2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
16 #include "include/kernel_functions.c"
17 #include "OpenCL/types_ocl.c"
18 #include "OpenCL/common.c"
20 #define COMPARE_S "OpenCL/check_single_comp4.c"
21 #define COMPARE_M "OpenCL/check_multi_comp4.c"
23 __constant u32 lotus_magic_table[256] =
25 0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
26 0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
27 0x30, 0x04, 0xb6, 0xdc, 0x7d, 0xdf, 0x32, 0x4b,
28 0xf7, 0xcb, 0x45, 0x9b, 0x31, 0xbb, 0x21, 0x5a,
29 0x41, 0x9f, 0xe1, 0xd9, 0x4a, 0x4d, 0x9e, 0xda,
30 0xa0, 0x68, 0x2c, 0xc3, 0x27, 0x5f, 0x80, 0x36,
31 0x3e, 0xee, 0xfb, 0x95, 0x1a, 0xfe, 0xce, 0xa8,
32 0x34, 0xa9, 0x13, 0xf0, 0xa6, 0x3f, 0xd8, 0x0c,
33 0x78, 0x24, 0xaf, 0x23, 0x52, 0xc1, 0x67, 0x17,
34 0xf5, 0x66, 0x90, 0xe7, 0xe8, 0x07, 0xb8, 0x60,
35 0x48, 0xe6, 0x1e, 0x53, 0xf3, 0x92, 0xa4, 0x72,
36 0x8c, 0x08, 0x15, 0x6e, 0x86, 0x00, 0x84, 0xfa,
37 0xf4, 0x7f, 0x8a, 0x42, 0x19, 0xf6, 0xdb, 0xcd,
38 0x14, 0x8d, 0x50, 0x12, 0xba, 0x3c, 0x06, 0x4e,
39 0xec, 0xb3, 0x35, 0x11, 0xa1, 0x88, 0x8e, 0x2b,
40 0x94, 0x99, 0xb7, 0x71, 0x74, 0xd3, 0xe4, 0xbf,
41 0x3a, 0xde, 0x96, 0x0e, 0xbc, 0x0a, 0xed, 0x77,
42 0xfc, 0x37, 0x6b, 0x03, 0x79, 0x89, 0x62, 0xc6,
43 0xd7, 0xc0, 0xd2, 0x7c, 0x6a, 0x8b, 0x22, 0xa3,
44 0x5b, 0x05, 0x5d, 0x02, 0x75, 0xd5, 0x61, 0xe3,
45 0x18, 0x8f, 0x55, 0x51, 0xad, 0x1f, 0x0b, 0x5e,
46 0x85, 0xe5, 0xc2, 0x57, 0x63, 0xca, 0x3d, 0x6c,
47 0xb4, 0xc5, 0xcc, 0x70, 0xb2, 0x91, 0x59, 0x0d,
48 0x47, 0x20, 0xc8, 0x4f, 0x58, 0xe0, 0x01, 0xe2,
49 0x16, 0x38, 0xc4, 0x6f, 0x3b, 0x0f, 0x65, 0x46,
50 0xbe, 0x7e, 0x2d, 0x7b, 0x82, 0xf9, 0x40, 0xb5,
51 0x1d, 0x73, 0xf8, 0xeb, 0x26, 0xc7, 0x87, 0x97,
52 0x25, 0x54, 0xb1, 0x28, 0xaa, 0x98, 0x9d, 0xa5,
53 0x64, 0x6d, 0x7a, 0xd4, 0x10, 0x81, 0x44, 0xef,
54 0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
55 0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
56 0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
59 #define BOX(S,i) (S)[(i)]
61 #define uint_to_hex_upper8(i) l_bin2asc[(i)]
63 static void lotus_mix (u32 *in, __local u32 s_lotus_magic_table[256])
67 for (int i = 0; i < 18; i++)
72 for (int j = 0; j < 12; j++)
77 p = (p + s--) & 0xff; p = ((tmp_in >> 0) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 0;
78 p = (p + s--) & 0xff; p = ((tmp_in >> 8) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 8;
79 p = (p + s--) & 0xff; p = ((tmp_in >> 16) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 16;
80 p = (p + s--) & 0xff; p = ((tmp_in >> 24) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 24;
87 static void lotus_transform_password (u32 in[4], u32 out[4], __local u32 s_lotus_magic_table[256])
93 //#pragma unroll // kernel fails if used
94 for (int i = 0; i < 4; i++)
96 t ^= (in[i] >> 0) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 0; t = ((out[i] >> 0) & 0xff);
97 t ^= (in[i] >> 8) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 8; t = ((out[i] >> 8) & 0xff);
98 t ^= (in[i] >> 16) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 16; t = ((out[i] >> 16) & 0xff);
99 t ^= (in[i] >> 24) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 24; t = ((out[i] >> 24) & 0xff);
103 static void pad (u32 w[4], const u32 len)
105 const u32 val = 16 - len;
107 const u32 mask1 = val << 24;
109 const u32 mask2 = val << 16
112 const u32 mask3 = val << 8
116 const u32 mask4 = val << 0
123 case 0: w[0] = mask4;
128 case 1: w[0] |= mask3;
133 case 2: w[0] |= mask2;
138 case 3: w[0] |= mask1;
143 case 4: w[1] = mask4;
147 case 5: w[1] |= mask3;
151 case 6: w[1] |= mask2;
155 case 7: w[1] |= mask1;
159 case 8: w[2] = mask4;
162 case 9: w[2] |= mask3;
165 case 10: w[2] |= mask2;
