2 * Author......: Jens Steube <jens.steube@gmail.com>
9 #include "include/constants.h"
10 #include "include/kernel_vendor.h"
29 #include "include/kernel_functions.c"
30 #include "types_amd.c"
31 #include "common_amd.c"
34 #define VECT_COMPARE_S "check_single_vect1_comp4_warp.c"
35 #define VECT_COMPARE_M "check_multi_vect1_comp4_warp.c"
39 #define VECT_COMPARE_S "check_single_vect2_comp4_warp.c"
40 #define VECT_COMPARE_M "check_multi_vect2_comp4_warp.c"
44 #define VECT_COMPARE_S "check_single_vect4_comp4_warp.c"
45 #define VECT_COMPARE_M "check_multi_vect4_comp4_warp.c"
48 __constant u32 lotus_magic_table[256] =
50 0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
51 0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
52 0x30, 0x04, 0xb6, 0xdc, 0x7d, 0xdf, 0x32, 0x4b,
53 0xf7, 0xcb, 0x45, 0x9b, 0x31, 0xbb, 0x21, 0x5a,
54 0x41, 0x9f, 0xe1, 0xd9, 0x4a, 0x4d, 0x9e, 0xda,
55 0xa0, 0x68, 0x2c, 0xc3, 0x27, 0x5f, 0x80, 0x36,
56 0x3e, 0xee, 0xfb, 0x95, 0x1a, 0xfe, 0xce, 0xa8,
57 0x34, 0xa9, 0x13, 0xf0, 0xa6, 0x3f, 0xd8, 0x0c,
58 0x78, 0x24, 0xaf, 0x23, 0x52, 0xc1, 0x67, 0x17,
59 0xf5, 0x66, 0x90, 0xe7, 0xe8, 0x07, 0xb8, 0x60,
60 0x48, 0xe6, 0x1e, 0x53, 0xf3, 0x92, 0xa4, 0x72,
61 0x8c, 0x08, 0x15, 0x6e, 0x86, 0x00, 0x84, 0xfa,
62 0xf4, 0x7f, 0x8a, 0x42, 0x19, 0xf6, 0xdb, 0xcd,
63 0x14, 0x8d, 0x50, 0x12, 0xba, 0x3c, 0x06, 0x4e,
64 0xec, 0xb3, 0x35, 0x11, 0xa1, 0x88, 0x8e, 0x2b,
65 0x94, 0x99, 0xb7, 0x71, 0x74, 0xd3, 0xe4, 0xbf,
66 0x3a, 0xde, 0x96, 0x0e, 0xbc, 0x0a, 0xed, 0x77,
67 0xfc, 0x37, 0x6b, 0x03, 0x79, 0x89, 0x62, 0xc6,
68 0xd7, 0xc0, 0xd2, 0x7c, 0x6a, 0x8b, 0x22, 0xa3,
69 0x5b, 0x05, 0x5d, 0x02, 0x75, 0xd5, 0x61, 0xe3,
70 0x18, 0x8f, 0x55, 0x51, 0xad, 0x1f, 0x0b, 0x5e,
71 0x85, 0xe5, 0xc2, 0x57, 0x63, 0xca, 0x3d, 0x6c,
72 0xb4, 0xc5, 0xcc, 0x70, 0xb2, 0x91, 0x59, 0x0d,
73 0x47, 0x20, 0xc8, 0x4f, 0x58, 0xe0, 0x01, 0xe2,
74 0x16, 0x38, 0xc4, 0x6f, 0x3b, 0x0f, 0x65, 0x46,
75 0xbe, 0x7e, 0x2d, 0x7b, 0x82, 0xf9, 0x40, 0xb5,
76 0x1d, 0x73, 0xf8, 0xeb, 0x26, 0xc7, 0x87, 0x97,
77 0x25, 0x54, 0xb1, 0x28, 0xaa, 0x98, 0x9d, 0xa5,
78 0x64, 0x6d, 0x7a, 0xd4, 0x10, 0x81, 0x44, 0xef,
79 0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
80 0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
81 0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
85 #define BOX(S,i) (u32x) ((S)[(i)])
89 #define BOX(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1])
93 #define BOX(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3])
97 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i)])
101 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1])
105 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3])
108 static void lotus_mix (u32x *in, __local u32 s_lotus_magic_table[256])
112 for (int i = 0; i < 18; i++)
117 for (int j = 0; j < 12; j++)
122 p = (p + s--) & 0xff; p = ((tmp_in >> 0) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 0;
123 p = (p + s--) & 0xff; p = ((tmp_in >> 8) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 8;
124 p = (p + s--) & 0xff; p = ((tmp_in >> 16) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 16;
125 p = (p + s--) & 0xff; p = ((tmp_in >> 24) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 24;
