2 * Author......: Jens Steube <jens.steube@gmail.com>
9 #include "include/constants.h"
10 #include "include/kernel_vendor.h"
25 #include "include/kernel_functions.c"
27 #include "common_nv.c"
30 #define VECT_COMPARE_S "check_single_vect1_comp4_warp.c"
31 #define VECT_COMPARE_M "check_multi_vect1_comp4_warp.c"
35 #define VECT_COMPARE_S "check_single_vect2_comp4_warp.c"
36 #define VECT_COMPARE_M "check_multi_vect2_comp4_warp.c"
40 #define VECT_COMPARE_S "check_single_vect4_comp4_warp.c"
41 #define VECT_COMPARE_M "check_multi_vect4_comp4_warp.c"
44 __device__ __constant__ u32x c_bfs[1024];
46 __device__ static void m02410m (u32 w[16], const u32 pw_len, const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x *words_buf_r, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
52 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
53 const u32 lid = threadIdx.x;
61 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
87 switch_buffer_by_offset (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
94 const u32 salt_len = (salt_bufs[salt_pos].salt_len < 4) ? salt_bufs[salt_pos].salt_len : 4;
96 const u32 pw_salt_len = pw_len + salt_len;
98 truncate_block (w, pw_salt_len);
111 const u32 F_w0c00 = 0 + MD5C00;
112 const u32 F_w1c01 = w[ 1] + MD5C01;
113 const u32 F_w2c02 = w[ 2] + MD5C02;
114 const u32 F_w3c03 = w[ 3] + MD5C03;
115 const u32 F_w4c04 = w[ 4] + MD5C04;
116 const u32 F_w5c05 = w[ 5] + MD5C05;
117 const u32 F_w6c06 = w[ 6] + MD5C06;
118 const u32 F_w7c07 = w[ 7] + MD5C07;
119 const u32 F_w8c08 = w[ 8] + MD5C08;
120 const u32 F_w9c09 = w[ 9] + MD5C09;
121 const u32 F_wac0a = w[10] + MD5C0a;
122 const u32 F_wbc0b = w[11] + MD5C0b;
123 const u32 F_wcc0c = w[12] + MD5C0c;
124 const u32 F_wdc0d = w[13] + MD5C0d;
125 const u32 F_wec0e = w[14] + MD5C0e;
126 const u32 F_wfc0f = w[15] + MD5C0f;
128 const u32 G_w1c10 = w[ 1] + MD5C10;
129 const u32 G_w6c11 = w[ 6] + MD5C11;
130 const u32 G_wbc12 = w[11] + MD5C12;
131 const u32 G_w0c13 = 0 + MD5C13;
132 const u32 G_w5c14 = w[ 5] + MD5C14;
133 const u32 G_wac15 = w[10] + MD5C15;
134 const u32 G_wfc16 = w[15] + MD5C16;
135 const u32 G_w4c17 = w[ 4] + MD5C17;
136 const u32 G_w9c18 = w[ 9] + MD5C18;
137 const u32 G_wec19 = w[14] + MD5C19;
138 const u32 G_w3c1a = w[ 3] + MD5C1a;
139 const u32 G_w8c1b = w[ 8] + MD5C1b;
140 const u32 G_wdc1c = w[13] + MD5C1c;
141 const u32 G_w2c1d = w[ 2] + MD5C1d;
142 const u32 G_w7c1e = w[ 7] + MD5C1e;
143 const u32 G_wcc1f = w[12] + MD5C1f;
145 const u32 H_w5c20 = w[ 5] + MD5C20;
146 const u32 H_w8c21 = w[ 8] + MD5C21;
147 const u32 H_wbc22 = w[11] + MD5C22;
148 const u32 H_wec23 = w[14] + MD5C23;
149 const u32 H_w1c24 = w[ 1] + MD5C24;
150 const u32 H_w4c25 = w[ 4] + MD5C25;
