2 * Author......: Jens Steube <jens.steube@gmail.com>
10 #include "include/constants.h"
11 #include "include/kernel_vendor.h"
18 #include "include/kernel_functions.c"
19 #include "OpenCL/types_ocl.c"
20 #include "OpenCL/common.c"
21 #include "OpenCL/simd.c"
23 static void md5_transform (const u32x w0[4], const u32x w1[4], const u32x w2[4], const u32x w3[4], u32x digest[4])
47 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
48 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
49 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
50 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
51 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
52 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
53 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
54 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
55 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
56 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
57 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
58 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
59 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
60 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
61 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
62 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
64 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
65 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
66 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
67 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
68 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
69 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
70 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
71 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
72 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
73 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
74 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
75 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
76 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
77 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
78 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
79 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
81 MD5_STEP (MD5_H , a, b, c, d, w5_t, MD5C20, MD5S20);
82 MD5_STEP (MD5_H , d, a, b, c, w8_t, MD5C21, MD5S21);
83 MD5_STEP (MD5_H , c, d, a, b, wb_t, MD5C22, MD5S22);
84 MD5_STEP (MD5_H , b, c, d, a, we_t, MD5C23, MD5S23);
85 MD5_STEP (MD5_H , a, b, c, d, w1_t, MD5C24, MD5S20);
86 MD5_STEP (MD5_H , d, a, b, c, w4_t, MD5C25, MD5S21);
87 MD5_STEP (MD5_H , c, d, a, b, w7_t, MD5C26, MD5S22);
88 MD5_STEP (MD5_H , b, c, d, a, wa_t, MD5C27, MD5S23);
89 MD5_STEP (MD5_H , a, b, c, d, wd_t, MD5C28, MD5S20);
90 MD5_STEP (MD5_H , d, a, b, c, w0_t, MD5C29, MD5S21);
91 MD5_STEP (MD5_H , c, d, a, b, w3_t, MD5C2a, MD5S22);
92 MD5_STEP (MD5_H , b, c, d, a, w6_t, MD5C2b, MD5S23);
93 MD5_STEP (MD5_H , a, b, c, d, w9_t, MD5C2c, MD5S20);
94 MD5_STEP (MD5_H , d, a, b, c, wc_t, MD5C2d, MD5S21);
95 MD5_STEP (MD5_H , c, d, a, b, wf_t, MD5C2e, MD5S22);
96 MD5_STEP (MD5_H , b, c, d, a, w2_t, MD5C2f, MD5S23);
98 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
99 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
100 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