132 static void lotus_transform_password (u32x *in, u32x *out, __local u32 s_lotus_magic_table[256])
134 u32x t = out[3] >> 24;
138 //#pragma unroll // kernel fails if used
139 for (int i = 0; i < 4; i++)
141 t ^= (in[i] >> 0) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 0; t = ((out[i] >> 0) & 0xff);
142 t ^= (in[i] >> 8) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 8; t = ((out[i] >> 8) & 0xff);
143 t ^= (in[i] >> 16) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 16; t = ((out[i] >> 16) & 0xff);
144 t ^= (in[i] >> 24) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 24; t = ((out[i] >> 24) & 0xff);
148 static void pad (u32 w[4], const u32 len)
150 const u32 val = 16 - len;
152 const u32 mask1 = val << 24;
154 const u32 mask2 = val << 16
157 const u32 mask3 = val << 8
161 const u32 mask4 = val << 0
168 case 0: w[0] = mask4;
173 case 1: w[0] |= mask3;
178 case 2: w[0] |= mask2;
183 case 3: w[0] |= mask1;
188 case 4: w[1] = mask4;
192 case 5: w[1] |= mask3;
196 case 6: w[1] |= mask2;
200 case 7: w[1] |= mask1;
204 case 8: w[2] = mask4;
207 case 9: w[2] |= mask3;
210 case 10: w[2] |= mask2;
213 case 11: w[2] |= mask1;
216 case 12: w[3] = mask4;
218 case 13: w[3] |= mask3;
220 case 14: w[3] |= mask2;
222 case 15: w[3] |= mask1;
227 static void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 s_lotus_magic_table[256])
239 x[ 8] = state[0] ^ block[0];
240 x[ 9] = state[1] ^ block[1];
241 x[10] = state[2] ^ block[2];
242 x[11] = state[3] ^ block[3];
244 lotus_mix (x, s_lotus_magic_table);
252 static void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 s_lotus_magic_table[256])
254 mdtransform_norecalc (state, block, s_lotus_magic_table);
256 lotus_transform_password (block, checksum, s_lotus_magic_table);
259 static void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 s_lotus_magic_table[256])
278 for (curpos = 0, idx = 0; curpos + 16 < size; curpos += 16, idx += 4)
280 block[0] = saved_key[idx + 0];
281 block[1] = saved_key[idx + 1];
282 block[2] = saved_key[idx + 2];
283 block[3] = saved_key[idx + 3];
285 mdtransform (state, checksum, block, s_lotus_magic_table);
288 block[0] = saved_key[idx + 0];
289 block[1] = saved_key[idx + 1];
290 block[2] = saved_key[idx + 2];
291 block[3] = saved_key[idx + 3];
293 mdtransform (state, checksum, block, s_lotus_magic_table);
295 mdtransform_norecalc (state, checksum, s_lotus_magic_table);
298 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 gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32x * 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)
304 const u32 gid = get_global_id (0);
305 const u32 lid = get_local_id (0);
313 pad (&w[ 0], pw_len & 0xf);
315 else if (pw_len < 32)
317 pad (&w[ 4], pw_len & 0xf);
319 else if (pw_len < 48)
321 pad (&w[ 8], pw_len & 0xf);
323 else if (pw_len < 64)
325 pad (&w[12], pw_len & 0xf);
332 const u32 salt0 = salt_bufs[salt_pos].salt_buf[0];
333 const u32 salt1 = salt_bufs[salt_pos].salt_buf[1] & 0xff | '(' << 8;
339 const u32 bf_loops = ceil ((float) bfs_cnt / VECT_DIV);
343 for (u32 il_pos = 0; il_pos < bf_loops; il_pos++)
345 const u32x w0r = words_buf_r[il_pos];
347 const u32x w0 = w0l | w0r;
375 domino_big_md (w_tmp, pw_len, state, s_lotus_magic_table);
377 const u32x w0_t = uint_to_hex_upper8 ((state[0] >> 0) & 255) << 0
