STEP0(MD5_Go, b, c, d, a, G_w4c17, MD5S13);
420 MD5_STEP0(MD5_Go, a, b, c, d, G_w9c18, MD5S10);
421 MD5_STEP0(MD5_Go, d, a, b, c, G_wec19, MD5S11);
422 MD5_STEP0(MD5_Go, c, d, a, b, G_w3c1a, MD5S12);
423 MD5_STEP0(MD5_Go, b, c, d, a, G_w8c1b, MD5S13);
424 MD5_STEP0(MD5_Go, a, b, c, d, G_wdc1c, MD5S10);
425 MD5_STEP0(MD5_Go, d, a, b, c, G_w2c1d, MD5S11);
426 MD5_STEP0(MD5_Go, c, d, a, b, G_w7c1e, MD5S12);
427 MD5_STEP0(MD5_Go, b, c, d, a, G_wcc1f, MD5S13);
429 MD5_STEP0(MD5_H1, a, b, c, d, H_w5c20, MD5S20);
430 MD5_STEP0(MD5_H2, d, a, b, c, H_w8c21, MD5S21);
431 MD5_STEP0(MD5_H1, c, d, a, b, H_wbc22, MD5S22);
432 MD5_STEP0(MD5_H2, b, c, d, a, H_wec23, MD5S23);
433 MD5_STEP0(MD5_H1, a, b, c, d, H_w1c24, MD5S20);
434 MD5_STEP0(MD5_H2, d, a, b, c, H_w4c25, MD5S21);
435 MD5_STEP0(MD5_H1, c, d, a, b, H_w7c26, MD5S22);
436 MD5_STEP0(MD5_H2, b, c, d, a, H_wac27, MD5S23);
437 MD5_STEP0(MD5_H1, a, b, c, d, H_wdc28, MD5S20);
438 MD5_STEP (MD5_H2, d, a, b, c, w0, H_w0c29, MD5S21);
439 MD5_STEP0(MD5_H1, c, d, a, b, H_w3c2a, MD5S22);
441 bool q_cond = allx (pre_c != c);
443 if (q_cond) continue;
445 MD5_STEP0(MD5_H2, b, c, d, a, H_w6c2b, MD5S23);
446 MD5_STEP0(MD5_H1, a, b, c, d, H_w9c2c, MD5S20);
447 MD5_STEP0(MD5_H2, d, a, b, c, H_wcc2d, MD5S21);
449 bool q_cond2 = allx (pre_d != d);
451 if (q_cond2) continue;
453 MD5_STEP0(MD5_H1, c, d, a, b, H_wfc2e, MD5S22);
454 MD5_STEP0(MD5_H2, b, c, d, a, H_w2c2f, MD5S23);
456 MD5_STEP (MD5_I , a, b, c, d, w0, I_w0c30, MD5S30);
457 MD5_STEP0(MD5_I , d, a, b, c, I_w7c31, MD5S31);
458 MD5_STEP0(MD5_I , c, d, a, b, I_wec32, MD5S32);
459 MD5_STEP0(MD5_I , b, c, d, a, I_w5c33, MD5S33);
460 MD5_STEP0(MD5_I , a, b, c, d, I_wcc34, MD5S30);
461 MD5_STEP0(MD5_I , d, a, b, c, I_w3c35, MD5S31);
462 MD5_STEP0(MD5_I , c, d, a, b, I_wac36, MD5S32);
463 MD5_STEP0(MD5_I , b, c, d, a, I_w1c37, MD5S33);
464 MD5_STEP0(MD5_I , a, b, c, d, I_w8c38, MD5S30);
465 MD5_STEP0(MD5_I , d, a, b, c, I_wfc39, MD5S31);
466 MD5_STEP0(MD5_I , c, d, a, b, I_w6c3a, MD5S32);
467 MD5_STEP0(MD5_I , b, c, d, a, I_wdc3b, MD5S33);
468 MD5_STEP0(MD5_I , a, b, c, d, I_w4c3c, MD5S30);
469 MD5_STEP0(MD5_I , d, a, b, c, I_wbc3d, MD5S31);
470 MD5_STEP0(MD5_I , c, d, a, b, I_w2c3e, MD5S32);
471 MD5_STEP0(MD5_I , b, c, d, a, I_w9c3f, MD5S33);
479 #include VECT_COMPARE_S
483 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_m04 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global 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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
489 const u32 gid = get_global_id (0);
491 if (gid >= gid_max) return;
495 w[ 0] = pws[gid].i[ 0];
496 w[ 1] = pws[gid].i[ 1];
497 w[ 2] = pws[gid].i[ 2];
498 w[ 3] = pws[gid].i[ 3];
509 w[14] = pws[gid].i[14];
512 const u32 pw_len = pws[gid].pw_len;
518 m00000m (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, bfs_cnt, digests_cnt, digests_offset);
521 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_m08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global 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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
527 const u32 gid = get_global_id (0);
529 if (gid >= gid_max) return;
533 w[ 0] = pws[gid].i[ 0];
534 w[ 1] = pws[gid].i[ 1];
535 w[ 2] = pws[gid].i[ 2];
536 w[ 3] = pws[gid].i[ 3];
537 w[ 4] = pws[gid].i[ 4];
