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"
48 #define SIPROUND(v0,v1,v2,v3) \
50 (v1) = rotl64 ((v1), 13); \
52 (v0) = as_ulong (as_uint2 ((v0)).s10); \
54 (v3) = rotl64 ((v3), 16); \
57 (v3) = rotl64 ((v3), 21); \
60 (v1) = rotl64 ((v1), 17); \
62 (v2) = as_ulong (as_uint2 ((v2)).s10);
64 #define SIPROUND(v0,v1,v2,v3) \
66 (v1) = rotl64 ((v1), 13); \
68 (v0) = rotl64 ((v0), 32); \
70 (v3) = rotl64 ((v3), 16); \
73 (v3) = rotl64 ((v3), 21); \
76 (v1) = rotl64 ((v1), 17); \
78 (v2) = rotl64 ((v2), 32);
81 __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)
87 const u32 lid = get_local_id (0);
93 const u32 gid = get_global_id (0);
95 if (gid >= gid_max) return;
99 wordl0[0] = pws[gid].i[ 0];
100 wordl0[1] = pws[gid].i[ 1];
101 wordl0[2] = pws[gid].i[ 2];
102 wordl0[3] = pws[gid].i[ 3];
106 wordl1[0] = pws[gid].i[ 4];
107 wordl1[1] = pws[gid].i[ 5];
108 wordl1[2] = pws[gid].i[ 6];
109 wordl1[3] = pws[gid].i[ 7];
125 const u32 pw_l_len = pws[gid].pw_len;
127 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
129 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
136 u64 v0p = SIPHASHM_0;
137 u64 v1p = SIPHASHM_1;
138 u64 v2p = SIPHASHM_2;
139 u64 v3p = SIPHASHM_3;
141 v0p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
142 v1p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
143 v2p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
144 v3p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
150 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
152 const u32 pw_r_len = combs_buf[il_pos].pw_len;
154 const u32 pw_len = pw_l_len + pw_r_len;
158 wordr0[0] = combs_buf[il_pos].i[0];
159 wordr0[1] = combs_buf[il_pos].i[1];
160 wordr0[2] = combs_buf[il_pos].i[2];
161 wordr0[3] = combs_buf[il_pos].i[3];
165 wordr1[0] = combs_buf[il_pos].i[4];
166 wordr1[1] = combs_buf[il_pos].i[5];
167 wordr1[2] = combs_buf[il_pos].i[6];
168 wordr1[3] = combs_buf[il_pos].i[7];
184 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
186 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
191 w[ 0] = wordl0[0] | wordr0[0];
192 w[ 1] = wordl0[1] | wordr0[1];
193 w[ 2] = wordl0[2] | wordr0[2];
194 w[ 3] = wordl0[3] | wordr0[3];
195 w[ 4] = wordl1[0] | wordr1[0];
196 w[ 5] = wordl1[1] | wordr1[1];
197 w[ 6] = wordl1[2] | wordr1[2];
198 w[ 7] = wordl1[3] | wordr1[3];
199 w[ 8] = wordl2[0] | wordr2[0];
200 w[ 9] = wordl2[1] | wordr2[1];
201 w[10] = wordl2[2] | wordr2[2];
202 w[11] = wordl2[3] | wordr2[3];
203 w[12] = wordl3[0] | wordr3[0];
204 w[13] = wordl3[1] | wordr3[1];
205 w[14] = wordl3[2] | wordr3[2];
206 w[15] = wordl3[3] | wordr3[3];
208 u64 *w_ptr = (u64 *) w;
210 w_ptr[pw_len / 8] |= (u64) pw_len << 56;
220 for (i = 0, j = 0; i <= pw_len; i += 8, j += 2)
222 u64x m = hl32_to_64 (w[j + 1], w[j + 0]);
226 SIPROUND (v0, v1, v2, v3);
227 SIPROUND (v0, v1, v2, v3);
234 SIPROUND (v0, v1, v2, v3);
235 SIPROUND (v0, v1, v2, v3);
236 SIPROUND (v0, v1, v2, v3);
237 SIPROUND (v0, v1, v2, v3);
239 const u64x v = v0 ^ v1 ^ v2 ^ v3;
241 const u32x a = l32_from_64 (v);
242 const u32x b = h32_from_64 (v);
249 #include VECT_COMPARE_M