151 const u32 H_w7c26 = w[ 7] + MD5C26;
152 const u32 H_wac27 = w[10] + MD5C27;
153 const u32 H_wdc28 = w[13] + MD5C28;
154 const u32 H_w0c29 = 0 + MD5C29;
155 const u32 H_w3c2a = w[ 3] + MD5C2a;
156 const u32 H_w6c2b = w[ 6] + MD5C2b;
157 const u32 H_w9c2c = w[ 9] + MD5C2c;
158 const u32 H_wcc2d = w[12] + MD5C2d;
159 const u32 H_wfc2e = w[15] + MD5C2e;
160 const u32 H_w2c2f = w[ 2] + MD5C2f;
162 const u32 I_w0c30 = 0 + MD5C30;
163 const u32 I_w7c31 = w[ 7] + MD5C31;
164 const u32 I_wec32 = w[14] + MD5C32;
165 const u32 I_w5c33 = w[ 5] + MD5C33;
166 const u32 I_wcc34 = w[12] + MD5C34;
167 const u32 I_w3c35 = w[ 3] + MD5C35;
168 const u32 I_wac36 = w[10] + MD5C36;
169 const u32 I_w1c37 = w[ 1] + MD5C37;
170 const u32 I_w8c38 = w[ 8] + MD5C38;
171 const u32 I_wfc39 = w[15] + MD5C39;
172 const u32 I_w6c3a = w[ 6] + MD5C3a;
173 const u32 I_wdc3b = w[13] + MD5C3b;
174 const u32 I_w4c3c = w[ 4] + MD5C3c;
175 const u32 I_wbc3d = w[11] + MD5C3d;
176 const u32 I_w2c3e = w[ 2] + MD5C3e;
177 const u32 I_w9c3f = w[ 9] + MD5C3f;
183 const u32 bf_loops = ceil ((float) bfs_cnt / VECT_DIV);
187 for (u32 il_pos = 0; il_pos < bf_loops; il_pos++)
189 const u32x w0r = c_bfs[il_pos];
191 const u32x w0 = w0l | w0r;
200 MD5_STEP (MD5_Fo, a, b, c, d, w0, F_w0c00, MD5S00);
201 MD5_STEP0(MD5_Fo, d, a, b, c, F_w1c01, MD5S01);
202 MD5_STEP0(MD5_Fo, c, d, a, b, F_w2c02, MD5S02);
203 MD5_STEP0(MD5_Fo, b, c, d, a, F_w3c03, MD5S03);
204 MD5_STEP0(MD5_Fo, a, b, c, d, F_w4c04, MD5S00);
205 MD5_STEP0(MD5_Fo, d, a, b, c, F_w5c05, MD5S01);
206 MD5_STEP0(MD5_Fo, c, d, a, b, F_w6c06, MD5S02);
207 MD5_STEP0(MD5_Fo, b, c, d, a, F_w7c07, MD5S03);
208 MD5_STEP0(MD5_Fo, a, b, c, d, F_w8c08, MD5S00);
209 MD5_STEP0(MD5_Fo, d, a, b, c, F_w9c09, MD5S01);
210 MD5_STEP0(MD5_Fo, c, d, a, b, F_wac0a, MD5S02);
211 MD5_STEP0(MD5_Fo, b, c, d, a, F_wbc0b, MD5S03);
212 MD5_STEP0(MD5_Fo, a, b, c, d, F_wcc0c, MD5S00);
213 MD5_STEP0(MD5_Fo, d, a, b, c, F_wdc0d, MD5S01);
214 MD5_STEP0(MD5_Fo, c, d, a, b, F_wec0e, MD5S02);
215 MD5_STEP0(MD5_Fo, b, c, d, a, F_wfc0f, MD5S03);
217 MD5_STEP0(MD5_Go, a, b, c, d, G_w1c10, MD5S10);
218 MD5_STEP0(MD5_Go, d, a, b, c, G_w6c11, MD5S11);
219 MD5_STEP0(MD5_Go, c, d, a, b, G_wbc12, MD5S12);
220 MD5_STEP (MD5_Go, b, c, d, a, w0, G_w0c13, MD5S13);
221 MD5_STEP0(MD5_Go, a, b, c, d, G_w5c14, MD5S10);
222 MD5_STEP0(MD5_Go, d, a, b, c, G_wac15, MD5S11);
223 MD5_STEP0(MD5_Go, c, d, a, b, G_wfc16, MD5S12);
224 MD5_STEP0(MD5_Go, b, c, d, a, G_w4c17, MD5S13);
225 MD5_STEP0(MD5_Go, a, b, c, d, G_w9c18, MD5S10);
226 MD5_STEP0(MD5_Go, d, a, b, c, G_wec19, MD5S11);
227 MD5_STEP0(MD5_Go, c, d, a, b, G_w3c1a, MD5S12);
228 MD5_STEP0(MD5_Go, b, c, d, a, G_w8c1b, MD5S13);
