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 "OpenCL/types_ocl.c"
18 #include "OpenCL/common.c"
20 #define COMPARE_S "OpenCL/check_single_comp4.c"
21 #define COMPARE_M "OpenCL/check_multi_comp4.c"
23 #define INITVAL 0x0101010101010101
25 #define SBOG_LPSti64 \
26 s_sbob_sl64[0][(t[0] >> (i * 8)) & 0xff] ^ \
27 s_sbob_sl64[1][(t[1] >> (i * 8)) & 0xff] ^ \
28 s_sbob_sl64[2][(t[2] >> (i * 8)) & 0xff] ^ \
29 s_sbob_sl64[3][(t[3] >> (i * 8)) & 0xff] ^ \
30 s_sbob_sl64[4][(t[4] >> (i * 8)) & 0xff] ^ \
31 s_sbob_sl64[5][(t[5] >> (i * 8)) & 0xff] ^ \
32 s_sbob_sl64[6][(t[6] >> (i * 8)) & 0xff] ^ \
33 s_sbob_sl64[7][(t[7] >> (i * 8)) & 0xff]
37 __constant u64 sbob_sl64[8][256] =
2105 __constant u64 sbob_rc64[12][8] =
2229 static void streebog_g (u64 h[8], const u64 m[8], __local u64 s_sbob_sl64[8][256])
2236 for (int i = 0; i < 8; i++)
2241 for (int i = 0; i < 8; i++)
2243 k[i] = SBOG_LPSti64;
2247 for (int i = 0; i < 8; i++)
2252 for (int r = 0; r < 12; r++)
2255 for (int i = 0; i < 8; i++)
2261 for (int i = 0; i < 8; i++)
2263 s[i] = SBOG_LPSti64;
2266 for (int i = 0; i < 8; i++)
2268 t[i] = k[i] ^ sbob_rc64[r][i];
2272 for (int i = 0; i < 8; i++)
2274 k[i] = SBOG_LPSti64;
2279 for (int i = 0; i < 8; i++)
2281 h[i] ^= s[i] ^ k[i] ^ m[i];
2285 __kernel void m11700_m04 (__global pw_t *pws, __global kernel_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)
2291 const u32 gid = get_global_id (0);
2292 const u32 lid = get_local_id (0);
2293 const u32 lsz = get_local_size (0);
2296 * shared lookup table
2299 __local u64 s_sbob_sl64[8][256];
2301 for (u32 i = lid; i < 256; i += lsz)
2303 s_sbob_sl64[0][i] = sbob_sl64[0][i];
2304 s_sbob_sl64[1][i] = sbob_sl64[1][i];
2305 s_sbob_sl64[2][i] = sbob_sl64[2][i];
2306 s_sbob_sl64[3][i] = sbob_sl64[3][i];
2307 s_sbob_sl64[4][i] = sbob_sl64[4][i];
2308 s_sbob_sl64[5][i] = sbob_sl64[5][i];
2309 s_sbob_sl64[6][i] = sbob_sl64[6][i];
2310 s_sbob_sl64[7][i] = sbob_sl64[7][i];
2313 barrier (CLK_LOCAL_MEM_FENCE);
2315 if (gid >= gid_max) return;
2323 wordl0[0] = pws[gid].i[ 0];
2324 wordl0[1] = pws[gid].i[ 1];
2325 wordl0[2] = pws[gid].i[ 2];
2326 wordl0[3] = pws[gid].i[ 3];
2330 wordl1[0] = pws[gid].i[ 4];
2331 wordl1[1] = pws[gid].i[ 5];
2332 wordl1[2] = pws[gid].i[ 6];
2333 wordl1[3] = pws[gid].i[ 7];
2349 const u32 pw_l_len = pws[gid].pw_len;
2351 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
2353 append_0x80_2x4 (wordl0, wordl1, pw_l_len);
2355 switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
2362 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
2364 const u32 pw_r_len = combs_buf[il_pos].pw_len;
2366 const u32 pw_len = pw_l_len + pw_r_len;
2370 wordr0[0] = combs_buf[il_pos].i[0];
2371 wordr0[1] = combs_buf[il_pos].i[1];
2372 wordr0[2] = combs_buf[il_pos].i[2];
2373 wordr0[3] = combs_buf[il_pos].i[3];
2377 wordr1[0] = combs_buf[il_pos].i[4];
2378 wordr1[1] = combs_buf[il_pos].i[5];
2379 wordr1[2] = combs_buf[il_pos].i[6];