378 | uint_to_hex_upper8 ((state[0] >> 8) & 255) << 16;
379 const u32x w1_t = uint_to_hex_upper8 ((state[0] >> 16) & 255) << 0
380 | uint_to_hex_upper8 ((state[0] >> 24) & 255) << 16;
381 const u32x w2_t = uint_to_hex_upper8 ((state[1] >> 0) & 255) << 0
382 | uint_to_hex_upper8 ((state[1] >> 8) & 255) << 16;
383 const u32x w3_t = uint_to_hex_upper8 ((state[1] >> 16) & 255) << 0
384 | uint_to_hex_upper8 ((state[1] >> 24) & 255) << 16;
385 const u32x w4_t = uint_to_hex_upper8 ((state[2] >> 0) & 255) << 0
386 | uint_to_hex_upper8 ((state[2] >> 8) & 255) << 16;
387 const u32x w5_t = uint_to_hex_upper8 ((state[2] >> 16) & 255) << 0
388 | uint_to_hex_upper8 ((state[2] >> 24) & 255) << 16;
389 const u32x w6_t = uint_to_hex_upper8 ((state[3] >> 0) & 255) << 0
390 | uint_to_hex_upper8 ((state[3] >> 8) & 255) << 16;
391 //const u32x w7_t = uint_to_hex_upper8 ((state[3] >> 16) & 255) << 0
392 // | uint_to_hex_upper8 ((state[3] >> 24) & 255) << 16;
394 const u32 pade = 0x0e0e0e0e;
397 w_tmp[ 1] = salt1 | w0_t << 16;
398 w_tmp[ 2] = w0_t >> 16 | w1_t << 16;
399 w_tmp[ 3] = w1_t >> 16 | w2_t << 16;
400 w_tmp[ 4] = w2_t >> 16 | w3_t << 16;
401 w_tmp[ 5] = w3_t >> 16 | w4_t << 16;
402 w_tmp[ 6] = w4_t >> 16 | w5_t << 16;
403 w_tmp[ 7] = w5_t >> 16 | w6_t << 16;
404 w_tmp[ 8] = w6_t >> 16 | pade << 16; // | w7_t << 8;
418 domino_big_md (w_tmp, 34, state, s_lotus_magic_table);
420 u32x a = state[0] & 0xffffffff;
421 u32x b = state[1] & 0xffffffff;
422 u32x c = state[2] & 0x000000ff;
423 u32x d = state[3] & 0x00000000;
430 #include VECT_COMPARE_M
434 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 gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32x * 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)
440 const u32 gid = get_global_id (0);
441 const u32 lid = get_local_id (0);
449 pad (&w[ 0], pw_len & 0xf);
451 else if (pw_len < 32)
453 pad (&w[ 4], pw_len & 0xf);
455 else if (pw_len < 48)
457 pad (&w[ 8], pw_len & 0xf);
459 else if (pw_len < 64)
461 pad (&w[12], pw_len & 0xf);
468 const u32 salt0 = salt_bufs[salt_pos].salt_buf[0];
469 const u32 salt1 = salt_bufs[salt_pos].salt_buf[1] & 0xff | '(' << 8;
475 const u32 search[4] =
477 digests_buf[digests_offset].digest_buf[DGST_R0],
478 digests_buf[digests_offset].digest_buf[DGST_R1],
479 digests_buf[digests_offset].digest_buf[DGST_R2],
480 digests_buf[digests_offset].digest_buf[DGST_R3]
487 const u32 bf_loops = ceil ((float) bfs_cnt / VECT_DIV);
491 for (u32 il_pos = 0; il_pos < bf_loops; il_pos++)
493 const u32x w0r = words_buf_r[il_pos];
495 const u32x w0 = w0l | w0r;
523 domino_big_md (w_tmp, pw_len, state, s_lotus_magic_table);
525 const u32x w0_t = uint_to_hex_upper8 ((state[0] >> 0) & 255) << 0
526 | uint_to_hex_upper8 ((state[0] >> 8) & 255) << 16;
527 const u32x w1_t = uint_to_hex_upper8 ((state[0] >> 16) & 255) << 0
528 | uint_to_hex_upper8 ((state[0] >> 24) & 255) << 16;
529 const u32x w2_t = uint_to_hex_upper8 ((state[1] >> 0) & 255) << 0
530 | uint_to_hex_upper8 ((state[1] >> 8) & 255) << 16;
531 const u32x w3_t = uint_to_hex_upper8 ((state[1] >> 16) & 255) << 0
532 | uint_to_hex_upper8 ((state[1] >> 24) & 255) << 16;
533 const u32x w4_t = uint_to_hex_upper8 ((state[2] >> 0) & 255) << 0
