2 * Authors.....: Jens Steube <jens.steube@gmail.com>
3 * Gabriele Gristina <matrix@hashcat.net>
4 * magnum <john.magnum@hushmail.com>
13 #include "include/constants.h"
14 #include "include/kernel_vendor.h"
21 #include "include/kernel_functions.c"
22 #include "OpenCL/types_ocl.c"
23 #include "OpenCL/common.c"
24 #include "OpenCL/simd.c"
26 __constant u32 lotus_magic_table[256] =
28 0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
29 0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
30 0x30, 0x04, 0xb6, 0xdc, 0x7d, 0xdf, 0x32, 0x4b,
31 0xf7, 0xcb, 0x45, 0x9b, 0x31, 0xbb, 0x21, 0x5a,
32 0x41, 0x9f, 0xe1, 0xd9, 0x4a, 0x4d, 0x9e, 0xda,
33 0xa0, 0x68, 0x2c, 0xc3, 0x27, 0x5f, 0x80, 0x36,
34 0x3e, 0xee, 0xfb, 0x95, 0x1a, 0xfe, 0xce, 0xa8,
35 0x34, 0xa9, 0x13, 0xf0, 0xa6, 0x3f, 0xd8, 0x0c,
36 0x78, 0x24, 0xaf, 0x23, 0x52, 0xc1, 0x67, 0x17,
37 0xf5, 0x66, 0x90, 0xe7, 0xe8, 0x07, 0xb8, 0x60,
38 0x48, 0xe6, 0x1e, 0x53, 0xf3, 0x92, 0xa4, 0x72,
39 0x8c, 0x08, 0x15, 0x6e, 0x86, 0x00, 0x84, 0xfa,
40 0xf4, 0x7f, 0x8a, 0x42, 0x19, 0xf6, 0xdb, 0xcd,
41 0x14, 0x8d, 0x50, 0x12, 0xba, 0x3c, 0x06, 0x4e,
42 0xec, 0xb3, 0x35, 0x11, 0xa1, 0x88, 0x8e, 0x2b,
43 0x94, 0x99, 0xb7, 0x71, 0x74, 0xd3, 0xe4, 0xbf,
44 0x3a, 0xde, 0x96, 0x0e, 0xbc, 0x0a, 0xed, 0x77,
45 0xfc, 0x37, 0x6b, 0x03, 0x79, 0x89, 0x62, 0xc6,
46 0xd7, 0xc0, 0xd2, 0x7c, 0x6a, 0x8b, 0x22, 0xa3,
47 0x5b, 0x05, 0x5d, 0x02, 0x75, 0xd5, 0x61, 0xe3,
48 0x18, 0x8f, 0x55, 0x51, 0xad, 0x1f, 0x0b, 0x5e,
49 0x85, 0xe5, 0xc2, 0x57, 0x63, 0xca, 0x3d, 0x6c,
50 0xb4, 0xc5, 0xcc, 0x70, 0xb2, 0x91, 0x59, 0x0d,
51 0x47, 0x20, 0xc8, 0x4f, 0x58, 0xe0, 0x01, 0xe2,
52 0x16, 0x38, 0xc4, 0x6f, 0x3b, 0x0f, 0x65, 0x46,
53 0xbe, 0x7e, 0x2d, 0x7b, 0x82, 0xf9, 0x40, 0xb5,
54 0x1d, 0x73, 0xf8, 0xeb, 0x26, 0xc7, 0x87, 0x97,
55 0x25, 0x54, 0xb1, 0x28, 0xaa, 0x98, 0x9d, 0xa5,
56 0x64, 0x6d, 0x7a, 0xd4, 0x10, 0x81, 0x44, 0xef,
57 0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
58 0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
59 0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
62 #define BOX(S,i) (S)[(i)]
65 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i)])
67 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1])
69 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3])
71 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7])
73 #define uint_to_hex_upper8(i) (u32x) (l_bin2asc[(i).s0], l_bin2asc[(i).s1], l_bin2asc[(i).s2], l_bin2asc[(i).s3], l_bin2asc[(i).s4], l_bin2asc[(i).s5], l_bin2asc[(i).s6], l_bin2asc[(i).s7], l_bin2asc[(i).s8], l_bin2asc[(i).s9], l_bin2asc[(i).sa], l_bin2asc[(i).sb], l_bin2asc[(i).sc], l_bin2asc[(i).sd], l_bin2asc[(i).se], l_bin2asc[(i).sf])
77 #define BOX1(S,i) (S)[(i)]
79 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1])
81 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3])
83 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7])
