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"
33 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
34 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
38 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
39 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
43 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
44 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
47 __constant u32 lotus_magic_table[256] =
49 0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
50 0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
51 0x30, 0x04, 0xb6, 0xdc, 0x7d, 0xdf, 0x32, 0x4b,
52 0xf7, 0xcb, 0x45, 0x9b, 0x31, 0xbb, 0x21, 0x5a,
53 0x41, 0x9f, 0xe1, 0xd9, 0x4a, 0x4d, 0x9e, 0xda,
54 0xa0, 0x68, 0x2c, 0xc3, 0x27, 0x5f, 0x80, 0x36,
55 0x3e, 0xee, 0xfb, 0x95, 0x1a, 0xfe, 0xce, 0xa8,
56 0x34, 0xa9, 0x13, 0xf0, 0xa6, 0x3f, 0xd8, 0x0c,
57 0x78, 0x24, 0xaf, 0x23, 0x52, 0xc1, 0x67, 0x17,
58 0xf5, 0x66, 0x90, 0xe7, 0xe8, 0x07, 0xb8, 0x60,
59 0x48, 0xe6, 0x1e, 0x53, 0xf3, 0x92, 0xa4, 0x72,
60 0x8c, 0x08, 0x15, 0x6e, 0x86, 0x00, 0x84, 0xfa,
61 0xf4, 0x7f, 0x8a, 0x42, 0x19, 0xf6, 0xdb, 0xcd,
62 0x14, 0x8d, 0x50, 0x12, 0xba, 0x3c, 0x06, 0x4e,
63 0xec, 0xb3, 0x35, 0x11, 0xa1, 0x88, 0x8e, 0x2b,
64 0x94, 0x99, 0xb7, 0x71, 0x74, 0xd3, 0xe4, 0xbf,
65 0x3a, 0xde, 0x96, 0x0e, 0xbc, 0x0a, 0xed, 0x77,
66 0xfc, 0x37, 0x6b, 0x03, 0x79, 0x89, 0x62, 0xc6,
67 0xd7, 0xc0, 0xd2, 0x7c, 0x6a, 0x8b, 0x22, 0xa3,
68 0x5b, 0x05, 0x5d, 0x02, 0x75, 0xd5, 0x61, 0xe3,
69 0x18, 0x8f, 0x55, 0x51, 0xad, 0x1f, 0x0b, 0x5e,
70 0x85, 0xe5, 0xc2, 0x57, 0x63, 0xca, 0x3d, 0x6c,
71 0xb4, 0xc5, 0xcc, 0x70, 0xb2, 0x91, 0x59, 0x0d,
72 0x47, 0x20, 0xc8, 0x4f, 0x58, 0xe0, 0x01, 0xe2,
73 0x16, 0x38, 0xc4, 0x6f, 0x3b, 0x0f, 0x65, 0x46,
74 0xbe, 0x7e, 0x2d, 0x7b, 0x82, 0xf9, 0x40, 0xb5,
75 0x1d, 0x73, 0xf8, 0xeb, 0x26, 0xc7, 0x87, 0x97,
76 0x25, 0x54, 0xb1, 0x28, 0xaa, 0x98, 0x9d, 0xa5,
77 0x64, 0x6d, 0x7a, 0xd4, 0x10, 0x81, 0x44, 0xef,
78 0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
79 0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
80 0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
84 #define BOX(S,i) u32x ((S)[(i)])
88 #define BOX(S,i) u32x ((S)[(i).s0], (S)[(i).s1])
92 #define BOX(S,i) u32x ((S)[(i).s0], (S)[(i).s1], (S)[(i).s2], (S)[(i).s3])
95 static void lotus_mix (u32x *in, __local u32 s_lotus_magic_table[256])
99 for (int i = 0; i < 18; i++)
104 for (int j = 0; j < 12; j++)
109 p = (p + s--) & 0xff; p = ((tmp_in >> 0) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 0;
110 p = (p + s--) & 0xff; p = ((tmp_in >> 8) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 8;
111 p = (p + s--) & 0xff; p = ((tmp_in >> 16) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 16;
112 p = (p + s--) & 0xff; p = ((tmp_in >> 24) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 24;
119 static void lotus_transform_password (u32x *in, u32x *out, __local u32 s_lotus_magic_table[256])
121 u32x t = out[3] >> 24;
126 for (int i = 0; i < 4; i++)
128 t ^= (in[i] >> 0) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 0; t = ((out[i] >> 0) & 0xff);
129 t ^= (in[i] >> 8) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 8; t = ((out[i] >> 8) & 0xff);
130 t ^= (in[i] >> 16) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 16; t = ((out[i] >> 16) & 0xff);
131 t ^= (in[i] >> 24) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 24; t = ((out[i] >> 24) & 0xff);
135 static void pad (u32x w[4], const u32 len)
137 const u32 val = 16 - len;
139 const u32 mask1 = val << 24;
141 const u32 mask2 = val << 16
144 const u32 mask3 = val << 8