168 case 11: w[2] |= mask1;
171 case 12: w[3] = mask4;
173 case 13: w[3] |= mask3;
175 case 14: w[3] |= mask2;
177 case 15: w[3] |= mask1;
182 static void mdtransform_norecalc (u32 state[4], u32 block[4], __local u32 s_lotus_magic_table[256])
194 x[ 8] = state[0] ^ block[0];
195 x[ 9] = state[1] ^ block[1];
196 x[10] = state[2] ^ block[2];
197 x[11] = state[3] ^ block[3];
199 lotus_mix (x, s_lotus_magic_table);
207 static void mdtransform (u32 state[4], u32 checksum[4], u32 block[4], __local u32 s_lotus_magic_table[256])
209 mdtransform_norecalc (state, block, s_lotus_magic_table);
211 lotus_transform_password (block, checksum, s_lotus_magic_table);
214 static void domino_big_md (const u32 saved_key[16], const u32 size, u32 state[4], __local u32 s_lotus_magic_table[256])
233 for (curpos = 0, idx = 0; curpos + 16 < size; curpos += 16, idx += 4)
235 block[0] = saved_key[idx + 0];
236 block[1] = saved_key[idx + 1];
237 block[2] = saved_key[idx + 2];
238 block[3] = saved_key[idx + 3];
240 mdtransform (state, checksum, block, s_lotus_magic_table);
243 block[0] = saved_key[idx + 0];
244 block[1] = saved_key[idx + 1];
245 block[2] = saved_key[idx + 2];
246 block[3] = saved_key[idx + 3];
248 mdtransform (state, checksum, block, s_lotus_magic_table);
250 mdtransform_norecalc (state, checksum, s_lotus_magic_table);
253 static void m08700m (__local u32 s_lotus_magic_table[256], __local u32 l_bin2asc[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32 * words_buf_r, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
259 const u32 gid = get_global_id (0);
260 const u32 lid = get_local_id (0);
268 pad (&w[ 0], pw_len & 0xf);
270 else if (pw_len < 32)
272 pad (&w[ 4], pw_len & 0xf);
274 else if (pw_len < 48)
276 pad (&w[ 8], pw_len & 0xf);
278 else if (pw_len < 64)
280 pad (&w[12], pw_len & 0xf);
287 const u32 salt0 = salt_bufs[salt_pos].salt_buf[0];
288 const u32 salt1 = salt_bufs[salt_pos].salt_buf[1] & 0xff | '(' << 8;
296 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
298 const u32 w0r = words_buf_r[il_pos];
300 const u32 w0 = w0l | w0r;
328 domino_big_md (w_tmp, pw_len, state, s_lotus_magic_table);
330 const u32 w0_t = uint_to_hex_upper8 ((state[0] >> 0) & 255) << 0
331 | uint_to_hex_upper8 ((state[0] >> 8) & 255) << 16;
332 const u32 w1_t = uint_to_hex_upper8 ((state[0] >> 16) & 255) << 0
333 | uint_to_hex_upper8 ((state[0] >> 24) & 255) << 16;
334 const u32 w2_t = uint_to_hex_upper8 ((state[1] >> 0) & 255) << 0
335 | uint_to_hex_upper8 ((state[1] >> 8) & 255) << 16;
336 const u32 w3_t = uint_to_hex_upper8 ((state[1] >> 16) & 255) << 0
337 | uint_to_hex_upper8 ((state[1] >> 24) & 255) << 16;
338 const u32 w4_t = uint_to_hex_upper8 ((state[2] >> 0) & 255) << 0
339 | uint_to_hex_upper8 ((state[2] >> 8) & 255) << 16;
340 const u32 w5_t = uint_to_hex_upper8 ((state[2] >> 16) & 255) << 0
341 | uint_to_hex_upper8 ((state[2] >> 24) & 255) << 16;
342 const u32 w6_t = uint_to_hex_upper8 ((state[3] >> 0) & 255) << 0
343 | uint_to_hex_upper8 ((state[3] >> 8) & 255) << 16;
344 //const u32 w7_t = uint_to_hex_upper8 ((state[3] >> 16) & 255) << 0
345 // | uint_to_hex_upper8 ((state[3] >> 24) & 255) << 16;
347 const u32 pade = 0x0e0e0e0e;
350 w_tmp[ 1] = salt1 | w0_t << 16;
351 w_tmp[ 2] = w0_t >> 16 | w1_t << 16;
352 w_tmp[ 3] = w1_t >> 16 | w2_t << 16;
353 w_tmp[ 4] = w2_t >> 16 | w3_t << 16;
354 w_tmp[ 5] = w3_t >> 16 | w4_t << 16;
355 w_tmp[ 6] = w4_t >> 16 | w5_t << 16;
356 w_tmp[ 7] = w5_t >> 16 | w6_t << 16;
357 w_tmp[ 8] = w6_t >> 16 | pade << 16; // | w7_t << 8;
371 domino_big_md (w_tmp, 34, state, s_lotus_magic_table);
373 u32 a = state[0] & 0xffffffff;
374 u32 b = state[1] & 0xffffffff;
375 u32 c = state[2] & 0x000000ff;