85 #define BOX1(S,i) (u32x) ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3], (S)[(i).s4], (S)[(i).s5], (S)[(i).s6], (S)[(i).s7], (S)[(i).s8], (S)[(i).s9], (S)[(i).sa], (S)[(i).sb], (S)[(i).sc], (S)[(i).sd], (S)[(i).se], (S)[(i).sf])
88 void lotus_mix (u32x *in, __local u32 *s_lotus_magic_table)
92 for (int i = 0; i < 18; i++)
99 for (int j = 0; j < 12; j++)
104 p = (p + s--) & 0xff; p = ((tmp_in >> 0) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 0;
105 p = (p + s--) & 0xff; p = ((tmp_in >> 8) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 8;
106 p = (p + s--) & 0xff; p = ((tmp_in >> 16) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 16;
107 p = (p + s--) & 0xff; p = ((tmp_in >> 24) & 0xff) ^ BOX1 (s_lotus_magic_table, p); tmp_out |= p << 24;
114 void lotus_transform_password (u32x in[4], u32x out[4], __local u32 *s_lotus_magic_table)
116 u32x t = out[3] >> 24;
123 for (int i = 0; i < 4; i++)
125 t ^= (in[i] >> 0) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 0; t = ((out[i] >> 0) & 0xff);
126 t ^= (in[i] >> 8) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 8; t = ((out[i] >> 8) & 0xff);
127 t ^= (in[i] >> 16) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 16; t = ((out[i] >> 16) & 0xff);
128 t ^= (in[i] >> 24) & 0xff; c = BOX1 (s_lotus_magic_table, t); out[i] ^= c << 24; t = ((out[i] >> 24) & 0xff);
132 void pad (u32 w[4], const u32 len)
134 const u32 val = 16 - len;
136 const u32 mask1 = val << 24;
138 const u32 mask2 = val << 16
141 const u32 mask3 = val << 8
145 const u32 mask4 = val << 0
152 case 0: w[0] = mask4;
157 case 1: w[0] |= mask3;
162 case 2: w[0] |= mask2;
167 case 3: w[0] |= mask1;
172 case 4: w[1] = mask4;
176 case 5: w[1] |= mask3;
180 case 6: w[1] |= mask2;
184 case 7: w[1] |= mask1;
188 case 8: w[2] = mask4;
191 case 9: w[2] |= mask3;
194 case 10: w[2] |= mask2;
197 case 11: w[2] |= mask1;
200 case 12: w[3] = mask4;
202 case 13: w[3] |= mask3;
204 case 14: w[3] |= mask2;
206 case 15: w[3] |= mask1;
211 void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 *s_lotus_magic_table)
223 x[ 8] = state[0] ^ block[0];
224 x[ 9] = state[1] ^ block[1];
225 x[10] = state[2] ^ block[2];
226 x[11] = state[3] ^ block[3];
228 lotus_mix (x, s_lotus_magic_table);
236 void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 *s_lotus_magic_table)
238 mdtransform_norecalc (state, block, s_lotus_magic_table);
240 lotus_transform_password (block, checksum, s_lotus_magic_table);
243 void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 *s_lotus_magic_table)
262 for (curpos = 0, idx = 0; curpos + 16 < size; curpos += 16, idx += 4)
264 block[0] = saved_key[idx + 0];
265 block[1] = saved_key[idx + 1];
266 block[2] = saved_key[idx + 2];
267 block[3] = saved_key[idx + 3];
269 mdtransform (state, checksum, block, s_lotus_magic_table);
272 block[0] = saved_key[idx + 0];
273 block[1] = saved_key[idx + 1];
274 block[2] = saved_key[idx + 2];
275 block[3] = saved_key[idx + 3];
277 mdtransform (state, checksum, block, s_lotus_magic_table);
