2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
24 #include "include/kernel_functions.c"
26 #include "common_nv.c"
29 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
30 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
34 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
35 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
38 __device__ __constant__ u32 lotus_magic_table[256] =
40 0xbd, 0x56, 0xea, 0xf2, 0xa2, 0xf1, 0xac, 0x2a,
41 0xb0, 0x93, 0xd1, 0x9c, 0x1b, 0x33, 0xfd, 0xd0,
42 0x30, 0x04, 0xb6, 0xdc, 0x7d, 0xdf, 0x32, 0x4b,
43 0xf7, 0xcb, 0x45, 0x9b, 0x31, 0xbb, 0x21, 0x5a,
44 0x41, 0x9f, 0xe1, 0xd9, 0x4a, 0x4d, 0x9e, 0xda,
45 0xa0, 0x68, 0x2c, 0xc3, 0x27, 0x5f, 0x80, 0x36,
46 0x3e, 0xee, 0xfb, 0x95, 0x1a, 0xfe, 0xce, 0xa8,
47 0x34, 0xa9, 0x13, 0xf0, 0xa6, 0x3f, 0xd8, 0x0c,
48 0x78, 0x24, 0xaf, 0x23, 0x52, 0xc1, 0x67, 0x17,
49 0xf5, 0x66, 0x90, 0xe7, 0xe8, 0x07, 0xb8, 0x60,
50 0x48, 0xe6, 0x1e, 0x53, 0xf3, 0x92, 0xa4, 0x72,
51 0x8c, 0x08, 0x15, 0x6e, 0x86, 0x00, 0x84, 0xfa,
52 0xf4, 0x7f, 0x8a, 0x42, 0x19, 0xf6, 0xdb, 0xcd,
53 0x14, 0x8d, 0x50, 0x12, 0xba, 0x3c, 0x06, 0x4e,
54 0xec, 0xb3, 0x35, 0x11, 0xa1, 0x88, 0x8e, 0x2b,
55 0x94, 0x99, 0xb7, 0x71, 0x74, 0xd3, 0xe4, 0xbf,
56 0x3a, 0xde, 0x96, 0x0e, 0xbc, 0x0a, 0xed, 0x77,
57 0xfc, 0x37, 0x6b, 0x03, 0x79, 0x89, 0x62, 0xc6,
58 0xd7, 0xc0, 0xd2, 0x7c, 0x6a, 0x8b, 0x22, 0xa3,
59 0x5b, 0x05, 0x5d, 0x02, 0x75, 0xd5, 0x61, 0xe3,
60 0x18, 0x8f, 0x55, 0x51, 0xad, 0x1f, 0x0b, 0x5e,
61 0x85, 0xe5, 0xc2, 0x57, 0x63, 0xca, 0x3d, 0x6c,
62 0xb4, 0xc5, 0xcc, 0x70, 0xb2, 0x91, 0x59, 0x0d,
63 0x47, 0x20, 0xc8, 0x4f, 0x58, 0xe0, 0x01, 0xe2,
64 0x16, 0x38, 0xc4, 0x6f, 0x3b, 0x0f, 0x65, 0x46,
65 0xbe, 0x7e, 0x2d, 0x7b, 0x82, 0xf9, 0x40, 0xb5,
66 0x1d, 0x73, 0xf8, 0xeb, 0x26, 0xc7, 0x87, 0x97,
67 0x25, 0x54, 0xb1, 0x28, 0xaa, 0x98, 0x9d, 0xa5,
68 0x64, 0x6d, 0x7a, 0xd4, 0x10, 0x81, 0x44, 0xef,
69 0x49, 0xd6, 0xae, 0x2e, 0xdd, 0x76, 0x5c, 0x2f,
70 0xa7, 0x1c, 0xc9, 0x09, 0x69, 0x9a, 0x83, 0xcf,
71 0x29, 0x39, 0xb9, 0xe9, 0x4c, 0xff, 0x43, 0xab,
75 #define BOX(S,i) u32x ((S)[(i)])
79 #define BOX(S,i) u32x ((S)[(i).x], (S)[(i).y])
82 __device__ static void lotus_mix (u32x *in, u32 s_lotus_magic_table[256])
86 for (int i = 0; i < 18; i++)
91 for (int j = 0; j < 12; j++)
96 p = (p + s--) & 0xff; p = ((tmp_in >> 0) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 0;
97 p = (p + s--) & 0xff; p = ((tmp_in >> 8) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 8;
98 p = (p + s--) & 0xff; p = ((tmp_in >> 16) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 16;
99 p = (p + s--) & 0xff; p = ((tmp_in >> 24) & 0xff) ^ BOX (s_lotus_magic_table, p); tmp_out |= p << 24;
106 __device__ static void lotus_transform_password (u32x *in, u32x *out, u32 s_lotus_magic_table[256])
108 u32x t = out[3] >> 24;
112 for (int i = 0; i < 4; i++)
114 t ^= (in[i] >> 0) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 0; t = ((out[i] >> 0) & 0xff);
115 t ^= (in[i] >> 8) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 8; t = ((out[i] >> 8) & 0xff);
116 t ^= (in[i] >> 16) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 16; t = ((out[i] >> 16) & 0xff);
117 t ^= (in[i] >> 24) & 0xff; c = BOX (s_lotus_magic_table, t); out[i] ^= c << 24; t = ((out[i] >> 24) & 0xff);
121 __device__ static void pad (u32x w[4], const u32 len)
123 const u32 val = 16 - len;
125 const u32 mask1 = val << 24;
127 const u32 mask2 = val << 16
130 const u32 mask3 = val << 8
134 const u32 mask4 = val << 0
141 case 0: w[0] = mask4;
146 case 1: w[0] |= mask3;
151 case 2: w[0] |= mask2;
156 case 3: w[0] |= mask1;