376 u32 d = state[3] & 0x00000000;
387 static void m08700s (__local u32 s_lotus_magic_table[256], __local u32 l_bin2asc[256], u32 w[16], const u32 pw_len, __global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32 * words_buf_r, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset)
393 const u32 gid = get_global_id (0);
394 const u32 lid = get_local_id (0);
402 pad (&w[ 0], pw_len & 0xf);
404 else if (pw_len < 32)
406 pad (&w[ 4], pw_len & 0xf);
408 else if (pw_len < 48)
410 pad (&w[ 8], pw_len & 0xf);
412 else if (pw_len < 64)
414 pad (&w[12], pw_len & 0xf);
421 const u32 salt0 = salt_bufs[salt_pos].salt_buf[0];
422 const u32 salt1 = salt_bufs[salt_pos].salt_buf[1] & 0xff | '(' << 8;
428 const u32 search[4] =
430 digests_buf[digests_offset].digest_buf[DGST_R0],
431 digests_buf[digests_offset].digest_buf[DGST_R1],
432 digests_buf[digests_offset].digest_buf[DGST_R2],
433 digests_buf[digests_offset].digest_buf[DGST_R3]
442 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos++)
444 const u32 w0r = words_buf_r[il_pos];
446 const u32 w0 = w0l | w0r;
474 domino_big_md (w_tmp, pw_len, state, s_lotus_magic_table);
476 const u32 w0_t = uint_to_hex_upper8 ((state[0] >> 0) & 255) << 0
477 | uint_to_hex_upper8 ((state[0] >> 8) & 255) << 16;
478 const u32 w1_t = uint_to_hex_upper8 ((state[0] >> 16) & 255) << 0
479 | uint_to_hex_upper8 ((state[0] >> 24) & 255) << 16;
480 const u32 w2_t = uint_to_hex_upper8 ((state[1] >> 0) & 255) << 0
481 | uint_to_hex_upper8 ((state[1] >> 8) & 255) << 16;
482 const u32 w3_t = uint_to_hex_upper8 ((state[1] >> 16) & 255) << 0
483 | uint_to_hex_upper8 ((state[1] >> 24) & 255) << 16;
484 const u32 w4_t = uint_to_hex_upper8 ((state[2] >> 0) & 255) << 0
485 | uint_to_hex_upper8 ((state[2] >> 8) & 255) << 16;
486 const u32 w5_t = uint_to_hex_upper8 ((state[2] >> 16) & 255) << 0
487 | uint_to_hex_upper8 ((state[2] >> 24) & 255) << 16;
488 const u32 w6_t = uint_to_hex_upper8 ((state[3] >> 0) & 255) << 0
489 | uint_to_hex_upper8 ((state[3] >> 8) & 255) << 16;
490 //const u32 w7_t = uint_to_hex_upper8 ((state[3] >> 16) & 255) << 0
491 // | uint_to_hex_upper8 ((state[3] >> 24) & 255) << 16;
493 const u32 pade = 0x0e0e0e0e;
496 w_tmp[ 1] = salt1 | w0_t << 16;
497 w_tmp[ 2] = w0_t >> 16 | w1_t << 16;
498 w_tmp[ 3] = w1_t >> 16 | w2_t << 16;
499 w_tmp[ 4] = w2_t >> 16 | w3_t << 16;
500 w_tmp[ 5] = w3_t >> 16 | w4_t << 16;
501 w_tmp[ 6] = w4_t >> 16 | w5_t << 16;
502 w_tmp[ 7] = w5_t >> 16 | w6_t << 16;
503 w_tmp[ 8] = w6_t >> 16 | pade << 16; // | w7_t << 8;
517 domino_big_md (w_tmp, 34, state, s_lotus_magic_table);
519 u32 a = state[0] & 0xffffffff;
520 u32 b = state[1] & 0xffffffff;
521 u32 c = state[2] & 0x000000ff;
522 u32 d = state[3] & 0x00000000;
533 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32 * words_buf_r, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
539 const u32 gid = get_global_id (0);
540 const u32 lid = get_local_id (0);
544 w[ 0] = pws[gid].i[ 0];
545 w[ 1] = pws[gid].i[ 1];
546 w[ 2] = pws[gid].i[ 2];
547 w[ 3] = pws[gid].i[ 3];
561 const u32 pw_len = pws[gid].pw_len;
567 __local u32 l_bin2asc[256];
569 const u32 lid4 = lid * 4;
571 const u32 lid40 = lid4 + 0;
572 const u32 lid41 = lid4 + 1;
573 const u32 lid42 = lid4 + 2;
574 const u32 lid43 = lid4 + 3;
576 const u32 v400 = (lid40 >> 0) & 15;
577 const u32 v401 = (lid40 >> 4) & 15;
578 const u32 v410 = (lid41 >> 0) & 15;
579 const u32 v411 = (lid41 >> 4) & 15;
580 const u32 v420 = (lid42 >> 0) & 15;
581 const u32 v421 = (lid42 >> 4) & 15;
582 const u32 v430 = (lid43 >> 0) & 15;
583 const u32 v431 = (lid43 >> 4) & 15;
585 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
