2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
28 #include "include/kernel_functions.c"
29 #include "types_amd.c"
30 #include "common_amd.c"
31 #include "include/rp_gpu.h"
35 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
36 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
40 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
41 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
45 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
46 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
49 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_m04 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
55 const u32 lid = get_local_id (0);
61 const u32 gid = get_global_id (0);
65 pw_buf0[0] = pws[gid].i[ 0];
66 pw_buf0[1] = pws[gid].i[ 1];
67 pw_buf0[2] = pws[gid].i[ 2];
68 pw_buf0[3] = pws[gid].i[ 3];
72 pw_buf1[0] = pws[gid].i[ 4];
73 pw_buf1[1] = pws[gid].i[ 5];
74 pw_buf1[2] = pws[gid].i[ 6];
75 pw_buf1[3] = pws[gid].i[ 7];
77 const u32 pw_len = pws[gid].pw_len;
85 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
86 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
87 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
88 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
92 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
93 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
94 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
95 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
111 const u32 salt_len = salt_bufs[salt_pos].salt_len;
117 __local u32 l_bin2asc[256];
119 const u32 lid4 = lid * 4;
121 const u32 lid40 = lid4 + 0;
122 const u32 lid41 = lid4 + 1;
123 const u32 lid42 = lid4 + 2;
124 const u32 lid43 = lid4 + 3;
126 const u32 v400 = (lid40 >> 0) & 15;
127 const u32 v401 = (lid40 >> 4) & 15;
128 const u32 v410 = (lid41 >> 0) & 15;
129 const u32 v411 = (lid41 >> 4) & 15;
130 const u32 v420 = (lid42 >> 0) & 15;
131 const u32 v421 = (lid42 >> 4) & 15;
132 const u32 v430 = (lid43 >> 0) & 15;
133 const u32 v431 = (lid43 >> 4) & 15;
135 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
136 | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
137 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
138 | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
139 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
140 | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
141 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
142 | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
144 barrier (CLK_LOCAL_MEM_FENCE);
146 if (gid >= gid_max) return;
152 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
182 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
217 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
219 w0_t[0] |= salt_buf0[0];
220 w0_t[1] |= salt_buf0[1];
221 w0_t[2] |= salt_buf0[2];
222 w0_t[3] |= salt_buf0[3];
223 w1_t[0] |= salt_buf1[0];
224 w1_t[1] |= salt_buf1[1];
225 w1_t[2] |= salt_buf1[2];
226 w1_t[3] |= salt_buf1[3];
227 w2_t[0] |= salt_buf2[0];
228 w2_t[1] |= salt_buf2[1];
229 w2_t[2] |= salt_buf2[2];
230 w2_t[3] |= salt_buf2[3];
231 w3_t[0] |= salt_buf3[0];
232 w3_t[1] |= salt_buf3[1];
233 w3_t[2] |= salt_buf3[2];
234 w3_t[3] |= salt_buf3[3];
242 s0[0] = salt_buf0[0];
243 s0[1] = salt_buf0[1];
244 s0[2] = salt_buf0[2];
245 s0[3] = salt_buf0[3];
249 s1[0] = salt_buf1[0];
250 s1[1] = salt_buf1[1];
251 s1[2] = salt_buf1[2];
252 s1[3] = salt_buf1[3];
268 switch_buffer_by_offset (s0, s1, s2, s3, salt_len + out_len);
287 const u32 pw_salt_len = salt_len + out_len + salt_len;
289 append_0x80_4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
291 w3_t[2] = pw_salt_len * 8;
302 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
303 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