229 MD5_STEP0(MD5_Go, a, b, c, d, G_wdc1c, MD5S10);
230 MD5_STEP0(MD5_Go, d, a, b, c, G_w2c1d, MD5S11);
231 MD5_STEP0(MD5_Go, c, d, a, b, G_w7c1e, MD5S12);
232 MD5_STEP0(MD5_Go, b, c, d, a, G_wcc1f, MD5S13);
234 MD5_STEP0(MD5_H1, a, b, c, d, H_w5c20, MD5S20);
235 MD5_STEP0(MD5_H2, d, a, b, c, H_w8c21, MD5S21);
236 MD5_STEP0(MD5_H1, c, d, a, b, H_wbc22, MD5S22);
237 MD5_STEP0(MD5_H2, b, c, d, a, H_wec23, MD5S23);
238 MD5_STEP0(MD5_H1, a, b, c, d, H_w1c24, MD5S20);
239 MD5_STEP0(MD5_H2, d, a, b, c, H_w4c25, MD5S21);
240 MD5_STEP0(MD5_H1, c, d, a, b, H_w7c26, MD5S22);
241 MD5_STEP0(MD5_H2, b, c, d, a, H_wac27, MD5S23);
242 MD5_STEP0(MD5_H1, a, b, c, d, H_wdc28, MD5S20);
243 MD5_STEP (MD5_H2, d, a, b, c, w0, H_w0c29, MD5S21);
244 MD5_STEP0(MD5_H1, c, d, a, b, H_w3c2a, MD5S22);
245 MD5_STEP0(MD5_H2, b, c, d, a, H_w6c2b, MD5S23);
246 MD5_STEP0(MD5_H1, a, b, c, d, H_w9c2c, MD5S20);
247 MD5_STEP0(MD5_H2, d, a, b, c, H_wcc2d, MD5S21);
248 MD5_STEP0(MD5_H1, c, d, a, b, H_wfc2e, MD5S22);
249 MD5_STEP0(MD5_H2, b, c, d, a, H_w2c2f, MD5S23);
251 MD5_STEP (MD5_I , a, b, c, d, w0, I_w0c30, MD5S30);
252 MD5_STEP0(MD5_I , d, a, b, c, I_w7c31, MD5S31);
253 MD5_STEP0(MD5_I , c, d, a, b, I_wec32, MD5S32);
254 MD5_STEP0(MD5_I , b, c, d, a, I_w5c33, MD5S33);
255 MD5_STEP0(MD5_I , a, b, c, d, I_wcc34, MD5S30);
256 MD5_STEP0(MD5_I , d, a, b, c, I_w3c35, MD5S31);
257 MD5_STEP0(MD5_I , c, d, a, b, I_wac36, MD5S32);
258 MD5_STEP0(MD5_I , b, c, d, a, I_w1c37, MD5S33);
259 MD5_STEP0(MD5_I , a, b, c, d, I_w8c38, MD5S30);
260 MD5_STEP0(MD5_I , d, a, b, c, I_wfc39, MD5S31);
261 MD5_STEP0(MD5_I , c, d, a, b, I_w6c3a, MD5S32);
262 MD5_STEP0(MD5_I , b, c, d, a, I_wdc3b, MD5S33);
263 MD5_STEP0(MD5_I , a, b, c, d, I_w4c3c, MD5S30);
264 MD5_STEP0(MD5_I , d, a, b, c, I_wbc3d, MD5S31);
265 MD5_STEP0(MD5_I , c, d, a, b, I_w2c3e, MD5S32);
266 MD5_STEP0(MD5_I , b, c, d, a, I_w9c3f, MD5S33);
278 #include VECT_COMPARE_M
282 __device__ static void m02410s (u32 w[16], const u32 pw_len, const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x *words_buf_r, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
288 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
289 const u32 lid = threadIdx.x;
297 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
323 switch_buffer_by_offset (salt_buf0, salt_buf1, salt_buf2, salt_buf3, pw_len);
325 w[0] |= salt_buf0[0];
326 w[1] |= salt_buf0[1];
327 w[2] |= salt_buf0[2];
328 w[3] |= salt_buf0[3];
330 const u32 salt_len = (salt_bufs[salt_pos].salt_len < 4) ? salt_bufs[salt_pos].salt_len : 4;
332 const u32 pw_salt_len = pw_len + salt_len;
334 truncate_block (w, pw_salt_len);
347 const u32 F_w0c00 = 0 + MD5C00;
348 const u32 F_w1c01 = w[ 1] + MD5C01;
