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) m00020_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 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
102 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
103 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
104 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
105 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
109 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
110 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
111 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
112 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
114 const u32 salt_len = salt_bufs[salt_pos].salt_len;
120 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
122 const u32 pw_r_len = c_combs[il_pos].pw_len;
124 const u32 pw_len = pw_l_len + pw_r_len;
131 wordr0[0] = c_combs[il_pos].i[0];
132 wordr0[1] = c_combs[il_pos].i[1];
133 wordr0[2] = c_combs[il_pos].i[2];
134 wordr0[3] = c_combs[il_pos].i[3];
135 wordr1[0] = c_combs[il_pos].i[4];
136 wordr1[1] = c_combs[il_pos].i[5];
137 wordr1[2] = c_combs[il_pos].i[6];
138 wordr1[3] = c_combs[il_pos].i[7];
148 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
150 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
158 w0[0] = wordl0[0] | wordr0[0];
159 w0[1] = wordl0[1] | wordr0[1];
160 w0[2] = wordl0[2] | wordr0[2];
161 w0[3] = wordl0[3] | wordr0[3];
162 w1[0] = wordl1[0] | wordr1[0];
163 w1[1] = wordl1[1] | wordr1[1];
164 w1[2] = wordl1[2] | wordr1[2];
165 w1[3] = wordl1[3] | wordr1[3];
166 w2[0] = wordl2[0] | wordr2[0];
167 w2[1] = wordl2[1] | wordr2[1];
168 w2[2] = wordl2[2] | wordr2[2];
169 w2[3] = wordl2[3] | wordr2[3];
170 w3[0] = wordl3[0] | wordr3[0];
171 w3[1] = wordl3[1] | wordr3[1];
172 w3[2] = wordl3[2] | wordr3[2];
173 w3[3] = wordl3[3] | wordr3[3];
179 const u32 pw_salt_len = pw_len + salt_len;
203 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
205 w0_t[0] |= salt_buf0[0];
206 w0_t[1] |= salt_buf0[1];
207 w0_t[2] |= salt_buf0[2];
208 w0_t[3] |= salt_buf0[3];
209 w1_t[0] |= salt_buf1[0];
210 w1_t[1] |= salt_buf1[1];
211 w1_t[2] |= salt_buf1[2];
212 w1_t[3] |= salt_buf1[3];
214 append_0x80_4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
216 w3_t[2] = pw_salt_len * 8;
229 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
230 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
231 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
232 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
233 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
234 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
235 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
236 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
237 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
238 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
239 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
240 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
241 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
242 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
243 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
244 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
246 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
247 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
248 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
249 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
250 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
251 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
252 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
253 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
254 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
255 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
256 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
257 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
258 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
259 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
260 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
261 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
263 MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
264 MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
265 MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
266 MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
267 MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
268 MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
269 MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
270 MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
271 MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
272 MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
273 MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
274 MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
275 MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
276 MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
277 MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
278 MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
280 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
281 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
282 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
283 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
284 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
285 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
286 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
287 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
288 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
289 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
290 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
291 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
292 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
293 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
294 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
295 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
303 #include VECT_COMPARE_M
307 extern "C" __global__ void __launch_bounds__ (256, 1) m00020_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)
311 extern "C" __global__ void __launch_bounds__ (256, 1) m00020_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)
315 extern "C" __global__ void __launch_bounds__ (256, 1) m00020_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)
321 const u32 lid = threadIdx.x;
327 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
329 if (gid >= gid_max) return;
333 wordl0[0] = pws[gid].i[ 0];
334 wordl0[1] = pws[gid].i[ 1];
335 wordl0[2] = pws[gid].i[ 2];
336 wordl0[3] = pws[gid].i[ 3];
340 wordl1[0] = pws[gid].i[ 4];
341 wordl1[1] = pws[gid].i[ 5];
