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"
19 #include "include/rp_gpu.h"
20 #include "OpenCL/rp.c"
22 #define COMPARE_S "OpenCL/check_single_comp4.c"
23 #define COMPARE_M "OpenCL/check_multi_comp4.c"
25 static void md5_transform (const u32 w0[4], const u32 w1[4], const u32 w2[4], const u32 w3[4], u32 digest[4])
51 MD5_STEP (MD5_Fo, a, b, c, d, w0_t, MD5C00, MD5S00);
52 MD5_STEP (MD5_Fo, d, a, b, c, w1_t, MD5C01, MD5S01);
53 MD5_STEP (MD5_Fo, c, d, a, b, w2_t, MD5C02, MD5S02);
54 MD5_STEP (MD5_Fo, b, c, d, a, w3_t, MD5C03, MD5S03);
55 MD5_STEP (MD5_Fo, a, b, c, d, w4_t, MD5C04, MD5S00);
56 MD5_STEP (MD5_Fo, d, a, b, c, w5_t, MD5C05, MD5S01);
57 MD5_STEP (MD5_Fo, c, d, a, b, w6_t, MD5C06, MD5S02);
58 MD5_STEP (MD5_Fo, b, c, d, a, w7_t, MD5C07, MD5S03);
59 MD5_STEP (MD5_Fo, a, b, c, d, w8_t, MD5C08, MD5S00);
60 MD5_STEP (MD5_Fo, d, a, b, c, w9_t, MD5C09, MD5S01);
61 MD5_STEP (MD5_Fo, c, d, a, b, wa_t, MD5C0a, MD5S02);
62 MD5_STEP (MD5_Fo, b, c, d, a, wb_t, MD5C0b, MD5S03);
63 MD5_STEP (MD5_Fo, a, b, c, d, wc_t, MD5C0c, MD5S00);
64 MD5_STEP (MD5_Fo, d, a, b, c, wd_t, MD5C0d, MD5S01);
65 MD5_STEP (MD5_Fo, c, d, a, b, we_t, MD5C0e, MD5S02);
66 MD5_STEP (MD5_Fo, b, c, d, a, wf_t, MD5C0f, MD5S03);
68 MD5_STEP (MD5_Go, a, b, c, d, w1_t, MD5C10, MD5S10);
69 MD5_STEP (MD5_Go, d, a, b, c, w6_t, MD5C11, MD5S11);
70 MD5_STEP (MD5_Go, c, d, a, b, wb_t, MD5C12, MD5S12);
71 MD5_STEP (MD5_Go, b, c, d, a, w0_t, MD5C13, MD5S13);
72 MD5_STEP (MD5_Go, a, b, c, d, w5_t, MD5C14, MD5S10);
73 MD5_STEP (MD5_Go, d, a, b, c, wa_t, MD5C15, MD5S11);
74 MD5_STEP (MD5_Go, c, d, a, b, wf_t, MD5C16, MD5S12);
75 MD5_STEP (MD5_Go, b, c, d, a, w4_t, MD5C17, MD5S13);
76 MD5_STEP (MD5_Go, a, b, c, d, w9_t, MD5C18, MD5S10);
77 MD5_STEP (MD5_Go, d, a, b, c, we_t, MD5C19, MD5S11);
78 MD5_STEP (MD5_Go, c, d, a, b, w3_t, MD5C1a, MD5S12);
79 MD5_STEP (MD5_Go, b, c, d, a, w8_t, MD5C1b, MD5S13);
80 MD5_STEP (MD5_Go, a, b, c, d, wd_t, MD5C1c, MD5S10);
81 MD5_STEP (MD5_Go, d, a, b, c, w2_t, MD5C1d, MD5S11);
82 MD5_STEP (MD5_Go, c, d, a, b, w7_t, MD5C1e, MD5S12);
83 MD5_STEP (MD5_Go, b, c, d, a, wc_t, MD5C1f, MD5S13);
85 MD5_STEP (MD5_H1, a, b, c, d, w5_t, MD5C20, MD5S20);
86 MD5_STEP (MD5_H2, d, a, b, c, w8_t, MD5C21, MD5S21);
87 MD5_STEP (MD5_H1, c, d, a, b, wb_t, MD5C22, MD5S22);
88 MD5_STEP (MD5_H2, b, c, d, a, we_t, MD5C23, MD5S23);
89 MD5_STEP (MD5_H1, a, b, c, d, w1_t, MD5C24, MD5S20);
90 MD5_STEP (MD5_H2, d, a, b, c, w4_t, MD5C25, MD5S21);
91 MD5_STEP (MD5_H1, c, d, a, b, w7_t, MD5C26, MD5S22);
92 MD5_STEP (MD5_H2, b, c, d, a, wa_t, MD5C27, MD5S23);
93 MD5_STEP (MD5_H1, a, b, c, d, wd_t, MD5C28, MD5S20);
94 MD5_STEP (MD5_H2, d, a, b, c, w0_t, MD5C29, MD5S21);
