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 "types_ocl.c"
20 #define COMPARE_S "check_single_comp4.c"
21 #define COMPARE_M "check_multi_comp4.c"
24 #define SIPROUND(v0,v1,v2,v3) \
26 (v1) = rotl64 ((v1), 13); \
28 (v0) = as_ulong (as_uint2 ((v0)).s10); \
30 (v3) = rotl64 ((v3), 16); \
33 (v3) = rotl64 ((v3), 21); \
36 (v1) = rotl64 ((v1), 17); \
38 (v2) = as_ulong (as_uint2 ((v2)).s10);
40 #define SIPROUND(v0,v1,v2,v3) \
42 (v1) = rotl64 ((v1), 13); \
44 (v0) = rotl64 ((v0), 32); \
46 (v3) = rotl64 ((v3), 16); \
49 (v3) = rotl64 ((v3), 21); \
52 (v1) = rotl64 ((v1), 17); \
54 (v2) = rotl64 ((v2), 32);
57 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10100_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)
63 const u32 lid = get_local_id (0);
69 const u32 gid = get_global_id (0);
71 if (gid >= gid_max) return;
75 wordl0[0] = pws[gid].i[ 0];
76 wordl0[1] = pws[gid].i[ 1];
77 wordl0[2] = pws[gid].i[ 2];
78 wordl0[3] = pws[gid].i[ 3];
82 wordl1[0] = pws[gid].i[ 4];
83 wordl1[1] = pws[gid].i[ 5];
84 wordl1[2] = pws[gid].i[ 6];
85 wordl1[3] = pws[gid].i[ 7];
101 const u32 pw_l_len = pws[gid].pw_len;
103 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
105 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
112 u64 v0p = SIPHASHM_0;
113 u64 v1p = SIPHASHM_1;
114 u64 v2p = SIPHASHM_2;
115 u64 v3p = SIPHASHM_3;
117 v0p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
118 v1p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
119 v2p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
120 v3p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
126 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
128 const u32 pw_r_len = combs_buf[il_pos].pw_len;
130 const u32 pw_len = pw_l_len + pw_r_len;
134 wordr0[0] = combs_buf[il_pos].i[0];
135 wordr0[1] = combs_buf[il_pos].i[1];
136 wordr0[2] = combs_buf[il_pos].i[2];
137 wordr0[3] = combs_buf[il_pos].i[3];
141 wordr1[0] = combs_buf[il_pos].i[4];
142 wordr1[1] = combs_buf[il_pos].i[5];
143 wordr1[2] = combs_buf[il_pos].i[6];
144 wordr1[3] = combs_buf[il_pos].i[7];
160 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
162 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
167 w[ 0] = wordl0[0] | wordr0[0];
168 w[ 1] = wordl0[1] | wordr0[1];
169 w[ 2] = wordl0[2] | wordr0[2];
170 w[ 3] = wordl0[3] | wordr0[3];
171 w[ 4] = wordl1[0] | wordr1[0];
172 w[ 5] = wordl1[1] | wordr1[1];
173 w[ 6] = wordl1[2] | wordr1[2];
174 w[ 7] = wordl1[3] | wordr1[3];
175 w[ 8] = wordl2[0] | wordr2[0];
176 w[ 9] = wordl2[1] | wordr2[1];
177 w[10] = wordl2[2] | wordr2[2];
178 w[11] = wordl2[3] | wordr2[3];
179 w[12] = wordl3[0] | wordr3[0];
180 w[13] = wordl3[1] | wordr3[1];
181 w[14] = wordl3[2] | wordr3[2];
182 w[15] = wordl3[3] | wordr3[3];
184 u64 *w_ptr = (u64 *) w;
186 w_ptr[pw_len / 8] |= (u64) pw_len << 56;
196 for (i = 0, j = 0; i <= pw_len; i += 8, j += 2)
198 u64 m = hl32_to_64 (w[j + 1], w[j + 0]);
202 SIPROUND (v0, v1, v2, v3);
203 SIPROUND (v0, v1, v2, v3);
210 SIPROUND (v0, v1, v2, v3);
211 SIPROUND (v0, v1, v2, v3);
212 SIPROUND (v0, v1, v2, v3);
213 SIPROUND (v0, v1, v2, v3);
215 const u64 v = v0 ^ v1 ^ v2 ^ v3;
217 const u32 a = l32_from_64 (v);
218 const u32 b = h32_from_64 (v);