586 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
587 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
588 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
589 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
590 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
591 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
592 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
594 barrier (CLK_LOCAL_MEM_FENCE);
600 __local u32 s_lotus_magic_table[256];
602 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
603 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
604 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
605 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
607 barrier (CLK_LOCAL_MEM_FENCE);
609 if (gid >= gid_max) return;
615 m08700m (s_lotus_magic_table, l_bin2asc, 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);
618 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_m08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32 * words_buf_r, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
624 const u32 gid = get_global_id (0);
625 const u32 lid = get_local_id (0);
629 w[ 0] = pws[gid].i[ 0];
630 w[ 1] = pws[gid].i[ 1];
631 w[ 2] = pws[gid].i[ 2];
632 w[ 3] = pws[gid].i[ 3];
633 w[ 4] = pws[gid].i[ 4];
634 w[ 5] = pws[gid].i[ 5];
635 w[ 6] = pws[gid].i[ 6];
636 w[ 7] = pws[gid].i[ 7];
646 const u32 pw_len = pws[gid].pw_len;
652 __local u32 l_bin2asc[256];
654 const u32 lid4 = lid * 4;
656 const u32 lid40 = lid4 + 0;
657 const u32 lid41 = lid4 + 1;
658 const u32 lid42 = lid4 + 2;
659 const u32 lid43 = lid4 + 3;
661 const u32 v400 = (lid40 >> 0) & 15;
662 const u32 v401 = (lid40 >> 4) & 15;
663 const u32 v410 = (lid41 >> 0) & 15;
664 const u32 v411 = (lid41 >> 4) & 15;
665 const u32 v420 = (lid42 >> 0) & 15;
666 const u32 v421 = (lid42 >> 4) & 15;
667 const u32 v430 = (lid43 >> 0) & 15;
668 const u32 v431 = (lid43 >> 4) & 15;
670 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
671 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
672 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
673 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
674 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
675 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
676 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
677 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
679 barrier (CLK_LOCAL_MEM_FENCE);
685 __local u32 s_lotus_magic_table[256];
687 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
688 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
689 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
690 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
692 barrier (CLK_LOCAL_MEM_FENCE);
694 if (gid >= gid_max) return;
700 m08700m (s_lotus_magic_table, l_bin2asc, 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);
703 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32 * words_buf_r, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
709 const u32 gid = get_global_id (0);
710 const u32 lid = get_local_id (0);
714 w[ 0] = pws[gid].i[ 0];
715 w[ 1] = pws[gid].i[ 1];
716 w[ 2] = pws[gid].i[ 2];
717 w[ 3] = pws[gid].i[ 3];
718 w[ 4] = pws[gid].i[ 4];
719 w[ 5] = pws[gid].i[ 5];
720 w[ 6] = pws[gid].i[ 6];
721 w[ 7] = pws[gid].i[ 7];
722 w[ 8] = pws[gid].i[ 8];
723 w[ 9] = pws[gid].i[ 9];
724 w[10] = pws[gid].i[10];
725 w[11] = pws[gid].i[11];
726 w[12] = pws[gid].i[12];
727 w[13] = pws[gid].i[13];
728 w[14] = pws[gid].i[14];
729 w[15] = pws[gid].i[15];
731 const u32 pw_len = pws[gid].pw_len;
737 __local u32 l_bin2asc[256];
739 const u32 lid4 = lid * 4;
741 const u32 lid40 = lid4 + 0;
742 const u32 lid41 = lid4 + 1;
743 const u32 lid42 = lid4 + 2;