253 __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)
257 __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)
261 __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)
267 const u32 lid = get_local_id (0);
273 const u32 gid = get_global_id (0);
275 if (gid >= gid_max) return;
279 wordl0[0] = pws[gid].i[ 0];
280 wordl0[1] = pws[gid].i[ 1];
281 wordl0[2] = pws[gid].i[ 2];
282 wordl0[3] = pws[gid].i[ 3];
286 wordl1[0] = pws[gid].i[ 4];
287 wordl1[1] = pws[gid].i[ 5];
288 wordl1[2] = pws[gid].i[ 6];
289 wordl1[3] = pws[gid].i[ 7];
305 const u32 pw_l_len = pws[gid].pw_len;
307 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
309 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
316 const u32 search[4] =
318 digests_buf[digests_offset].digest_buf[DGST_R0],
319 digests_buf[digests_offset].digest_buf[DGST_R1],
320 digests_buf[digests_offset].digest_buf[DGST_R2],
321 digests_buf[digests_offset].digest_buf[DGST_R3]
328 u64 v0p = SIPHASHM_0;
329 u64 v1p = SIPHASHM_1;
330 u64 v2p = SIPHASHM_2;
331 u64 v3p = SIPHASHM_3;
333 v0p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
334 v1p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
335 v2p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[1], salt_bufs[salt_pos].salt_buf[0]);
336 v3p ^= hl32_to_64 (salt_bufs[salt_pos].salt_buf[3], salt_bufs[salt_pos].salt_buf[2]);
342 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
344 const u32 pw_r_len = combs_buf[il_pos].pw_len;
346 const u32 pw_len = pw_l_len + pw_r_len;
350 wordr0[0] = combs_buf[il_pos].i[0];
351 wordr0[1] = combs_buf[il_pos].i[1];
352 wordr0[2] = combs_buf[il_pos].i[2];
353 wordr0[3] = combs_buf[il_pos].i[3];
357 wordr1[0] = combs_buf[il_pos].i[4];
358 wordr1[1] = combs_buf[il_pos].i[5];
359 wordr1[2] = combs_buf[il_pos].i[6];
360 wordr1[3] = combs_buf[il_pos].i[7];
376 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
378 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
383 w[ 0] = wordl0[0] | wordr0[0];
384 w[ 1] = wordl0[1] | wordr0[1];
385 w[ 2] = wordl0[2] | wordr0[2];
386 w[ 3] = wordl0[3] | wordr0[3];
387 w[ 4] = wordl1[0] | wordr1[0];
388 w[ 5] = wordl1[1] | wordr1[1];
389 w[ 6] = wordl1[2] | wordr1[2];
390 w[ 7] = wordl1[3] | wordr1[3];
391 w[ 8] = wordl2[0] | wordr2[0];
392 w[ 9] = wordl2[1] | wordr2[1];
393 w[10] = wordl2[2] | wordr2[2];
394 w[11] = wordl2[3] | wordr2[3];
395 w[12] = wordl3[0] | wordr3[0];
396 w[13] = wordl3[1] | wordr3[1];
397 w[14] = wordl3[2] | wordr3[2];
398 w[15] = wordl3[3] | wordr3[3];
400 u64 *w_ptr = (u64 *) w;
402 w_ptr[pw_len / 8] |= (u64) pw_len << 56;
412 for (i = 0, j = 0; i <= pw_len; i += 8, j += 2)
414 u64x m = hl32_to_64 (w[j + 1], w[j + 0]);
418 SIPROUND (v0, v1, v2, v3);
419 SIPROUND (v0, v1, v2, v3);
426 SIPROUND (v0, v1, v2, v3);
427 SIPROUND (v0, v1, v2, v3);
428 SIPROUND (v0, v1, v2, v3);
429 SIPROUND (v0, v1, v2, v3);
431 const u64x v = v0 ^ v1 ^ v2 ^ v3;
433 const u32x a = l32_from_64 (v);
434 const u32x b = h32_from_64 (v);
441 #include VECT_COMPARE_S
445 __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)
449 __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)