538 w[ 5] = pws[gid].i[ 5];
539 w[ 6] = pws[gid].i[ 6];
540 w[ 7] = pws[gid].i[ 7];
547 w[14] = pws[gid].i[14];
550 const u32 pw_len = pws[gid].pw_len;
556 m00000m (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, bfs_cnt, digests_cnt, digests_offset);
559 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_m16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global 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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
565 const u32 gid = get_global_id (0);
567 if (gid >= gid_max) return;
571 w[ 0] = pws[gid].i[ 0];
572 w[ 1] = pws[gid].i[ 1];
573 w[ 2] = pws[gid].i[ 2];
574 w[ 3] = pws[gid].i[ 3];
575 w[ 4] = pws[gid].i[ 4];
576 w[ 5] = pws[gid].i[ 5];
577 w[ 6] = pws[gid].i[ 6];
578 w[ 7] = pws[gid].i[ 7];
579 w[ 8] = pws[gid].i[ 8];
580 w[ 9] = pws[gid].i[ 9];
581 w[10] = pws[gid].i[10];
582 w[11] = pws[gid].i[11];
583 w[12] = pws[gid].i[12];
584 w[13] = pws[gid].i[13];
585 w[14] = pws[gid].i[14];
586 w[15] = pws[gid].i[15];
588 const u32 pw_len = pws[gid].pw_len;
594 m00000m (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, bfs_cnt, digests_cnt, digests_offset);
597 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_s04 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global 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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
603 const u32 gid = get_global_id (0);
605 if (gid >= gid_max) return;
609 w[ 0] = pws[gid].i[ 0];
610 w[ 1] = pws[gid].i[ 1];
611 w[ 2] = pws[gid].i[ 2];
612 w[ 3] = pws[gid].i[ 3];
623 w[14] = pws[gid].i[14];
626 const u32 pw_len = pws[gid].pw_len;
632 m00000s (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, bfs_cnt, digests_cnt, digests_offset);
635 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_s08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global 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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
641 const u32 gid = get_global_id (0);
643 if (gid >= gid_max) return;
647 w[ 0] = pws[gid].i[ 0];
648 w[ 1] = pws[gid].i[ 1];
649 w[ 2] = pws[gid].i[ 2];
650 w[ 3] = pws[gid].i[ 3];
651 w[ 4] = pws[gid].i[ 4];
652 w[ 5] = pws[gid].i[ 5];
653 w[ 6] = pws[gid].i[ 6];
654 w[ 7] = pws[gid].i[ 7];
661 w[14] = pws[gid].i[14];
664 const u32 pw_len = pws[gid].pw_len;
670 m00000s (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, bfs_cnt, digests_cnt, digests_offset);
673 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00000_s16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global 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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
679 const u32 gid = get_global_id (0);
681 if (gid >= gid_max) return;
685 w[ 0] = pws[gid].i[ 0];
686 w[ 1] = pws[gid].i[ 1];
687 w[ 2] = pws[gid].i[ 2];
688 w[ 3] = pws[gid].i[ 3];
689 w[ 4] = pws[gid].i[ 4];
690 w[ 5] = pws[gid].i[ 5];
691 w[ 6] = pws[gid].i[ 6];
692 w[ 7] = pws[gid].i[ 7];
693 w[ 8] = pws[gid].i[ 8];
694 w[ 9] = pws[gid].i[ 9];
695 w[10] = pws[gid].i[10];
696 w[11] = pws[gid].i[11];
697 w[12] = pws[gid].i[12];
698 w[13] = pws[gid].i[13];
699 w[14] = pws[gid].i[14];
700 w[15] = pws[gid].i[15];
702 const u32 pw_len = pws[gid].pw_len;
708 m00000s (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, bfs_cnt, digests_cnt, digests_offset);