2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
24 #include "include/kernel_functions.c"
26 #include "common_nv.c"
29 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
30 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
34 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
35 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
39 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
40 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
43 __device__ __constant__ comb_t c_combs[1024];
45 extern "C" __global__ void __launch_bounds__ (256, 1) m01100_m04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
51 const u32 lid = threadIdx.x;
57 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
59 if (gid >= gid_max) return;
63 wordl0[0] = pws[gid].i[ 0];
64 wordl0[1] = pws[gid].i[ 1];
65 wordl0[2] = pws[gid].i[ 2];
66 wordl0[3] = pws[gid].i[ 3];
70 wordl1[0] = pws[gid].i[ 4];
71 wordl1[1] = pws[gid].i[ 5];
72 wordl1[2] = pws[gid].i[ 6];
73 wordl1[3] = pws[gid].i[ 7];
89 const u32 pw_l_len = pws[gid].pw_len;
91 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
93 append_0x80_2 (wordl0, wordl1, pw_l_len);
95 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
104 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
105 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
106 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
107 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
111 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
112 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
113 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
114 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
118 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
119 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
123 const u32 salt_len = salt_bufs[salt_pos].salt_len;
129 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
131 const u32 pw_r_len = c_combs[il_pos].pw_len;
133 const u32 pw_len = pw_l_len + pw_r_len;
137 wordr0[0] = c_combs[il_pos].i[0];
138 wordr0[1] = c_combs[il_pos].i[1];
139 wordr0[2] = c_combs[il_pos].i[2];
140 wordr0[3] = c_combs[il_pos].i[3];
144 wordr1[0] = c_combs[il_pos].i[4];
145 wordr1[1] = c_combs[il_pos].i[5];
146 wordr1[2] = c_combs[il_pos].i[6];
147 wordr1[3] = c_combs[il_pos].i[7];
163 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
165 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
170 w0[0] = wordl0[0] | wordr0[0];
171 w0[1] = wordl0[1] | wordr0[1];
172 w0[2] = wordl0[2] | wordr0[2];
173 w0[3] = wordl0[3] | wordr0[3];
177 w1[0] = wordl1[0] | wordr1[0];
178 w1[1] = wordl1[1] | wordr1[1];
179 w1[2] = wordl1[2] | wordr1[2];
180 w1[3] = wordl1[3] | wordr1[3];
201 make_unicode (w0, w0_t, w1_t);
202 make_unicode (w1, w2_t, w3_t);
204 w3_t[2] = pw_len * 8 * 2;
211 MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00);
212 MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01);
213 MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02);
214 MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03);
215 MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00);
216 MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01);
217 MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02);
218 MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03);
219 MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00);
220 MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01);
221 MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02);
222 MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03);
223 MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00);
224 MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01);
225 MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02);
226 MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03);
228 MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
229 MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
230 MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
231 MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
232 MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
233 MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
234 MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
235 MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
236 MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
237 MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
238 MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
239 MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
240 MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
241 MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
242 MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
243 MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
245 MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
246 MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
247 MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
248 MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
249 MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
250 MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
251 MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
252 MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
253 MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
254 MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
255 MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
256 MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
257 MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
258 MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
259 MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
260 MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
271 w1_t[0] = salt_buf0[0];
272 w1_t[1] = salt_buf0[1];
273 w1_t[2] = salt_buf0[2];
274 w1_t[3] = salt_buf0[3];
275 w2_t[0] = salt_buf1[0];
276 w2_t[1] = salt_buf1[1];
277 w2_t[2] = salt_buf1[2];
278 w2_t[3] = salt_buf1[3];
279 w3_t[0] = salt_buf2[0];
280 w3_t[1] = salt_buf2[1];
281 w3_t[2] = (16 + salt_len) * 8;
289 MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00);
290 MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01);
291 MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02);
292 MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03);
293 MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00);
294 MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01);
295 MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02);
296 MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03);
297 MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00);
298 MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01);
299 MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02);
300 MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03);
301 MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00);
302 MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01);
303 MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02);
304 MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03);
306 MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
307 MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
308 MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
309 MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
310 MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
311 MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
312 MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
313 MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
314 MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
315 MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
316 MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
317 MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
318 MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
319 MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
320 MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
321 MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
323 MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
324 MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
325 MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
326 MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
327 MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
328 MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
329 MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
330 MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
331 MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
332 MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
333 MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
334 MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
335 MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
336 MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
337 MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
338 MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
345 #include VECT_COMPARE_M
349 extern "C" __global__ void __launch_bounds__ (256, 1) m01100_m08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
353 extern "C" __global__ void __launch_bounds__ (256, 1) m01100_m16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