349 const u32 F_w2c02 = w[ 2] + MD5C02;
350 const u32 F_w3c03 = w[ 3] + MD5C03;
351 const u32 F_w4c04 = w[ 4] + MD5C04;
352 const u32 F_w5c05 = w[ 5] + MD5C05;
353 const u32 F_w6c06 = w[ 6] + MD5C06;
354 const u32 F_w7c07 = w[ 7] + MD5C07;
355 const u32 F_w8c08 = w[ 8] + MD5C08;
356 const u32 F_w9c09 = w[ 9] + MD5C09;
357 const u32 F_wac0a = w[10] + MD5C0a;
358 const u32 F_wbc0b = w[11] + MD5C0b;
359 const u32 F_wcc0c = w[12] + MD5C0c;
360 const u32 F_wdc0d = w[13] + MD5C0d;
361 const u32 F_wec0e = w[14] + MD5C0e;
362 const u32 F_wfc0f = w[15] + MD5C0f;
364 const u32 G_w1c10 = w[ 1] + MD5C10;
365 const u32 G_w6c11 = w[ 6] + MD5C11;
366 const u32 G_wbc12 = w[11] + MD5C12;
367 const u32 G_w0c13 = 0 + MD5C13;
368 const u32 G_w5c14 = w[ 5] + MD5C14;
369 const u32 G_wac15 = w[10] + MD5C15;
370 const u32 G_wfc16 = w[15] + MD5C16;
371 const u32 G_w4c17 = w[ 4] + MD5C17;
372 const u32 G_w9c18 = w[ 9] + MD5C18;
373 const u32 G_wec19 = w[14] + MD5C19;
374 const u32 G_w3c1a = w[ 3] + MD5C1a;
375 const u32 G_w8c1b = w[ 8] + MD5C1b;
376 const u32 G_wdc1c = w[13] + MD5C1c;
377 const u32 G_w2c1d = w[ 2] + MD5C1d;
378 const u32 G_w7c1e = w[ 7] + MD5C1e;
379 const u32 G_wcc1f = w[12] + MD5C1f;
381 const u32 H_w5c20 = w[ 5] + MD5C20;
382 const u32 H_w8c21 = w[ 8] + MD5C21;
383 const u32 H_wbc22 = w[11] + MD5C22;
384 const u32 H_wec23 = w[14] + MD5C23;
385 const u32 H_w1c24 = w[ 1] + MD5C24;
386 const u32 H_w4c25 = w[ 4] + MD5C25;
387 const u32 H_w7c26 = w[ 7] + MD5C26;
388 const u32 H_wac27 = w[10] + MD5C27;
389 const u32 H_wdc28 = w[13] + MD5C28;
390 const u32 H_w0c29 = 0 + MD5C29;
391 const u32 H_w3c2a = w[ 3] + MD5C2a;
392 const u32 H_w6c2b = w[ 6] + MD5C2b;
393 const u32 H_w9c2c = w[ 9] + MD5C2c;
394 const u32 H_wcc2d = w[12] + MD5C2d;
395 const u32 H_wfc2e = w[15] + MD5C2e;
396 const u32 H_w2c2f = w[ 2] + MD5C2f;
398 const u32 I_w0c30 = 0 + MD5C30;
399 const u32 I_w7c31 = w[ 7] + MD5C31;
400 const u32 I_wec32 = w[14] + MD5C32;
401 const u32 I_w5c33 = w[ 5] + MD5C33;
402 const u32 I_wcc34 = w[12] + MD5C34;
403 const u32 I_w3c35 = w[ 3] + MD5C35;
404 const u32 I_wac36 = w[10] + MD5C36;
405 const u32 I_w1c37 = w[ 1] + MD5C37;
406 const u32 I_w8c38 = w[ 8] + MD5C38;
407 const u32 I_wfc39 = w[15] + MD5C39;
408 const u32 I_w6c3a = w[ 6] + MD5C3a;
409 const u32 I_wdc3b = w[13] + MD5C3b;
410 const u32 I_w4c3c = w[ 4] + MD5C3c;
411 const u32 I_wbc3d = w[11] + MD5C3d;
412 const u32 I_w2c3e = w[ 2] + MD5C3e;
413 const u32 I_w9c3f = w[ 9] + MD5C3f;
419 const u32 search[4] =
421 digests_buf[digests_offset].digest_buf[DGST_R0],
422 digests_buf[digests_offset].digest_buf[DGST_R1],
423 digests_buf[digests_offset].digest_buf[DGST_R2],
424 digests_buf[digests_offset].digest_buf[DGST_R3]