101 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
102 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
103 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
104 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
105 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
106 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
107 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
108 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
109 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
110 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
111 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
112 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
113 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
121 static void hmac_md5_pad (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4])
123 w0[0] = w0[0] ^ 0x36363636;
124 w0[1] = w0[1] ^ 0x36363636;
125 w0[2] = w0[2] ^ 0x36363636;
126 w0[3] = w0[3] ^ 0x36363636;
127 w1[0] = w1[0] ^ 0x36363636;
128 w1[1] = w1[1] ^ 0x36363636;
129 w1[2] = w1[2] ^ 0x36363636;
130 w1[3] = w1[3] ^ 0x36363636;
131 w2[0] = w2[0] ^ 0x36363636;
132 w2[1] = w2[1] ^ 0x36363636;
133 w2[2] = w2[2] ^ 0x36363636;
134 w2[3] = w2[3] ^ 0x36363636;
135 w3[0] = w3[0] ^ 0x36363636;
136 w3[1] = w3[1] ^ 0x36363636;
137 w3[2] = w3[2] ^ 0x36363636;
138 w3[3] = w3[3] ^ 0x36363636;
145 md5_transform (w0, w1, w2, w3, ipad);
147 w0[0] = w0[0] ^ 0x6a6a6a6a;
148 w0[1] = w0[1] ^ 0x6a6a6a6a;
149 w0[2] = w0[2] ^ 0x6a6a6a6a;
150 w0[3] = w0[3] ^ 0x6a6a6a6a;
151 w1[0] = w1[0] ^ 0x6a6a6a6a;
152 w1[1] = w1[1] ^ 0x6a6a6a6a;
153 w1[2] = w1[2] ^ 0x6a6a6a6a;
154 w1[3] = w1[3] ^ 0x6a6a6a6a;
155 w2[0] = w2[0] ^ 0x6a6a6a6a;
156 w2[1] = w2[1] ^ 0x6a6a6a6a;
157 w2[2] = w2[2] ^ 0x6a6a6a6a;
158 w2[3] = w2[3] ^ 0x6a6a6a6a;
159 w3[0] = w3[0] ^ 0x6a6a6a6a;
160 w3[1] = w3[1] ^ 0x6a6a6a6a;
161 w3[2] = w3[2] ^ 0x6a6a6a6a;
162 w3[3] = w3[3] ^ 0x6a6a6a6a;
169 md5_transform (w0, w1, w2, w3, opad);
172 static void hmac_md5_run (u32x w0[4], u32x w1[4], u32x w2[4], u32x w3[4], u32x ipad[4], u32x opad[4], u32x digest[4])
179 md5_transform (w0, w1, w2, w3, digest);
195 w3[2] = (64 + 16) * 8;
203 md5_transform (w0, w1, w2, w3, digest);
206 static void m05300m (__local u32 w_s[16], u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 ikepsk_t *ikepsk_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 s_msg_buf[128])
212 const u32 gid = get_global_id (0);
213 const u32 lid = get_local_id (0);
219 const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
220 const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
228 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
230 const u32x w0r = w0r_create_bft (bfs_buf, il_pos);
232 const u32x w0lr = w0l | w0r;
269 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
285 w3_t[2] = (64 + nr_len) * 8;
290 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
309 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
314 for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
316 w0_t[0] = s_msg_buf[off + 0];
317 w0_t[1] = s_msg_buf[off + 1];
318 w0_t[2] = s_msg_buf[off + 2];
319 w0_t[3] = s_msg_buf[off + 3];
320 w1_t[0] = s_msg_buf[off + 4];
321 w1_t[1] = s_msg_buf[off + 5];
322 w1_t[2] = s_msg_buf[off + 6];
323 w1_t[3] = s_msg_buf[off + 7];