148 const u32 mask4 = val << 0
155 case 0: w[0] = mask4;
160 case 1: w[0] |= mask3;
165 case 2: w[0] |= mask2;
170 case 3: w[0] |= mask1;
175 case 4: w[1] = mask4;
179 case 5: w[1] |= mask3;
183 case 6: w[1] |= mask2;
187 case 7: w[1] |= mask1;
191 case 8: w[2] = mask4;
194 case 9: w[2] |= mask3;
197 case 10: w[2] |= mask2;
200 case 11: w[2] |= mask1;
203 case 12: w[3] = mask4;
205 case 13: w[3] |= mask3;
207 case 14: w[3] |= mask2;
209 case 15: w[3] |= mask1;
214 static void mdtransform_norecalc (u32x state[4], u32x block[4], __local u32 s_lotus_magic_table[256])
226 x[ 8] = state[0] ^ block[0];
227 x[ 9] = state[1] ^ block[1];
228 x[10] = state[2] ^ block[2];
229 x[11] = state[3] ^ block[3];
231 lotus_mix (x, s_lotus_magic_table);
239 static void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], __local u32 s_lotus_magic_table[256])
241 mdtransform_norecalc (state, block, s_lotus_magic_table);
243 lotus_transform_password (block, checksum, s_lotus_magic_table);
246 static void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], __local u32 s_lotus_magic_table[256])
257 block[0] = saved_key[0];
258 block[1] = saved_key[1];
259 block[2] = saved_key[2];
260 block[3] = saved_key[3];
262 mdtransform (state, checksum, block, s_lotus_magic_table);
264 mdtransform_norecalc (state, checksum, s_lotus_magic_table);
267 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08600_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
273 const u32 lid = get_local_id (0);
279 const u32 gid = get_global_id (0);
285 const u32 lid4 = lid * 4;
287 __local u32 s_lotus_magic_table[256];
289 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
290 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
291 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
292 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
294 barrier (CLK_LOCAL_MEM_FENCE);
296 if (gid >= gid_max) return;
300 wordl0[0] = pws[gid].i[ 0];
301 wordl0[1] = pws[gid].i[ 1];
302 wordl0[2] = pws[gid].i[ 2];
303 wordl0[3] = pws[gid].i[ 3];
307 wordl1[0] = pws[gid].i[ 4];
308 wordl1[1] = pws[gid].i[ 5];
309 wordl1[2] = pws[gid].i[ 6];
310 wordl1[3] = pws[gid].i[ 7];
326 const u32 pw_l_len = pws[gid].pw_len;
328 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
330 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
337 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
339 const u32 pw_r_len = combs_buf[il_pos].pw_len;
341 u32 pw_len = pw_l_len + pw_r_len;
345 wordr0[0] = combs_buf[il_pos].i[0];
346 wordr0[1] = combs_buf[il_pos].i[1];
347 wordr0[2] = combs_buf[il_pos].i[2];
348 wordr0[3] = combs_buf[il_pos].i[3];
352 wordr1[0] = combs_buf[il_pos].i[4];
353 wordr1[1] = combs_buf[il_pos].i[5];
354 wordr1[2] = combs_buf[il_pos].i[6];
355 wordr1[3] = combs_buf[il_pos].i[7];
371 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
373 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
378 w[ 0] = wordl0[0] | wordr0[0];
379 w[ 1] = wordl0[1] | wordr0[1];
380 w[ 2] = wordl0[2] | wordr0[2];
381 w[ 3] = wordl0[3] | wordr0[3];
382 w[ 4] = wordl1[0] | wordr1[0];
383 w[ 5] = wordl1[1] | wordr1[1];
384 w[ 6] = wordl1[2] | wordr1[2];
385 w[ 7] = wordl1[3] | wordr1[3];
386 w[ 8] = wordl2[0] | wordr2[0];
387 w[ 9] = wordl2[1] | wordr2[1];
388 w[10] = wordl2[2] | wordr2[2];
389 w[11] = wordl2[3] | wordr2[3];
390 w[12] = wordl3[0] | wordr3[0];
391 w[13] = wordl3[1] | wordr3[1];
392 w[14] = wordl3[2] | wordr3[2];
393 w[15] = wordl3[3] | wordr3[3];
408 pad (&w[ 0], pw_len & 0xf);
410 else if (pw_len < 32)
412 pad (&w[ 4], pw_len & 0xf);
414 else if (pw_len < 48)
416 pad (&w[ 8], pw_len & 0xf);