2380 wordr1[3] = combs_buf[il_pos].i[7];
2396 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
2398 switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
2403 w[ 0] = wordl0[0] | wordr0[0];
2404 w[ 1] = wordl0[1] | wordr0[1];
2405 w[ 2] = wordl0[2] | wordr0[2];
2406 w[ 3] = wordl0[3] | wordr0[3];
2407 w[ 4] = wordl1[0] | wordr1[0];
2408 w[ 5] = wordl1[1] | wordr1[1];
2409 w[ 6] = wordl1[2] | wordr1[2];
2410 w[ 7] = wordl1[3] | wordr1[3];
2411 w[ 8] = wordl2[0] | wordr2[0];
2412 w[ 9] = wordl2[1] | wordr2[1];
2413 w[10] = wordl2[2] | wordr2[2];
2414 w[11] = wordl2[3] | wordr2[3];
2415 w[12] = wordl3[0] | wordr3[0];
2416 w[13] = wordl3[1] | wordr3[1];
2417 w[14] = wordl3[1] | wordr3[1];
2418 w[15] = wordl3[1] | wordr3[1];
2420 append_0x01_4x4 (&w[0], &w[1], &w[2], &w[3], pw_len);
2423 * reverse message block
2428 m[0] = hl32_to_64 (w[15], w[14]);
2429 m[1] = hl32_to_64 (w[13], w[12]);
2430 m[2] = hl32_to_64 (w[11], w[10]);
2431 m[3] = hl32_to_64 (w[ 9], w[ 8]);
2432 m[4] = hl32_to_64 (w[ 7], w[ 6]);
2433 m[5] = hl32_to_64 (w[ 5], w[ 4]);
2434 m[6] = hl32_to_64 (w[ 3], w[ 2]);
2435 m[7] = hl32_to_64 (w[ 1], w[ 0]);
2437 m[0] = swap64 (m[0]);
2438 m[1] = swap64 (m[1]);
2439 m[2] = swap64 (m[2]);
2440 m[3] = swap64 (m[3]);
2441 m[4] = swap64 (m[4]);
2442 m[5] = swap64 (m[5]);
2443 m[6] = swap64 (m[6]);
2444 m[7] = swap64 (m[7]);
2446 // state buffer (hash)
2459 streebog_g (h, m, s_sbob_sl64);
2470 z[7] = swap64 ((u64) (pw_len * 8));
2472 streebog_g (h, z, s_sbob_sl64);
2473 streebog_g (h, m, s_sbob_sl64);
2475 const u32 r0 = l32_from_64 (h[0]);
2476 const u32 r1 = h32_from_64 (h[0]);
2477 const u32 r2 = l32_from_64 (h[1]);
2478 const u32 r3 = h32_from_64 (h[1]);
2484 __kernel void m11700_m08 (__global pw_t *pws, __global kernel_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)
2488 __kernel void m11700_m16 (__global pw_t *pws, __global kernel_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)
2492 __kernel void m11700_s04 (__global pw_t *pws, __global kernel_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)
2498 const u32 gid = get_global_id (0);
2499 const u32 lid = get_local_id (0);
2500 const u32 lsz = get_local_size (0);
2503 * shared lookup table
2506 __local u64 s_sbob_sl64[8][256];
2508 for (u32 i = lid; i < 256; i += lsz)
2510 s_sbob_sl64[0][i] = sbob_sl64[0][i];
2511 s_sbob_sl64[1][i] = sbob_sl64[1][i];
2512 s_sbob_sl64[2][i] = sbob_sl64[2][i];
2513 s_sbob_sl64[3][i] = sbob_sl64[3][i];
2514 s_sbob_sl64[4][i] = sbob_sl64[4][i];
2515 s_sbob_sl64[5][i] = sbob_sl64[5][i];
2516 s_sbob_sl64[6][i] = sbob_sl64[6][i];
2517 s_sbob_sl64[7][i] = sbob_sl64[7][i];
2520 barrier (CLK_LOCAL_MEM_FENCE);
2522 if (gid >= gid_max) return;
2528 const u32 search[4] =
2530 digests_buf[digests_offset].digest_buf[DGST_R0],
2531 digests_buf[digests_offset].digest_buf[DGST_R1],
2532 digests_buf[digests_offset].digest_buf[DGST_R2],
2533 digests_buf[digests_offset].digest_buf[DGST_R3]