324 w2_t[0] = s_msg_buf[off + 8];
325 w2_t[1] = s_msg_buf[off + 9];
326 w2_t[2] = s_msg_buf[off + 10];
327 w2_t[3] = s_msg_buf[off + 11];
328 w3_t[0] = s_msg_buf[off + 12];
329 w3_t[1] = s_msg_buf[off + 13];
330 w3_t[2] = s_msg_buf[off + 14];
331 w3_t[3] = s_msg_buf[off + 15];
333 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
336 w0_t[0] = s_msg_buf[off + 0];
337 w0_t[1] = s_msg_buf[off + 1];
338 w0_t[2] = s_msg_buf[off + 2];
339 w0_t[3] = s_msg_buf[off + 3];
340 w1_t[0] = s_msg_buf[off + 4];
341 w1_t[1] = s_msg_buf[off + 5];
342 w1_t[2] = s_msg_buf[off + 6];
343 w1_t[3] = s_msg_buf[off + 7];
344 w2_t[0] = s_msg_buf[off + 8];
345 w2_t[1] = s_msg_buf[off + 9];
346 w2_t[2] = s_msg_buf[off + 10];
347 w2_t[3] = s_msg_buf[off + 11];
348 w3_t[0] = s_msg_buf[off + 12];
349 w3_t[1] = s_msg_buf[off + 13];
350 w3_t[2] = (64 + msg_len) * 8;
353 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
355 COMPARE_M_SIMD (digest[0], digest[3], digest[2], digest[1]);
359 static void m05300s (__local u32 w_s[16], u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], const u32 pw_len, __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 ikepsk_t *ikepsk_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, __local u32 s_msg_buf[128])
365 const u32 gid = get_global_id (0);
366 const u32 lid = get_local_id (0);
372 const u32 nr_len = ikepsk_bufs[salt_pos].nr_len;
373 const u32 msg_len = ikepsk_bufs[salt_pos].msg_len;
379 const u32 search[4] =
381 digests_buf[digests_offset].digest_buf[DGST_R0],
382 digests_buf[digests_offset].digest_buf[DGST_R1],
383 digests_buf[digests_offset].digest_buf[DGST_R2],
384 digests_buf[digests_offset].digest_buf[DGST_R3]
393 for (u32 il_pos = 0; il_pos < bfs_cnt; il_pos += VECT_SIZE)
395 const u32x w0r = w0r_create_bft (bfs_buf, il_pos);
397 const u32x w0lr = w0l | w0r;
434 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
450 w3_t[2] = (64 + nr_len) * 8;
455 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
474 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
479 for (left = ikepsk_bufs[salt_pos].msg_len, off = 0; left >= 56; left -= 64, off += 16)
481 w0_t[0] = s_msg_buf[off + 0];
482 w0_t[1] = s_msg_buf[off + 1];
483 w0_t[2] = s_msg_buf[off + 2];
484 w0_t[3] = s_msg_buf[off + 3];
485 w1_t[0] = s_msg_buf[off + 4];
486 w1_t[1] = s_msg_buf[off + 5];
487 w1_t[2] = s_msg_buf[off + 6];
488 w1_t[3] = s_msg_buf[off + 7];
489 w2_t[0] = s_msg_buf[off + 8];
490 w2_t[1] = s_msg_buf[off + 9];
491 w2_t[2] = s_msg_buf[off + 10];
492 w2_t[3] = s_msg_buf[off + 11];
493 w3_t[0] = s_msg_buf[off + 12];
494 w3_t[1] = s_msg_buf[off + 13];
495 w3_t[2] = s_msg_buf[off + 14];
496 w3_t[3] = s_msg_buf[off + 15];
498 md5_transform (w0_t, w1_t, w2_t, w3_t, ipad);
501 w0_t[0] = s_msg_buf[off + 0];
502 w0_t[1] = s_msg_buf[off + 1];
503 w0_t[2] = s_msg_buf[off + 2];
504 w0_t[3] = s_msg_buf[off + 3];
505 w1_t[0] = s_msg_buf[off + 4];
506 w1_t[1] = s_msg_buf[off + 5];
507 w1_t[2] = s_msg_buf[off + 6];
508 w1_t[3] = s_msg_buf[off + 7];
509 w2_t[0] = s_msg_buf[off + 8];
510 w2_t[1] = s_msg_buf[off + 9];
511 w2_t[2] = s_msg_buf[off + 10];
512 w2_t[3] = s_msg_buf[off + 11];
513 w3_t[0] = s_msg_buf[off + 12];
514 w3_t[1] = s_msg_buf[off + 13];
515 w3_t[2] = (64 + msg_len) * 8;