95 MD5_STEP (MD5_H1, c, d, a, b, w3_t, MD5C2a, MD5S22);
96 MD5_STEP (MD5_H2, b, c, d, a, w6_t, MD5C2b, MD5S23);
97 MD5_STEP (MD5_H1, a, b, c, d, w9_t, MD5C2c, MD5S20);
98 MD5_STEP (MD5_H2, d, a, b, c, wc_t, MD5C2d, MD5S21);
99 MD5_STEP (MD5_H1, c, d, a, b, wf_t, MD5C2e, MD5S22);
100 MD5_STEP (MD5_H2, b, c, d, a, w2_t, MD5C2f, MD5S23);
102 MD5_STEP (MD5_I , a, b, c, d, w0_t, MD5C30, MD5S30);
103 MD5_STEP (MD5_I , d, a, b, c, w7_t, MD5C31, MD5S31);
104 MD5_STEP (MD5_I , c, d, a, b, we_t, MD5C32, MD5S32);
105 MD5_STEP (MD5_I , b, c, d, a, w5_t, MD5C33, MD5S33);
106 MD5_STEP (MD5_I , a, b, c, d, wc_t, MD5C34, MD5S30);
107 MD5_STEP (MD5_I , d, a, b, c, w3_t, MD5C35, MD5S31);
108 MD5_STEP (MD5_I , c, d, a, b, wa_t, MD5C36, MD5S32);
109 MD5_STEP (MD5_I , b, c, d, a, w1_t, MD5C37, MD5S33);
110 MD5_STEP (MD5_I , a, b, c, d, w8_t, MD5C38, MD5S30);
111 MD5_STEP (MD5_I , d, a, b, c, wf_t, MD5C39, MD5S31);
112 MD5_STEP (MD5_I , c, d, a, b, w6_t, MD5C3a, MD5S32);
113 MD5_STEP (MD5_I , b, c, d, a, wd_t, MD5C3b, MD5S33);
114 MD5_STEP (MD5_I , a, b, c, d, w4_t, MD5C3c, MD5S30);
115 MD5_STEP (MD5_I , d, a, b, c, wb_t, MD5C3d, MD5S31);
116 MD5_STEP (MD5_I , c, d, a, b, w2_t, MD5C3e, MD5S32);
117 MD5_STEP (MD5_I , b, c, d, a, w9_t, MD5C3f, MD5S33);
125 static void hmac_md5_pad (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4])
127 w0[0] = w0[0] ^ 0x36363636;
128 w0[1] = w0[1] ^ 0x36363636;
129 w0[2] = w0[2] ^ 0x36363636;
130 w0[3] = w0[3] ^ 0x36363636;
131 w1[0] = w1[0] ^ 0x36363636;
132 w1[1] = w1[1] ^ 0x36363636;
133 w1[2] = w1[2] ^ 0x36363636;
134 w1[3] = w1[3] ^ 0x36363636;
135 w2[0] = w2[0] ^ 0x36363636;
136 w2[1] = w2[1] ^ 0x36363636;
137 w2[2] = w2[2] ^ 0x36363636;
138 w2[3] = w2[3] ^ 0x36363636;
139 w3[0] = w3[0] ^ 0x36363636;
140 w3[1] = w3[1] ^ 0x36363636;
141 w3[2] = w3[2] ^ 0x36363636;
142 w3[3] = w3[3] ^ 0x36363636;
149 md5_transform (w0, w1, w2, w3, ipad);
151 w0[0] = w0[0] ^ 0x6a6a6a6a;
152 w0[1] = w0[1] ^ 0x6a6a6a6a;
153 w0[2] = w0[2] ^ 0x6a6a6a6a;
154 w0[3] = w0[3] ^ 0x6a6a6a6a;
155 w1[0] = w1[0] ^ 0x6a6a6a6a;
156 w1[1] = w1[1] ^ 0x6a6a6a6a;
157 w1[2] = w1[2] ^ 0x6a6a6a6a;
158 w1[3] = w1[3] ^ 0x6a6a6a6a;
159 w2[0] = w2[0] ^ 0x6a6a6a6a;
160 w2[1] = w2[1] ^ 0x6a6a6a6a;
161 w2[2] = w2[2] ^ 0x6a6a6a6a;
162 w2[3] = w2[3] ^ 0x6a6a6a6a;
163 w3[0] = w3[0] ^ 0x6a6a6a6a;
164 w3[1] = w3[1] ^ 0x6a6a6a6a;
165 w3[2] = w3[2] ^ 0x6a6a6a6a;
166 w3[3] = w3[3] ^ 0x6a6a6a6a;
173 md5_transform (w0, w1, w2, w3, opad);
176 static void hmac_md5_run (u32 w0[4], u32 w1[4], u32 w2[4], u32 w3[4], u32 ipad[4], u32 opad[4], u32 digest[4])
183 md5_transform (w0, w1, w2, w3, digest);
199 w3[2] = (64 + 16) * 8;
207 md5_transform (w0, w1, w2, w3, digest);