744 const u32 lid43 = lid4 + 3;
746 const u32 v400 = (lid40 >> 0) & 15;
747 const u32 v401 = (lid40 >> 4) & 15;
748 const u32 v410 = (lid41 >> 0) & 15;
749 const u32 v411 = (lid41 >> 4) & 15;
750 const u32 v420 = (lid42 >> 0) & 15;
751 const u32 v421 = (lid42 >> 4) & 15;
752 const u32 v430 = (lid43 >> 0) & 15;
753 const u32 v431 = (lid43 >> 4) & 15;
755 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
756 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
757 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
758 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
759 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
760 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
761 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
762 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
764 barrier (CLK_LOCAL_MEM_FENCE);
770 __local u32 s_lotus_magic_table[256];
772 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
773 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
774 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
775 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
777 barrier (CLK_LOCAL_MEM_FENCE);
779 if (gid >= gid_max) return;
785 m08700m (s_lotus_magic_table, l_bin2asc, 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);
788 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32 * words_buf_r, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
794 const u32 gid = get_global_id (0);
795 const u32 lid = get_local_id (0);
799 w[ 0] = pws[gid].i[ 0];
800 w[ 1] = pws[gid].i[ 1];
801 w[ 2] = pws[gid].i[ 2];
802 w[ 3] = pws[gid].i[ 3];
816 const u32 pw_len = pws[gid].pw_len;
822 __local u32 l_bin2asc[256];
824 const u32 lid4 = lid * 4;
826 const u32 lid40 = lid4 + 0;
827 const u32 lid41 = lid4 + 1;
828 const u32 lid42 = lid4 + 2;
829 const u32 lid43 = lid4 + 3;
831 const u32 v400 = (lid40 >> 0) & 15;
832 const u32 v401 = (lid40 >> 4) & 15;
833 const u32 v410 = (lid41 >> 0) & 15;
834 const u32 v411 = (lid41 >> 4) & 15;
835 const u32 v420 = (lid42 >> 0) & 15;
836 const u32 v421 = (lid42 >> 4) & 15;
837 const u32 v430 = (lid43 >> 0) & 15;
838 const u32 v431 = (lid43 >> 4) & 15;
840 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
841 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
842 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
843 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
844 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
845 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
846 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
847 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
849 barrier (CLK_LOCAL_MEM_FENCE);
855 __local u32 s_lotus_magic_table[256];
857 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
858 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
859 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
860 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
862 barrier (CLK_LOCAL_MEM_FENCE);
864 if (gid >= gid_max) return;
870 m08700s (s_lotus_magic_table, l_bin2asc, 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);
873 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_s08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32 * words_buf_r, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
879 const u32 gid = get_global_id (0);
880 const u32 lid = get_local_id (0);
884 w[ 0] = pws[gid].i[ 0];
885 w[ 1] = pws[gid].i[ 1];
886 w[ 2] = pws[gid].i[ 2];
887 w[ 3] = pws[gid].i[ 3];
888 w[ 4] = pws[gid].i[ 4];
889 w[ 5] = pws[gid].i[ 5];
890 w[ 6] = pws[gid].i[ 6];
891 w[ 7] = pws[gid].i[ 7];
901 const u32 pw_len = pws[gid].pw_len;
907 __local u32 l_bin2asc[256];
909 const u32 lid4 = lid * 4;