431 const u32 bf_loops = ceil ((float) bfs_cnt / VECT_DIV);
435 for (u32 il_pos = 0; il_pos < bf_loops; il_pos++)
437 const u32x w0r = c_bfs[il_pos];
439 const u32x w0 = w0l | w0r;
448 MD5_STEP (MD5_Fo, a, b, c, d, w0, F_w0c00, MD5S00);
449 MD5_STEP0(MD5_Fo, d, a, b, c, F_w1c01, MD5S01);
450 MD5_STEP0(MD5_Fo, c, d, a, b, F_w2c02, MD5S02);
451 MD5_STEP0(MD5_Fo, b, c, d, a, F_w3c03, MD5S03);
452 MD5_STEP0(MD5_Fo, a, b, c, d, F_w4c04, MD5S00);
453 MD5_STEP0(MD5_Fo, d, a, b, c, F_w5c05, MD5S01);
454 MD5_STEP0(MD5_Fo, c, d, a, b, F_w6c06, MD5S02);
455 MD5_STEP0(MD5_Fo, b, c, d, a, F_w7c07, MD5S03);
456 MD5_STEP0(MD5_Fo, a, b, c, d, F_w8c08, MD5S00);
457 MD5_STEP0(MD5_Fo, d, a, b, c, F_w9c09, MD5S01);
458 MD5_STEP0(MD5_Fo, c, d, a, b, F_wac0a, MD5S02);
459 MD5_STEP0(MD5_Fo, b, c, d, a, F_wbc0b, MD5S03);
460 MD5_STEP0(MD5_Fo, a, b, c, d, F_wcc0c, MD5S00);
461 MD5_STEP0(MD5_Fo, d, a, b, c, F_wdc0d, MD5S01);
462 MD5_STEP0(MD5_Fo, c, d, a, b, F_wec0e, MD5S02);
463 MD5_STEP0(MD5_Fo, b, c, d, a, F_wfc0f, MD5S03);
465 MD5_STEP0(MD5_Go, a, b, c, d, G_w1c10, MD5S10);
466 MD5_STEP0(MD5_Go, d, a, b, c, G_w6c11, MD5S11);
467 MD5_STEP0(MD5_Go, c, d, a, b, G_wbc12, MD5S12);
468 MD5_STEP (MD5_Go, b, c, d, a, w0, G_w0c13, MD5S13);
469 MD5_STEP0(MD5_Go, a, b, c, d, G_w5c14, MD5S10);
470 MD5_STEP0(MD5_Go, d, a, b, c, G_wac15, MD5S11);
471 MD5_STEP0(MD5_Go, c, d, a, b, G_wfc16, MD5S12);
472 MD5_STEP0(MD5_Go, b, c, d, a, G_w4c17, MD5S13);
473 MD5_STEP0(MD5_Go, a, b, c, d, G_w9c18, MD5S10);
474 MD5_STEP0(MD5_Go, d, a, b, c, G_wec19, MD5S11);
475 MD5_STEP0(MD5_Go, c, d, a, b, G_w3c1a, MD5S12);
476 MD5_STEP0(MD5_Go, b, c, d, a, G_w8c1b, MD5S13);
477 MD5_STEP0(MD5_Go, a, b, c, d, G_wdc1c, MD5S10);
478 MD5_STEP0(MD5_Go, d, a, b, c, G_w2c1d, MD5S11);
479 MD5_STEP0(MD5_Go, c, d, a, b, G_w7c1e, MD5S12);
480 MD5_STEP0(MD5_Go, b, c, d, a, G_wcc1f, MD5S13);
482 MD5_STEP0(MD5_H1, a, b, c, d, H_w5c20, MD5S20);
483 MD5_STEP0(MD5_H2, d, a, b, c, H_w8c21, MD5S21);
484 MD5_STEP0(MD5_H1, c, d, a, b, H_wbc22, MD5S22);
485 MD5_STEP0(MD5_H2, b, c, d, a, H_wec23, MD5S23);
486 MD5_STEP0(MD5_H1, a, b, c, d, H_w1c24, MD5S20);
487 MD5_STEP0(MD5_H2, d, a, b, c, H_w4c25, MD5S21);
488 MD5_STEP0(MD5_H1, c, d, a, b, H_w7c26, MD5S22);
489 MD5_STEP0(MD5_H2, b, c, d, a, H_wac27, MD5S23);
490 MD5_STEP0(MD5_H1, a, b, c, d, H_wdc28, MD5S20);
491 MD5_STEP (MD5_H2, d, a, b, c, w0, H_w0c29, MD5S21);
492 MD5_STEP0(MD5_H1, c, d, a, b, H_w3c2a, MD5S22);
493 MD5_STEP0(MD5_H2, b, c, d, a, H_w6c2b, MD5S23);
494 MD5_STEP0(MD5_H1, a, b, c, d, H_w9c2c, MD5S20);
495 MD5_STEP0(MD5_H2, d, a, b, c, H_wcc2d, MD5S21);
496 MD5_STEP0(MD5_H1, c, d, a, b, H_wfc2e, MD5S22);
497 MD5_STEP0(MD5_H2, b, c, d, a, H_w2c2f, MD5S23);