210 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00050_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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
216 const u32 lid = get_local_id (0);
222 const u32 gid = get_global_id (0);
224 if (gid >= gid_max) return;
228 pw_buf0[0] = pws[gid].i[ 0];
229 pw_buf0[1] = pws[gid].i[ 1];
230 pw_buf0[2] = pws[gid].i[ 2];
231 pw_buf0[3] = pws[gid].i[ 3];
235 pw_buf1[0] = pws[gid].i[ 4];
236 pw_buf1[1] = pws[gid].i[ 5];
237 pw_buf1[2] = pws[gid].i[ 6];
238 pw_buf1[3] = pws[gid].i[ 7];
240 const u32 pw_len = pws[gid].pw_len;
248 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
249 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
250 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
251 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
255 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
256 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
257 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
258 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
262 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
263 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
264 salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
265 salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
269 salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
270 salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
271 salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
272 salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
274 const u32 salt_len = salt_bufs[salt_pos].salt_len;
280 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
310 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
347 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
349 w0_t[0] = salt_buf0[0];
350 w0_t[1] = salt_buf0[1];
351 w0_t[2] = salt_buf0[2];
352 w0_t[3] = salt_buf0[3];
353 w1_t[0] = salt_buf1[0];
354 w1_t[1] = salt_buf1[1];
355 w1_t[2] = salt_buf1[2];
356 w1_t[3] = salt_buf1[3];
357 w2_t[0] = salt_buf2[0];
358 w2_t[1] = salt_buf2[1];
359 w2_t[2] = salt_buf2[2];
360 w2_t[3] = salt_buf2[3];
361 w3_t[0] = salt_buf3[0];
362 w3_t[1] = salt_buf3[1];
363 w3_t[2] = (64 + salt_len) * 8;
368 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
370 const u32 r0 = digest[0];
371 const u32 r1 = digest[3];
372 const u32 r2 = digest[2];
373 const u32 r3 = digest[1];
379 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00050_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)
383 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00050_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)