161 case 4: w[1] = mask4;
165 case 5: w[1] |= mask3;
169 case 6: w[1] |= mask2;
173 case 7: w[1] |= mask1;
177 case 8: w[2] = mask4;
180 case 9: w[2] |= mask3;
183 case 10: w[2] |= mask2;
186 case 11: w[2] |= mask1;
189 case 12: w[3] = mask4;
191 case 13: w[3] |= mask3;
193 case 14: w[3] |= mask2;
195 case 15: w[3] |= mask1;
200 __device__ static void mdtransform_norecalc (u32x state[4], u32x block[4], u32 s_lotus_magic_table[256])
212 x[ 8] = state[0] ^ block[0];
213 x[ 9] = state[1] ^ block[1];
214 x[10] = state[2] ^ block[2];
215 x[11] = state[3] ^ block[3];
217 lotus_mix (x, s_lotus_magic_table);
225 __device__ static void mdtransform (u32x state[4], u32x checksum[4], u32x block[4], u32 s_lotus_magic_table[256])
227 mdtransform_norecalc (state, block, s_lotus_magic_table);
229 lotus_transform_password (block, checksum, s_lotus_magic_table);
232 __device__ static void domino_big_md (const u32x saved_key[16], const u32 size, u32x state[4], u32 s_lotus_magic_table[256])
243 block[0] = saved_key[0];
244 block[1] = saved_key[1];
245 block[2] = saved_key[2];
246 block[3] = saved_key[3];
248 mdtransform (state, checksum, block, s_lotus_magic_table);
250 mdtransform_norecalc (state, checksum, s_lotus_magic_table);
253 __device__ __constant__ comb_t c_combs[1024];
255 extern "C" __global__ void __launch_bounds__ (256, 1) m08600_m04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
261 const u32 lid = threadIdx.x;
267 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
273 __shared__ u32 s_lotus_magic_table[256];
275 s_lotus_magic_table[lid] = lotus_magic_table[lid];
279 if (gid >= gid_max) return;
283 wordl0[0] = pws[gid].i[ 0];
284 wordl0[1] = pws[gid].i[ 1];
285 wordl0[2] = pws[gid].i[ 2];
286 wordl0[3] = pws[gid].i[ 3];
290 wordl1[0] = pws[gid].i[ 4];
291 wordl1[1] = pws[gid].i[ 5];
292 wordl1[2] = pws[gid].i[ 6];
293 wordl1[3] = pws[gid].i[ 7];
309 const u32 pw_l_len = pws[gid].pw_len;
311 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
313 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
320 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
322 const u32 pw_r_len = c_combs[il_pos].pw_len;
324 u32 pw_len = pw_l_len + pw_r_len;
328 wordr0[0] = c_combs[il_pos].i[0];
329 wordr0[1] = c_combs[il_pos].i[1];
330 wordr0[2] = c_combs[il_pos].i[2];
331 wordr0[3] = c_combs[il_pos].i[3];
335 wordr1[0] = c_combs[il_pos].i[4];
336 wordr1[1] = c_combs[il_pos].i[5];
337 wordr1[2] = c_combs[il_pos].i[6];
338 wordr1[3] = c_combs[il_pos].i[7];
354 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
356 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
361 w[ 0] = wordl0[0] | wordr0[0];
362 w[ 1] = wordl0[1] | wordr0[1];
363 w[ 2] = wordl0[2] | wordr0[2];
364 w[ 3] = wordl0[3] | wordr0[3];
365 w[ 4] = wordl1[0] | wordr1[0];
366 w[ 5] = wordl1[1] | wordr1[1];
367 w[ 6] = wordl1[2] | wordr1[2];
368 w[ 7] = wordl1[3] | wordr1[3];
369 w[ 8] = wordl2[0] | wordr2[0];
370 w[ 9] = wordl2[1] | wordr2[1];
371 w[10] = wordl2[2] | wordr2[2];
372 w[11] = wordl2[3] | wordr2[3];
373 w[12] = wordl3[0] | wordr3[0];
374 w[13] = wordl3[1] | wordr3[1];
375 w[14] = wordl3[2] | wordr3[2];
376 w[15] = wordl3[3] | wordr3[3];
391 pad (&w[ 0], pw_len & 0xf);
393 else if (pw_len < 32)
395 pad (&w[ 4], pw_len & 0xf);
397 else if (pw_len < 48)
399 pad (&w[ 8], pw_len & 0xf);
401 else if (pw_len < 64)
403 pad (&w[12], pw_len & 0xf);
406 domino_big_md (w, pw_len, state, s_lotus_magic_table);
408 const u32x r0 = state[0];
409 const u32x r1 = state[1];
410 const u32x r2 = state[2];
411 const u32x r3 = state[3];
413 #include VECT_COMPARE_M