518 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
520 COMPARE_S_SIMD (digest[0], digest[3], digest[2], digest[1]);
524 __kernel void m05300_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 ikepsk_t *ikepsk_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
530 const u32 gid = get_global_id (0);
531 const u32 lid = get_local_id (0);
532 const u32 lsz = get_local_size (0);
540 for (u32 i = lid; i < 16; i += lsz)
542 w_s[i] = ikepsk_bufs[salt_pos].nr_buf[i];
545 barrier (CLK_LOCAL_MEM_FENCE);
547 __local u32 s_msg_buf[128];
549 for (u32 i = lid; i < 128; i += lsz)
551 s_msg_buf[i] = ikepsk_bufs[salt_pos].msg_buf[i];
554 barrier (CLK_LOCAL_MEM_FENCE);
556 if (gid >= gid_max) return;
564 w0[0] = pws[gid].i[ 0];
565 w0[1] = pws[gid].i[ 1];
566 w0[2] = pws[gid].i[ 2];
567 w0[3] = pws[gid].i[ 3];
590 const u32 pw_len = pws[gid].pw_len;
596 m05300m (w_s, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, ikepsk_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_msg_buf);
599 __kernel void m05300_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 ikepsk_t *ikepsk_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
605 const u32 gid = get_global_id (0);
606 const u32 lid = get_local_id (0);
607 const u32 lsz = get_local_size (0);
615 for (u32 i = lid; i < 16; i += lsz)
617 w_s[i] = ikepsk_bufs[salt_pos].nr_buf[i];
620 barrier (CLK_LOCAL_MEM_FENCE);
622 __local u32 s_msg_buf[128];
624 for (u32 i = lid; i < 128; i += lsz)
626 s_msg_buf[i] = ikepsk_bufs[salt_pos].msg_buf[i];
629 barrier (CLK_LOCAL_MEM_FENCE);
631 if (gid >= gid_max) return;
639 w0[0] = pws[gid].i[ 0];
640 w0[1] = pws[gid].i[ 1];
641 w0[2] = pws[gid].i[ 2];
642 w0[3] = pws[gid].i[ 3];
646 w1[0] = pws[gid].i[ 4];
647 w1[1] = pws[gid].i[ 5];
648 w1[2] = pws[gid].i[ 6];
649 w1[3] = pws[gid].i[ 7];
665 const u32 pw_len = pws[gid].pw_len;
671 m05300m (w_s, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, ikepsk_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_msg_buf);
674 __kernel void m05300_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 ikepsk_t *ikepsk_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
680 const u32 gid = get_global_id (0);
681 const u32 lid = get_local_id (0);
682 const u32 lsz = get_local_size (0);
690 for (u32 i = lid; i < 16; i += lsz)
692 w_s[i] = ikepsk_bufs[salt_pos].nr_buf[i];
695 barrier (CLK_LOCAL_MEM_FENCE);
697 __local u32 s_msg_buf[128];
699 for (u32 i = lid; i < 128; i += lsz)
701 s_msg_buf[i] = ikepsk_bufs[salt_pos].msg_buf[i];
704 barrier (CLK_LOCAL_MEM_FENCE);
706 if (gid >= gid_max) return;
714 w0[0] = pws[gid].i[ 0];
715 w0[1] = pws[gid].i[ 1];
716 w0[2] = pws[gid].i[ 2];
717 w0[3] = pws[gid].i[ 3];
721 w1[0] = pws[gid].i[ 4];
722 w1[1] = pws[gid].i[ 5];
723 w1[2] = pws[gid].i[ 6];
724 w1[3] = pws[gid].i[ 7];
728 w2[0] = pws[gid].i[ 8];
729 w2[1] = pws[gid].i[ 9];
730 w2[2] = pws[gid].i[10];
731 w2[3] = pws[gid].i[11];
735 w3[0] = pws[gid].i[12];
736 w3[1] = pws[gid].i[13];
740 const u32 pw_len = pws[gid].pw_len;
746 m05300m (w_s, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, ikepsk_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_msg_buf);
749 __kernel void m05300_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 ikepsk_t *ikepsk_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