534 | uint_to_hex_upper8 ((state[2] >> 8) & 255) << 16;
535 const u32x w5_t = uint_to_hex_upper8 ((state[2] >> 16) & 255) << 0
536 | uint_to_hex_upper8 ((state[2] >> 24) & 255) << 16;
537 const u32x w6_t = uint_to_hex_upper8 ((state[3] >> 0) & 255) << 0
538 | uint_to_hex_upper8 ((state[3] >> 8) & 255) << 16;
539 //const u32x w7_t = uint_to_hex_upper8 ((state[3] >> 16) & 255) << 0
540 // | uint_to_hex_upper8 ((state[3] >> 24) & 255) << 16;
542 const u32 pade = 0x0e0e0e0e;
545 w_tmp[ 1] = salt1 | w0_t << 16;
546 w_tmp[ 2] = w0_t >> 16 | w1_t << 16;
547 w_tmp[ 3] = w1_t >> 16 | w2_t << 16;
548 w_tmp[ 4] = w2_t >> 16 | w3_t << 16;
549 w_tmp[ 5] = w3_t >> 16 | w4_t << 16;
550 w_tmp[ 6] = w4_t >> 16 | w5_t << 16;
551 w_tmp[ 7] = w5_t >> 16 | w6_t << 16;
552 w_tmp[ 8] = w6_t >> 16 | pade << 16; // | w7_t << 8;
566 domino_big_md (w_tmp, 34, state, s_lotus_magic_table);
568 u32x a = state[0] & 0xffffffff;
569 u32x b = state[1] & 0xffffffff;
570 u32x c = state[2] & 0x000000ff;
571 u32x d = state[3] & 0x00000000;
578 #include VECT_COMPARE_S
582 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_m04 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32x * 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)
588 const u32 gid = get_global_id (0);
589 const u32 lid = get_local_id (0);
593 w[ 0] = pws[gid].i[ 0];
594 w[ 1] = pws[gid].i[ 1];
595 w[ 2] = pws[gid].i[ 2];
596 w[ 3] = pws[gid].i[ 3];
610 const u32 pw_len = pws[gid].pw_len;
616 __local u32 l_bin2asc[256];
618 const u32 lid4 = lid * 4;
620 const u32 lid40 = lid4 + 0;
621 const u32 lid41 = lid4 + 1;
622 const u32 lid42 = lid4 + 2;
623 const u32 lid43 = lid4 + 3;
625 const u32 v400 = (lid40 >> 0) & 15;
626 const u32 v401 = (lid40 >> 4) & 15;
627 const u32 v410 = (lid41 >> 0) & 15;
628 const u32 v411 = (lid41 >> 4) & 15;
629 const u32 v420 = (lid42 >> 0) & 15;
630 const u32 v421 = (lid42 >> 4) & 15;
631 const u32 v430 = (lid43 >> 0) & 15;
632 const u32 v431 = (lid43 >> 4) & 15;
634 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
635 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
636 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
637 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
638 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
639 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
640 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
641 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
643 barrier (CLK_LOCAL_MEM_FENCE);
649 __local u32 s_lotus_magic_table[256];
651 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
652 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
653 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
654 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
656 barrier (CLK_LOCAL_MEM_FENCE);
658 if (gid >= gid_max) return;
664 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);
667 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_m08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32x * 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)
673 const u32 gid = get_global_id (0);
674 const u32 lid = get_local_id (0);
678 w[ 0] = pws[gid].i[ 0];
679 w[ 1] = pws[gid].i[ 1];
680 w[ 2] = pws[gid].i[ 2];
681 w[ 3] = pws[gid].i[ 3];