418 else if (pw_len < 64)
420 pad (&w[12], pw_len & 0xf);
423 domino_big_md (w, pw_len, state, s_lotus_magic_table);
425 const u32x r0 = state[0];
426 const u32x r1 = state[1];
427 const u32x r2 = state[2];
428 const u32x r3 = state[3];
430 #include VECT_COMPARE_M
434 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08600_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)
438 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08600_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)
442 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08600_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 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
448 const u32 lid = get_local_id (0);
454 const u32 gid = get_global_id (0);
460 const u32 lid4 = lid * 4;
462 __local u32 s_lotus_magic_table[256];
464 s_lotus_magic_table[lid4 + 0] = lotus_magic_table[lid4 + 0];
465 s_lotus_magic_table[lid4 + 1] = lotus_magic_table[lid4 + 1];
466 s_lotus_magic_table[lid4 + 2] = lotus_magic_table[lid4 + 2];
467 s_lotus_magic_table[lid4 + 3] = lotus_magic_table[lid4 + 3];
469 barrier (CLK_LOCAL_MEM_FENCE);
471 if (gid >= gid_max) return;
475 wordl0[0] = pws[gid].i[ 0];
476 wordl0[1] = pws[gid].i[ 1];
477 wordl0[2] = pws[gid].i[ 2];
478 wordl0[3] = pws[gid].i[ 3];
482 wordl1[0] = pws[gid].i[ 4];
483 wordl1[1] = pws[gid].i[ 5];
484 wordl1[2] = pws[gid].i[ 6];
485 wordl1[3] = pws[gid].i[ 7];
501 const u32 pw_l_len = pws[gid].pw_len;
503 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
505 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
512 const u32 search[4] =
514 digests_buf[digests_offset].digest_buf[DGST_R0],
515 digests_buf[digests_offset].digest_buf[DGST_R1],
516 digests_buf[digests_offset].digest_buf[DGST_R2],
517 digests_buf[digests_offset].digest_buf[DGST_R3]
524 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
526 const u32 pw_r_len = combs_buf[il_pos].pw_len;
528 u32 pw_len = pw_l_len + pw_r_len;
532 wordr0[0] = combs_buf[il_pos].i[0];
533 wordr0[1] = combs_buf[il_pos].i[1];
534 wordr0[2] = combs_buf[il_pos].i[2];
535 wordr0[3] = combs_buf[il_pos].i[3];
539 wordr1[0] = combs_buf[il_pos].i[4];
540 wordr1[1] = combs_buf[il_pos].i[5];
541 wordr1[2] = combs_buf[il_pos].i[6];
542 wordr1[3] = combs_buf[il_pos].i[7];
558 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
560 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
565 w[ 0] = wordl0[0] | wordr0[0];
566 w[ 1] = wordl0[1] | wordr0[1];
567 w[ 2] = wordl0[2] | wordr0[2];
568 w[ 3] = wordl0[3] | wordr0[3];
569 w[ 4] = wordl1[0] | wordr1[0];
570 w[ 5] = wordl1[1] | wordr1[1];
571 w[ 6] = wordl1[2] | wordr1[2];
572 w[ 7] = wordl1[3] | wordr1[3];
573 w[ 8] = wordl2[0] | wordr2[0];
574 w[ 9] = wordl2[1] | wordr2[1];
575 w[10] = wordl2[2] | wordr2[2];
576 w[11] = wordl2[3] | wordr2[3];
577 w[12] = wordl3[0] | wordr3[0];
578 w[13] = wordl3[1] | wordr3[1];
579 w[14] = wordl3[2] | wordr3[2];
580 w[15] = wordl3[3] | wordr3[3];
595 pad (&w[ 0], pw_len & 0xf);
597 else if (pw_len < 32)
599 pad (&w[ 4], pw_len & 0xf);
601 else if (pw_len < 48)
603 pad (&w[ 8], pw_len & 0xf);
605 else if (pw_len < 64)
607 pad (&w[12], pw_len & 0xf);
610 domino_big_md (w, pw_len, state, s_lotus_magic_table);
612 const u32x r0 = state[0];
613 const u32x r1 = state[1];
614 const u32x r2 = state[2];
615 const u32x r3 = state[3];
617 #include VECT_COMPARE_S
621 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08600_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)
625 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m08600_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)