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"
23 __constant u64 keccakf_rndc[24] =
25 0x0000000000000001, 0x0000000000008082, 0x800000000000808a,
26 0x8000000080008000, 0x000000000000808b, 0x0000000080000001,
27 0x8000000080008081, 0x8000000000008009, 0x000000000000008a,
28 0x0000000000000088, 0x0000000080008009, 0x000000008000000a,
29 0x000000008000808b, 0x800000000000008b, 0x8000000000008089,
30 0x8000000000008003, 0x8000000000008002, 0x8000000000000080,
31 0x000000000000800a, 0x800000008000000a, 0x8000000080008081,
32 0x8000000000008080, 0x0000000080000001, 0x8000000080008008
35 __constant u32 keccakf_rotc[24] =
37 1, 3, 6, 10, 15, 21, 28, 36, 45, 55, 2, 14,
38 27, 41, 56, 8, 25, 43, 62, 18, 39, 61, 20, 44
41 __constant u32 keccakf_piln[24] =
43 10, 7, 11, 17, 18, 3, 5, 16, 8, 21, 24, 4,
44 15, 23, 19, 13, 12, 2, 20, 14, 22, 9, 6, 1
48 #define KECCAK_ROUNDS 24
51 #define Theta1(s) (st[0 + s] ^ st[5 + s] ^ st[10 + s] ^ st[15 + s] ^ st[20 + s])
64 u32 j = keccakf_piln[s]; \
65 u32 k = keccakf_rotc[s]; \
67 st[j] = rotl64 (t, k); \
78 st[0 + s] ^= ~bc1 & bc2; \
79 st[1 + s] ^= ~bc2 & bc3; \
80 st[2 + s] ^= ~bc3 & bc4; \
81 st[3 + s] ^= ~bc4 & bc0; \
82 st[4 + s] ^= ~bc0 & bc1; \
85 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05000_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)
91 const u32 lid = get_local_id (0);
97 const u32 gid = get_global_id (0);
99 if (gid >= gid_max) return;
103 wordl0[0] = pws[gid].i[ 0];
104 wordl0[1] = pws[gid].i[ 1];
105 wordl0[2] = pws[gid].i[ 2];
106 wordl0[3] = pws[gid].i[ 3];
110 wordl1[0] = pws[gid].i[ 4];
111 wordl1[1] = pws[gid].i[ 5];
112 wordl1[2] = pws[gid].i[ 6];
113 wordl1[3] = pws[gid].i[ 7];
129 const u32 pw_l_len = pws[gid].pw_len;
131 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
133 append_0x01_2 (wordl0, wordl1, pw_l_len);
135 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
139 * 0x80 keccak, very special
142 const u32 mdlen = salt_bufs[salt_pos].keccak_mdlen;
144 const u32 rsiz = 200 - (2 * mdlen);
146 const u32 add80w = (rsiz - 1) / 8;
152 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
154 const u32 pw_r_len = combs_buf[il_pos].pw_len;
156 const u32 pw_len = pw_l_len + pw_r_len;
160 wordr0[0] = combs_buf[il_pos].i[0];
161 wordr0[1] = combs_buf[il_pos].i[1];
162 wordr0[2] = combs_buf[il_pos].i[2];
163 wordr0[3] = combs_buf[il_pos].i[3];
167 wordr1[0] = combs_buf[il_pos].i[4];
168 wordr1[1] = combs_buf[il_pos].i[5];
169 wordr1[2] = combs_buf[il_pos].i[6];
170 wordr1[3] = combs_buf[il_pos].i[7];
186 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
188 append_0x01_2 (wordr0, wordr1, pw_r_len);
190 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
195 w0[0] = wordl0[0] | wordr0[0];