304 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
305 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
306 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
307 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
308 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
309 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
310 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
311 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
312 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
313 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
314 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
315 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
316 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
317 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
319 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
320 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
321 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
322 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
323 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
324 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
325 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
326 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
327 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
328 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
329 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
330 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
331 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
332 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
333 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
334 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
336 MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
337 MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
338 MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
339 MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
340 MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
341 MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
342 MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
343 MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
344 MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
345 MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
346 MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
347 MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
348 MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
349 MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
350 MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
351 MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
353 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
354 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
355 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
356 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
357 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
358 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
359 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
360 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
361 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
362 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
363 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
364 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
365 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
367 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
368 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
369 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
376 #include VECT_COMPARE_M
380 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_m08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
384 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_m16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
388 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s04 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
394 const u32 lid = get_local_id (0);
400 const u32 gid = get_global_id (0);
404 pw_buf0[0] = pws[gid].i[ 0];
405 pw_buf0[1] = pws[gid].i[ 1];
406 pw_buf0[2] = pws[gid].i[ 2];
407 pw_buf0[3] = pws[gid].i[ 3];
411 pw_buf1[0] = pws[gid].i[ 4];
412 pw_buf1[1] = pws[gid].i[ 5];
413 pw_buf1[2] = pws[gid].i[ 6];
414 pw_buf1[3] = pws[gid].i[ 7];
416 const u32 pw_len = pws[gid].pw_len;
424 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
425 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
426 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
427 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
431 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
432 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
433 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
434 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
450 const u32 salt_len = salt_bufs[salt_pos].salt_len;
456 const u32 search[4] =
458 digests_buf[digests_offset].digest_buf[DGST_R0],
459 digests_buf[digests_offset].digest_buf[DGST_R1],
460 digests_buf[digests_offset].digest_buf[DGST_R2],
461 digests_buf[digests_offset].digest_buf[DGST_R3]
468 __local u32 l_bin2asc[256];
470 const u32 lid4 = lid * 4;
472 const u32 lid40 = lid4 + 0;
473 const u32 lid41 = lid4 + 1;
474 const u32 lid42 = lid4 + 2;
475 const u32 lid43 = lid4 + 3;
477 const u32 v400 = (lid40 >> 0) & 15;
478 const u32 v401 = (lid40 >> 4) & 15;
479 const u32 v410 = (lid41 >> 0) & 15;
480 const u32 v411 = (lid41 >> 4) & 15;
481 const u32 v420 = (lid42 >> 0) & 15;
482 const u32 v421 = (lid42 >> 4) & 15;
483 const u32 v430 = (lid43 >> 0) & 15;
484 const u32 v431 = (lid43 >> 4) & 15;
486 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'a' - 10 + v400) << 8
487 | ((v401 < 10) ? '0' + v401 : 'a' - 10 + v401) << 0;
488 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'a' - 10 + v410) << 8
489 | ((v411 < 10) ? '0' + v411 : 'a' - 10 + v411) << 0;
490 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'a' - 10 + v420) << 8
491 | ((v421 < 10) ? '0' + v421 : 'a' - 10 + v421) << 0;
492 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'a' - 10 + v430) << 8
493 | ((v431 < 10) ? '0' + v431 : 'a' - 10 + v431) << 0;
495 barrier (CLK_LOCAL_MEM_FENCE);
497 if (gid >= gid_max) return;
503 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
533 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
568 switch_buffer_by_offset (w0_t, w1_t, w2_t, w3_t, salt_len);
570 w0_t[0] |= salt_buf0[0];
571 w0_t[1] |= salt_buf0[1];
572 w0_t[2] |= salt_buf0[2];
573 w0_t[3] |= salt_buf0[3];
574 w1_t[0] |= salt_buf1[0];
575 w1_t[1] |= salt_buf1[1];
576 w1_t[2] |= salt_buf1[2];
577 w1_t[3] |= salt_buf1[3];
578 w2_t[0] |= salt_buf2[0];
579 w2_t[1] |= salt_buf2[1];
580 w2_t[2] |= salt_buf2[2];
581 w2_t[3] |= salt_buf2[3];
582 w3_t[0] |= salt_buf3[0];
583 w3_t[1] |= salt_buf3[1];
584 w3_t[2] |= salt_buf3[2];
585 w3_t[3] |= salt_buf3[3];
593 s0[0] = salt_buf0[0];
594 s0[1] = salt_buf0[1];
595 s0[2] = salt_buf0[2];
596 s0[3] = salt_buf0[3];
600 s1[0] = salt_buf1[0];
601 s1[1] = salt_buf1[1];
602 s1[2] = salt_buf1[2];
603 s1[3] = salt_buf1[3];
619 switch_buffer_by_offset (s0, s1, s2, s3, salt_len + out_len);
638 const u32 pw_salt_len = salt_len + out_len + salt_len;
640 append_0x80_4 (w0_t, w1_t, w2_t, w3_t, pw_salt_len);
642 w3_t[2] = pw_salt_len * 8;
653 MD5_STEP (MD5_Fo, a, b, c, d, w0_t[0], MD5C00, MD5S00);
654 MD5_STEP (MD5_Fo, d, a, b, c, w0_t[1], MD5C01, MD5S01);
655 MD5_STEP (MD5_Fo, c, d, a, b, w0_t[2], MD5C02, MD5S02);
656 MD5_STEP (MD5_Fo, b, c, d, a, w0_t[3], MD5C03, MD5S03);
657 MD5_STEP (MD5_Fo, a, b, c, d, w1_t[0], MD5C04, MD5S00);