279 mdtransform_norecalc (state, checksum, s_lotus_magic_table);
282 void m08700m (__local u32 *s_lotus_magic_table, __local u32 *l_bin2asc, u32 w[16], const u32 pw_len, __global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant 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 il_cnt, const u32 digests_cnt, const u32 digests_offset)
288 const u32 gid = get_global_id (0);
289 const u32 lid = get_local_id (0);
297 pad (&w[ 0], pw_len & 0xf);
299 else if (pw_len < 32)
301 pad (&w[ 4], pw_len & 0xf);
303 else if (pw_len < 48)
305 pad (&w[ 8], pw_len & 0xf);
307 else if (pw_len < 64)
309 pad (&w[12], pw_len & 0xf);
316 const u32 salt0 = salt_bufs[salt_pos].salt_buf[0];
317 const u32 salt1 = (salt_bufs[salt_pos].salt_buf[1] & 0xff) | '(' << 8;
325 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
327 const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
329 const u32x w0 = w0l | w0r;
361 domino_big_md (w_tmp, pw_len, state, s_lotus_magic_table);
363 const u32x w0_t = uint_to_hex_upper8 ((state[0] >> 0) & 255) << 0
364 | uint_to_hex_upper8 ((state[0] >> 8) & 255) << 16;
365 const u32x w1_t = uint_to_hex_upper8 ((state[0] >> 16) & 255) << 0
366 | uint_to_hex_upper8 ((state[0] >> 24) & 255) << 16;
367 const u32x w2_t = uint_to_hex_upper8 ((state[1] >> 0) & 255) << 0
368 | uint_to_hex_upper8 ((state[1] >> 8) & 255) << 16;
369 const u32x w3_t = uint_to_hex_upper8 ((state[1] >> 16) & 255) << 0
370 | uint_to_hex_upper8 ((state[1] >> 24) & 255) << 16;
371 const u32x w4_t = uint_to_hex_upper8 ((state[2] >> 0) & 255) << 0
372 | uint_to_hex_upper8 ((state[2] >> 8) & 255) << 16;
373 const u32x w5_t = uint_to_hex_upper8 ((state[2] >> 16) & 255) << 0
374 | uint_to_hex_upper8 ((state[2] >> 24) & 255) << 16;
375 const u32x w6_t = uint_to_hex_upper8 ((state[3] >> 0) & 255) << 0
376 | uint_to_hex_upper8 ((state[3] >> 8) & 255) << 16;
377 //const u32x w7_t = uint_to_hex_upper8 ((state[3] >> 16) & 255) << 0
378 // | uint_to_hex_upper8 ((state[3] >> 24) & 255) << 16;
380 const u32x pade = 0x0e0e0e0e;
383 w_tmp[ 1] = salt1 | w0_t << 16;
384 w_tmp[ 2] = w0_t >> 16 | w1_t << 16;
385 w_tmp[ 3] = w1_t >> 16 | w2_t << 16;
386 w_tmp[ 4] = w2_t >> 16 | w3_t << 16;
387 w_tmp[ 5] = w3_t >> 16 | w4_t << 16;
388 w_tmp[ 6] = w4_t >> 16 | w5_t << 16;
389 w_tmp[ 7] = w5_t >> 16 | w6_t << 16;
390 w_tmp[ 8] = w6_t >> 16 | pade << 16; // | w7_t << 8;
404 domino_big_md (w_tmp, 34, state, s_lotus_magic_table);
406 u32x a = state[0] & 0xffffffff;
407 u32x b = state[1] & 0xffffffff;
408 u32x c = state[2] & 0x000000ff;
409 u32x d = state[3] & 0x00000000;
411 COMPARE_M_SIMD (a, b, c, d);
415 void m08700s (__local u32 *s_lotus_magic_table, __local u32 *l_bin2asc, u32 w[16], const u32 pw_len, __global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant 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 il_cnt, const u32 digests_cnt, const u32 digests_offset)
421 const u32 gid = get_global_id (0);
