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) m03800_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];
128 const u32 salt_len = salt_bufs[salt_pos].salt_len;
134 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
136 const u32 pw_r_len = c_combs[il_pos].pw_len;
138 const u32 pw_len = pw_l_len + pw_r_len;
142 wordr0[0] = c_combs[il_pos].i[0];
143 wordr0[1] = c_combs[il_pos].i[1];
144 wordr0[2] = c_combs[il_pos].i[2];
145 wordr0[3] = c_combs[il_pos].i[3];
149 wordr1[0] = c_combs[il_pos].i[4];
150 wordr1[1] = c_combs[il_pos].i[5];
151 wordr1[2] = c_combs[il_pos].i[6];
152 wordr1[3] = c_combs[il_pos].i[7];
168 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
170 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
175 w0[0] = wordl0[0] | wordr0[0];
176 w0[1] = wordl0[1] | wordr0[1];
177 w0[2] = wordl0[2] | wordr0[2];
178 w0[3] = wordl0[3] | wordr0[3];
182 w1[0] = wordl1[0] | wordr1[0];
183 w1[1] = wordl1[1] | wordr1[1];
184 w1[2] = wordl1[2] | wordr1[2];
185 w1[3] = wordl1[3] | wordr1[3];
189 w2[0] = wordl2[0] | wordr2[0];
190 w2[1] = wordl2[1] | wordr2[1];
191 w2[2] = wordl2[2] | wordr2[2];
192 w2[3] = wordl2[3] | wordr2[3];
196 w3[0] = wordl3[0] | wordr3[0];
197 w3[1] = wordl3[1] | wordr3[1];
233 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
235 w0_t[0] |= salt_buf0[0];
236 w0_t[1] |= salt_buf0[1];
237 w0_t[2] |= salt_buf0[2];
238 w0_t[3] |= salt_buf0[3];
239 w1_t[0] |= salt_buf1[0];
240 w1_t[1] |= salt_buf1[1];
241 w1_t[2] |= salt_buf1[2];
242 w1_t[3] |= salt_buf1[3];
243 w2_t[0] |= salt_buf2[0];
244 w2_t[1] |= salt_buf2[1];
245 w2_t[2] |= salt_buf2[2];
246 w2_t[3] |= salt_buf2[3];
247 w3_t[0] |= salt_buf3[0];
248 w3_t[1] |= salt_buf3[1];
249 w3_t[2] |= salt_buf3[2];
250 w3_t[3] |= salt_buf3[3];
258 s0[0] = salt_buf0[0];
259 s0[1] = salt_buf0[1];
260 s0[2] = salt_buf0[2];
261 s0[3] = salt_buf0[3];
265 s1[0] = salt_buf1[0];
266 s1[1] = salt_buf1[1];
267 s1[2] = salt_buf1[2];
268 s1[3] = salt_buf1[3];
284 switch_buffer_by_offset (s0, s1, s2, s3, salt_len + pw_len);
303 const u32 pw_salt_len = salt_len + pw_len + salt_len;
305 append_0x80_4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
307 w3_t[2] = pw_salt_len * 8;
318 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
319 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
320 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
321 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
322 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
323 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
324 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
325 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
326 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
327 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
328 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
329 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
330 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
331 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
332 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
333 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
335 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
336 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
337 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
338 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
339 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
340 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
341 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
342 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
343 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
344 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
345 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
346 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
347 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
348 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
349 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
350 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
352 MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
353 MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
354 MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
355 MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
356 MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
357 MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
358 MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
359 MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
360 MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
361 MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
362 MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
363 MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
364 MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
365 MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
366 MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
367 MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
369 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
370 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
371 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
372 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
373 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
374 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
375 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
376 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
377 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
378 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
379 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
380 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
381 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
383 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
384 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
385 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
392 #include VECT_COMPARE_M
396 extern "C" __global__ void __launch_bounds__ (256, 1) m03800_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)
400 extern "C" __global__ void __launch_bounds__ (256, 1) m03800_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)
404 extern "C" __global__ void __launch_bounds__ (256, 1) m03800_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)
410 const u32 lid = threadIdx.x;
416 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
418 if (gid >= gid_max) return;
422 wordl0[0] = pws[gid].i[ 0];
423 wordl0[1] = pws[gid].i[ 1];
424 wordl0[2] = pws[gid].i[ 2];
425 wordl0[3] = pws[gid].i[ 3];
429 wordl1[0] = pws[gid].i[ 4];
430 wordl1[1] = pws[gid].i[ 5];
431 wordl1[2] = pws[gid].i[ 6];
432 wordl1[3] = pws[gid].i[ 7];
448 const u32 pw_l_len = pws[gid].pw_len;
450 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