658 MD5_STEP (MD5_Fo, d, a, b, c, w1_t[1], MD5C05, MD5S01);
659 MD5_STEP (MD5_Fo, c, d, a, b, w1_t[2], MD5C06, MD5S02);
660 MD5_STEP (MD5_Fo, b, c, d, a, w1_t[3], MD5C07, MD5S03);
661 MD5_STEP (MD5_Fo, a, b, c, d, w2_t[0], MD5C08, MD5S00);
662 MD5_STEP (MD5_Fo, d, a, b, c, w2_t[1], MD5C09, MD5S01);
663 MD5_STEP (MD5_Fo, c, d, a, b, w2_t[2], MD5C0a, MD5S02);
664 MD5_STEP (MD5_Fo, b, c, d, a, w2_t[3], MD5C0b, MD5S03);
665 MD5_STEP (MD5_Fo, a, b, c, d, w3_t[0], MD5C0c, MD5S00);
666 MD5_STEP (MD5_Fo, d, a, b, c, w3_t[1], MD5C0d, MD5S01);
667 MD5_STEP (MD5_Fo, c, d, a, b, w3_t[2], MD5C0e, MD5S02);
668 MD5_STEP (MD5_Fo, b, c, d, a, w3_t[3], MD5C0f, MD5S03);
670 MD5_STEP (MD5_Go, a, b, c, d, w0_t[1], MD5C10, MD5S10);
671 MD5_STEP (MD5_Go, d, a, b, c, w1_t[2], MD5C11, MD5S11);
672 MD5_STEP (MD5_Go, c, d, a, b, w2_t[3], MD5C12, MD5S12);
673 MD5_STEP (MD5_Go, b, c, d, a, w0_t[0], MD5C13, MD5S13);
674 MD5_STEP (MD5_Go, a, b, c, d, w1_t[1], MD5C14, MD5S10);
675 MD5_STEP (MD5_Go, d, a, b, c, w2_t[2], MD5C15, MD5S11);
676 MD5_STEP (MD5_Go, c, d, a, b, w3_t[3], MD5C16, MD5S12);
677 MD5_STEP (MD5_Go, b, c, d, a, w1_t[0], MD5C17, MD5S13);
678 MD5_STEP (MD5_Go, a, b, c, d, w2_t[1], MD5C18, MD5S10);
679 MD5_STEP (MD5_Go, d, a, b, c, w3_t[2], MD5C19, MD5S11);
680 MD5_STEP (MD5_Go, c, d, a, b, w0_t[3], MD5C1a, MD5S12);
681 MD5_STEP (MD5_Go, b, c, d, a, w2_t[0], MD5C1b, MD5S13);
682 MD5_STEP (MD5_Go, a, b, c, d, w3_t[1], MD5C1c, MD5S10);
683 MD5_STEP (MD5_Go, d, a, b, c, w0_t[2], MD5C1d, MD5S11);
684 MD5_STEP (MD5_Go, c, d, a, b, w1_t[3], MD5C1e, MD5S12);
685 MD5_STEP (MD5_Go, b, c, d, a, w3_t[0], MD5C1f, MD5S13);
687 MD5_STEP (MD5_H , a, b, c, d, w1_t[1], MD5C20, MD5S20);
688 MD5_STEP (MD5_H , d, a, b, c, w2_t[0], MD5C21, MD5S21);
689 MD5_STEP (MD5_H , c, d, a, b, w2_t[3], MD5C22, MD5S22);
690 MD5_STEP (MD5_H , b, c, d, a, w3_t[2], MD5C23, MD5S23);
691 MD5_STEP (MD5_H , a, b, c, d, w0_t[1], MD5C24, MD5S20);
692 MD5_STEP (MD5_H , d, a, b, c, w1_t[0], MD5C25, MD5S21);
693 MD5_STEP (MD5_H , c, d, a, b, w1_t[3], MD5C26, MD5S22);
694 MD5_STEP (MD5_H , b, c, d, a, w2_t[2], MD5C27, MD5S23);
695 MD5_STEP (MD5_H , a, b, c, d, w3_t[1], MD5C28, MD5S20);
696 MD5_STEP (MD5_H , d, a, b, c, w0_t[0], MD5C29, MD5S21);
697 MD5_STEP (MD5_H , c, d, a, b, w0_t[3], MD5C2a, MD5S22);
698 MD5_STEP (MD5_H , b, c, d, a, w1_t[2], MD5C2b, MD5S23);
699 MD5_STEP (MD5_H , a, b, c, d, w2_t[1], MD5C2c, MD5S20);
700 MD5_STEP (MD5_H , d, a, b, c, w3_t[0], MD5C2d, MD5S21);
701 MD5_STEP (MD5_H , c, d, a, b, w3_t[3], MD5C2e, MD5S22);
702 MD5_STEP (MD5_H , b, c, d, a, w0_t[2], MD5C2f, MD5S23);
704 MD5_STEP (MD5_I , a, b, c, d, w0_t[0], MD5C30, MD5S30);
705 MD5_STEP (MD5_I , d, a, b, c, w1_t[3], MD5C31, MD5S31);
706 MD5_STEP (MD5_I , c, d, a, b, w3_t[2], MD5C32, MD5S32);
707 MD5_STEP (MD5_I , b, c, d, a, w1_t[1], MD5C33, MD5S33);
708 MD5_STEP (MD5_I , a, b, c, d, w3_t[0], MD5C34, MD5S30);
709 MD5_STEP (MD5_I , d, a, b, c, w0_t[3], MD5C35, MD5S31);
710 MD5_STEP (MD5_I , c, d, a, b, w2_t[2], MD5C36, MD5S32);
711 MD5_STEP (MD5_I , b, c, d, a, w0_t[1], MD5C37, MD5S33);
712 MD5_STEP (MD5_I , a, b, c, d, w2_t[0], MD5C38, MD5S30);
713 MD5_STEP (MD5_I , d, a, b, c, w3_t[3], MD5C39, MD5S31);
714 MD5_STEP (MD5_I , c, d, a, b, w1_t[2], MD5C3a, MD5S32);
715 MD5_STEP (MD5_I , b, c, d, a, w3_t[1], MD5C3b, MD5S33);
716 MD5_STEP (MD5_I , a, b, c, d, w1_t[0], MD5C3c, MD5S30);
717 MD5_STEP (MD5_I , d, a, b, c, w2_t[3], MD5C3d, MD5S31);
718 MD5_STEP (MD5_I , c, d, a, b, w0_t[2], MD5C3e, MD5S32);
719 MD5_STEP (MD5_I , b, c, d, a, w2_t[1], MD5C3f, MD5S33);
726 #include VECT_COMPARE_S
730 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
734 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m03800_s16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)