422 const u32 lid = get_local_id (0);
430 pad (&w[ 0], pw_len & 0xf);
432 else if (pw_len < 32)
434 pad (&w[ 4], pw_len & 0xf);
436 else if (pw_len < 48)
438 pad (&w[ 8], pw_len & 0xf);
440 else if (pw_len < 64)
442 pad (&w[12], pw_len & 0xf);
449 const u32 salt0 = salt_bufs[salt_pos].salt_buf[0];
450 const u32 salt1 = (salt_bufs[salt_pos].salt_buf[1] & 0xff) | '(' << 8;
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]
470 for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
472 const u32x w0r = words_buf_r[il_pos / VECT_SIZE];
474 const u32x w0 = w0l | w0r;
506 domino_big_md (w_tmp, pw_len, state, s_lotus_magic_table);
508 const u32x w0_t = uint_to_hex_upper8 ((state[0] >> 0) & 255) << 0
509 | uint_to_hex_upper8 ((state[0] >> 8) & 255) << 16;
510 const u32x w1_t = uint_to_hex_upper8 ((state[0] >> 16) & 255) << 0
511 | uint_to_hex_upper8 ((state[0] >> 24) & 255) << 16;
512 const u32x w2_t = uint_to_hex_upper8 ((state[1] >> 0) & 255) << 0
513 | uint_to_hex_upper8 ((state[1] >> 8) & 255) << 16;
514 const u32x w3_t = uint_to_hex_upper8 ((state[1] >> 16) & 255) << 0
515 | uint_to_hex_upper8 ((state[1] >> 24) & 255) << 16;
516 const u32x w4_t = uint_to_hex_upper8 ((state[2] >> 0) & 255) << 0
517 | uint_to_hex_upper8 ((state[2] >> 8) & 255) << 16;
518 const u32x w5_t = uint_to_hex_upper8 ((state[2] >> 16) & 255) << 0
519 | uint_to_hex_upper8 ((state[2] >> 24) & 255) << 16;
520 const u32x w6_t = uint_to_hex_upper8 ((state[3] >> 0) & 255) << 0
521 | uint_to_hex_upper8 ((state[3] >> 8) & 255) << 16;
522 //const u32x w7_t = uint_to_hex_upper8 ((state[3] >> 16) & 255) << 0
523 // | uint_to_hex_upper8 ((state[3] >> 24) & 255) << 16;
525 const u32x pade = 0x0e0e0e0e;
528 w_tmp[ 1] = salt1 | w0_t << 16;
529 w_tmp[ 2] = w0_t >> 16 | w1_t << 16;
530 w_tmp[ 3] = w1_t >> 16 | w2_t << 16;
531 w_tmp[ 4] = w2_t >> 16 | w3_t << 16;
532 w_tmp[ 5] = w3_t >> 16 | w4_t << 16;
533 w_tmp[ 6] = w4_t >> 16 | w5_t << 16;
534 w_tmp[ 7] = w5_t >> 16 | w6_t << 16;
535 w_tmp[ 8] = w6_t >> 16 | pade << 16; // | w7_t << 8;
549 domino_big_md (w_tmp, 34, state, s_lotus_magic_table);
551 u32x a = state[0] & 0xffffffff;
552 u32x b = state[1] & 0xffffffff;
553 u32x c = state[2] & 0x000000ff;
554 u32x d = state[3] & 0x00000000;
556 COMPARE_S_SIMD (a, b, c, d);
560 __kernel void m08700_m04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
566 const u32 gid = get_global_id (0);
567 const u32 lid = get_local_id (0);
568 const u32 lsz = get_local_size (0);
574 __local u32 s_lotus_magic_table[256];
576 for (u32 i = lid; i < 256; i += lsz)
578 s_lotus_magic_table[i] = lotus_magic_table[i];
581 __local u32 l_bin2asc[256];
583 for (u32 i = lid; i < 256; i += lsz)
585 const u32 i0 = (i >> 0) & 15;
586 const u32 i1 = (i >> 4) & 15;
588 l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'A' - 10 + i0) << 8
589 | ((i1 < 10) ? '0' + i1 : 'A' - 10 + i1) << 0;
592 barrier (CLK_LOCAL_MEM_FENCE);
594 if (gid >= gid_max) return;