2542 wordl0[0] = pws[gid].i[ 0];
2543 wordl0[1] = pws[gid].i[ 1];
2544 wordl0[2] = pws[gid].i[ 2];
2545 wordl0[3] = pws[gid].i[ 3];
2549 wordl1[0] = pws[gid].i[ 4];
2550 wordl1[1] = pws[gid].i[ 5];
2551 wordl1[2] = pws[gid].i[ 6];
2552 wordl1[3] = pws[gid].i[ 7];
2568 const u32 pw_l_len = pws[gid].pw_len;
2570 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
2572 append_0x80_2x4 (wordl0, wordl1, pw_l_len);
2574 switch_buffer_by_offset_le (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
2581 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
2583 const u32 pw_r_len = combs_buf[il_pos].pw_len;
2585 const u32 pw_len = pw_l_len + pw_r_len;
2589 wordr0[0] = combs_buf[il_pos].i[0];
2590 wordr0[1] = combs_buf[il_pos].i[1];
2591 wordr0[2] = combs_buf[il_pos].i[2];
2592 wordr0[3] = combs_buf[il_pos].i[3];
2596 wordr1[0] = combs_buf[il_pos].i[4];
2597 wordr1[1] = combs_buf[il_pos].i[5];
2598 wordr1[2] = combs_buf[il_pos].i[6];
2599 wordr1[3] = combs_buf[il_pos].i[7];
2615 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
2617 switch_buffer_by_offset_le (wordr0, wordr1, wordr2, wordr3, pw_l_len);
2622 w[ 0] = wordl0[0] | wordr0[0];
2623 w[ 1] = wordl0[1] | wordr0[1];
2624 w[ 2] = wordl0[2] | wordr0[2];
2625 w[ 3] = wordl0[3] | wordr0[3];
2626 w[ 4] = wordl1[0] | wordr1[0];
2627 w[ 5] = wordl1[1] | wordr1[1];
2628 w[ 6] = wordl1[2] | wordr1[2];
2629 w[ 7] = wordl1[3] | wordr1[3];
2630 w[ 8] = wordl2[0] | wordr2[0];
2631 w[ 9] = wordl2[1] | wordr2[1];
2632 w[10] = wordl2[2] | wordr2[2];
2633 w[11] = wordl2[3] | wordr2[3];
2634 w[12] = wordl3[0] | wordr3[0];
2635 w[13] = wordl3[1] | wordr3[1];
2636 w[14] = wordl3[1] | wordr3[1];
2637 w[15] = wordl3[1] | wordr3[1];
2639 append_0x01_4x4 (&w[0], &w[1], &w[2], &w[3], pw_len);
2642 * reverse message block
2647 m[0] = hl32_to_64 (w[15], w[14]);
2648 m[1] = hl32_to_64 (w[13], w[12]);
2649 m[2] = hl32_to_64 (w[11], w[10]);
2650 m[3] = hl32_to_64 (w[ 9], w[ 8]);
2651 m[4] = hl32_to_64 (w[ 7], w[ 6]);
2652 m[5] = hl32_to_64 (w[ 5], w[ 4]);
2653 m[6] = hl32_to_64 (w[ 3], w[ 2]);
2654 m[7] = hl32_to_64 (w[ 1], w[ 0]);
2656 m[0] = swap64 (m[0]);
2657 m[1] = swap64 (m[1]);
2658 m[2] = swap64 (m[2]);
2659 m[3] = swap64 (m[3]);
2660 m[4] = swap64 (m[4]);
2661 m[5] = swap64 (m[5]);
2662 m[6] = swap64 (m[6]);
2663 m[7] = swap64 (m[7]);
2665 // state buffer (hash)
2678 streebog_g (h, m, s_sbob_sl64);
2689 z[7] = swap64 ((u64) (pw_len * 8));
2691 streebog_g (h, z, s_sbob_sl64);
2692 streebog_g (h, m, s_sbob_sl64);
2694 const u32 r0 = l32_from_64 (h[0]);
2695 const u32 r1 = h32_from_64 (h[0]);
2696 const u32 r2 = l32_from_64 (h[1]);
2697 const u32 r3 = h32_from_64 (h[1]);
2703 __kernel void m11700_s08 (__global pw_t *pws, __global kernel_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)
2707 __kernel void m11700_s16 (__global pw_t *pws, __global kernel_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)