417 extern "C" __global__ void __launch_bounds__ (256, 1) m08600_m08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
421 extern "C" __global__ void __launch_bounds__ (256, 1) m08600_m16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
425 extern "C" __global__ void __launch_bounds__ (256, 1) m08600_s04 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, 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)
431 const u32 lid = threadIdx.x;
437 const u32 gid = (blockIdx.x * blockDim.x) + threadIdx.x;
443 __shared__ u32 s_lotus_magic_table[256];
445 s_lotus_magic_table[lid] = lotus_magic_table[lid];
449 if (gid >= gid_max) return;
453 wordl0[0] = pws[gid].i[ 0];
454 wordl0[1] = pws[gid].i[ 1];
455 wordl0[2] = pws[gid].i[ 2];
456 wordl0[3] = pws[gid].i[ 3];
460 wordl1[0] = pws[gid].i[ 4];
461 wordl1[1] = pws[gid].i[ 5];
462 wordl1[2] = pws[gid].i[ 6];
463 wordl1[3] = pws[gid].i[ 7];
479 const u32 pw_l_len = pws[gid].pw_len;
481 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
483 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, c_combs[0].pw_len);
490 const u32 search[4] =
492 digests_buf[digests_offset].digest_buf[DGST_R0],
493 digests_buf[digests_offset].digest_buf[DGST_R1],
494 digests_buf[digests_offset].digest_buf[DGST_R2],
495 digests_buf[digests_offset].digest_buf[DGST_R3]
502 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
504 const u32 pw_r_len = c_combs[il_pos].pw_len;
506 u32 pw_len = pw_l_len + pw_r_len;
510 wordr0[0] = c_combs[il_pos].i[0];
511 wordr0[1] = c_combs[il_pos].i[1];
512 wordr0[2] = c_combs[il_pos].i[2];
513 wordr0[3] = c_combs[il_pos].i[3];
517 wordr1[0] = c_combs[il_pos].i[4];
518 wordr1[1] = c_combs[il_pos].i[5];
519 wordr1[2] = c_combs[il_pos].i[6];
520 wordr1[3] = c_combs[il_pos].i[7];
536 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
538 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
543 w[ 0] = wordl0[0] | wordr0[0];
544 w[ 1] = wordl0[1] | wordr0[1];
545 w[ 2] = wordl0[2] | wordr0[2];
546 w[ 3] = wordl0[3] | wordr0[3];
547 w[ 4] = wordl1[0] | wordr1[0];
548 w[ 5] = wordl1[1] | wordr1[1];
549 w[ 6] = wordl1[2] | wordr1[2];
550 w[ 7] = wordl1[3] | wordr1[3];
551 w[ 8] = wordl2[0] | wordr2[0];
552 w[ 9] = wordl2[1] | wordr2[1];
553 w[10] = wordl2[2] | wordr2[2];
554 w[11] = wordl2[3] | wordr2[3];
555 w[12] = wordl3[0] | wordr3[0];
556 w[13] = wordl3[1] | wordr3[1];
557 w[14] = wordl3[2] | wordr3[2];
558 w[15] = wordl3[3] | wordr3[3];
573 pad (&w[ 0], pw_len & 0xf);
575 else if (pw_len < 32)
577 pad (&w[ 4], pw_len & 0xf);
579 else if (pw_len < 48)
581 pad (&w[ 8], pw_len & 0xf);
583 else if (pw_len < 64)
585 pad (&w[12], pw_len & 0xf);
588 domino_big_md (w, pw_len, state, s_lotus_magic_table);
590 const u32x r0 = state[0];
591 const u32x r1 = state[1];
592 const u32x r2 = state[2];
593 const u32x r3 = state[3];
595 #include VECT_COMPARE_S
599 extern "C" __global__ void __launch_bounds__ (256, 1) m08600_s08 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
603 extern "C" __global__ void __launch_bounds__ (256, 1) m08600_s16 (const pw_t *pws, const gpu_rule_t *rules_buf, const comb_t *combs_buf, const bf_t *bfs_buf, const void *tmps, void *hooks, const u32 *bitmaps_buf_s1_a, const u32 *bitmaps_buf_s1_b, const u32 *bitmaps_buf_s1_c, const u32 *bitmaps_buf_s1_d, const u32 *bitmaps_buf_s2_a, const u32 *bitmaps_buf_s2_b, const u32 *bitmaps_buf_s2_c, const u32 *bitmaps_buf_s2_d, plain_t *plains_buf, const digest_t *digests_buf, u32 *hashes_shown, const salt_t *salt_bufs, const void *esalt_bufs, u32 *d_return_buf, u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)