682 w[ 4] = pws[gid].i[ 4];
683 w[ 5] = pws[gid].i[ 5];
684 w[ 6] = pws[gid].i[ 6];
685 w[ 7] = pws[gid].i[ 7];
695 const u32 pw_len = pws[gid].pw_len;
701 __local u32 l_bin2asc[256];
703 const u32 lid4 = lid * 4;
705 const u32 lid40 = lid4 + 0;
706 const u32 lid41 = lid4 + 1;
707 const u32 lid42 = lid4 + 2;
708 const u32 lid43 = lid4 + 3;
710 const u32 v400 = (lid40 >> 0) & 15;
711 const u32 v401 = (lid40 >> 4) & 15;
712 const u32 v410 = (lid41 >> 0) & 15;
713 const u32 v411 = (lid41 >> 4) & 15;
714 const u32 v420 = (lid42 >> 0) & 15;
715 const u32 v421 = (lid42 >> 4) & 15;
716 const u32 v430 = (lid43 >> 0) & 15;
717 const u32 v431 = (lid43 >> 4) & 15;
719 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
720 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
721 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
722 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
723 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
724 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
725 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
726 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
728 barrier (CLK_LOCAL_MEM_FENCE);
734 __local u32 s_lotus_magic_table[256];
736 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
737 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
738 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
739 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
741 barrier (CLK_LOCAL_MEM_FENCE);
743 if (gid >= gid_max) return;
749 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);
752 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_m16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32x * 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)
758 const u32 gid = get_global_id (0);
759 const u32 lid = get_local_id (0);
763 w[ 0] = pws[gid].i[ 0];
764 w[ 1] = pws[gid].i[ 1];
765 w[ 2] = pws[gid].i[ 2];
766 w[ 3] = pws[gid].i[ 3];
767 w[ 4] = pws[gid].i[ 4];
768 w[ 5] = pws[gid].i[ 5];
769 w[ 6] = pws[gid].i[ 6];
770 w[ 7] = pws[gid].i[ 7];
771 w[ 8] = pws[gid].i[ 8];
772 w[ 9] = pws[gid].i[ 9];
773 w[10] = pws[gid].i[10];
774 w[11] = pws[gid].i[11];
775 w[12] = pws[gid].i[12];
776 w[13] = pws[gid].i[13];
777 w[14] = pws[gid].i[14];
778 w[15] = pws[gid].i[15];
780 const u32 pw_len = pws[gid].pw_len;
786 __local u32 l_bin2asc[256];
788 const u32 lid4 = lid * 4;
790 const u32 lid40 = lid4 + 0;
791 const u32 lid41 = lid4 + 1;
792 const u32 lid42 = lid4 + 2;
793 const u32 lid43 = lid4 + 3;
795 const u32 v400 = (lid40 >> 0) & 15;
796 const u32 v401 = (lid40 >> 4) & 15;
797 const u32 v410 = (lid41 >> 0) & 15;
798 const u32 v411 = (lid41 >> 4) & 15;
799 const u32 v420 = (lid42 >> 0) & 15;
800 const u32 v421 = (lid42 >> 4) & 15;
801 const u32 v430 = (lid43 >> 0) & 15;
802 const u32 v431 = (lid43 >> 4) & 15;
804 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
805 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
806 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
807 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
808 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
809 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
810 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
811 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