602 w[ 0] = pws[gid].i[ 0];
603 w[ 1] = pws[gid].i[ 1];
604 w[ 2] = pws[gid].i[ 2];
605 w[ 3] = pws[gid].i[ 3];
619 const u32 pw_len = pws[gid].pw_len;
625 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, il_cnt, digests_cnt, digests_offset);
628 __kernel void m08700_m08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
634 const u32 gid = get_global_id (0);
635 const u32 lid = get_local_id (0);
636 const u32 lsz = get_local_size (0);
642 __local u32 s_lotus_magic_table[256];
644 for (u32 i = lid; i < 256; i += lsz)
646 s_lotus_magic_table[i] = lotus_magic_table[i];
649 __local u32 l_bin2asc[256];
651 for (u32 i = lid; i < 256; i += lsz)
653 const u32 i0 = (i >> 0) & 15;
654 const u32 i1 = (i >> 4) & 15;
656 l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'A' - 10 + i0) << 8
657 | ((i1 < 10) ? '0' + i1 : 'A' - 10 + i1) << 0;
660 barrier (CLK_LOCAL_MEM_FENCE);
662 if (gid >= gid_max) return;
670 w[ 0] = pws[gid].i[ 0];
671 w[ 1] = pws[gid].i[ 1];
672 w[ 2] = pws[gid].i[ 2];
673 w[ 3] = pws[gid].i[ 3];
674 w[ 4] = pws[gid].i[ 4];
675 w[ 5] = pws[gid].i[ 5];
676 w[ 6] = pws[gid].i[ 6];
677 w[ 7] = pws[gid].i[ 7];
687 const u32 pw_len = pws[gid].pw_len;
693 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, il_cnt, digests_cnt, digests_offset);
696 __kernel void m08700_m16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
702 const u32 gid = get_global_id (0);
703 const u32 lid = get_local_id (0);
704 const u32 lsz = get_local_size (0);
710 __local u32 s_lotus_magic_table[256];
712 for (u32 i = lid; i < 256; i += lsz)
714 s_lotus_magic_table[i] = lotus_magic_table[i];
717 __local u32 l_bin2asc[256];
719 for (u32 i = lid; i < 256; i += lsz)
721 const u32 i0 = (i >> 0) & 15;
722 const u32 i1 = (i >> 4) & 15;
724 l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'A' - 10 + i0) << 8
725 | ((i1 < 10) ? '0' + i1 : 'A' - 10 + i1) << 0;
728 barrier (CLK_LOCAL_MEM_FENCE);
730 if (gid >= gid_max) return;
738 w[ 0] = pws[gid].i[ 0];
739 w[ 1] = pws[gid].i[ 1];
740 w[ 2] = pws[gid].i[ 2];
741 w[ 3] = pws[gid].i[ 3];
742 w[ 4] = pws[gid].i[ 4];
743 w[ 5] = pws[gid].i[ 5];
744 w[ 6] = pws[gid].i[ 6];
745 w[ 7] = pws[gid].i[ 7];
746 w[ 8] = pws[gid].i[ 8];
747 w[ 9] = pws[gid].i[ 9];
748 w[10] = pws[gid].i[10];
749 w[11] = pws[gid].i[11];
750 w[12] = pws[gid].i[12];
751 w[13] = pws[gid].i[13];
752 w[14] = pws[gid].i[14];
753 w[15] = pws[gid].i[15];
755 const u32 pw_len = pws[gid].pw_len;
761 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, il_cnt, digests_cnt, digests_offset);
764 __kernel void m08700_s04 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
770 const u32 gid = get_global_id (0);
771 const u32 lid = get_local_id (0);
772 const u32 lsz = get_local_size (0);
778 __local u32 s_lotus_magic_table[256];
780 for (u32 i = lid; i < 256; i += lsz)
782 s_lotus_magic_table[i] = lotus_magic_table[i];
785 __local u32 l_bin2asc[256];
787 for (u32 i = lid; i < 256; i += lsz)