755 const u32 gid = get_global_id (0);
756 const u32 lid = get_local_id (0);
757 const u32 lsz = get_local_size (0);
765 for (u32 i = lid; i < 16; i += lsz)
767 w_s[i] = ikepsk_bufs[salt_pos].nr_buf[i];
770 barrier (CLK_LOCAL_MEM_FENCE);
772 __local u32 s_msg_buf[128];
774 for (u32 i = lid; i < 128; i += lsz)
776 s_msg_buf[i] = ikepsk_bufs[salt_pos].msg_buf[i];
779 barrier (CLK_LOCAL_MEM_FENCE);
781 if (gid >= gid_max) return;
789 w0[0] = pws[gid].i[ 0];
790 w0[1] = pws[gid].i[ 1];
791 w0[2] = pws[gid].i[ 2];
792 w0[3] = pws[gid].i[ 3];
815 const u32 pw_len = pws[gid].pw_len;
821 m05300s (w_s, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, ikepsk_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_msg_buf);
824 __kernel void m05300_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 ikepsk_t *ikepsk_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
830 const u32 gid = get_global_id (0);
831 const u32 lid = get_local_id (0);
832 const u32 lsz = get_local_size (0);
840 for (u32 i = lid; i < 16; i += lsz)
842 w_s[i] = ikepsk_bufs[salt_pos].nr_buf[i];
845 barrier (CLK_LOCAL_MEM_FENCE);
847 __local u32 s_msg_buf[128];
849 for (u32 i = lid; i < 128; i += lsz)
851 s_msg_buf[i] = ikepsk_bufs[salt_pos].msg_buf[i];
854 barrier (CLK_LOCAL_MEM_FENCE);
856 if (gid >= gid_max) return;
864 w0[0] = pws[gid].i[ 0];
865 w0[1] = pws[gid].i[ 1];
866 w0[2] = pws[gid].i[ 2];
867 w0[3] = pws[gid].i[ 3];
871 w1[0] = pws[gid].i[ 4];
872 w1[1] = pws[gid].i[ 5];
873 w1[2] = pws[gid].i[ 6];
874 w1[3] = pws[gid].i[ 7];
890 const u32 pw_len = pws[gid].pw_len;
896 m05300s (w_s, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, ikepsk_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_msg_buf);
899 __kernel void m05300_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 ikepsk_t *ikepsk_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 bfs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
905 const u32 gid = get_global_id (0);
906 const u32 lid = get_local_id (0);
907 const u32 lsz = get_local_size (0);
915 for (u32 i = lid; i < 16; i += lsz)
917 w_s[i] = ikepsk_bufs[salt_pos].nr_buf[i];
920 barrier (CLK_LOCAL_MEM_FENCE);
922 __local u32 s_msg_buf[128];
924 for (u32 i = lid; i < 128; i += lsz)
926 s_msg_buf[i] = ikepsk_bufs[salt_pos].msg_buf[i];
929 barrier (CLK_LOCAL_MEM_FENCE);
931 if (gid >= gid_max) return;
939 w0[0] = pws[gid].i[ 0];
940 w0[1] = pws[gid].i[ 1];
941 w0[2] = pws[gid].i[ 2];
942 w0[3] = pws[gid].i[ 3];
946 w1[0] = pws[gid].i[ 4];
947 w1[1] = pws[gid].i[ 5];
948 w1[2] = pws[gid].i[ 6];
949 w1[3] = pws[gid].i[ 7];
953 w2[0] = pws[gid].i[ 8];
954 w2[1] = pws[gid].i[ 9];
955 w2[2] = pws[gid].i[10];
956 w2[3] = pws[gid].i[11];
960 w3[0] = pws[gid].i[12];
961 w3[1] = pws[gid].i[13];
965 const u32 pw_len = pws[gid].pw_len;
971 m05300s (w_s, w0, w1, w2, w3, pw_len, pws, rules_buf, combs_buf, bfs_buf, tmps, hooks, bitmaps_buf_s1_a, bitmaps_buf_s1_b, bitmaps_buf_s1_c, bitmaps_buf_s1_d, bitmaps_buf_s2_a, bitmaps_buf_s2_b, bitmaps_buf_s2_c, bitmaps_buf_s2_d, plains_buf, digests_buf, hashes_shown, salt_bufs, ikepsk_bufs, d_return_buf, d_scryptV_buf, bitmap_mask, bitmap_shift1, bitmap_shift2, salt_pos, loop_pos, loop_cnt, bfs_cnt, digests_cnt, digests_offset, s_msg_buf);