813 barrier (CLK_LOCAL_MEM_FENCE);
819 __local u32 s_lotus_magic_table[256];
821 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
822 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
823 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
824 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
826 barrier (CLK_LOCAL_MEM_FENCE);
828 if (gid >= gid_max) return;
834 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);
837 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_s04 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32x * 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)
843 const u32 gid = get_global_id (0);
844 const u32 lid = get_local_id (0);
848 w[ 0] = pws[gid].i[ 0];
849 w[ 1] = pws[gid].i[ 1];
850 w[ 2] = pws[gid].i[ 2];
851 w[ 3] = pws[gid].i[ 3];
865 const u32 pw_len = pws[gid].pw_len;
871 __local u32 l_bin2asc[256];
873 const u32 lid4 = lid * 4;
875 const u32 lid40 = lid4 + 0;
876 const u32 lid41 = lid4 + 1;
877 const u32 lid42 = lid4 + 2;
878 const u32 lid43 = lid4 + 3;
880 const u32 v400 = (lid40 >> 0) & 15;
881 const u32 v401 = (lid40 >> 4) & 15;
882 const u32 v410 = (lid41 >> 0) & 15;
883 const u32 v411 = (lid41 >> 4) & 15;
884 const u32 v420 = (lid42 >> 0) & 15;
885 const u32 v421 = (lid42 >> 4) & 15;
886 const u32 v430 = (lid43 >> 0) & 15;
887 const u32 v431 = (lid43 >> 4) & 15;
889 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
890 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
891 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
892 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
893 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
894 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
895 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
896 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
898 barrier (CLK_LOCAL_MEM_FENCE);
904 __local u32 s_lotus_magic_table[256];
906 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
907 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
908 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
909 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
911 barrier (CLK_LOCAL_MEM_FENCE);
913 if (gid >= gid_max) return;
919 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);
922 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_s08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32x * 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)
928 const u32 gid = get_global_id (0);
929 const u32 lid = get_local_id (0);
933 w[ 0] = pws[gid].i[ 0];
934 w[ 1] = pws[gid].i[ 1];
935 w[ 2] = pws[gid].i[ 2];
936 w[ 3] = pws[gid].i[ 3];
937 w[ 4] = pws[gid].i[ 4];
938 w[ 5] = pws[gid].i[ 5];
939 w[ 6] = pws[gid].i[ 6];
940 w[ 7] = pws[gid].i[ 7];
950 const u32 pw_len = pws[gid].pw_len;
956 __local u32 l_bin2asc[256];
958 const u32 lid4 = lid * 4;
960 const u32 lid40 = lid4 + 0;
961 const u32 lid41 = lid4 + 1;
962 const u32 lid42 = lid4 + 2;
963 const u32 lid43 = lid4 + 3;
965 const u32 v400 = (lid40 >> 0) & 15;
966 const u32 v401 = (lid40 >> 4) & 15;
967 const u32 v410 = (lid41 >> 0) & 15;
968 const u32 v411 = (lid41 >> 4) & 15;
969 const u32 v420 = (lid42 >> 0) & 15;
970 const u32 v421 = (lid42 >> 4) & 15;