789 const u32 i0 = (i >> 0) & 15;
790 const u32 i1 = (i >> 4) & 15;
792 l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'A' - 10 + i0) << 8
793 | ((i1 < 10) ? '0' + i1 : 'A' - 10 + i1) << 0;
796 barrier (CLK_LOCAL_MEM_FENCE);
798 if (gid >= gid_max) return;
806 w[ 0] = pws[gid].i[ 0];
807 w[ 1] = pws[gid].i[ 1];
808 w[ 2] = pws[gid].i[ 2];
809 w[ 3] = pws[gid].i[ 3];
823 const u32 pw_len = pws[gid].pw_len;
829 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, il_cnt, digests_cnt, digests_offset);
832 __kernel void m08700_s08 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
838 const u32 gid = get_global_id (0);
839 const u32 lid = get_local_id (0);
840 const u32 lsz = get_local_size (0);
846 __local u32 s_lotus_magic_table[256];
848 for (u32 i = lid; i < 256; i += lsz)
850 s_lotus_magic_table[i] = lotus_magic_table[i];
853 __local u32 l_bin2asc[256];
855 for (u32 i = lid; i < 256; i += lsz)
857 const u32 i0 = (i >> 0) & 15;
858 const u32 i1 = (i >> 4) & 15;
860 l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'A' - 10 + i0) << 8
861 | ((i1 < 10) ? '0' + i1 : 'A' - 10 + i1) << 0;
864 barrier (CLK_LOCAL_MEM_FENCE);
866 if (gid >= gid_max) return;
874 w[ 0] = pws[gid].i[ 0];
875 w[ 1] = pws[gid].i[ 1];
876 w[ 2] = pws[gid].i[ 2];
877 w[ 3] = pws[gid].i[ 3];
878 w[ 4] = pws[gid].i[ 4];
879 w[ 5] = pws[gid].i[ 5];
880 w[ 6] = pws[gid].i[ 6];
881 w[ 7] = pws[gid].i[ 7];
891 const u32 pw_len = pws[gid].pw_len;
897 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, il_cnt, digests_cnt, digests_offset);
900 __kernel void m08700_s16 (__global pw_t *pws, __global kernel_rule_t *rules_buf, __global comb_t *combs_buf, __constant 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 il_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
906 const u32 gid = get_global_id (0);
907 const u32 lid = get_local_id (0);
908 const u32 lsz = get_local_size (0);
914 __local u32 s_lotus_magic_table[256];
916 for (u32 i = lid; i < 256; i += lsz)
918 s_lotus_magic_table[i] = lotus_magic_table[i];
921 __local u32 l_bin2asc[256];
923 for (u32 i = lid; i < 256; i += lsz)
925 const u32 i0 = (i >> 0) & 15;
926 const u32 i1 = (i >> 4) & 15;
928 l_bin2asc[i] = ((i0 < 10) ? '0' + i0 : 'A' - 10 + i0) << 8
929 | ((i1 < 10) ? '0' + i1 : 'A' - 10 + i1) << 0;
932 barrier (CLK_LOCAL_MEM_FENCE);
934 if (gid >= gid_max) return;
942 w[ 0] = pws[gid].i[ 0];
943 w[ 1] = pws[gid].i[ 1];
944 w[ 2] = pws[gid].i[ 2];
945 w[ 3] = pws[gid].i[ 3];
946 w[ 4] = pws[gid].i[ 4];
947 w[ 5] = pws[gid].i[ 5];
948 w[ 6] = pws[gid].i[ 6];
949 w[ 7] = pws[gid].i[ 7];
950 w[ 8] = pws[gid].i[ 8];
951 w[ 9] = pws[gid].i[ 9];
952 w[10] = pws[gid].i[10];
953 w[11] = pws[gid].i[11];
954 w[12] = pws[gid].i[12];
955 w[13] = pws[gid].i[13];
956 w[14] = pws[gid].i[14];
957 w[15] = pws[gid].i[15];
959 const u32 pw_len = pws[gid].pw_len;
965 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, il_cnt, digests_cnt, digests_offset);