229 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10100_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)
233 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10100_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)
237 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10100_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)
243 const u32 lid = get_local_id (0);
249 const u32 gid = get_global_id (0);
251 if (gid >= gid_max) return;
255 wordl0[0] = pws[gid].i[ 0];
256 wordl0[1] = pws[gid].i[ 1];
257 wordl0[2] = pws[gid].i[ 2];
258 wordl0[3] = pws[gid].i[ 3];
262 wordl1[0] = pws[gid].i[ 4];
263 wordl1[1] = pws[gid].i[ 5];
264 wordl1[2] = pws[gid].i[ 6];
265 wordl1[3] = pws[gid].i[ 7];
281 const u32 pw_l_len = pws[gid].pw_len;
283 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
285 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
292 const u32 search[4] =
294 digests_buf[digests_offset].digest_buf[DGST_R0],
295 digests_buf[digests_offset].digest_buf[DGST_R1],
296 digests_buf[digests_offset].digest_buf[DGST_R2],
297 digests_buf[digests_offset].digest_buf[DGST_R3]
304 u64 v0p = SIPHASHM_0;
305 u64 v1p = SIPHASHM_1;
306 u64 v2p = SIPHASHM_2;
307 u64 v3p = SIPHASHM_3;
309 v0p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
310 v1p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
311 v2p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
312 v3p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
318 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
320 const u32 pw_r_len = combs_buf[il_pos].pw_len;
322 const u32 pw_len = pw_l_len + pw_r_len;
326 wordr0[0] = combs_buf[il_pos].i[0];
327 wordr0[1] = combs_buf[il_pos].i[1];
328 wordr0[2] = combs_buf[il_pos].i[2];
329 wordr0[3] = combs_buf[il_pos].i[3];
333 wordr1[0] = combs_buf[il_pos].i[4];
334 wordr1[1] = combs_buf[il_pos].i[5];
335 wordr1[2] = combs_buf[il_pos].i[6];
336 wordr1[3] = combs_buf[il_pos].i[7];
352 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
354 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
359 w[ 0] = wordl0[0] | wordr0[0];
360 w[ 1] = wordl0[1] | wordr0[1];
361 w[ 2] = wordl0[2] | wordr0[2];
362 w[ 3] = wordl0[3] | wordr0[3];
363 w[ 4] = wordl1[0] | wordr1[0];
364 w[ 5] = wordl1[1] | wordr1[1];
365 w[ 6] = wordl1[2] | wordr1[2];
366 w[ 7] = wordl1[3] | wordr1[3];
367 w[ 8] = wordl2[0] | wordr2[0];
368 w[ 9] = wordl2[1] | wordr2[1];
369 w[10] = wordl2[2] | wordr2[2];
370 w[11] = wordl2[3] | wordr2[3];
371 w[12] = wordl3[0] | wordr3[0];
372 w[13] = wordl3[1] | wordr3[1];
373 w[14] = wordl3[2] | wordr3[2];
374 w[15] = wordl3[3] | wordr3[3];
376 u64 *w_ptr = (u64 *) w;
378 w_ptr[pw_len / 8] |= (u64) pw_len << 56;
388 for (i = 0, j = 0; i <= pw_len; i += 8, j += 2)
390 u64 m = hl32_to_64 (w[j + 1], w[j + 0]);
394 SIPROUND (v0, v1, v2, v3);
395 SIPROUND (v0, v1, v2, v3);
402 SIPROUND (v0, v1, v2, v3);
403 SIPROUND (v0, v1, v2, v3);
404 SIPROUND (v0, v1, v2, v3);
405 SIPROUND (v0, v1, v2, v3);
407 const u64 v = v0 ^ v1 ^ v2 ^ v3;
409 const u32 a = l32_from_64 (v);
410 const u32 b = h32_from_64 (v);
421 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10100_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)
425 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10100_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)