499 MD5_STEP (MD5_I , a, b, c, d, w0, I_w0c30, MD5S30);
500 MD5_STEP0(MD5_I , d, a, b, c, I_w7c31, MD5S31);
501 MD5_STEP0(MD5_I , c, d, a, b, I_wec32, MD5S32);
502 MD5_STEP0(MD5_I , b, c, d, a, I_w5c33, MD5S33);
503 MD5_STEP0(MD5_I , a, b, c, d, I_wcc34, MD5S30);
504 MD5_STEP0(MD5_I , d, a, b, c, I_w3c35, MD5S31);
505 MD5_STEP0(MD5_I , c, d, a, b, I_wac36, MD5S32);
506 MD5_STEP0(MD5_I , b, c, d, a, I_w1c37, MD5S33);
507 MD5_STEP0(MD5_I , a, b, c, d, I_w8c38, MD5S30);
508 MD5_STEP0(MD5_I , d, a, b, c, I_wfc39, MD5S31);
509 MD5_STEP0(MD5_I , c, d, a, b, I_w6c3a, MD5S32);
510 MD5_STEP0(MD5_I , b, c, d, a, I_wdc3b, MD5S33);
511 MD5_STEP0(MD5_I , a, b, c, d, I_w4c3c, MD5S30);
513 bool q_cond = ((a & 0x00ffffff) != search[0]);
515 if (q_cond) continue;
517 MD5_STEP0(MD5_I , d, a, b, c, I_wbc3d, MD5S31);
518 MD5_STEP0(MD5_I , c, d, a, b, I_w2c3e, MD5S32);
519 MD5_STEP0(MD5_I , b, c, d, a, I_w9c3f, MD5S33);
531 #include VECT_COMPARE_S
535 extern "C" __global__ void __launch_bounds__ (256, 1) m02410_m04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x *words_buf_r, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
541 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
543 if (gid >= gid_max) return;
547 w[ 0] = pws[gid].i[ 0];
548 w[ 1] = pws[gid].i[ 1];
549 w[ 2] = pws[gid].i[ 2];
550 w[ 3] = pws[gid].i[ 3];
564 const u32 pw_len = pws[gid].pw_len;
570 m02410m (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);
573 extern "C" __global__ void __launch_bounds__ (256, 1) m02410_m08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x *words_buf_r, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
577 extern "C" __global__ void __launch_bounds__ (256, 1) m02410_m16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x *words_buf_r, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
581 extern "C" __global__ void __launch_bounds__ (256, 1) m02410_s04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x *words_buf_r, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
587 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
589 if (gid >= gid_max) return;
593 w[ 0] = pws[gid].i[ 0];
594 w[ 1] = pws[gid].i[ 1];
595 w[ 2] = pws[gid].i[ 2];
596 w[ 3] = pws[gid].i[ 3];
610 const u32 pw_len = pws[gid].pw_len;
616 m02410s (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);
619 extern "C" __global__ void __launch_bounds__ (256, 1) m02410_s08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x *words_buf_r, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
623 extern "C" __global__ void __launch_bounds__ (256, 1) m02410_s16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const u32x *words_buf_r, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)