971 const u32 v430 = (lid43 >> 0) & 15;
972 const u32 v431 = (lid43 >> 4) & 15;
974 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
975 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
976 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
977 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
978 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
979 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
980 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
981 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
983 barrier (CLK_LOCAL_MEM_FENCE);
989 __local u32 s_lotus_magic_table[256];
991 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
992 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
993 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
994 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
996 barrier (CLK_LOCAL_MEM_FENCE);
998 if (gid >= gid_max) return;
1004 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);
1007 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08700_s16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global u32x * 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)
1013 const u32 gid = get_global_id (0);
1014 const u32 lid = get_local_id (0);
1018 w[ 0] = pws[gid].i[ 0];
1019 w[ 1] = pws[gid].i[ 1];
1020 w[ 2] = pws[gid].i[ 2];
1021 w[ 3] = pws[gid].i[ 3];
1022 w[ 4] = pws[gid].i[ 4];
1023 w[ 5] = pws[gid].i[ 5];
1024 w[ 6] = pws[gid].i[ 6];
1025 w[ 7] = pws[gid].i[ 7];
1026 w[ 8] = pws[gid].i[ 8];
1027 w[ 9] = pws[gid].i[ 9];
1028 w[10] = pws[gid].i[10];
1029 w[11] = pws[gid].i[11];
1030 w[12] = pws[gid].i[12];
1031 w[13] = pws[gid].i[13];
1032 w[14] = pws[gid].i[14];
1033 w[15] = pws[gid].i[15];
1035 const u32 pw_len = pws[gid].pw_len;
1041 __local u32 l_bin2asc[256];
1043 const u32 lid4 = lid * 4;
1045 const u32 lid40 = lid4 + 0;
1046 const u32 lid41 = lid4 + 1;
1047 const u32 lid42 = lid4 + 2;
1048 const u32 lid43 = lid4 + 3;
1050 const u32 v400 = (lid40 >> 0) & 15;
1051 const u32 v401 = (lid40 >> 4) & 15;
1052 const u32 v410 = (lid41 >> 0) & 15;
1053 const u32 v411 = (lid41 >> 4) & 15;
1054 const u32 v420 = (lid42 >> 0) & 15;
1055 const u32 v421 = (lid42 >> 4) & 15;
1056 const u32 v430 = (lid43 >> 0) & 15;
1057 const u32 v431 = (lid43 >> 4) & 15;
1059 l_bin2asc[lid40] = ((v400 < 10) ? '0' + v400 : 'A' - 10 + v400) << 8
1060 | ((v401 < 10) ? '0' + v401 : 'A' - 10 + v401) << 0;
1061 l_bin2asc[lid41] = ((v410 < 10) ? '0' + v410 : 'A' - 10 + v410) << 8
1062 | ((v411 < 10) ? '0' + v411 : 'A' - 10 + v411) << 0;
1063 l_bin2asc[lid42] = ((v420 < 10) ? '0' + v420 : 'A' - 10 + v420) << 8
1064 | ((v421 < 10) ? '0' + v421 : 'A' - 10 + v421) << 0;
1065 l_bin2asc[lid43] = ((v430 < 10) ? '0' + v430 : 'A' - 10 + v430) << 8
1066 | ((v431 < 10) ? '0' + v431 : 'A' - 10 + v431) << 0;
1068 barrier (CLK_LOCAL_MEM_FENCE);
1074 __local u32 s_lotus_magic_table[256];
1076 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
1077 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
1078 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
1079 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
1081 barrier (CLK_LOCAL_MEM_FENCE);
1083 if (gid >= gid_max) return;
1089 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);