196 w0[1] = wordl0[1] | wordr0[1];
197 w0[2] = wordl0[2] | wordr0[2];
198 w0[3] = wordl0[3] | wordr0[3];
202 w1[0] = wordl1[0] | wordr1[0];
203 w1[1] = wordl1[1] | wordr1[1];
204 w1[2] = wordl1[2] | wordr1[2];
205 w1[3] = wordl1[3] | wordr1[3];
209 w2[0] = wordl2[0] | wordr2[0];
210 w2[1] = wordl2[1] | wordr2[1];
211 w2[2] = wordl2[2] | wordr2[2];
212 w2[3] = wordl2[3] | wordr2[3];
216 w3[0] = wordl3[0] | wordr3[0];
217 w3[1] = wordl3[1] | wordr3[1];
224 st[ 0] = (u64x) (w0[0]) | (u64x) (w0[1]) << 32;
225 st[ 1] = (u64x) (w0[2]) | (u64x) (w0[3]) << 32;
226 st[ 2] = (u64x) (w1[0]) | (u64x) (w1[1]) << 32;
227 st[ 3] = (u64x) (w1[2]) | (u64x) (w1[3]) << 32;
231 st[ 0] = (u64x) (w0[0].s0, w0[0].s1) | (u64x) (w0[1].s0, w0[1].s1) << 32;
232 st[ 1] = (u64x) (w0[2].s0, w0[2].s1) | (u64x) (w0[3].s0, w0[3].s1) << 32;
233 st[ 2] = (u64x) (w1[0].s0, w1[0].s1) | (u64x) (w1[1].s0, w1[1].s1) << 32;
234 st[ 3] = (u64x) (w1[2].s0, w1[2].s1) | (u64x) (w1[3].s0, w1[3].s1) << 32;
259 st[add80w] |= 0x8000000000000000;
263 for (round = 0; round < KECCAK_ROUNDS; round++)
267 u64 bc0 = Theta1 (0);
268 u64 bc1 = Theta1 (1);
269 u64 bc2 = Theta1 (2);
270 u64 bc3 = Theta1 (3);
271 u64 bc4 = Theta1 (4);
275 t = bc4 ^ rotl64 (bc1, 1); Theta2 (0);
276 t = bc0 ^ rotl64 (bc2, 1); Theta2 (1);
277 t = bc1 ^ rotl64 (bc3, 1); Theta2 (2);
278 t = bc2 ^ rotl64 (bc4, 1); Theta2 (3);
279 t = bc3 ^ rotl64 (bc0, 1); Theta2 (4);
320 st[0] ^= keccakf_rndc[round];
323 const u32 r0 = l32_from_64 (st[1]);
324 const u32 r1 = h32_from_64 (st[1]);
325 const u32 r2 = l32_from_64 (st[2]);
326 const u32 r3 = h32_from_64 (st[2]);
332 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05000_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)
336 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05000_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)
340 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05000_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)
346 const u32 lid = get_local_id (0);
352 const u32 gid = get_global_id (0);
354 if (gid >= gid_max) return;
358 wordl0[0] = pws[gid].i[ 0];
359 wordl0[1] = pws[gid].i[ 1];
360 wordl0[2] = pws[gid].i[ 2];
361 wordl0[3] = pws[gid].i[ 3];
365 wordl1[0] = pws[gid].i[ 4];
366 wordl1[1] = pws[gid].i[ 5];
367 wordl1[2] = pws[gid].i[ 6];
368 wordl1[3] = pws[gid].i[ 7];
384 const u32 pw_l_len = pws[gid].pw_len;
386 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
388 append_0x01_2 (wordl0, wordl1, pw_l_len);
390 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
397 const u32 search[4] =
399 digests_buf[digests_offset].digest_buf[DGST_R0],
400 digests_buf[digests_offset].digest_buf[DGST_R1],
401 digests_buf[digests_offset].digest_buf[DGST_R2],