342 wordl1[2] = pws[gid].i[ 6];
343 wordl1[3] = pws[gid].i[ 7];
359 const u32 pw_l_len = pws[gid].pw_len;
361 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
363 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
372 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[0];
373 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[1];
374 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[2];
375 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[3];
379 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[4];
380 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[5];
381 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[6];
382 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[7];
384 const u32 salt_len = salt_bufs[salt_pos].salt_len;
390 const u32 search[4] =
392 digests_buf[digests_offset].digest_buf[DGST_R0],
393 digests_buf[digests_offset].digest_buf[DGST_R1],
394 digests_buf[digests_offset].digest_buf[DGST_R2],
395 digests_buf[digests_offset].digest_buf[DGST_R3]
402 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
404 const u32 pw_r_len = c_combs[il_pos].pw_len;
406 const u32 pw_len = pw_l_len + pw_r_len;
413 wordr0[0] = c_combs[il_pos].i[0];
414 wordr0[1] = c_combs[il_pos].i[1];
415 wordr0[2] = c_combs[il_pos].i[2];
416 wordr0[3] = c_combs[il_pos].i[3];
417 wordr1[0] = c_combs[il_pos].i[4];
418 wordr1[1] = c_combs[il_pos].i[5];
419 wordr1[2] = c_combs[il_pos].i[6];
420 wordr1[3] = c_combs[il_pos].i[7];
430 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
432 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
440 w0[0] = wordl0[0] | wordr0[0];
441 w0[1] = wordl0[1] | wordr0[1];
442 w0[2] = wordl0[2] | wordr0[2];
443 w0[3] = wordl0[3] | wordr0[3];
444 w1[0] = wordl1[0] | wordr1[0];
445 w1[1] = wordl1[1] | wordr1[1];
446 w1[2] = wordl1[2] | wordr1[2];
447 w1[3] = wordl1[3] | wordr1[3];
448 w2[0] = wordl2[0] | wordr2[0];
449 w2[1] = wordl2[1] | wordr2[1];
450 w2[2] = wordl2[2] | wordr2[2];
451 w2[3] = wordl2[3] | wordr2[3];
452 w3[0] = wordl3[0] | wordr3[0];
453 w3[1] = wordl3[1] | wordr3[1];
454 w3[2] = wordl3[2] | wordr3[2];
455 w3[3] = wordl3[3] | wordr3[3];
461 const u32 pw_salt_len = pw_len + salt_len;
485 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
487 w3_t[2] = pw_salt_len * 8;
489 w0_t[0] |= salt_buf0[0];
490 w0_t[1] |= salt_buf0[1];
491 w0_t[2] |= salt_buf0[2];
492 w0_t[3] |= salt_buf0[3];
493 w1_t[0] |= salt_buf1[0];
494 w1_t[1] |= salt_buf1[1];
495 w1_t[2] |= salt_buf1[2];
496 w1_t[3] |= salt_buf1[3];
498 append_0x80_4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
500 w3_t[2] = pw_salt_len * 8;
513 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
514 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
515 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
516 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
517 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
518 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
519 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
520 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
521 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
522 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
523 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
524 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
525 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
526 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
527 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
528 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
530 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
531 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
532 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
533 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
534 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
535 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
536 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
537 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
538 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
539 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
540 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
541 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
542 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
543 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
544 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
545 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
547 MD5_STEP (MD5_H1, a, b, c, d, w1_t[1], MD5C20, MD5S20);
548 MD5_STEP (MD5_H2, d, a, b, c, w2_t[0], MD5C21, MD5S21);
549 MD5_STEP (MD5_H1, c, d, a, b, w2_t[3], MD5C22, MD5S22);
550 MD5_STEP (MD5_H2, b, c, d, a, w3_t[2], MD5C23, MD5S23);
551 MD5_STEP (MD5_H1, a, b, c, d, w0_t[1], MD5C24, MD5S20);
552 MD5_STEP (MD5_H2, d, a, b, c, w1_t[0], MD5C25, MD5S21);
553 MD5_STEP (MD5_H1, c, d, a, b, w1_t[3], MD5C26, MD5S22);
554 MD5_STEP (MD5_H2, b, c, d, a, w2_t[2], MD5C27, MD5S23);
555 MD5_STEP (MD5_H1, a, b, c, d, w3_t[1], MD5C28, MD5S20);
556 MD5_STEP (MD5_H2, d, a, b, c, w0_t[0], MD5C29, MD5S21);
557 MD5_STEP (MD5_H1, c, d, a, b, w0_t[3], MD5C2a, MD5S22);
558 MD5_STEP (MD5_H2, b, c, d, a, w1_t[2], MD5C2b, MD5S23);
559 MD5_STEP (MD5_H1, a, b, c, d, w2_t[1], MD5C2c, MD5S20);
560 MD5_STEP (MD5_H2, d, a, b, c, w3_t[0], MD5C2d, MD5S21);
561 MD5_STEP (MD5_H1, c, d, a, b, w3_t[3], MD5C2e, MD5S22);
562 MD5_STEP (MD5_H2, b, c, d, a, w0_t[2], MD5C2f, MD5S23);
564 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
565 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
566 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
567 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
568 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
569 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
570 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
571 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
572 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
573 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
574 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
575 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
576 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
578 bool q_cond = (search[0] != a);
580 if (q_cond) continue;
582 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
583 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
584 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
592 #include VECT_COMPARE_S
596 extern "C" __global__ void __launch_bounds__ (256, 1) m00020_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)
600 extern "C" __global__ void __launch_bounds__ (256, 1) m00020_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)