387 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00050_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 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
393 const u32 lid = get_local_id (0);
399 const u32 gid = get_global_id (0);
401 if (gid >= gid_max) return;
405 pw_buf0[0] = pws[gid].i[ 0];
406 pw_buf0[1] = pws[gid].i[ 1];
407 pw_buf0[2] = pws[gid].i[ 2];
408 pw_buf0[3] = pws[gid].i[ 3];
412 pw_buf1[0] = pws[gid].i[ 4];
413 pw_buf1[1] = pws[gid].i[ 5];
414 pw_buf1[2] = pws[gid].i[ 6];
415 pw_buf1[3] = pws[gid].i[ 7];
417 const u32 pw_len = pws[gid].pw_len;
425 salt_buf0[0] = salt_bufs[salt_pos].salt_buf[ 0];
426 salt_buf0[1] = salt_bufs[salt_pos].salt_buf[ 1];
427 salt_buf0[2] = salt_bufs[salt_pos].salt_buf[ 2];
428 salt_buf0[3] = salt_bufs[salt_pos].salt_buf[ 3];
432 salt_buf1[0] = salt_bufs[salt_pos].salt_buf[ 4];
433 salt_buf1[1] = salt_bufs[salt_pos].salt_buf[ 5];
434 salt_buf1[2] = salt_bufs[salt_pos].salt_buf[ 6];
435 salt_buf1[3] = salt_bufs[salt_pos].salt_buf[ 7];
439 salt_buf2[0] = salt_bufs[salt_pos].salt_buf[ 8];
440 salt_buf2[1] = salt_bufs[salt_pos].salt_buf[ 9];
441 salt_buf2[2] = salt_bufs[salt_pos].salt_buf[10];
442 salt_buf2[3] = salt_bufs[salt_pos].salt_buf[11];
446 salt_buf3[0] = salt_bufs[salt_pos].salt_buf[12];
447 salt_buf3[1] = salt_bufs[salt_pos].salt_buf[13];
448 salt_buf3[2] = salt_bufs[salt_pos].salt_buf[14];
449 salt_buf3[3] = salt_bufs[salt_pos].salt_buf[15];
451 const u32 salt_len = salt_bufs[salt_pos].salt_len;
457 const u32 search[4] =
459 digests_buf[digests_offset].digest_buf[DGST_R0],
460 digests_buf[digests_offset].digest_buf[DGST_R1],
461 digests_buf[digests_offset].digest_buf[DGST_R2],
462 digests_buf[digests_offset].digest_buf[DGST_R3]
469 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
499 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
536 hmac_md5_pad (w0_t, w1_t, w2_t, w3_t, ipad, opad);
538 w0_t[0] = salt_buf0[0];
539 w0_t[1] = salt_buf0[1];
540 w0_t[2] = salt_buf0[2];
541 w0_t[3] = salt_buf0[3];
542 w1_t[0] = salt_buf1[0];
543 w1_t[1] = salt_buf1[1];
544 w1_t[2] = salt_buf1[2];
545 w1_t[3] = salt_buf1[3];
546 w2_t[0] = salt_buf2[0];
547 w2_t[1] = salt_buf2[1];
548 w2_t[2] = salt_buf2[2];
549 w2_t[3] = salt_buf2[3];
550 w3_t[0] = salt_buf3[0];
551 w3_t[1] = salt_buf3[1];
552 w3_t[2] = (64 + salt_len) * 8;
557 hmac_md5_run (w0_t, w1_t, w2_t, w3_t, ipad, opad, digest);
559 const u32 r0 = digest[0];
560 const u32 r1 = digest[3];
561 const u32 r2 = digest[2];
562 const u32 r3 = digest[1];
568 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00050_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)
572 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m00050_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)