357 extern "C" __global__ void __launch_bounds__ (256, 1) m01100_s04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, 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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
363 const u32 lid = threadIdx.x;
369 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
371 if (gid >= gid_max) return;
375 wordl0[0] = pws[gid].i[ 0];
376 wordl0[1] = pws[gid].i[ 1];
377 wordl0[2] = pws[gid].i[ 2];
378 wordl0[3] = pws[gid].i[ 3];
382 wordl1[0] = pws[gid].i[ 4];
383 wordl1[1] = pws[gid].i[ 5];
384 wordl1[2] = pws[gid].i[ 6];
385 wordl1[3] = pws[gid].i[ 7];
401 const u32 pw_l_len = pws[gid].pw_len;
403 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
405 append_0x80_2 (wordl0, wordl1, pw_l_len);
407 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
414 const u32 search[4] =
416 digests_buf[digests_offset].digest_buf[DGST_R0],
417 digests_buf[digests_offset].digest_buf[DGST_R1],
418 digests_buf[digests_offset].digest_buf[DGST_R2],
419 digests_buf[digests_offset].digest_buf[DGST_R3]
428 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
429 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
430 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
431 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
435 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
436 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
437 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
438 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
442 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
443 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
447 const u32 salt_len = salt_bufs[salt_pos].salt_len;
453 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
455 const u32 pw_r_len = c_combs[il_pos].pw_len;
457 const u32 pw_len = pw_l_len + pw_r_len;
461 wordr0[0] = c_combs[il_pos].i[0];
462 wordr0[1] = c_combs[il_pos].i[1];
463 wordr0[2] = c_combs[il_pos].i[2];
464 wordr0[3] = c_combs[il_pos].i[3];
468 wordr1[0] = c_combs[il_pos].i[4];
469 wordr1[1] = c_combs[il_pos].i[5];
470 wordr1[2] = c_combs[il_pos].i[6];
471 wordr1[3] = c_combs[il_pos].i[7];
487 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
489 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
494 w0[0] = wordl0[0] | wordr0[0];
495 w0[1] = wordl0[1] | wordr0[1];
496 w0[2] = wordl0[2] | wordr0[2];
497 w0[3] = wordl0[3] | wordr0[3];
501 w1[0] = wordl1[0] | wordr1[0];
502 w1[1] = wordl1[1] | wordr1[1];
503 w1[2] = wordl1[2] | wordr1[2];
504 w1[3] = wordl1[3] | wordr1[3];
525 make_unicode (w0, w0_t, w1_t);
526 make_unicode (w1, w2_t, w3_t);
528 w3_t[2] = pw_len * 8 * 2;
535 MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00);
536 MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01);
537 MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02);
538 MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03);
539 MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00);
540 MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01);
541 MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02);
542 MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03);
543 MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00);
544 MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01);
545 MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02);
546 MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03);
547 MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00);
548 MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01);
549 MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02);
550 MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03);
552 MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
553 MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
554 MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
555 MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
556 MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
557 MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
558 MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
559 MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
560 MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
561 MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
562 MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
563 MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
564 MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
565 MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
566 MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
567 MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
569 MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
570 MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
571 MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
572 MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
573 MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
574 MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
575 MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
576 MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
577 MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
578 MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
579 MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
580 MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
581 MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
582 MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
583 MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
584 MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
595 w1_t[0] = salt_buf0[0];
596 w1_t[1] = salt_buf0[1];
597 w1_t[2] = salt_buf0[2];
598 w1_t[3] = salt_buf0[3];
599 w2_t[0] = salt_buf1[0];
600 w2_t[1] = salt_buf1[1];
601 w2_t[2] = salt_buf1[2];
602 w2_t[3] = salt_buf1[3];
603 w3_t[0] = salt_buf2[0];
604 w3_t[1] = salt_buf2[1];
605 w3_t[2] = (16 + salt_len) * 8;
613 MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00);
614 MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01);
615 MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02);
616 MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03);
617 MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00);
618 MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01);
619 MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02);
620 MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03);
621 MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00);
622 MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01);
623 MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02);
624 MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03);
625 MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00);
626 MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01);
627 MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02);
628 MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03);
630 MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
631 MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
632 MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
633 MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
634 MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
635 MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
636 MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
637 MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
638 MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
639 MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
640 MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
641 MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
642 MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
643 MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
644 MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
645 MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
647 MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
648 MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
649 MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
650 MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
651 MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
652 MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
653 MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
654 MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
655 MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
656 MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
657 MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
658 MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
659 MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
661 bool q_cond = (search[0] != a);
663 if (q_cond) continue;
665 MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
666 MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
667 MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
674 #include VECT_COMPARE_S
678 extern "C" __global__ void __launch_bounds__ (256, 1) m01100_s08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
682 extern "C" __global__ void __launch_bounds__ (256, 1) m01100_s16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, 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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)