911 const u32 lid40 = lid4 + 0;
912 const u32 lid41 = lid4 + 1;
913 const u32 lid42 = lid4 + 2;
914 const u32 lid43 = lid4 + 3;
916 const u32 v400 = (lid40 >> 0) & 15;
917 const u32 v401 = (lid40 >> 4) & 15;
918 const u32 v410 = (lid41 >> 0) & 15;
919 const u32 v411 = (lid41 >> 4) & 15;
920 const u32 v420 = (lid42 >> 0) & 15;
921 const u32 v421 = (lid42 >> 4) & 15;
922 const u32 v430 = (lid43 >> 0) & 15;
923 const u32 v431 = (lid43 >> 4) & 15;
925 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
926 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
927 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
928 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
929 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
930 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
931 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
932 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
934 barrier (CLK_LOCAL_MEM_FENCE);
940 __local u32 s_lotus_magic_table[256];
942 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
943 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
944 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
945 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
947 barrier (CLK_LOCAL_MEM_FENCE);
949 if (gid >= gid_max) return;
955 m08700s (s_lotus_magic_table, l_bin2asc, 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);
958 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_s16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant u32 * words_buf_r, __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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
964 const u32 gid = get_global_id (0);
965 const u32 lid = get_local_id (0);
969 w[ 0] = pws[gid].i[ 0];
970 w[ 1] = pws[gid].i[ 1];
971 w[ 2] = pws[gid].i[ 2];
972 w[ 3] = pws[gid].i[ 3];
973 w[ 4] = pws[gid].i[ 4];
974 w[ 5] = pws[gid].i[ 5];
975 w[ 6] = pws[gid].i[ 6];
976 w[ 7] = pws[gid].i[ 7];
977 w[ 8] = pws[gid].i[ 8];
978 w[ 9] = pws[gid].i[ 9];
979 w[10] = pws[gid].i[10];
980 w[11] = pws[gid].i[11];
981 w[12] = pws[gid].i[12];
982 w[13] = pws[gid].i[13];
983 w[14] = pws[gid].i[14];
984 w[15] = pws[gid].i[15];
986 const u32 pw_len = pws[gid].pw_len;
992 __local u32 l_bin2asc[256];
994 const u32 lid4 = lid * 4;
996 const u32 lid40 = lid4 + 0;
997 const u32 lid41 = lid4 + 1;
998 const u32 lid42 = lid4 + 2;
999 const u32 lid43 = lid4 + 3;
1001 const u32 v400 = (lid40 >> 0) & 15;
1002 const u32 v401 = (lid40 >> 4) & 15;
1003 const u32 v410 = (lid41 >> 0) & 15;
1004 const u32 v411 = (lid41 >> 4) & 15;
1005 const u32 v420 = (lid42 >> 0) & 15;
1006 const u32 v421 = (lid42 >> 4) & 15;
1007 const u32 v430 = (lid43 >> 0) & 15;
1008 const u32 v431 = (lid43 >> 4) & 15;
1010 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
1011 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
1012 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
1013 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
1014 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
1015 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
1016 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
1017 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
1019 barrier (CLK_LOCAL_MEM_FENCE);
1025 __local u32 s_lotus_magic_table[256];
1027 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
1028 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
1029 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
1030 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
1032 barrier (CLK_LOCAL_MEM_FENCE);
1034 if (gid >= gid_max) return;
1040 m08700s (s_lotus_magic_table, l_bin2asc, 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);