402 digests_buf[digests_offset].digest_buf[DGST_R3]
406 * 0x80 keccak, very special
409 const u32 mdlen = salt_bufs[salt_pos].keccak_mdlen;
411 const u32 rsiz = 200 - (2 * mdlen);
413 const u32 add80w = (rsiz - 1) / 8;
419 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
421 const u32 pw_r_len = combs_buf[il_pos].pw_len;
423 const u32 pw_len = pw_l_len + pw_r_len;
427 wordr0[0] = combs_buf[il_pos].i[0];
428 wordr0[1] = combs_buf[il_pos].i[1];
429 wordr0[2] = combs_buf[il_pos].i[2];
430 wordr0[3] = combs_buf[il_pos].i[3];
434 wordr1[0] = combs_buf[il_pos].i[4];
435 wordr1[1] = combs_buf[il_pos].i[5];
436 wordr1[2] = combs_buf[il_pos].i[6];
437 wordr1[3] = combs_buf[il_pos].i[7];
453 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
455 append_0x01_2 (wordr0, wordr1, pw_r_len);
457 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
462 w0[0] = wordl0[0] | wordr0[0];
463 w0[1] = wordl0[1] | wordr0[1];
464 w0[2] = wordl0[2] | wordr0[2];
465 w0[3] = wordl0[3] | wordr0[3];
469 w1[0] = wordl1[0] | wordr1[0];
470 w1[1] = wordl1[1] | wordr1[1];
471 w1[2] = wordl1[2] | wordr1[2];
472 w1[3] = wordl1[3] | wordr1[3];
476 w2[0] = wordl2[0] | wordr2[0];
477 w2[1] = wordl2[1] | wordr2[1];
478 w2[2] = wordl2[2] | wordr2[2];
479 w2[3] = wordl2[3] | wordr2[3];
483 w3[0] = wordl3[0] | wordr3[0];
484 w3[1] = wordl3[1] | wordr3[1];
491 st[ 0] = (u64x) (w0[0]) | (u64x) (w0[1]) << 32;
492 st[ 1] = (u64x) (w0[2]) | (u64x) (w0[3]) << 32;
493 st[ 2] = (u64x) (w1[0]) | (u64x) (w1[1]) << 32;
494 st[ 3] = (u64x) (w1[2]) | (u64x) (w1[3]) << 32;
498 st[ 0] = (u64x) (w0[0].s0, w0[0].s1) | (u64x) (w0[1].s0, w0[1].s1) << 32;
499 st[ 1] = (u64x) (w0[2].s0, w0[2].s1) | (u64x) (w0[3].s0, w0[3].s1) << 32;
500 st[ 2] = (u64x) (w1[0].s0, w1[0].s1) | (u64x) (w1[1].s0, w1[1].s1) << 32;
501 st[ 3] = (u64x) (w1[2].s0, w1[2].s1) | (u64x) (w1[3].s0, w1[3].s1) << 32;
526 st[add80w] |= 0x8000000000000000;
530 for (round = 0; round < KECCAK_ROUNDS; round++)
534 u64 bc0 = Theta1 (0);
535 u64 bc1 = Theta1 (1);
536 u64 bc2 = Theta1 (2);
537 u64 bc3 = Theta1 (3);
538 u64 bc4 = Theta1 (4);
542 t = bc4 ^ rotl64 (bc1, 1); Theta2 (0);
543 t = bc0 ^ rotl64 (bc2, 1); Theta2 (1);
544 t = bc1 ^ rotl64 (bc3, 1); Theta2 (2);
545 t = bc2 ^ rotl64 (bc4, 1); Theta2 (3);
546 t = bc3 ^ rotl64 (bc0, 1); Theta2 (4);
587 st[0] ^= keccakf_rndc[round];
590 const u32 r0 = l32_from_64 (st[1]);
591 const u32 r1 = h32_from_64 (st[1]);
592 const u32 r2 = l32_from_64 (st[2]);
593 const u32 r3 = h32_from_64 (st[2]);
599 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05000_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)
603 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05000_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)