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_vect4_comp4.c"
39 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
42 __constant u32 padding[8] =
62 static void swap (__local RC4_KEY *rc4_key, const u8 i, const u8 j)
67 rc4_key->S[i] = rc4_key->S[j];
71 static void rc4_init_16 (__local RC4_KEY *rc4_key, const u32 data[4])
76 __local u32 *ptr = (__local u32 *) rc4_key->S;
79 for (u32 i = 0; i < 64; i++)
84 const u32 d0 = data[0] >> 0;
85 const u32 d1 = data[0] >> 8;
86 const u32 d2 = data[0] >> 16;
87 const u32 d3 = data[0] >> 24;
88 const u32 d4 = data[1] >> 0;
93 for (u32 i = 0; i < 255; i += 5)
95 j += rc4_key->S[i + 0] + d0; swap (rc4_key, i + 0, j);
96 j += rc4_key->S[i + 1] + d1; swap (rc4_key, i + 1, j);
97 j += rc4_key->S[i + 2] + d2; swap (rc4_key, i + 2, j);
98 j += rc4_key->S[i + 3] + d3; swap (rc4_key, i + 3, j);
99 j += rc4_key->S[i + 4] + d4; swap (rc4_key, i + 4, j);
102 j += rc4_key->S[255] + d0; swap (rc4_key, 255, j);
105 static u8 rc4_next_16 (__local RC4_KEY *rc4_key, u8 i, u8 j, __constant u32 in[4], u32 out[4])
107 for (u32 k = 0; k < 4; k++)
116 swap (rc4_key, i, j);
118 idx = rc4_key->S[i] + rc4_key->S[j];
120 xor4 |= rc4_key->S[idx] << 0;
125 swap (rc4_key, i, j);
127 idx = rc4_key->S[i] + rc4_key->S[j];
129 xor4 |= rc4_key->S[idx] << 8;
134 swap (rc4_key, i, j);
136 idx = rc4_key->S[i] + rc4_key->S[j];
138 xor4 |= rc4_key->S[idx] << 16;
143 swap (rc4_key, i, j);
145 idx = rc4_key->S[i] + rc4_key->S[j];
147 xor4 |= rc4_key->S[idx] << 24;
149 out[k] = in[k] ^ xor4;
155 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10410_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 pdf_t *pdf_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)
161 const u32 lid = get_local_id (0);
167 const u32 gid = get_global_id (0);
169 if (gid >= gid_max) return;
173 wordl0[0] = pws[gid].i[ 0];
174 wordl0[1] = pws[gid].i[ 1];
175 wordl0[2] = pws[gid].i[ 2];
176 wordl0[3] = pws[gid].i[ 3];
180 wordl1[0] = pws[gid].i[ 4];
181 wordl1[1] = pws[gid].i[ 5];
182 wordl1[2] = pws[gid].i[ 6];
183 wordl1[3] = pws[gid].i[ 7];
199 const u32 pw_l_len = pws[gid].pw_len;
201 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
203 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
210 __local RC4_KEY rc4_keys[64];
211 __local RC4_KEY *rc4_key = &rc4_keys[lid];
217 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
221 wordr0[0] = combs_buf[il_pos].i[0];
222 wordr0[1] = combs_buf[il_pos].i[1];
223 wordr0[2] = combs_buf[il_pos].i[2];
224 wordr0[3] = combs_buf[il_pos].i[3];
228 wordr1[0] = combs_buf[il_pos].i[4];
229 wordr1[1] = combs_buf[il_pos].i[5];
230 wordr1[2] = combs_buf[il_pos].i[6];
231 wordr1[3] = combs_buf[il_pos].i[7];
247 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
249 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
254 w0[0] = wordl0[0] | wordr0[0];
255 w0[1] = wordl0[1] | wordr0[1];
266 rc4_init_16 (rc4_key, key);
270 rc4_next_16 (rc4_key, 0, 0, padding, out);
272 const u32x r0 = out[0];
273 const u32x r1 = out[1];
274 const u32x r2 = out[2];
275 const u32x r3 = out[3];
277 #include VECT_COMPARE_M
281 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10410_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 pdf_t *pdf_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)
285 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10410_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 pdf_t *pdf_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)
289 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10410_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 pdf_t *pdf_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)
295 const u32 lid = get_local_id (0);
301 const u32 gid = get_global_id (0);
303 if (gid >= gid_max) return;
307 wordl0[0] = pws[gid].i[ 0];
308 wordl0[1] = pws[gid].i[ 1];
309 wordl0[2] = pws[gid].i[ 2];
310 wordl0[3] = pws[gid].i[ 3];
314 wordl1[0] = pws[gid].i[ 4];
315 wordl1[1] = pws[gid].i[ 5];
316 wordl1[2] = pws[gid].i[ 6];
317 wordl1[3] = pws[gid].i[ 7];
333 const u32 pw_l_len = pws[gid].pw_len;
335 if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
337 switch_buffer_by_offset (wordl0, wordl1, wordl2, wordl3, combs_buf[0].pw_len);
344 const u32 search[4] =
346 digests_buf[digests_offset].digest_buf[DGST_R0],
347 digests_buf[digests_offset].digest_buf[DGST_R1],
348 digests_buf[digests_offset].digest_buf[DGST_R2],
349 digests_buf[digests_offset].digest_buf[DGST_R3]
356 __local RC4_KEY rc4_keys[64];
357 __local RC4_KEY *rc4_key = &rc4_keys[lid];
363 for (u32 il_pos = 0; il_pos < combs_cnt; il_pos++)
367 wordr0[0] = combs_buf[il_pos].i[0];
368 wordr0[1] = combs_buf[il_pos].i[1];
369 wordr0[2] = combs_buf[il_pos].i[2];
370 wordr0[3] = combs_buf[il_pos].i[3];
374 wordr1[0] = combs_buf[il_pos].i[4];
375 wordr1[1] = combs_buf[il_pos].i[5];
376 wordr1[2] = combs_buf[il_pos].i[6];
377 wordr1[3] = combs_buf[il_pos].i[7];
393 if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
395 switch_buffer_by_offset (wordr0, wordr1, wordr2, wordr3, pw_l_len);
400 w0[0] = wordl0[0] | wordr0[0];
401 w0[1] = wordl0[1] | wordr0[1];
412 rc4_init_16 (rc4_key, key);
416 rc4_next_16 (rc4_key, 0, 0, padding, out);
418 const u32x r0 = out[0];
419 const u32x r1 = out[1];
420 const u32x r2 = out[2];
421 const u32x r3 = out[3];
423 #include VECT_COMPARE_S
427 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10410_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 pdf_t *pdf_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)
431 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m10410_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 pdf_t *pdf_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)