452 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
461 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
462 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
463 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
464 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
468 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
469 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
470 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
471 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
487 const u32 salt_len = salt_bufs[salt_pos].salt_len;
493 const u32 search[4] =
495 digests_buf[digests_offset].digest_buf[DGST_R0],
496 digests_buf[digests_offset].digest_buf[DGST_R1],
497 digests_buf[digests_offset].digest_buf[DGST_R2],
498 digests_buf[digests_offset].digest_buf[DGST_R3]
505 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
507 const u32 pw_r_len = c_combs[il_pos].pw_len;
509 const u32 pw_len = pw_l_len + pw_r_len;
513 wordr0[0] = c_combs[il_pos].i[0];
514 wordr0[1] = c_combs[il_pos].i[1];
515 wordr0[2] = c_combs[il_pos].i[2];
516 wordr0[3] = c_combs[il_pos].i[3];
520 wordr1[0] = c_combs[il_pos].i[4];
521 wordr1[1] = c_combs[il_pos].i[5];
522 wordr1[2] = c_combs[il_pos].i[6];
523 wordr1[3] = c_combs[il_pos].i[7];
539 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
541 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
546 w0[0] = wordl0[0] | wordr0[0];
547 w0[1] = wordl0[1] | wordr0[1];
548 w0[2] = wordl0[2] | wordr0[2];
549 w0[3] = wordl0[3] | wordr0[3];
553 w1[0] = wordl1[0] | wordr1[0];
554 w1[1] = wordl1[1] | wordr1[1];
555 w1[2] = wordl1[2] | wordr1[2];
556 w1[3] = wordl1[3] | wordr1[3];
560 w2[0] = wordl2[0] | wordr2[0];
561 w2[1] = wordl2[1] | wordr2[1];
562 w2[2] = wordl2[2] | wordr2[2];
563 w2[3] = wordl2[3] | wordr2[3];
567 w3[0] = wordl3[0] | wordr3[0];
568 w3[1] = wordl3[1] | wordr3[1];
604 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
606 w0_t[0] |= salt_buf0[0];
607 w0_t[1] |= salt_buf0[1];
608 w0_t[2] |= salt_buf0[2];
609 w0_t[3] |= salt_buf0[3];
610 w1_t[0] |= salt_buf1[0];
611 w1_t[1] |= salt_buf1[1];
612 w1_t[2] |= salt_buf1[2];
613 w1_t[3] |= salt_buf1[3];
614 w2_t[0] |= salt_buf2[0];
615 w2_t[1] |= salt_buf2[1];
616 w2_t[2] |= salt_buf2[2];
617 w2_t[3] |= salt_buf2[3];
618 w3_t[0] |= salt_buf3[0];
619 w3_t[1] |= salt_buf3[1];
620 w3_t[2] |= salt_buf3[2];
621 w3_t[3] |= salt_buf3[3];
629 s0[0] = salt_buf0[0];
630 s0[1] = salt_buf0[1];
631 s0[2] = salt_buf0[2];
632 s0[3] = salt_buf0[3];
636 s1[0] = salt_buf1[0];
637 s1[1] = salt_buf1[1];
638 s1[2] = salt_buf1[2];
639 s1[3] = salt_buf1[3];
655 switch_buffer_by_offset (s0, s1, s2, s3, salt_len + pw_len);
674 const u32 pw_salt_len = salt_len + pw_len + salt_len;
676 append_0x80_4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
678 w3_t[2] = pw_salt_len * 8;
689 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
690 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
691 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
692 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
693 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
694 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
695 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
696 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
697 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
698 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
699 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
700 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
701 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
702 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
703 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
704 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
706 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
707 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
708 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
709 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
710 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
711 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
712 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
713 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
714 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
715 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
716 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
717 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
718 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
719 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
720 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
721 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
723 MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
724 MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
725 MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
726 MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
727 MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
728 MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
729 MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
730 MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
731 MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
732 MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
733 MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
734 MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
735 MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
736 MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
737 MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
738 MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
740 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
741 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
742 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
743 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
744 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
745 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
746 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
747 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
748 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
749 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
750 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
751 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
752 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
753 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
754 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
755 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
762 #include VECT_COMPARE_S
766 extern "C" __global__ void __launch_bounds__ (256, 1) m03800_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)
770 extern "C" __global__ void __launch_bounds__ (256, 1) m03800_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)